input
stringlengths
1
18.7k
output
stringlengths
1
18.7k
Submitted issue here <https://github.com/flyteorg/flyte/issues/1242> Thanks Prafulla Mahindrakar for your help :hugging_face:
I think this issue exists?
I think this issue exists?
is this related to what Haytham filed with the duplicate subworkflow node ids?
is this related to what Haytham filed with the duplicate subworkflow node ids?
I think so
I think so
<https://github.com/flyteorg/flyte/issues/1085>
<https://github.com/flyteorg/flyte/issues/1085>
It looks like a different bug. In other case, node ids were changed
It looks like a different bug. In other case, node ids were changed
Jason Porter
Jason Porter
I looked into what’s returned from flyteadmin, and there was no field that could fit Actually I looked into UI, and it doesn’t have #1085
hey team, is there a way to parameterize the admin host in flytectl with command line flags entirely in lieu of the configure file? `--admin.endpoint string` doesn’t do anything for me
yes you can . Can you upgrade to the latest version of flytectl
yes you can . Can you upgrade to the latest version of flytectl
just reran `curl -s <https://raw.githubusercontent.com/lyft/flytectl/master/install.sh> | bash`
just reran `curl -s <https://raw.githubusercontent.com/lyft/flytectl/master/install.sh> | bash`
can you check the version if its &gt;=<https://github.com/flyteorg/flytectl/releases/tag/v0.2.5|v0.2.5> `flytectl version`
can you check the version if its &gt;=<https://github.com/flyteorg/flytectl/releases/tag/v0.2.5|v0.2.5> `flytectl version`
oop `"0.1.9` whats recommended way fetch latest
oop `"0.1.9` whats recommended way fetch latest
Kenny Workman are you on osx? if so can you use brew?
Kenny Workman are you on osx? if so can you use brew?
sure, but i’m also looking for a method that will port well to containers/*nix flavored systems well will the install script fetch latest on a fresh system?
sure, but i’m also looking for a method that will port well to containers/*nix flavored systems well will the install script fetch latest on a fresh system?
you have to move it manually move the binaries to your prev location
you have to move it manually move the binaries to your prev location
thats fine, aw I see it doesn’t actually do something if there’s an existing binary on the system just wiped and fetched latest the `--admin.endpoint` flag stil does not work with `0.2.14` - is there a new flag?
thats fine, aw I see it doesn’t actually do something if there’s an existing binary on the system just wiped and fetched latest the `--admin.endpoint` flag stil does not work with `0.2.14` - is there a new flag?
hey Kenny Workman I just did this and it worked. ```bin/flytectl get projects --admin.endpoint dns:///demo.nuclyde.io ``` you have to give it in subcommand context . i.e the command should have ```flytectl verb noun --admin.endpoint``` It wont work in ```flytecl --admin.endpoint``` or ```flytectl verb --admin.endpoint`...
hey Kenny Workman I just did this and it worked. ```bin/flytectl get projects --admin.endpoint dns:///demo.nuclyde.io ``` you have to give it in subcommand context . i.e the command should have ```flytectl verb noun --admin.endpoint``` It wont work in ```flytecl --admin.endpoint``` or ```flytectl verb --admin.endpoint`...
Prafulla Mahindrakar, thanks but this still doesn’t work for me ```./bin/flytectl version { "App": "flytectl", "Build": "5969b18", "Version": "0.2.14", "BuildTime": "2021-07-15 10:38:01.613502 -0700 PDT m=+0.017129500" }%``` ```./bin/flytectl version 9 ~$ ./bin/flytectl get projects --admin.endpoint "flyte....
Prafulla Mahindrakar, thanks but this still doesn’t work for me ```./bin/flytectl version { "App": "flytectl", "Build": "5969b18", "Version": "0.2.14", "BuildTime": "2021-07-15 10:38:01.613502 -0700 PDT m=+0.017129500" }%``` ```./bin/flytectl version 9 ~$ ./bin/flytectl get projects --admin.endpoint "flyte....
