Kubernetes SIG Cluster Lifecycle, 6 Feb 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Kubernetes SIG Cluster Lifecycle 20190206 - Cluster API

Description

Meeting Notes: https://docs.google.com/document/d/1Ys-DOR5UsgbMEeciuG0HOgDQc8kZsaWIWJeKJ1-UfbY/edit#heading=h.1knsflv47tux

A

All right, hello and welcome to the Wednesday February 6th edition of the cluster API project office hours. If you have any topics that you want to bring up today, please go ahead and add them to the agenda. I will go ahead and put a copy of the link to that in the chat.

A

All right so start with I have a PR up for the initial work for adding a pivot face to cluster cuddle. It's still very much work in progress. There are broken tests right now and there's some additional testing. I need to do, but I think I've solved most of the issues with the current pivot workflows. So I would definitely appreciate some eyes on there to you know, get some initial feedback before I continue closing away at it.

A

For example, one of the some of the things that I've addressed in now migrates machines, machine objects and cluster objects for all namespaces. Instead of just a single namespace.

A

It also deletes objects from the source cluster after they've been successfully migrated over to the target cluster and to do that. I kind of follow the links to make sure that we're only deleting it after all, of the dependent objects are migrated over and I've also fixed, some ordering issues that I saw because the current pivot workflow it attempts to migrate machine deployments before it migrates machine sets before migrates machines.

A

So if you do have machine deployments or machine sets in there, there's a potential that, while those migrations are happening, that you're going to get duplicated objects created before the dependent objects are migrated over. So the pivot PR kind of fixes the ordering there to make sure the machines are migrated over before the Machine sets before the Machine deployments.

B

Just a quick question: how would that work in terms of older reps.

A

That's a good question that I haven't fully verified. We could actually probably clear out the owner reps and have them recreated because of the label links, but the other thing, too is right. Now, I am scaling down the stateful sets in the source cluster. One thing we could potentially do is scale down the staple sets in the source cluster and the target cluster, while we're doing the migration and then only re-enable are only scaling back up in the target cluster.

A

After all of the objects have been migrated over and then that would solve the ordering issues with the next okay.

B

I'll try to speak up yeahthere's if I recall there should be some adoption logic that you could rely on admissions that if you decided to go that way, yep.

A

But this is one of the reasons why I was seeking feedback to make sure that I'm not missing things like that for sure.

A

All right next item on the agenda is Vince, who has a PR out there for creating a more concrete, but still relatively soft link between machines in a cluster yeah.

C

So on the PR addresses the issue number 41, which asked for a stronger link between machine start to a cluster. Some folks, like I, think preferred the label approach with which is the one implement the PR I know. There was some discussion between labels versus imitations, I love, some more feedback.

C

The goal of the PR is to solve specifically the power that we're having right now of creating multiple clusters and in the main, in the same namespace, which is not allowed at this point, and also allowing still like some providers to not like specify any any labels and assume it. The cluster will be nil when they win the machine, I trade or run.

C

That's it for me.

D

So we have a quick question, so there was earlier a PR that was merged specifically to allow the machine controller to function like without the need of a cluster object. Being there I mean that was like I think was like about 20 days ago or so, and then just looking at your code change. One of the things that you have altered here is shown, for example, in the in the machine controller, a check where you see where it was only logging that the cluster not found now you've modified that to basically return error.

D

That was one of the precisely the thing that was actually changed in that previous chain, and that was actually the intent of that. So to me like this PR and what was most like 20 days ago, you kind of in two different opposite direction: I, don't know like maybe so yeah. We have some respect around that. Please so.

C

The logic actually moved in that cluster at the guy closer we'll look. If there is a label and if there isn't a watch, will return nil cluster and a new error which, like retain the same purpose as before yeah I, think the baby stays the same as well.

C

Does that make sense.

A

All right any other questions on that topic before we move on.

E

All right, next up, sorry I was just meant I, just wanted to say, like I, think Liam Booker, just fine I'm, not a big fan of how it's being implemented right. Now, though, because it is asking for a specific cluster label to exist, cost is equal to something like if we were to rely on labels. I would expect us to rely on something like collectors.

C

Yeah I think the first thing that I would like to kind of decide it's like if we want to like use labels or annotations, because you had some thoughts around that I'm fine using the label selector, that's like a much nicer approach to, but if we go down the annotation route, I would like to know if, like what everyone thinks about I.

E

Have I like label approach personally I, don't like the annotation approach, but if anyone else has any objection so.

C

There was one one comment about the label: I said: like users could potentially change the label. I think this particular issue could be solved later on with, like a validation, web book approach and I mean it's like. We can just open an issue and I could solve that problem. Later. That's fine and.

