Kubernetes SIG Network, 12 Jul 2018

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Kubernetes SIG Network 2018-07-12

Description

Kubernetes SIG Network meeting for July 12, 2018

A

So recording is on this is the Signet work meeting for Thursday July, 12 2018 and si will do really quick run-through on the test grid. Dashboard, I.

A

Will share my screen and we can run through that really quick.

A

Can everybody see my screen and the test grid dashboard?

A

Yes, all right excellent, so the only update I have for the Q ADM ones was that I need to find a way or find out how criminate find out how to get qu Bernays anywhere updated in those tests? Does anybody from Google know how to get an updated, cognize anywhere deployed on some of those CI tests.

A

Away or anybody else.

A

If not I'll pursue that off the call does anybody know what some of the new tests that have showed up are, for example, the container d1, actually that one's not as new but I haven't.

B

A

The cos beta and no s net. Oh, oh hi! Sorry, we were fighting the audio, no problem.

C

To the question of communities anywhere, unfortunately, I think the answer is not what you wanted. I don't have any context on it, I, don't not even sure who I want to talk to. Maybe the test infrastructure team. Here, okay,.

A

Well, I noticed that you had approved my pull request that supposedly fixed the cube ATM once or at least got thrown further, but I don't know how to get that actually deployed into the CI tests. Can.

C

You contact Jeff, Crafton, XD, I, xdy, sure I say that because I'm gonna be out for while so.

A

Xdy, you said my.

C

A

Why there we go alright, alright can do for any of the gke and other tests. Is there any status on those guys.

D

So the neg tests are failing because it's API, that's being enabled enable I, think yeah look at the other ones, oh yeah, which specific ones I think the Signet work in Gretzky's ii ii ii ii ii. I want to do to an API, that's rolling out. Okay,.

A

Mark that down really quick come on lets me select it. It is not letting me select it. Okay, sig network ingress. What is it.

D

Okay ABI supposed to roll out, but in the so we'll ignore those big Network UK.

A

Yep, that's up. It looks like just ingress release: okay, okay, alright! Well, that's hopefully good! So that basically means that we have.

A

Okay, so basically the cube ATM ones are because of Cabernets anywhere. This one is the same thing and no idea about this one. So, basically the what you're saying is that the only test that we really have that is failing for a reason we don't yet know. Is this one right, because the cube, ATM ones are companies anywhere and two container de 2 e and g CI, GC ingress and significant EK e and signet work? Ingress is all due to that AP I update right.

E

Jing is this this uh no snack.

D

Is this you know any contact, no.

D

Okay, then, we should just yeah.

D

Yeah we should you actually just give annotate and see.

A

Okay well track that down: okay, I'm, just gonna. Take a really quick note here.

A

Okay, I guess: that's it for the test grid.

A

It actually looks a lot better than it has in the past. Even though there's a lot of red, at least that's what it seems like to me.

A

A

Alright, so fun I think you're up on SRO, VD, PC and I design. Sure.

B

So about a month ago there was this demo on SRV networking motto and one of the comments made was you know that the the communication method that was implemented to communicate the device ID between device bugging and CNI was too messy right because it was a gr, PC, server and client and they talk directly with each other. So no I just wanted to solicit some feedback on. You know what, if we change that to a annotation or prospect kind of based communication, so there is sick. Node.

B

Has this new proposal for device ID field in impulse? Back with that change? You know, Triplett will be able to. You know populate that field when we, when we call device plug in and then see and I, can invoke with you know, with either that information from cubelet or we can go to record, read the pots back from the CNI itself. So would that be kind of reasonable design for for this SRV to work, I.

C

Apologize I, don't know which proposal or which change you're talking about it and it may have gone through without me. Looking at it can, can you tell me more what that actually is sure.

B

I have to posted a link to the agenda, but I can post a in the chat as well. This is here it is.

C

A

Well, as I understood it Fung, um you know when we had talked about it the other day as well right. The proposal adds device ID to the pod spec, and the issue here is that if you have an SR, io v or some other kind of CNI plug-in that wants to use that information, how should that plug-in? Get the information from the pod spec? Is that more or less a summary duh? Yes?