Kenny Workman You also need `--admin.insecure` flag. try this ```flytectl get projects --admin.endpoint 'dns:///127.0.0.1:30081' --admin.insecure ````
Kenny Workman You also need `--admin.insecure` flag. try this ```flytectl get projects --admin.endpoint 'dns:///127.0.0.1:30081' --admin.insecure ````
bingo, it worked, thanks!
seeing an odd issue with propeller Ketan Umare: ```{"json":{"src":"controller.go:137"},"level":"info","msg":"==\u003e Enqueueing workflow [pineapple-development/faa7934a6dd0a4c7d9c9]","ts":"2021-07-15T20:23:59Z"} {"json":{"exec_id":"faa7934a6dd0a4c7d9c9","ns":"pineapple-development","routine":"worker-21","src":"handler...
why do you think its completed Also the snippet of the log you shared, is that repeating again and again?
why do you think its completed Also the snippet of the log you shared, is that repeating again and again?
yea it is
yea it is
n0 is running and n1 is not even started
n0 is running and n1 is not even started
Anooj Patel said all the pods are in completed phase
Anooj Patel said all the pods are in completed phase
its a workflowNode ``` Workflow Node Status: Execution Error: &lt;nil&gt; Phase: 1``` so its running another workflow - launch plan? is that other workflow running?
its a workflowNode ``` Workflow Node Status: Execution Error: &lt;nil&gt; Phase: 1``` so its running another workflow - launch plan? is that other workflow running?
hmm
hmm
The dynamic task launched a bunch of tasks, and a of the ones I’ve check via kubectl from the list, listed as running, were already completed
The dynamic task launched a bunch of tasks, and a of the ones I’ve check via kubectl from the list, listed as running, were already completed
good call Anooj Patel: there appear to be 2 containers in `CreateContainerError`?
good call Anooj Patel: there appear to be 2 containers in `CreateContainerError`?
yea just saw it on refresh
yea just saw it on refresh
i think this is a confluence of errors lol one is that there are containers failing, but the flyteworkflow yaml is also giant
i think this is a confluence of errors lol one is that there are containers failing, but the flyteworkflow yaml is also giant
looks like its running again? A few tasks at least
looks like its running again? A few tasks at least
i think propeller is struggling to update the workflow object ```{"json":{"exec_id":"fqsqptwy","node":"n1/fucdnhta","ns":"pineapple-development","res_ver":"1151478865","routine":"worker-18","src":"executor.go:124","wf":"pineapple:development:pineappleflow.core.flyte.training.search_workflow.OptimizedSearchWorkflow"},"l...
i think propeller is struggling to update the workflow object ```{"json":{"exec_id":"fqsqptwy","node":"n1/fucdnhta","ns":"pineapple-development","res_ver":"1151478865","routine":"worker-18","src":"executor.go:124","wf":"pineapple:development:pineappleflow.core.flyte.training.search_workflow.OptimizedSearchWorkflow"},"l...
Jeev B it does that, because it is unable to record state to etcD and so it does not that it did this previously
Jeev B it does that, because it is unable to record state to etcD and so it does not that it did this previously
ok that makes sense that would fit with the giant `FlyteWorkflow` CRD problem Anooj Patel: we really need to start using map tasks :confused: our API server is falling apart :stuck_out_tongue:
ok that makes sense that would fit with the giant `FlyteWorkflow` CRD problem Anooj Patel: we really need to start using map tasks :confused: our API server is falling apart :stuck_out_tongue:
Jeev B also - cc Katrina Rogan / Haytham Abuelfutuh we need to auto create the map task :disappointed:
Jeev B also - cc Katrina Rogan / Haytham Abuelfutuh we need to auto create the map task :disappointed:
like replace these dynamic tasks under the hood with map tasks you mean? could you elaborate Ketan Umare?
like replace these dynamic tasks under the hood with map tasks you mean? could you elaborate Ketan Umare?
