Kubernetes SIG Node, 9 May 2023

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: SIG Node Sidecar WG 2023-05-09

Description

Meeting notes and agenda: https://docs.google.com/document/d/1E1guvFJ5KBQIGcjCrQqFywU9_cBQHRtHvjuqcVbCXvU/edit#heading=h.m8xoiv5t6qma

GMT20230509-160346_Recording_1582x1120.mp4

A

This meeting is being.

B

C

um Hello, hello, it's um May, 9th 2023. It's a signal outside car working group. Welcome everybody um on agenda today. I think uh first item we want to discuss is this um new discoveries that uh we went through PR review, assuming that new Fields will not be accepted by old kublets, but it turns out it wasn't the correct assumption. So old coupon doesn't care about any new fields and basically you can accept like random Json as long as it has some like a little bit of resemblance to spot.

C

um So I, don't know like I mean if you change PR. This PR will read this uh like. If you have a new like port with this, really started policy. Always then this container will stack on initialization and will never go past initialization stage and I think I mean I'm debasing to myself.

C

um Is it normal behavior? Do we want to just let it be or we need to change something and what we can change if there is any opinions.

A

uh I I just think that uh disabling the feature date, while you learning the part, should not be a load, so I think we have to drain all parts before disabling the feature date so so that we don't have to worry about those cases.

A

It it should be a problem if, if we just if we just make those part filled.

A

So I think we have to remove those parts anyway,.

C

We still have a problem right like even if you drain the ports and then disable the feature, gate and restart the cupboard. The spot will be scheduled again and now it will. It will have this flag, but Kubota will not react on this flag, but it will still accept this uh portrait.

A

Oh no! No, if we disable the feature, gate and and.

A

The new part will be created without.

B

A

Fields, so you don't have to worry about that I think because cubic server.

C

Separate feature game yes, but it has separate feature Gates, so it helps. It may have feature gate enabled like well cool but doesn't have it enabled.

A

um uh Yes, that could happen. We.

C

Can add the functionality rejected of that like.

D

If you're doing like a so cubelet supports like two nodes or two versions back, so you upgrade your API server and you start creating this and now you've got this old cube with it. Doesn't you know, existed before this cap was even written right and you just end up with a knit container that runs forever never finishes, and the pods hung essentially I, don't know what you can do about that. Really, though, it's.

C

Yeah I mean what I didn't test is uh if you change some existing field like and add the new properties that didn't exist before, like restart policy, uh a sidecar like with a new value side car for the entire port. Will it will accept it or not?.

D

Coming on this issue, I thought part of the question: was you have the option um with 128 to do something different if the feature case disabled, like rejecting the admission of the Pod at cubelet um and with the older versions, it could be a lot of you don't have that option, because the feature a doesn't exist really, and so would you want to on a 12a cube? Let's say: hey.

D

This feature gets disabled I'm, going to reject this pod entirely and then, if you're, using an older, Cuba, that'll just get accepted and hang and initializing forever see it's kind of weird that the behavior differs. But it seems better that Cube will just reject the pot entirely. I, don't know.

D

I, don't think that there has to be like some prior uh some precedence for things like this I'm, not familiar with that.

D

Is the future able to pod really no really no longer works on that version?.

C

Okay, yeah um I think many fields that I looked before they kind of benign four reports like um the agent new functionality. That is not necessarily required, so theoretically pot will still run if you would like some configuration setting or something.

C

um In our case we will probably break the port because it will not start any longer. Yeah um I.

A

Mean maybe you.

C

Can get away with some detection mechanism saying like this is how you detect the sports and document it somehow foreign.

A

I already made that in your your branch.

C

Is it what all.

A

Right, could you see this PR.

A

I just made a fix to just reject the part, with the the part having container with the restart policy and just make that part failed.

A

Okay and I think this is what we we said in tab in our chat before.

C

Yeah, that's what you wanted and I thought it's a default Behavior.

A

Yes, I think that this is a valid option.

