=== 21WAAD3MF is now known as wallyworld [00:52] anyone knows his way around state/open.go ? [01:24] perrito666: it depends on what you want to know, i might be able to help [01:29] cmars: hah gues you figured i based my plugin off yours :) [01:29] axw: mornin'. can we have a hangout now instead of in an hour? [01:29] was gonna give you credit once i got something working [01:29] morning wallyworld. sure thing, just give me a moment [01:31] wallyworld: erm, my sound isn't working. gotta fix that first... [01:31] ok [01:36] wallyworld: tx, sadly my head is falling on the kb so I better hit the bed before I introduce a bug instead of fixing the current one [01:37] perrito666: np, i'm on call anyway now. if you have a question, feel free to email to the list or ask again later [01:46] wallyworld: ok, I am more curious about fixing this bug than about going to sleep :p so her I go [01:47] I am trying to fix the restore functionality [01:48] now, at some point the restore calls state.Open(), I tried to replace it by using juju.NetConn and NewConnFromName and in all cases, it timeouts at mgo.DialWithInfo thile trying to make Ping() [01:49] perrito666: ok. there may also be someone else looking into that from juju-core [01:49] "that" being? [01:50] i think Horacio Durán [01:51] sadly that would be me [01:51] he's started to fix some of the backup bugs and was also going to look at restore [01:51] oh [01:51] hi [01:51] hi [01:51] i didn't realise! [01:52] perrito666: give me a couple of minutes to finish this call [01:52] sure [02:00] perrito666: sorry, back now [02:00] i'm not across the restore stuff specifically [02:01] wallyworld: I think the restore part of my explanation can be safely ignored [02:01] wallyworld: gotta go to the shops for a little while, bbs [02:01] I just provided it for context [02:01] axw: sure, np [02:02] perrito666: so you are looking to, in general, replace calls to state.Open() with juju.NewConn ? [02:02] to use the api [02:03] so you definitely have a state server running? [02:03] api server even [02:05] wallyworld: well I am pretty sure I do, I try to query mongo by hand and it responds, yet when juju tries to dial it just timeouts [02:05] mongo != api server though [02:06] the api server listens on port 17070 [02:06] true, altough I am pretty sure this breaks before getting to state [02:06] what code are you changing? [02:07] well, current existing code calls open, open in time calls DialWithInfo [02:07] which file? [02:08] DialWithInfo creates a session [02:08] ah sorry [02:08] state/open.go [02:08] sure, but the caller to that [02:08] which caller of state.Open() is being replaced? [02:08] cmd/plugins/juju-restore/restore.go [02:09] around :187 [02:09] so at the time restore runs, is there a bootstrap node running? [02:09] i don't think there is [02:10] ah there may be [02:10] cause looks like it calls rebootstrap() [02:10] there is [02:10] but you might find that it is just that the api server has not started yet [02:11] cause it can take a while to spin up the bootstrap node and then start the services [02:12] maybe to see if that's the issue, pause the restore script or add in a big attempt loop to see if it just needs more time [02:12] wallyworld: mm I tried looping on that [02:12] I waited 30 mins total [02:12] that is a lot [02:12] can you do a juju status when it fails? [02:12] ie does juju status work? [02:12] that would need an api server connection [02:13] mm, it does not [02:13] so if juju status is broken also, then there's an issue with the bootstrap node [02:13] you would need to ssh in and look at the log file [02:14] cause it could be the node itself starts but then the juju services fail to start [02:15] mm, the service seems to be running, I even restarted it by hand [02:15] in what port should the state server be listening? [02:15] 37017 [02:15] 17070 [02:15] 37017 is ongo [02:15] mongo [02:16] perrito666: when you say you restarted the state service by hand, that doesn't make sense to me because the state service runs inside the machine agent - did you start jujud? [02:17] wallyworld: yes [02:17] and the machine log file is good? [02:18] and yet juju status fails also [02:18] there's gotta be something logged which shows the problem [02:18] until something like juju status is happy, then the code changes