Jeev B yes, like we used to do in older flytekit if you run a for-loop with the same task, we will just replace it with a map task
Jeev B yes, like we used to do in older flytekit if you run a for-loop with the same task, we will just replace it with a map task
oh inside the dynamic node?
oh inside the dynamic node?
the problem is for non symmetric inputs ya
the problem is for non symmetric inputs ya
oh yea i rmb that but only for python tasks not sidecar tbh we should just start using map tasks...we will
oh yea i rmb that but only for python tasks not sidecar tbh we should just start using map tasks...we will
Ketan Umare do we have an issue tracking this? we should file one
Ketan Umare do we have an issue tracking this? we should file one
what issue Katrina Rogan? auto creating map tasks? from for loops ? If so there is probably some questions there but ya, lets create the issue and we can think over it
what issue Katrina Rogan? auto creating map tasks? from for loops ? If so there is probably some questions there but ya, lets create the issue and we can think over it
yeah that one exactly just so we don't lose track, don't necessarily need to spec it all out now
Hi! We're using a `ContainerTask` to run some of our native things. Some of the inputs that the container receives are ~500mb. When running the Task, the `flyte-copilot-downloader` container in the pod fails with: ```{"json":{},"level":"error","msg":"Failed to persist [&lt;INPUT_NAME&gt;], err failed to read Blob from...
Dug into the go code a bit, and I managed to fix this issue by adding ``` limits: maxDownloadMBs: 100000``` to all storage configs on kubernetes I could find. Flyte should probably catch this error though and make the task fail I think.
Dug into the go code a bit, and I managed to fix this issue by adding ``` limits: maxDownloadMBs: 100000``` to all storage configs on kubernetes I could find. Flyte should probably catch this error though and make the task fail I think.
Ohh this error is only in copilot, so copilot should be allowed to download as much as needed Cc Yuvraj (union.ai) Samuel Kogler / Bernhard Stadlbauer Sorry for the delay in response. What you have done for now is fine - you updated the maxDownloadMbs: large number. You only need to update this for flyte-data-config i...
Ohh this error is only in copilot, so copilot should be allowed to download as much as needed Cc Yuvraj (union.ai) Samuel Kogler / Bernhard Stadlbauer Sorry for the delay in response. What you have done for now is fine - you updated the maxDownloadMbs: large number. You only need to update this for flyte-data-config i...
Thank you Ketan, that is great to hear! No worries, this was still a very fast reply. We'll watch the issues you created and test when we upgrade our Flyte version after they are resolved. For now, our fix is working fine, so not a big deal.
Thank you Ketan, that is great to hear! No worries, this was still a very fast reply. We'll watch the issues you created and test when we upgrade our Flyte version after they are resolved. For now, our fix is working fine, so not a big deal.
cc <@U01QWFALUE9>
cc <@U01QWFALUE9>
Referenced in GitHub Discussions <https://github.com/flyteorg/flyte/discussions/1287>
Hi! I have a question about resource limits on task pods. In flytekit-java, the `reources` in task template.container is not set. The result of this is the Pods for tasks all get default as ```resources: limits: cpu: "1" memory: 1Gi requests: cpu: "1" memory: 1Gi``` I think t...
These should be limits from admin ideally, I think but yes this is the last line of defense Cc Katrina Rogan
These should be limits from admin ideally, I think but yes this is the last line of defense Cc Katrina Rogan
CC <@U024KQ3GL59>
CC <@U024KQ3GL59>
ah thank you Nian Tang, we should remove the last check and initialize resource fields for containers when they don't exist
ah thank you Nian Tang, we should remove the last check and initialize resource fields for containers when they don't exist
:+1: . Thanks for the confirmation. We can fix it
to compare Flyte to Dagster (since that is a sort of meaningful evaluation/comparison in my company): <https://docs.dagster.io/integrations/great-expectations> and <https://docs.dagster.io/integrations/dbt> I know DBT and Great Expectations (GE) are heavily used. In theory, there is nothing prohibiting Flyte integrati...