F

Also, the the key for the label if we go with a label approach, ideally, would be namespaced, so cluster api to Kate's, dot, io, /, something or other and users should hopefully understand that if they see a label that they don't really recognize or understand what it's for, they should leave it alone and if they modify it, whether it's accidental or on purpose, it's basically user error. At that point,.

B

I haven't looked too closely at this yet, but isn't my understanding that this is just going to be a single label that points to a cluster for it.

B

G

B

What Mason was trying to point out, then, is usually.

G

H

Should be a labels like.

B

The cluster object would have labels selector field, giving a label expressions that it would match against. If it's going to be a single label always would and a controller ref make more sense.

C

F

You're saying invert the the direction then so that you put the label selector on the cluster and you know if something matches great. If not it doesn't, but then.

B

Electors is usually have a controlling outfit with a selector of labels, the things it wants to control and the objects have the labels. But if we're going with a single label, that's always there then maybe consider an controller reference.

B

The machines back to the.

C

Could you put this thoughts in the PR that I can look at it? I'll fly down.

B

Okay, this is just a suggestion. You don't have to go with it. That's way if it still makes sense to do believe, I'm just throwing my two cents in no.

C

Of course I think I think would be. It would be great so like we can discuss offline yeah.

A

Moving on to the next topic, Pablo standard controller implementation in the repository.

I

Yes, my question is because in asking me or ask that to be in a couple of open office meetings- and there was none like a consensus of the state of this- a little ID code is there, but at the same time the worst some person here called me and was like creating a new machine controller was talking life foots crash implementing all the control a lawyer, so I wanted it to the repository I. Look, there's a lot of stuff there for me.

I

Make sense what I'm seeing but to be honest, I, don't know what the state of that I've been seen, some activity not quite recently, but recently so before.

I

Jumping into that, my question is I'm personally interested mostly, we started implementing a machine provider so to be more specific, the kind what she provided I'm uh she provided for testing environments that we discussing that would be kinda project and but I don't know if I had to start implementing the controller or most of the logic on the controller is already depository see if I can just take a look and try to figure out myself but Ebola. We appreciate is your opinion of the state of this implementation.

A

So I think a lot of this should be covered within the get book documentation as far as what you need to implement to get a provider going and but you know the kind of step-by-step how to get started and.

J

Wants to do it like more of a ingrained into cluster API proper itself. At least that was a conversation piece, but I, don't I, don't know. If timing is accurate, like you can you could absolutely take a look at this? No.

C

J

I don't know if the if we'd want to entertainment, refactoring work. Yet until we get past the D 1 alpha 1 phase, as you mentioned in the kind of column, I, think that you could probably do the work, and there are people here who could help guide that in a in a proper LZ but I, don't think it'll be landed, landed until after the actual release of okay.

I

But listen! It's like the code is already repository. We'd make reference to the book by the way. Is that actual this code is intended to be cheap, except involved for implementing the controllers, because what I understand is that the difference between the controller they basically keep track of the state of the API related options. Machine said machines and there are the actuator of his name given we should be want to actually go and, for instance, provision one specific machine. My understanding is that the controller part should be.

I

You read it, because it's acting from the API options and the provider or actuator to act in the specific instantiation of that, but and I like the way. This is split and uh tempore. My question in the is: if this is split, this kind of part of the proposal we spoke to the provider just implement everything. I mean there is any any because for me, make no sense. I mean if I want to command provider, for whatever bathroom, very much be able to provider. I would focus only on providing the Machine.

I

The the creating an undestructible, no stupid machine I will respect the generator. The Lord is reconciliation logic to be donated.

I

That's my understanding and that's what I understand that what I see the bull but now my question is: like oh cool I, take this code from the repository I start working on that, assuming that that is going to be I, don't know, maintain or some intention of part of the scope of this project to offer that kind of basic loyally, so that it provides only focus on the platform specific form because I remember that in another meeting, so one patient at our Factory temple. But they don't know what the factory temple is. The factoring.

I

What English, which.

K

I, don't think you're obligated to use the actuator I think everyone that is right as far as I know, everyone is is using it today.

K

I, don't know if there's anything implicit in the actuator that sort of defines any behavior that we haven't defined formally I I think there is a lot of behavior in the machine set controller, but that is generic anyway. I personally might be sense. I'd use the actuator, but there's also not a ton of it. So if you're finding the actuators a pain, then for let us know why, and but it's not a big deal to write your own controller, but I would write. I would use the actuator and, let's give a reason not to okay.

A

All right anything.

C

A

That before we move on.

A

Yeah, alright, next one is a general PSA out there for the AWS provider. We want to kind of reset our version, numbers and kind of throw away our previous releases.

A

We wanted to give folks here- and this is a larger former forum than our office hours meeting. So we want to kind of advertise it out there to make sure that that's not going to be a major problem for anybody's, been consuming VI AWS provider code today.

