Kubernetes SIG Testing, 13 Mar 2018

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Kubernetes Sig Testing 2018-03-13

Description

https://bit.ly/k8s-sig-testing-notes

A

Hello, everybody. This is Tuesday March 13th. This is the weekly suggesting meeting I'm, going to paste the link to the meeting notes into the chat. If you have anything that you want to put on the agenda, go ahead and suggest it there.

A

The first item: Tim, you bet the floor, talking about cross di jobs and bill dependencies.

B

So I don't quite understand the dependency graph between some of the tests, there's a cross builder, which publishes a bunch of artifacts into their separate GCS bucket and then some other tests leverage it that's part of what we're seeing on side of the cuvette DM release blocking jobs for the one nine to one 10 upgrade, but the it doesn't look like they've actually been working for a long time. So I don't exactly know who owns it or maintains it. But I do know that the problem that we are facing.

B

The reason why we have this weird one person shuffles a bit off to the other location that the other person depends upon. It's, because all the build artifacts aren't produced from a single canonical location, and so there's two parts to this question. One is who owns the cross, see I, build jobs, and the second one is: where is basil at four cross builds.

C

Owning the jobless kind of both release and test what, as far the dependency graph code other than debatable jobs, jobs normally pick up the CIO job, not the CI cross, build job and we published like a latest tag file which they consume its most jobs. I.

B

Again, I inherited this so I, don't exactly know, what's sane and what's not saying and why these.

C

Little cross building is not being run in CI locally. You can cross build the pure go images, I think now so binaries that don't depend on C go but C go while possible.

C

We need to inform it of the C tool chains. What.

B

What build art effects depend upon? C go home.

C

I'm not sure off the top of my head, I also I, think Jeff is looking into this. It's the get up, but I'm, not sure. If anyone else's and I don't know what the plan is, there I think we should probably be making depend on the CI build job.

C

It's clear how much effort that will be in where that.

B

Should happen, should we login issue somewhere, so at least we can track it. I don't know if it would go into test. Infra I probably seems like we're going to test in front there's.

C

An issue in testing for around the inconsistency with the cross drugs. It's not clear to me what the correct answer is. It looks like there's a different path for those depending on if it's a master release, job or not, which is interesting.

B

So how are you going to track resolution on this and does it we can test asynchronous of this right like we can? We can do this by hand to validate, and we will probably do that anyways just to be careful. But ideally this automation looks like it's been broken for a really long time and no one was paying attention until we got to release so I.

C

Think this sounds like a test him for issue to me and there is one for the crossbow, but.

C

It's not obvious what the build publishing should look like for the previous branches outside of the release bulb well,.

B

If we link the issues in the meeting notes with a set of issues, then I'll just follow up with them and loop in the sig release. Folks, because they keep on poking me and I'm poking back and I'm like I. Don't know who is this so there's.

C

Also, just a large mess around using Korean aids anywhere is actually taking in a couple of independent versions. So the moment the cube read them, jobs are actually using the CI, build the basil CI build and the CI cross build, which is it's.

B

Crazy yeah and then my plan in in the 1:11 cycle. It ain't talking with the cluster API focus long term. Google wants to get rid of all of the cluster directory yeah in getting rid of all of the cluster directory and adding cluster API into the mix and I would love to have folks tag-team in setting up blocking, build jobs that get rid of qu, better communities anywhere and use the cluster API and the 111 cycle.

C

Here is the cross.

A

Okay, cool so we'll take that discussion on to the issue next item on the agenda cool. You could just jump it hearing the state of the title up. It looks like we were kind of putting in the last pieces here and we might be close to rolling tide out for a KK as well.

D

Yeah so I since I'm, going on vacation on Friday I, don't know if I'm going to be able to run around or I, don't know if it's a good idea for me to roll it out before I go unless someone else is willing to own that, otherwise I can roll it out as soon as I get back.

D

We have the will that'll, give it some time to soak I guess so that could be a good thing to do just because we just rolled out the new status changes. That being said, they're, probably not a particularly volatile thing or something that could cause problems really, so we we could roll it out soon and.

A

So I guess the one question I had it and this kind of touches on the last point on the list as well. There was a conversation about how many tokens we were about to be using after codes law and then also I, know that, like what the status change implementation for tide there's a lot of conversations about whether or not that would use too many tokens like. Are we worried about running out of quota there yeah.