C

Yeah I think great um I think we need to go with that, but for older kublets. What do you write in PR? Like? Do we just say um you're out of luck, here's a way to detect the situations and if you detect the situation, um just fix it and don't use a spot.

C

B

C

B

It is only an alpha feature right, so it is. There is an expectation that you are not running production workloads on it and that you, um you read, you read the fine print on what happens. If you decide to turn it off again,.

C

Okay, so you think it may.

B

Be okay, I think for Alpha right, like once once we get to Beta And GA promotions, but then we'll also have a history of kublets that actually support the functionality. I mean it's it's hard, because it's an alpha.

E

Yeah, if we say we need this as beta know, that it was discussed as coupon, that uh we would like to have like three versions of skew. Instead of two, we will have to wait like one year before moving to Beta what.

E

Like to make sure that even the all the three plus one uh three minus one, release of the cubelet will have this check enabled. But if, if it has to be like this.

B

Yeah, it's a really good point.

C

Yeah I don't see any any.

D

Better Solutions.

E

Let's do it, okay,.

C

Yeah I think the voice PR reviewer for us and uh he's still count replied. So maybe um he has a chance. You can share some thoughts but uh yeah. Hopefully he just he can just approve it. Okay, let's go to the next touch. End item get pause.

A

Yeah I I I just want to discuss about the way to display the part with the size color container. So for the second page, uh let's say there is a part with one unique container and three regular containers.

A

Then two two Cube control get pause, just display the state for the waiting, State waiting phase. It just shows like this. There are one in a container to initialize and it is just okay and for running phase.

A

There are three regular containers to run, so it is just okay and the next page, with cycle container feature for the current implementation.

A

We have to make those part we have to convert that part to have one in a container and two cycle containers. So there will be three any containers, including two sidecar, and it will show the initialization phase correctly. I think there are no problem here, but for running phase we don't have to monitor outside clear containers are running so I think that we have to include the site per container and regular containers to the second column.

A

You can see the next page.

A

Yes, yes, like this, so I think this is the way we should do if there are no objection.

D

C

Seems like yeah sounds good.

C

So what is this I will send that filtering happens somewhere and we don't show any containers any longer or or just proposing to show restarted policy.

A

C

I didn't get what what is changing. Oh.

A

The the change is that we have to the current implementation just shows the regular containers in learning phase. So that is the reason there is just one container in running phase, but I think that we have to display one regular container and two cycle container in running phase.

C

You say something: I didn't get it.

B

Oh sorry, yeah I was just catching up with yeah. It's highlighted in red. It's the counts of three three of ready.

C

Yeah I think it's fine I, just don't know where this filtering happens. Is it on site or somewhere else.

B

It should be in the print the server side, printer code right, you.

A

D

C

It sounds very good for me: I I'm glad you found this.

A

B

A

Then my question is: do we have to? Is it? Is it okay to include this in the main PR or we have to make the scene follow up here?.

B

I think it's fine in the main, PR I just added a comment earlier, where, even if you disable the sidecar feature, you might still want to report it this way, because that's kind of the reality of what's Happening until something changes, I think it belongs.

C

No I think the question Joe is is not whether we wanted into release, but because we want to complicate this PRN when.

B

ah Yeah, oh, no, that's a good question! um No I! Don't I! Don't feel strongly about that follow-up. Pr would be fine with me.

B

It's important.

C

Yeah I mean if we can mention that this website here is coming in XPR. If approvers will be uh asking for that, we can include it sure.

C

Okay with that yeah, that sounds great.

C

uh So uh split NPR: do you want to discuss how to split PR? We discussed it last meeting and we didn't find a very good way to split.

E

It no no, no, it I think he doesn't want to split the pr he wants to group group. The commits yeah split.

C

E

Split because now we have like I, don't know like 20 commits in.

C

E

B

C

He probably just two.

E

Two commits or something.

B

Yeah is that you, you do the handout you. You have one commit for the hand, edits that add the API changes and then you have another commit for all the generated code for all the API changes, and then you have another one for implementation or more than one of the implementation. If you need so it's probably a minimum of three.

