Kubernetes SIG Arch - KEP Reading Club, 14 Mar 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Kubernetes SIG Arch - KEP Reading Club 20220314

Description

KEPs discussed today:
- Beta APIs off by default:
https://github.com/kubernetes/enhancements/tree/master/keps/sig-architecture/3136-beta-apis-off-by-default

- Sub-second probes
https://github.com/kubernetes/enhancements/pull/3067

Slack link with discussion: https://kubernetes.slack.com/archives/C5P3FE08M/p1647273031940579

A

Oh awesome hi um just as a prelude to every community meeting. uh This is a kubernetes community meeting, so we do follow the kubernetes and cncf code of conduct, which boils down to be excellent to each other. So um please adhere to that and this was also being recorded. So don't uh please don't say or do anything that you don't don't want posted on the internet. Okay, that having that that being said, uh I think we can get started with the first cap just in case folks. Don't have the link with them.

A

This is the link to it. um I will start a 10 minute timer we can extend uh depending on if we need more time. So that's totally cool like.

B

A

Feel the need to like rush and uh like finish it in 10 minutes. Just take your time. It's cool, okay, awesome! So uh that's the that's the cap. I am going to start the timer in three two and start.

A

Okay, that's ten minutes.

B

Does anyone just.

A

Require like a couple of extra minutes, just let me know otherwise we can start discussion.

A

It's cool like if you need two more minutes, maybe.

A

Okay, okay, seeing seeing none, uh we can start so just like right off the bat. Does anyone have any questions.

C

Yeah, I wonder, did we consider uh any other configuration alternative? Besides just runtime flags, it can be difficult to uh manage those to things like cube, am not not overwhelmingly difficult, but uh just makes me wonder.

A

Yeah, that might make things a little complicated, you're right.

C

That's certainly fair.

B

Okay, some any other questions.

D

uh I have a question: can people access this prr survey that is linked in this gap because I can't see the results and uh that would be interesting, see that.

A

uh So by the pierre survey, do you mean the.

E

F

A

That that's mentioned.

D

Yeah so there's a link to pr survey.

C

I think it's this uh prod readiness review.

D

Yes, yes uh and there's a link in the risks and mitigations uh section.

B

D

uh I can't access this report. uh I don't know like quest permissions, but I'm wondering if there's like any public version of this report or anything.

D

Okay, I'm trying to paste this link in the zoom.

D

Yeah, it would be interesting to say.

B

Okay, I will bring it up on slack after this thanks for pointing that out.

B

Awesome: okay,.

A

So one observation that I had was- and this is mostly like a misconception on my part- I had seen this cap a few weeks back as well, uh but my understanding uh that got clarified right now mostly was along that beta features themselves are going to be disabled by default and not just beta apis, which was like a misconception.

A

uh Beta features continue to be enabled by default. It's just that the beta apis are going to be disabled uh by default. So- and this also is an interesting consideration, because there's another feature in api machinery that I'm currently working on for the next release along enabling field level versioning for customer resources, um so over there like it'll, be interesting to know now consider uh how to treat beta feature types over there now because of this feature like do we keep it the same.

A

uh Do we keep it independent of this of this feature that that's going to be an interesting consideration and conversation to have um okay, so uh any other questions, or else we can maybe move.

B

C

Maybe just uh follow on to what you're saying is there ever cases where beta features will rely on beta apis and if so, how would you test around that to make sure that you don't run into issues.

A

uh Could you elaborate on that.

C

Oh, I just don't know enough about api machinery to know if there are chances for beta features to rely on beta apis.

A

So, uh as far as I know, it is possible that that could happen, and it's very possible that that also happens so far in today's status, at least uh so right now before this step is implemented. If you were to test it out, it would be along. The lines of api is enabled by default. Feature is enabled by default, and things would work as it. The things would work as expected and, like you could test it out as as expected, uh my understanding now around.

A

How would you test it out if any, if the feature of the beta feature gate is a beta feature but like if the beta api is disabled by default, but the beta feature is enabled by default. uh That's a valid question. Let me just write that down.

F

Yeah that that just kind of feels like, like potentially a bad user experience where something's like half, turned on right so like the features enabled but like the thing, I need to call to use it isn't and then I need to go back and like have someone spin it up, or you know, enable it on the fly or something like that.

F

So it I'm not saying that like this is a bad approach it, but in cases where that's a possibility, like be good to be very clear about that saying, like hey the feature is enabled, but in order before you can start using, it you'll need to turn this thing on.

A

Awesome thanks uh I've written it down. I will uh start a thread on slack after this call just.

F

A

The the channel name where this thread would be started is uh the architecture just in case I don't remember all of your handle names um so just like keep an eye out from there. Okay, um yeah, that's helpful. That's that's helpful to clear that thanks. Okay, uh any other last questions for this skip before we move on to the next one.