D

So I've actually been monitoring this, like just over the past couple hours because and just rolled out the the status updates, and it really should not be using that many tokens or using a few hundred tokens I guess the first time to set all the statuses. And if you look at the the API token usage graphs on Velodrome, you can see that there's a small spike, but it is it's not super significant.

D

When we do deploy for the main kubernetes repo, it will only set statuses for some of the PRS at first and then it will stop. So we have several rate limiting for that now. So we shouldn't run into problems. It'll just be it'll, take a couple hours essentially for all the statuses to be set for kubernetes, but we won't run out of tokens. So I think we should be pretty okay, like I, think everything will token wise, I think we'll be we're ready for kubernetes.

A

Are we actively feeling any pain from remaining on the budget until you get back from vacation.

D

A

Know but then you did a redeploy, a couple yeah we could go that was kind of frustrating.

C

Mangia github is always frustrating yeah.

D

So like when we do code I'm, assuming that we're going to thought to continue to use the submit queue and not go immediately to tide from that, so once that's done anyways there shouldn't be any change at all like the deployments won't be touched, I assume for a week after that.

A

Okay, so we're closed and I guess it just depends on if somebody must own a gun, I'm.

C

Gonna browning, it I'm not sure, if we're quite ready to make the switch over I think it might be a good idea to be linking to be ready to be linking to the PR Status page. Okay,.

A

They're not also a tracker issue which is way old in the meeting notes. There was a couple other things that I don't know. If we want to consider blockers for the actual implementation, there was a request to be able to ignore specific services that came in a couple weeks ago and then I can't quit all I. Think then there's another requester with the tight status has to be required and so I guess. If you want to also say hey, let's, let's make sure we can link the PR status.

A

Let's, let's try to put all that into the tracker, so we can at least keep tabs on them. I, don't.

C

Think we have to get the tied status to required to roll up for good days. I. Also, don't think we have to be able to ignore contexts for communities since we have like skin and weed pests and things, and we generally don't work for tests that are walking right. So I guess, if it's just going out.

A

To the repo and lt, or that should be fine, yeah.

D

It's not actually a important for the tied status to be required at all, there's no race condition or anything there. It's yeah it that's more like a UI thing, and it almost makes sense for not to be required. If people are going to manually merge, you know if there's gonna still be a few people, but I'm not sure mother doesn't make a big difference. Okay, those other points, those should be quick fixes like making or allowing for other context to be skipped. That's a that's an easy one to add.

A

A

The last thing on the agenda that I just put in here really quickly I'm just to do an update of the public utility work that I'm doing I updated the trekking issue to have a more like detailed breakdown, and so we're finished with the first phase up up streaming, all the utilities that we were using and then now I also refactor the majority of the jobs downstream and up at your origin to use the utilities from test and fry and so they're getting battle, tested and soaked and we're finding little bugs going there.

A

I am right now working on a pod spec that will integrate all these things with ena containers. The way that we expect them to we just removed.

A

We can now. We can now put that pod spec directly into the prod config and have Planck just use it verbatim, and so that will be the next step, and so once that's verified, we can start by treating about how the config should look.

C

Actually quick note there, son and I have been talking about the fact that Communities is proud. Config is still this yeah and it it's difficult to do things like make breaking changes to the config I'm hoping kind of in the next quarter. Here we'll put out a plan for some proposals around redesigning the config fit. There have been some other interesting ideas: okay, Joe kicked around the idea of inverting it and having repos.

B

C

Instead of implying them some things like that, possibly versioning the types- hopefully, we can make that part of rolling out the pod utilities. That was his original.

A

Or and I know, this was another thing that I think had been considered, but I'm not sure we've decided to go with it. If we wanted to break out the plank configuration and just be able to.

A

Silo, like a smaller plank with a more restrictive label selector but yeah that if this use original but.

C

We have a real pain point right now around like. If we change the config format, we have to learn every one and then we have to roll it out. So we're looking towards, like maybe using API versions like organize, does for the actual program components.

C

I'm going to write a dock after I finish, we have a lot of writing to do here at Google that cycle going on. Okay, okay,.

A

That's everything that we had on the agenda for this week. Do we well I, guess Ryan asked and check if type check is blocking, keep your daddy's yet I've.

C

Been waiting for lazy consensus and I'm putting on a pair today, it seems like lazy consensus. Let's do it. Okay,.

A

Cool, does anybody also anything they'll ever have.

A

Awesome 15 minutes early have a good weekend.