B

um That's really nice, because the code, the generated code should be like basically in its own, commit.

E

Now, first is API change.

A

But do we have to make make commit for generating code or.

B

Using the way I do it is I'll I'll get all the changes into a workspace and then I'll just get add all of the API change files get commit that git, add all of the generated code files get commit that and then get at everything else and get commit. So I, usually basically just do a reset soft to get all the changes into my workspace and then I. You don't have to actually run the regenerator. You just break them into the case.

C

Yeah I wonder if some of the startup Pro logic can be separated as well uh in a sort commit, but.

B

I think it's interesting to.

C

Multiple but it's interconnected with other code as we enable it like API, maybe change us digitized by, but this is a fine new method like we will name the methods and we look at new, unique containers, so it will be one commit changing code of another commit. It may not be ideal.

E

No but I I think if we have like just the three commits like this, that that's totally doable no I.

B

Don't see the issue yeah yeah, usually it takes about 10 minutes really that set up it's pretty cool.

C

Okay, let's try this and see if it works.

C

um Eventually, I I forgot to ask you, but uh is it, can you do the Cinch okay.

B

C

Could um no I'm asking gunjo yeah, oh.

B

Yeah sorry, oh.

A

I think I can just squash a comments here, like maybe three or four.

E

Squishing is probably like more difficult. uh uh uh Yes,.

A

E

I I, the way I would do it is I would like um get get reset to the first.

D

E

Having all the other uh or just before the first commit, then you have like all the changes and then, as as Joe says, you just click on some files commit click. Another files.

D

Commit and then the.

E

Rest commit yes,.

A

E

You have issues I can I can I can also help if you, if you want, but uh it's.

D

E

Code, it makes sense that it's you, your commit.

A

A

E

A

Let you know if there is a problem.

E

And if you are unsure uh you, you can like push to uh to a different branch, then we can review if you want and before you as you wish.

A

Okay, I think I prefer just to the sidecar branch. Yes,.

B

I just posted how I roughly do it just gotta put the right files in there. Please.

A

B

You're, basically, just splitting by file there's grouping files in the commensation pretty quickly.

B

E

A

E

Gonna say I'm, just gonna think the the sidecar Branch locally just in case, so we have it somewhere.

B

Yeah it's if you're scared, it's always good to have this another workspace I'm, um a bunch of us can pull it too. Before you make the change.

B

It's pretty mechanical.

B

D

B

Squash first, if you just want to squash down to one one, that means you're like we'll, lose you and then your reset. It's basically one.

C

Okay, so last item is testing piano, um John I reached out and she suggested he will test. uh I think you'll have some kind and I think it's a good idea. I I told him that I asked him about the resinous probes and he says that raisins props are okay, not to have it for now. Zero test doesn't require them, um so yeah I think uh we may have some PR tested by maybe this week, and maybe it will give more confidence to approvers to approve the pr.

E

Great that's great.

C

Okay, yeah and I looked at the pr yesterday there were no new comments. I I think Ronald just reviewed it to today morning, right.

B

Yeah I I was pretty happy with the last stuff. I looked at it too, so I think it's ready for any any approvers we can get on. It start getting emerged.

E

And did you hear from Lyft because I remember, you talked together with them last week, yeah.

D

I talked to uh that was Tom um for a little bit. He was asking about where the discussion for this uh helps going on I put into the PR, but nothing nothing about actually using it, though,.

E

Well, if we have history already.

C

Yeah I think we still might be enough for now. I.

D

Can think more.

C

There's more than merrier, but uh yeah I think we just uh increasing confidence for approvers to get it.

C

ah Okay. All right, yeah, so excited on me, is to change the pr and make sure that we are proving PR with the new business modification I'm, going to uh testing code like like changing code uh into split, separate, commits and yeah everything else. This is on track, I! Think, okay! Thank you!

C

A

C

E

Bye. Thank you very much bye-bye. Thank you.