to restore.go won't work either [02:19] wallyworld: interesting though, restore is trying to open a state server n 37017 [02:20] the current restore using state.open()? [02:20] it will because it connects straight to mongo [02:20] the new juju.NewConn() methods instead go via the api server on port 17070 [02:21] aghh, juju.NewConn fails just as Open, so someting is definitely broken in my recently restored node [02:22] wallyworld: is that in trunk yet? [02:22] my logs show NewConnFromName accessing mongo directly on 37017 [02:22] stokachu: the api server stuff? [02:22] yea [02:22] yes, been there since 1.16 [02:22] used universally since 1.18 [02:23] perrito666: i'd be surprised and sad if the log files on that node didn't show what was wrong [02:24] * perrito666 run the extremely tedious setup script [02:25] perrito666: it will still be waiting for you tomorrow after you get some sleep :-) [02:26] wallyworld: certainly but now its personal [02:26] lol [02:26] feel free to pastebin logs files if you want some more eyes [02:26] * perrito666 paints canonical logo on his face and yells mel gibson style [02:26] woot i actually a juju plugin to do something in go [02:28] stokachu: I sense a verb missing there :p [02:29] would have been funnier if you said "i a missing verb there" :-) [02:29] hah [02:29] back... [02:29] to much time looking at juju core code [02:30] wallyworld: axw: I'm here for standup [02:30] wallyworld: my wife is watching tv in spanish next to me, when 2 lang module enabled in my head I loose capacity for witty sentences in both languages [02:30] waigani: huh? i thought you were on holidays so we had it early :-) [02:30] waigani: we already had it early, weren't expecting you [02:30] but we can have another [02:30] :( [02:31] I'm in auk airport [02:31] okay, maybe I can talk through what I'm doing? [02:31] waigani: sure, i'm in the hangout [02:31] brt [02:41] wallyworld: https://pastebin.canonical.com/108967/ [02:44] perrito666: looking, sorry was otp [02:44] on the same note https://pastebin.canonical.com/108968/ [02:47] perrito666: is there any more in machine-0.log? [02:52] wallyworld: well, there is before that altough I am not sure if I can distinguish between pre/post restore (restore is a particularly ugly thing) [02:53] perrito666: what i mean is, after the output you logged. that log looks ok i think. there was one timeout with the api client connecting but thatcan happen and it appeared to be ok after that but i wanted to be sure by looking at subsequent logging [02:55] nope, after that it just loops with https://pastebin.canonical.com/108969/ [02:56] hmmm, ok. so that says there is an issue with the api server [02:57] you may need to enable trace level logging and/or add extra logging to see why it's failing. i wonder if netstat shows the port as open [02:57] tcp 0 1 10.140.171.13:59925 10.150.60.153:17070 SYN_SENT 4001/jujud [02:58] that's a different ip address to what is being dialled [02:58] oh no [02:58] it's not [02:59] nope, just without the dns nae [02:59] yeah [02:59] if it were me, i'd have to add lots of extra debug logging at this point to see what's happening as i'm out of ideas [03:00] but you can see even internally the machine agent api client can't start [03:00] so there's a core issue with starting the api server itself [03:01] axw: local provider is sorta ok. it doesn't like starting precise containers on trusty although it used to. and if i start a precise container first and it fails, subsequent trusty containers also fail, but starting a trusty container first works [03:01] wallyworld: well, I think the restore step is actually breaking the state api server [03:01] since it works right before [03:01] likely [03:01] (restore bootstraps a machine and then untars the backup on top of it) [03:01] roger wrote all that so i have no insight off the top of my head as to what might be wrong [03:02] wallyworld: ah ok. there have been a few bugs flying around about host vs. container series mismatch not working [03:03] axw: yeah, i'm going to try explicitly setting default series to see if i can get precise to work. but precise failing should not also then kill trusty :-( [03:04] wallyworld: I think there might be something wrong with the backup, tomorrow I will strip one into pieces and see what is wrong, as