austin Great you asked - <https://github.com/flyteorg/flytekit/pull/495> We have a WIP (almost complete - waiting for a PR) with GE. We think our integration has better ergonomics for DBT we can easily write an integration
austin Great you asked - <https://github.com/flyteorg/flytekit/pull/495> We have a WIP (almost complete - waiting for a PR) with GE. We think our integration has better ergonomics for DBT we can easily write an integration
Ah, ya, I actually remember seeing the GE one. And, I'm not sure how critical... but now I can comment that it is doable
Ah, ya, I actually remember seeing the GE one. And, I'm not sure how critical... but now I can comment that it is doable
Cool
Cool
:+1: :pray:
Hey! Curious if there have been any initial load tests done around map tasks and their scalability (outside of normal k8s limitations)? If so, can someone please share? Also any idea how it may behave with spot instances?
we (Freenome) will be testing this further in the coming weeks with some reasonably large fanouts (&gt;3k) Miguel Toledo. If i understand correctly, these shouldnt work any differently than regular tasks from an operational standpoint. its still 1 pod / task atm, and should be rescheduled when a spot node is pulled. A...
we (Freenome) will be testing this further in the coming weeks with some reasonably large fanouts (&gt;3k) Miguel Toledo. If i understand correctly, these shouldnt work any differently than regular tasks from an operational standpoint. its still 1 pod / task atm, and should be rescheduled when a spot node is pulled. A...
Was curious if it was treated as an independent task or a collection. Definitely let me know about the findings! Looking into to doing our own testing in the coming weeks as well
Was curious if it was treated as an independent task or a collection. Definitely let me know about the findings! Looking into to doing our own testing in the coming weeks as well
yea would be great to share war stories there :stuck_out_tongue:
yea would be great to share war stories there :stuck_out_tongue:
If it works well, we're hoping to drop some of our spark tasks in favor of map tasks Will definitely share! Especially the spot tasks as that may differ from our infrastructure
If it works well, we're hoping to drop some of our spark tasks in favor of map tasks Will definitely share! Especially the spot tasks as that may differ from our infrastructure
yeah so far the map_tasks i have tested; they usually scheduled on spot instance or preempt in GCS. Thus some of the map tasks will just fail and the work restarts. i havent enabled the caching yet so it just restarts the whole array. So if i enable caching, it should only restart the evicted ones at least thats my hun...
yeah so far the map_tasks i have tested; they usually scheduled on spot instance or preempt in GCS. Thus some of the map tasks will just fail and the work restarts. i havent enabled the caching yet so it just restarts the whole array. So if i enable caching, it should only restart the evicted ones at least thats my hun...
Jay Ganbat: that's interesting. is this a userspace failure or some infra issue?
Jay Ganbat: that's interesting. is this a userspace failure or some infra issue?
Jay Ganbat by restart the entire array do you mean restart every task? For example, I have map tasks 1-10 and task 2 fails, are you seeing all 1-10 retry or just task 2?
Jay Ganbat by restart the entire array do you mean restart every task? For example, I have map tasks 1-10 and task 2 fails, are you seeing all 1-10 retry or just task 2?
maybe it would help to get some clarification about intended behavior on this. Ketan Umare Katrina Rogan
maybe it would help to get some clarification about intended behavior on this. Ketan Umare Katrina Rogan
yeah all 10 retries, but i think thats because im not setting cache_version in the Metadata
yeah all 10 retries, but i think thats because im not setting cache_version in the Metadata
Ok, I believe there were optimizations done to propeller and caching where it doesn't actually schedule the pod if the cache outputs are there so it kind of implicitly is only retrying the subtasks that fail? but maybe I'm missing something
Ok, I believe there were optimizations done to propeller and caching where it doesn't actually schedule the pod if the cache outputs are there so it kind of implicitly is only retrying the subtasks that fail? but maybe I'm missing something
that works as intended Miguel Toledo for regular and dynamic tasks at least. we haven't tried that with map tasks yet.
that works as intended Miguel Toledo for regular and dynamic tasks at least. we haven't tried that with map tasks yet.