K

Well, do you want to give a little bit of background on why we're coming backwards.

A

Well, the main reason why we're going backwards is is because you know in we started.

C

A

The V 10.0 about alpha dot X numbering and we want to kind of go back to a more sane kind of pre, 1 dot, o version, because we are implementing many breaking changes over time. So until we get into a much more stable release, we'd like to be you know, semantic version was in a better place.

K

That's reasonable and explains a better thing for.

J

What it's worth I opened an issue: I don't want a bike shed here now, but if folks have thoughts on it with regards to how are we going to build inversion across providers once we actually hit a 1 alpha one and how we want to manage and track and maintain that that's a long conversation which we can probably start to bike on the issue and then maybe next week we can. We can discuss that.

L

So are we following like like a standard, versioning approach across all providers in case users want to use our users use multiple providers? That's.

J

Exactly the topic I want to bring up in that issue. Okay, how are we going to manage this? How are we going to version it? How are we going to build it? How we're gonna track it there's a lot of details, logistics there, you.

A

Moving on to the next topic, we have nighttime about integration test. Yes,.

H

I, basically just the one you fall out, this issue I think it's me bringing up pretty free before so I. Look at the code. The repo I since I cannot find any existing integration, testing, the main repo and basically to implement this kind of passed. It's it's required. Non-Trivial.

H

Working to implement the I traitor stop maybe, and so my question is, since we are doing our integration testing different providers, do we still wanted to add this for their for the main repo.

A

So, in my opinion, I think there is value in having integration tests in the main repo to ensure that, as new releases are cut and they're consumed by the downstream providers that we have some type of assurance that there aren't any major breakage there. I.

K

Wonder if we can write ain't, no op machine actually or that doesn't actually do anything so.

L

M

J

Kind of fits into that kind. Space word like I'd like to have a test gene in the main cluster API repository, eventually, where you could actually test the full lifecycle of the actuators. Yes,.

N

A

N

Which is having like a you know, no cloud provider or dependency actuator just for tests harnesses will be really useful for this sort of integration test just to sort of test logic, it's a little bit larger than a unit test, and then that could eventually turn into kind. If we wanted to.

M

It would also be useful for the get booked instead of asking people to create files. If we actually had a debilitation entry, they can just copy directory and have a an actuator.

H

Okay, I'll continue to look into this, so I assume you. This is still a ve one alpha one issue: we want to pursue.

A

I I think it would still be a v1 alpha one issue: unless there's anybody that objects there.

A

The next item is uh intro and an interface config and I, don't necessarily want to mispronounce your username Gary I've, no.

O

A

O

John bel Marek I'm at Google I, just getting involved here with Austria PI stuff I recognized a few of you from Seattle from from Elaine evening there, but anyway, I just wanted to introduce myself I'll be showing up from for a while now and one of the things I was going to look at first I noticed that the the interface config and particularly IPAM functionality, is currently embedded in the provider spec and I'd like to make a proposal to bubble that up, so that we can use something like a common mutating webpop to allocate IP addresses and things like that and every provider doesn't have to integrate with different ipad solutions.

O

A

Know the dear had previously done a little bit of work in that area as well. So you might have a little bit of thought input as well.

O

Great yeah, so I guess I'll start I'll create an issue and then maybe people can comment, I'll put my thoughts in it and then we can start to have.

A

A discussion I think that sounds great and welcome. Thank you.

A

Moving on Chuck relaying a message for Illya yeah.

P

He asked me just to share the the issue link that he created today. It's generally around a high-level discussion about cluster API focused anybody had thoughts regarding that in response to that or feedback there. He would appreciate that.

A

All right, so this probably ties a little bit into the proposal that Vince and Robby started putting together as well. So while we're probably while it doesn't make sense to focus on it for v1 alpha, one I think we definitely want to probably rally around finding some kind of consensus between Ilyas proposal and the proposal. Defense and Robbie are working on to try to get to a better place supposed to be one alpha. One.

P

Yep definitely opposed to view enough one.

A

So you know, if you get a chance, please do look at the issue, provide some feedback, but at some point after v1 alpha one where we'll probably have to have a proposal period and and start hashing out kind of a direction from the multiple proposals.

A

A

Next item, Vance, with the pull request, yeah.

N

I snuck this on here on Vince's behalf I'm.

C

Saying I'm gonna.

N

Write this I think this. This pull request is fine. I just wanted to point out that we're starting to modify some of the like that generated the ml files in a few places- and you know in this case it cleans up our code like this is a good change, but it's gonna make the repository a bit more fragile, because people are going to inadvertently start reverting some of these changes when they run make after they're generating the API types, so I've been suggested.