for me I am now officially out or tomorrow I will be sleeping on the kn at the standup [03:04] kb* [03:04] np, good night :-) [03:04] wallyworld: oh I didn't see that bit... weird [03:04] yeah [03:04] wallyworld: I think you can also bootstrap --series=trusty,precise to get it to work [03:05] not sure why trying precise would fail trusty tho [03:05] ta, will try that also to try and get a handle on it [03:05] * wallyworld -> food === wallyworld_ is now known as wallyworld [03:43] wallyworld: I just pasted the output I see from destroy-environment with manual [03:43] wallyworld: it's as I expected [03:43] axw: i missed it as my laptop got disconnected [03:43] wallyworld: I mean I pasted it in the bug [03:43] ah, looking [03:43] #1306357 [03:43] <_mup_> Bug #1306357: destroy environment fails for manual provider

[03:45] axw: clearly then i need to get my eyes tested as i had thought i included it all, sorry :-( [03:45] although i wish the last error was first [03:45] wallyworld: nps. it does kinda get lost down there... [03:45] as it would read much nicer that way [03:46] ie root cause, followed by option to fix === vladk|offline is now known as vladk [04:06] wallyworld: I'm going to look at fixing these openstack tests. If you do have any spare time, it would still be useful if you could review the placement CL [04:06] but if you're busy then that's okay [04:07] axw: funny you should mention that - just finished another review and am looking right now [04:07] wallyworld: cool :) [04:17] axw: this is a personal view, but i tend to think that if a method returning a (value, error) returns a err != nil, then the value should be considered invalid. so this bit irks me: [04:17] if c.Placement != nil && err == instance.ErrPlacementScopeMissing { [04:17] i would use an out of band signal like a bool or something [04:17] wallyworld: err was originally nil, that was something william wanted [04:18] I suppose I could change it to reutrn a nil placement, and have the caller construct one [04:18] hmmm. is there value in adding a bool to the return values [04:18] or something [04:19] I don't really think so, then you may as well just check if the scope has a non-empty scope [04:19] i sorta think that err != nil meaning the value is bad is kinda idiomatic Go [04:20] yeah... probably should have just left it as it was [04:20] change it since he isn't here :-) [04:22] wallyworld: I think I will just change it to return a nil Placement, and hten the caller will create a Placement with empty scope and the input string as the directive field [04:22] ok [04:22] i think that sounds good [04:22] the caller needs to know the rule anyway, at least this way it's the usual case of nil value iff error [04:22] sorta best of both worlds [04:23] ta [04:28] axw: with these lines in addmachine [04:29] if params.IsCodeNotImplemented(err) { [04:29] [04:29] 135 if c.Placement != nil { [04:29] is there any point trying again if c.Placement is nil? [04:29] should it just be a single if ... && ... ? [04:29] wallyworld: yes we should try again, because we're calling a new API method [04:30] wallyworld: client.AddMachines now calls a new API method by default [04:30] wallyworld: and client.AddMachines1dot18 calls the old one [04:30] oh,right. hadn't go to that bit yet, i recalled it was the same api from earlier review [04:30] it was, I fixed it :) [04:30] but i guess versioning [04:30] wish we had it [04:30] indeed [04:32] do i have to invoke "scp" with the ssh.Copy function in utils/ssh? [04:34] stokachu: the openssh client impl will delegate to scp, if that's what you're asking [04:34] https://github.com/battlemidget/juju-sos/blob/master/main.go#L89-L94 [04:34] so im trying to replicate juju scp within my plugin [04:35] this is my log output : http://paste.ubuntu.com/7312090/ [04:35] i think my actual copyStr is incorrect as i was following was is required by juju scp [04:35] * axw looks [04:35] what is* [04:36] stokachu: I think you want the target and source in separate args [04:36] im a newb with golang as well so if i got stupid stuff in there [04:37] lemme try that [04:37] stokachu: i.e. a length-2 slice [04:37] ok lemme see if i can make that happen [04:38] axw: is there a reason why we store placement as a string and not a parsed object. and hence precheck take s a string and not