A

B

Well and I think the bigger question also is well so that is in particular. If we say this is the right design, then then the question is how I see and I read it right, but but but more broadly, you know. Is this whole device ID proposal itself reasonable for a first time.

C

Proposal so I can't say whether it's reasonable or not just reading the first paragraph, I'm gonna read it here just reading the first paragraph it seems like the use case on its face is valuable. I will have to go off, but if it does what you seem to think it does, it seems like a much better answer than a G RPC service right.

B

Right, we've actually implemented a PLC that actually does this right. Of course we use the annotation, but really it's the same kind idea, and it seems okay and and that's why and it seems a simpler and, and you know less messy and then the G RPC, so just something to to take a look, and we do want to move this forward in some fashion right. So that's that's the first question and really the second thing. The second comment that was made was regarding topology information for SRV, Nicks and I.

B

Think that, with the new with the new resource cost proposal, we can possibly solve that in some way. Of course, we still need an external system to actually apply those labels, but you know at least there is a you know, theoretical way that we can do do it. But the question I wanted to ask: is you know for for use cases that do not require this feature, meaning you know, I have a data center that has the same connectivity everywhere.

B

Would it be okay to say you know as a first step that we do not supply this connectivity information that we that we make the assumption that everything is connected the same way and as long as I can grab a SRV device? I can't assume the same connectivity and make that kind of version. One of this work I.

C

Would love to say yes, it's hard for me to know how applicable that is to the light of the world, like everybody.

E

A

Here, who has these use.

C

Cases does that apply to your use case, I.

B

Would love to hear some feedback if people have some use case? We've heard from some people at least right that that their data center is a uniform, but it would be great to hear what everybody else thinks, but.

B

That's it that's what I have cool.

C

Is there a thread, there's not a thread on the mailing list on this? Is there there's not okay, you you want to start what and that'll be a place to gather or feedback. Can people can then share it around sure? Well,.

F

There's probably also worth noting, we do have an alternative approach to this kind of problem with network service mash, in the way that it handles the provision of network services from hardware exists.

A

All right can you follow up to phung's mail Edie when you posted the list and maybe give a pointer to that or shorter.

F

Absolutely absolutely.

A

Again, probably posed it's a good idea, all right thanks. So next up is Dane ipv6, dual stack.

A

Or maybe he just wanted to say that FYI, a new version of it has been pushed and that we should all go check that out and comment on it.

C

So I recommend anybody who's interested in these six and dual stack to go. Read the proposal I. Think. Overall, the proposal is good. There's a really interesting api compatibility discussion within there that I've escalated to stake architecture to get broader visibility on, but otherwise I think it's a really great proposal and then we please go.

E

A

Were there any other agenda items or things anybody wanted to bring up.

C

Now that we're in the 1:12 cycle there's a lot of proposals that I want to try to get back to that have been sort of tabled for a while.

C

That said, a summer vacation season is really in full swing here and I'm gonna be out or not at my desk anyway for the next three weeks, so I'm gonna try to get to whatever proposals I can between now and when I leave on Tuesday and otherwise you know people please feel free to iterate and whatever, while I'm gone, don't block on me, but I'm gonna try to get to what I can there that's my commitment? No time.

A

What do you think the timeline is for the dual stack cap? Is that something that you think, given the current state, that proposal could happen in the 112 cycle? It could.

C

I think if the team implementing has the bandwidth to do it, I'm not sure what the result of the city architecture discussion will be I, don't expect it to be dramatically different change, but you know you never know what he's okay.

A

And I'm gonna call it den Winship here for a second. Was there anything network policy related that was slated for 112? That people should take a look at.

E

B

E

Can think of okay just.

A

A

All right any other items before we close early.

F

One quick one: the does anyone have a pointer to how to run the tests from the dashboard offline I would like to I would love to be able to poke at those a bit.

C

Most of the tests print that command line if she should need to be able to run the tests in the test hole. Awesome I'll, go follow-up about that. Thank you.

A

All right, that's it thanks everybody and see you in two weeks, Tim.

C