I see, yeah it'll be interesting to see how it works with map tasks. If it does reschedule each job then that's likely gonna be a blocker for us and was the main reason we had to go with spark tasks
I see, yeah it'll be interesting to see how it works with map tasks. If it does reschedule each job then that's likely gonna be a blocker for us and was the main reason we had to go with spark tasks
i feel like it should retry only the failed ones, regardless of whether caching is enabled... within the same execution of the workflow.
i feel like it should retry only the failed ones, regardless of whether caching is enabled... within the same execution of the workflow.
+1
+1
i believe we should be writing individual subtasks to discovery just as we did formerly for pre-map, array tasks
i believe we should be writing individual subtasks to discovery just as we did formerly for pre-map, array tasks
Katrina Rogan is there anywhere to learn or can you comment are what are the main differences in the backend between array tasks and map tasks?
Katrina Rogan is there anywhere to learn or can you comment are what are the main differences in the backend between array tasks and map tasks?
there's very few differences, functionally they should be the same just changes in how we serialize a few bits (success criteria, input locations, etc)
there's very few differences, functionally they should be the same just changes in how we serialize a few bits (success criteria, input locations, etc)
&gt; Ok, I believe there were optimizations done to propeller and caching where it doesn't actually schedule the pod if the cache outputs are there so it kind of implicitly is only retrying the subtasks that fail? Katrina Rogan ok thanks! 1 last question. is my statement above valid? as in it should implicitly retry on...
&gt; Ok, I believe there were optimizations done to propeller and caching where it doesn't actually schedule the pod if the cache outputs are there so it kind of implicitly is only retrying the subtasks that fail? Katrina Rogan ok thanks! 1 last question. is my statement above valid? as in it should implicitly retry on...
yes i think so!
yes i think so!
Awesome! And 1 last, last question haha. Is the retry at the sub task level or overall array level?
Awesome! And 1 last, last question haha. Is the retry at the sub task level or overall array level?
oh shoot sorry missed this. i think the retry is at the overall array level unfortunately
oh shoot sorry missed this. i think the retry is at the overall array level unfortunately
Katrina Rogan: would memoization help here? its a bit painful to retry 3k tasks if 1 fails...
Katrina Rogan: would memoization help here? its a bit painful to retry 3k tasks if 1 fails...
Yes memorization will help, and we should retry at the task level too
Yes memorization will help, and we should retry at the task level too
Should I create a github issue for this?
Should I create a github issue for this?
Miguel Toledo please do, i am not sure, why is it not possible to re-run individual tasks on k8s Katrina Rogan?
Miguel Toledo please do, i am not sure, why is it not possible to re-run individual tasks on k8s Katrina Rogan?
it should be doable, i'm just not sure that we are doing it... but i could be wrong
it should be doable, i'm just not sure that we are doing it... but i could be wrong
got it, i want to understand your thought process and probably come up with a conclusion
got it, i want to understand your thought process and probably come up with a conclusion
Github issue: <https://github.com/flyteorg/flyte/issues/1276>
There was a change to how branch-node sub-nodes work in <https://github.com/flyteorg/flytepropeller/pull/285>. I think it broke workflows that use sub-nodes as upstream nodes, they don’t compile anymore, for instance (using java sdk) ```@AutoService(SdkWorkflow.class) public class BranchNodeAndSubNodeWorkflow extends S...