A

Okay, seeing none, I think we can move on to the next one thanks again for all the feedback. uh I will let folks know.

A

Awesome, okay, so the next one isn't really so it this isn't a merged cap yet uh and it's not a ready pr as well. It's a work in progress pr, but this was suggested by like another community member and it seems really really interesting and impactful.

A

So if this is something that you use uh at work or you could potentially use at work, this is a great time to get an early feedback, uh because the pr itself is work in progress right. So uh this is the link in case uh you didn't have it and.

B

A

A

Okay, this is a relatively lengthy cap, so, like we can take a, we can take some more minutes. So it's cool okay. So uh I'm just going to start the recording, not recording the timer in three two and one.

A

Okay, does good folks want some more some more time couple minutes or.

B

Can we start.

B

A

Saying no requests for additional time, uh so anyone have any questions about for.

B

C

I guess the only thing that comes to mind is the implementation around the offset. um I know somebody had chimed in and said it was kind of unintuitive. um I would want to. I guess, I'd ask, did we consider perhaps doing period units and then taking whichever is greater if both period units and periods seconds is defined.

B

A

But can we consider maybe using kdb units.

B

A

Would you just like mind taking a look at the dark ones and augmenting it like like just whatever you had in mind more accurately,.

A

Okay, any other questions so first.

F

Yeah, I think that the only thing that like really screamed out to me- but I don't think that I saw an answer and I've gone through any of the comments. So maybe the conversations in there uh is once you get down to like milliseconds of things like you, you can kind of like accidentally like overrun yourselves with like probes right so like ddos yourself, just because things are checking in so quickly.

F

I guess what what is the plan? What might be a plan to put in place to like prevent users from accidentally doing that to themselves? Because you know one team may have just a handful of these things and another team may have just another handful. But multiple teams having just a handful, adds up to a mountain of these things that can quickly overwhelm you know the api server. So I I think I'd just like to know a bit more about what the what plans might be put in place to to keep that from happening.

A

Yeah, that's that's a better question, uh so I've written it down, but for the time being um there is this under the risks and mitigation part. Like the first point. uh I think they are being mindful about that, but they might still need to add a few more details around it to make it as clear as possible.

A

uh So yeah thanks for the feedback. That's that's! That's a valid point.

E

Oh, I agree actually and in the recent meditation the author mentions, that uh they set a minimum value that can't be surpassed and they've, given a heuristic as 100 milliseconds, I think it'd be if possible. I think it'd be good to uh figure out and get a more precise value. That has, I mean a little more backing than just a heuristic.

B

E

To see what works best.

B

B

Okay, awesome thanks! Oh second, I wrote that down.

B

A

Any other questions.

D

So I saw there was one comment uh by john about a suggestion to add, like period, milliseconds, etc, fields and, to sum them together with period seconds to like, define the values instead of subtract offset from existing values.

D

uh The response to that was that clients that don't recognize the new field would be liable to trigger much too early, um and this is a problem and.

F

Honestly, this.

D

This comment that john wrote was my first thought when I read about offsets like uh just some adding to values. Seconds and milliseconds feels much more natural to me than subtracting some offsets, um and I see there is the problem with with clients that clients would need to uh all clients would need to update, be updated to recognize this new value uh to support um like new timeouts and uh don't trigger too early. I was wondering if like because it feels like, um of course, kubernetes product.

D

I guess that doesn't have like can't control all the clients, but I don't know if there are any like um conversations about how that updates in would be implemented in in clients that and if anything could be done to like um promote implementation of this new feature in clients.

A

And if you don't mind, would you mind jotting down those points in the dark like I try to keep up uh yeah.

D

A

To accurately capture it sorry about that.

D

Okay, I I was writing yeah.

B

A

Okay, awesome uh any other comments or questions or concerns at all.

A

Also, this kept is part of a sig node, so if at all, you are interested in maybe like helping out uh with the kept on like helping out the author with, like the various things surrounding it, uh please feel free to reach out there. um If you don't know who the author is, I think it's mentioned or at the top of the cap or like uh it's mike brown.

A

Just like, I think his handle is m brown on slack uh he's he's an amazing person like feel free to reach out to him uh if at all, you're interested in helping out.

A

Okay, awesome uh last call for questions comments otherwise, like we can call it early and give you back 15 minutes of your day.

A

Okay, great awesome, uh thank you so much for joining everyone uh see you next month. Also, if you have gaps that you know for a fact that you would like to discuss, feel free to reach out on on slack and like, let me know for the next month's meeting so that we can. uh We can plan ahead and I I can reach out to the authors and also get them on call.

A

Okay, thanks again, bye folks.

C

Thanks everybody thanks ben.