N

Adding you know, presubmit check as we've done for another one of these changes that we made but I think. Maybe this is some feedback. We should maybe try to take back to the q builder folks, which is you know? How would they suggest that we handle these sorts of changes right like we know that there's some tweaks and customizations? We need to make to the CRT definitions to make them correct and they keep getting clobbered right and I.

N

Think in general, Q builder tries to use a scaffolding approach and, in this case, they're overriding instead of respecting our scaffolding. So.

A

I think in this case, with the PR that you linked, that's actually adding the generated descriptions that didn't make it in in a previous PR that mattered that field.

A

But we did have an issue recently where a change was reverted after it was merged in because of that exact reason, though, because Q builder kind of regenerated it and moved.

N

Okay, maybe I misunderstood what Vince is comment up here. I thought that this won't do the same thing. If this is just regeneration and that's even even easier to approve so.

C

Does actually the provider the PIO was merged? I think last week didn't regenerate the seer IDs, which now, like the three pair they open like this week. I actually had this change, all three of them and people were getting confused like why we're changing something that, like it wasn't related to it. So I probably like remove this from the spirit and open a new PR and fix it in the CI, so that, like once here, the needs to be changed. It's caught in CI and my computer dog emerged.

C

The other issue was the we reverted back some manual CID changes, which I think we will probably need to open up an issue upstream with your builder yeah. Yes, that specific use case, it's not supported so yeah I feel about reverted and like we never fixed it.

A

Yeah and that was specifically related to queue builder, not supporting a validation annotation that the CRD spec actually allows for open API validations, yeah I.

C

I guess: look at this bringing up the question like should we have like changed this year do manually in any case, or should we bring the support to cook builder so that, like I mean we can regenerate, see release it's a broader question, but.

N

Yeah I think, ideally we would be able to do things upstream so that we're not fighting with the tooling but I. Think it's fine until the tooling is updated to put the workarounds in place that unblocked us.

C

Okay, maybe then we should use to Mai's to to do that. I guess, yeah, what a bunch of fashion script.

Q

Yeah I think maybe customizing some basil, goodness tea, but those changes back after we need controller Jen.

C

C

A

Any other discussion on that topic.

A

All right next item is from steven augustus. How do we know that we've reached me 1 alpha 1 and I will defer other to robert attempt go into that.

J

We're tracking against a milestone, our milestones were the repository and once copy reaches the milestone, we'll probably cut it, and then downstream providers can then track and match and then try to update so that released semantics of how we're planning to sort of coordinate that across different providers. This isn't open-ended question to deal with for for the actual release, so the TLDR is that there's a milestone. Okay,.

R

So the milestone seems squishy to me. So I wasn't sure if it was like, we had finitely defined all the things that were going into it or or what we've.

J

Done a good job of putting everything out that it's not important, it is somewhat squishy and the fact that, as we get closer and closer, we still find things, but it hasn't grown. It's actually been tracking downwards for the last couple of weeks. Okay, perfect thanks.

R

N

As Tim mentioned, this topic, I think, is a common one. People ask we mentioned. We discussed it I think two weeks ago in a meeting there's like a section of the notes. That literally says like how do we know when we're done with the Alpha, milestone and I? Think since then I think four, so issues have been closed.

N

I'd have like 24, which is which is pretty decent in terms of velocity, and there are a number of things in the milestone that we left in there that we said we'd be willing to kick out if we got towards the end. So yeah I agree it's a little bit squishy because we over trained on leaving things and to remind us that we might want to do them, instead of being particularly harsh, so we'll be as we get towards the the end date: sweet Thanks. We.

J

Have a soft date targeted for around the same time as 114 disease.

C

G

Minus 114, it should be in the last Oh actual date against nausea, yeah.

R

End of March.

A

All right- and that brings us to the end of our agenda- are there any other topics? Questions concerns that anybody has.

R

Are there any so I'm for context, I'm working on essentially refactoring the cluster API, Fraser or Pepsi I'm affectionately, calling it now I'm wondering if there are any things that I anything that I'm pretty much modeling a lot of the stuff that I'm doing off of the AWS work. So I'm curious. If there are any things that I need to be watching out for that are coming up.

A

So I think the one thing would be the PR that Vince has out there for changing how we're linking clusters and machines that could potentially affect providers.

A

But outside of that I'm not currently aware of anything and and we're specifically booting kind of major breaking changes outside of v1 alpha one yeah.

N

I would say if you're looking at, like the longer view the documents and I started, it would be great to get your feedback on that because that's gonna be a larger sort of reshuffling of the API. Sorry, where is that? Is that linked in the notes? Sorry, if I'm missing it, it's definitely linked from last week, and it is linked from the bottom of the issue that was referenced earlier, that Ilya opened cool as well.

A

Any other topics.

A

All right, thank you. Thank you, everyone for joining and have a good rest of your day.

A

Q