Can you share the generated - workflow spec Gleb Kanterov the change was intentional, we found an issue in the compiler in highly nested scenarios and this also would lead to a perf penalty. We have now simplified and I can share the expectation for branch node
Can you share the generated - workflow spec Gleb Kanterov the change was intentional, we found an issue in the compiler in highly nested scenarios and this also would lead to a perf penalty. We have now simplified and I can share the expectation for branch node
is proto file enough? I need to add an option to export it because registration doesn’t happen I can’t see it in admin
is proto file enough? I need to add an option to export it because registration doesn’t happen I can’t see it in admin
proto file is enough btw, here are all the test cases - <https://github.com/flyteorg/flytepropeller/tree/master/pkg/compiler/test/testdata/branch> i have a meeting till 9:30 then i can help with this But, I do agree there was a change in behavior. It simplifies the branch specification. But, IIRC, you did bring up an ...
proto file is enough btw, here are all the test cases - <https://github.com/flyteorg/flytepropeller/tree/master/pkg/compiler/test/testdata/branch> i have a meeting till 9:30 then i can help with this But, I do agree there was a change in behavior. It simplifies the branch specification. But, IIRC, you did bring up an ...
yeah… the problem that there is no other way to achieve the same behaviour, because branch nodes only allow to have 1 node as statement for conditional case
yeah… the problem that there is no other way to achieve the same behaviour, because branch nodes only allow to have 1 node as statement for conditional case
you can create a subworkflow - Gleb Kanterov? the node can be a subworkflow
you can create a subworkflow - Gleb Kanterov? the node can be a subworkflow
Yes, it’s possible, then UI will look strange The workflow that uses conditional is already sub-workflow Do you think using sub-workflow is workaround for this issue, or new behaviour is how it’s intended to work?
Yes, it’s possible, then UI will look strange The workflow that uses conditional is already sub-workflow Do you think using sub-workflow is workaround for this issue, or new behaviour is how it’s intended to work?
ya you can use subworkflows for this. So the idea is that conditionals exist and downstream can only consume outputs of conditionals though the error is weird so if you send me the protobuf, i can fix the error to be nicer Gleb Kanterov also the reason why this is wrong, is because the output of a node in conditional i...
ya you can use subworkflows for this. So the idea is that conditionals exist and downstream can only consume outputs of conditionals though the error is weird so if you send me the protobuf, i can fix the error to be nicer Gleb Kanterov also the reason why this is wrong, is because the output of a node in conditional i...
I see. I need to implement serialization to proto for Java SDK to share proto :)
I see. I need to implement serialization to proto for Java SDK to share proto :)
Let me know, if you want to talk at some point
Let me know, if you want to talk at some point
Here are proto files, not sure I implemented serialization correctly, didn’t check it Workflow is 08_… (also attached text version)
Hi I could use help from anyone that’s deployed Flyte to GCP. I’ve been trying to use `helm` for this purpose based off of <https://github.com/flyteorg/flyte/blob/master/helm/values-gcp.yaml|this example>. I’ve overcome a few obstacles (which I’d like to document in a later post) with the default GCP helm example, and ...
Jeev B / Sören Brunk
Jeev B / Sören Brunk
Nicholas LoFaso add the /console path to the URL but port forwarding just console won’t suffice you need to use ingress since console makes calls to flyteadmin
Nicholas LoFaso add the /console path to the URL but port forwarding just console won’t suffice you need to use ingress since console makes calls to flyteadmin
using port forwarding to `localhost:8080/console` yielded I will try using the public ingress
using port forwarding to `localhost:8080/console` yielded I will try using the public ingress
right that’s because it can’t make any other calls see the chrome dev tools for the failing calls
right that’s because it can’t make any other calls see the chrome dev tools for the failing calls
Nicholas LoFaso Ingress options depend a bit on which ingress controller you want to use. Are you using the managed GKE ingress controller?
Nicholas LoFaso Ingress options depend a bit on which ingress controller you want to use. Are you using the managed GKE ingress controller?
Hi Sören Brunk yes that’s what we’re using
Hi Sören Brunk yes that’s what we’re using
Nicholas LoFaso: we use a small envoy proxy internally for our sandboxes that doesnt require ingress or ingress controller and can be port-forwarded to via `kubectl` . i can share the config if you want. Ketan Umare Sören Brunk: that might be worth adding anyway for just ease of getting started / troubleshooting?