a parsed struct etc. i would normally look to parse on the way in and then pass around the parsed struct etc so we fail as close to the system boundary as possible. am i missing a design decision? [04:39] sweet, gotten farther http://paste.ubuntu.com/7312102/ [04:39] wallyworld: originally I did that, william wanted it changed. it should not get to the environment if the scope doesn't match [04:39] though maybe i should be using the instance.SelectPublicAddress of machine? [04:39] axw: hmmmm. ok. i disagree with william here then :-( [04:40] stokachu: cool. ahh, "juju scp" does the magic of converting machine IDs to addresses [04:40] wallyworld: why? the environment should not need the scope [04:40] ive got a execssh that i borrowed from someone that uses instance.selectpublicaddress [04:40] going ot try that [04:41] axw: what i mean is that the string should be parsed into whatever internal representation makes sense at the system boundary ie a struct of some sort, possibly different to what is used on the client ie minus the scope [04:41] stokachu: see juju-core/cmd/juju/scp.go, hostFromTarget -- that's where it maps machine IDs to addresses [04:41] and internal apis should then use that typed struct [04:42] axw: ahh i see that now [04:42] not an "untyped" string [04:42] but, doesn't matter, it's already been changed to get approval [04:42] to bad expandArgs isnt public [04:42] wallyworld: the directive string is free-form, so how are you going to do that? [04:43] wallyworld: it's up to the provider to decide what makes sense in directives [04:43] axw: ah bollocks, i was thinking there was more to it than just a string. but you are saying that by the time it's stored, it represents a mass name or whatever [04:44] that makes more sense. i hadn't fully re-groked the implementation [04:44] wallyworld: as far as the infrastructure is concerned, it's an opaque blob of bytes. the provider will interpret it. provider/maas will interpret it as maas-name to start with [04:45] ok [04:45] we may converge on some convention, like thing=value [04:45] az=uswest-1 or whatever [04:46] stokachu: it's also worth noting that some providers (e.g. azure) require proxying through machine 0 [04:46] stokachu: so you may want to just shell out to "juju scp" if you can... [04:47] axw: ah good point [04:47] cleaner than what im doing [04:47] is there a shell function in juju-core thats exposed? [04:47] or should i just use os.Exec [04:48] stokachu: os/exec is as good as anything [04:48] axw: good deal [04:48] ill do that instead [04:48] there are some utils in juju, but I don't think they'd be useful [04:48] cool no worries [04:48] axw: yeah, i'm a fan of a little more structure. but none the less, land that f*cker [04:49] hazmat: fwiw the first line that api-endpoints returns is the one that we last connected to, so if you just do "head -n1" you can get the same output we used to give [04:50] wallyworld: thanks [04:50] np. sorry if i went over old ground [04:50] nope, that's cool [04:51] jam: i was going to get your opinion on that bug - i'd like to close now as "invalid" or whatever given the other ifx has landed [04:51] wallyworld: sorry, which bug? [04:51] jam: the one you just remarked on above [04:51] bug 1311227 [04:51] <_mup_> Bug #1311227: juju api-endpoints cli regression on trunk/1.19

[04:52] wallyworld: localhost shouldn't be in the output [04:52] and I would be fine pruning ipv6 by default [04:53] jam: it can be for local provider since localhost is the public address for local provider [04:53] jam: martin's branch does prune ip6 by default [04:53] wallyworld: sure, I'm not saying don't print localhost when that's the address, but *don't* print localhost for ec2 [04:54] we shouldn't have localhost for ec2, but we would have 127.0.0.1 and that'll get pruned [04:54] jam: martin's branch probably ensures that's the case, since for ec2 localhost is machinelocal isn't it? [04:54] wallyworld: hmmm... I don't know that Martin's patch is *quite* right. I'd rather still cache IPv6, but just not display them on api-endpoints [04:54] we don't use any scope heuristics for hostnames [04:54] wallyworld: right, I think his patch is what we want, and we do want to be caching the network scope data instead of just addrs [04:54] jam: it's ok for now i think since we don't need/use ip6 yet [04:55] jam: so, i think then that kapil's bug has 2 bits 1. the ip6/127.0.0.1 stuff which martin's bug fixes, and 2. the multiple api address thing which is new and intended [04:56] so therefore we can mark the bug as invalid [04:56] right ? [04:57] wallyworld: so I still think there are bits that we can evolve on api-endpoints. Namely, to change what we cache from just addrs to being the full HostPort content (which includes network scope), and then api-endpoints can grow flags to do --network-scope=public [04:57] wallyworld: so while I think we've addressed the regression today [04:57] I don't think the bug is "just closed" [04:57] sure, but that's not the bug as described [04:58] we can get it off 1.19.1 at least [04:58] wallyworld: right, i think the *regression* portion is stuff that we intend (multiple addresses, even per server), because we think they might be routable [04:58] and we don't save enough information (yet) to be able to provide --network-scope [04:58] yep, i don't see any regression at all [04:58] (and then default it to public) [04:59] wallyworld: giving private addresses in api-endpoints by default is wrong [04:59] but "good enough" for now. [04:59] And hazmat has a point about actually grouping the data by server, so you have a feeling for what machine is a fallback [04:59] ok, so let's retarget off 1.19.1 then [05:00] SGTM [05:00] jam: 2.0 or 1.20? [05:00] 2.0 i guess? [05:03] I'd be ok with 2.0 [05:04] axw: when I use restore with patchValue I get this error: http://pastebin.ubuntu.com/7312196/ [05:05] so heres my latest change using juju scp https://github.com/battlemidget/juju-sos/blob/master/main.go#L89-L96 [05:05] and the error output http://paste.ubuntu.com/7312200/ [05:05] i verified that juju ssh 1 and /tmp/sosreport*xz exists on the machine [05:07] anyway, I need to go catch a plane [05:08] waigani: sorry, need more context. show me in vegas :) [05:09] axw: -r doesn't work with machine num it seems [05:09] juju scp 1:/tmp/test . works [05:09] but juju scp -r 1:/tmp/test* . fails [05:09] stokachu: you need to separate the command out into individual args [05:09] stokachu: i.e. "juju", "scp", ... [05:09] this is manually running the command from the shell [05:10] stokachu: there are some limitations with juju scp, I forget exactly how to pass extra args... lemme see [05:10] http://paste.ubuntu.com/7312211/ [05:10] thats what ive tested manually [05:13] stokachu: stick "--" before -r [05:14] axw: you da man [05:14] axw: is that juju 1.16? as 1.18 is a bit broken wrt scp [05:14] stokachu: in 1.18 (for a while until it gets fixed) args for just scp must come at the end and be grouped [05:15] jam: well I'm on trunk... I forget which versions do what wrt scp [05:15] so: juju scp 1:foo 2:bar "-r -o SSH SpecialSauc" [05:15] jam: what I just described does work on trunk, so presumably on 1.18 too? [05:15] ah [05:15] jam: i.e. I just tested "juju scp -- -r 0:/tmp/foo /tmp/bar" [05:16] axw: https://bugs.launchpad.net/juju-core/+bug/1306208 was fixed in 1.18.1 I guess [05:16] <_mup_> Bug #1306208: juju scp no longer allows multiple extra arguments to pass throug

[05:16] axw: trunk just lets you pass everything, and you shouldn't need "--" I thought [05:16] you do need --, otherwise juju tries to interpret the args [05:17] axw: fairy nuff [05:20] yea i had to use -- with 1.18.1-trusty [05:20] axw: that worked :D:D [05:21] stokachu: cool :) [05:29] jam: morning [05:30] morning vladk, its early for you, isn't it ? [05:30] well, early for you to be on IRC :) [05:42] good mornings [05:50] fwereade: morning :) [05:53] morning fwereade, we've missed you [05:54] waigani, jam: it's nice to be back :) [05:54] heh, easter holiday? [05:55] brb [05:57] hey fwereade [05:58] fwereade: I was about to approve https://codereview.appspot.com/85040046 (placement directives) - do you want another look first? [05:58] axw, I'll cast a quick eye over it :) [05:58] okey dokey [06:01] axw, ok, based on a quick read of your responses I think I'm fine -- my only question is exactly what happens with the internal API change as we upgrade [06:02] fwereade: the provisioner will be unhappy until it has upgraded [06:02] axw, I *think* that it's fine, given that the environment provisioner only runs on the leader state server, and therefore the upgrade happens in lockstep [06:02] axw, but other provisioners? [06:02] axw, hm, I have a little bit of a concern about error messages during upgrade [06:02] fwereade: it will be the same for the container provisioners, I think [06:02] back [06:02] * axw checks [06:02] axw, *we* might know they're fine [06:03] axw, but people who read our logs don't get quite such a sunny prospect of our general competence [06:03] axw: so we talked about having EnsureAvailability with a value of say 0 just preserve the existing desired num of servers [06:03] AFAICT, we never *record* the desired number of servers [06:03] we just have a number of things that are running. [06:04] jam: it's implied by what's in stateServerInfo [06:04] and we have stuff like WantsVote() but I can't see anywhere that sets NoVote=true to indicate that we no longer want to be votiing. [06:04] jam: len(VotingStateMachineIds) [06:04] jam: that's done in EnsureAvailability, in state/addmachine.go [06:04] axw: sure, but isn't that the actual ones that are voting? I guess it would be an availability check? [06:04] axw, this must ofc be balanced against the hassle of maintaining the multiple code paths [06:05] jam: VotingMachineIds is really the ones that *want* to vote [06:05] fwereade: just checking still, sorry [06:05] axw, np [06:06] axw, what I did with the unit agent the other day was just to leave it blocking until the state server it's connected to *does* understand the message, and then continue as usual [06:06] fwereade: yeah, this is common to all provisioners - it will cause an error on upgrade for container provisioners [06:06] hmm ok [06:06] I'll take a look at that code [06:06] fwereade: worker/uniter? [06:06] axw, it's not the best code in the world but it seemed to work [06:06] just a sec yeah somewhere there [06:07] fwereade: got it I think [06:07] logger.Infof("waiting for state server to be upgraded") [06:07] yeah okay, I can add that in [06:07] axw, cool [06:08] * axw senses another need for API versioning imminently [06:08] although I suppose we can just see that fields are zero values... [06:09] fwereade: yuck, this means threading the tomb all the way through... oh well. [06:09] I suppose it's for the best [06:09] * fwereade glances pointedly at jam re API versioning [06:09] * jam ducks and pretends to catch a plane [06:10] * fwereade does understand [06:10] fwereade: I made sure it was in the topics list [06:10] jam, great, thanks :) [06:10] jam: sorry, back to ensure-ha: if you just send 0 or -1 to state.EnsureAvailability, then it can load st.StateServerInfo() and set numStateServers=len(VotingMachineIds) [06:12] axw: I'm going to use 0, because it isn't otherwise valid, and we don't have to woryr about negative numbers. [06:12] sounds good [06:12] axw: I was thinking to do that originaly, but trying to verify the actual meaning of the various values was ... tricky [06:12] oh I don't have to thread the tomb, hooray [06:13] jam: it's not super clear, I agree [06:13] axw: I was reading through the code and trying to figure out what the actual invariants are [06:13] axw: I was really surprised that ensureAvailabilityIntentions doesn't take into account the new request [06:13] so we end up with 2 passes at it [06:14] also, the WantsVote vs HasVote split is confusing. Probably necessary, but very confusing [06:14] jam: yeah, we need to know what the existing ones want to do [06:15] jam: we certainly could do with some developer docs on this [06:15] I don't understand what the peergrouper does, haven't looked at it at all [06:16] I know what EnsureAvailability does, but it's easy to forget :) [06:16] axw: one advantage of "-1" is that it is odd :) [06:17] heh [06:17] axw: I took out the <= 0 and it still failed, and had to remember 0 is even [06:20] axw: non-negative or nonnegative ? [06:21] our error message currently says >0 [06:21] and "greater than or equal to 0" is long [06:21] jam: non-negative looks good to me [06:21] though non-math people won't get non-negative, I guess [06:21] really? [06:21] number of state servers must be odd >= 0 [06:21] number of state servers must be odd and >= 0 [06:21] ? [06:22] will non-math people understand >= ? ;) sure, I guess so [06:22] axw: non-engineering/scientists sort of people don't distinguish "positive" from "nonnegative" [06:23]