Kubernetes SIG Windows, 20 Aug 2019

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Kubernetes SIG Windows 20190820

Description

Kubernetes SIG Windows 20190820

A

Everybody and welcome to another sig windows, meetup I. Thank you all for attending it's the 20th of August and, as always, please adhere to the kubernetes and CN CF code of conduct, since this is a recorded meeting. I'm still catching up personally on a few things from the community. Patrick had sent me quite a few things in the last couple weeks, so I'm catching up so we're just gonna go through the agenda.

A

There's a few items here that I wanted to bring up based on items that me and Patrick had discussed in the last couple of weeks, but we haven't had a chance to kind of disclose them to the general community. I'll start with those, and some of these will spill over to subsequent weeks as well. But you know just to kind of give you guys a heads up the first one, there's two PRS that AWS team has pulled for 1.14 and 1.15 I linked them in the document and I put them under 827.

A

Essentially, what do we want to do? Is we want to hub probably deep talk to the AWS guys and see if the EBS support that they will enable, if that's something that can be handled under the CSI work, that they be spearheading so I? Think that's, there's going to be an important another important support for storage that you can add so just putting them in here as a heads up and give her some time this weekend can look at it, otherwise we'll discuss in next week and see how can move forward from there.

A

The second item is a ticket that I guess. Ben was one last one. The commented around the empty deer and some of the volume supporting windows misrepresenting the amount of storage.

A

Do we know who who's gonna work on that band? Is that there's Ginsu there.

B

Is a PR already open for that one I think there were just some comments about the implementation: okay,.

A

Cool so I'm, assuming we'll fix it for 116, then it looks like right. I think.

B

A

Yeah, nice cool, that's a good catch all right! So deep I guess we're pushing to you know the CSI books. Yes, Cassie ice I'm, assuming that's yours right.

C

Yeah, actually that was opened by Jing a little bit so yeah go ahead. Oh.

D

You can go ahead actually so I think we are just physically trying to they find api's force as a proxy working on and that API should include support, ice, cassie and me exactly and we want to have kind of cover. The api cannot cover all the cases, so we want to find for this stage the minimum sets of api. We should support and/or ice cassie case. I think there is a link I post there, that's the most basic ones, I think.

D

Currently we should support I, think it that lays also just makes one API remove ice ice, Cassie connection, don't someone mention that and I just want to see because I'm not expert on ice Cassie part, but we just want to make sure anyone has some concerns or feedback about it. That part.

C

Yeah and just wanted to also add the context that so basically the link you posted is also from John's repo who posted the PR in CSI proxy as well yeah, so he's basically incorporating any John is also on the call. So you can just speak to her John.

E

E

That PR is really they are the media API. You want for the CSI proxy for the alpha version. So really the welcome.

D

So one point we have is here is thing: ESCA see is quite company good, so we just want to. Is it in a row? What's windows line use exactly the same API as Windows, so don't change anything.

E

Then, let's try get your question in your point.

D

E

I understand your your point to your question. Oh.

D

I was just saying where I just tried to marrow the windows API for I scuzzy.

E

Yes, what's correct, I'm, seeing it's a good thing or a bad thing? That's it no.

D

No I think it's. It's sounds reasonable. It's reasonable, yeah, okay,.

C

And I think just for clarity also like this is the PR that I just shared in the screen is the one that's adding the definitions in the CSI proxies API put above definitions, I. Think there's been a few reviews like a few folks, already took a look like Peter dead, so Jing, it's I think I want to do this as well. If you can take a look at some point.

C

F

C

I think this is what's going to end up in the proxy and will be kinda interfacing with the the I scuzzy bindings that show.

D

And by the way we are like Eve and I thinking to have CSI windows kind of working group, so we can work together. This is a big project and to have regular meetings, I plan to set it up- probably Tuesday Friday 30 minutes, and if anyone interest enjoying this meeting, please let me know I help you I just want to help to set it up. Yeah, oh yeah,.

E

I'd love to you, I'd love to doing that.

D

So could you just send me your email songs, like yeah.

G

D

C

Yeah great that'd be that'd, be a good forum to kind of move this along thanks Shane thank.

D

A

Did I guess you can continue with your GMS a better study since we've got you on there on the mic? Okay,.

C

Sure so, basically, we were planning to take GMS a to beta in 1/16 that was kind of the target. So we have a couple of outstanding items. One is, let me start with the second one. In fact, this is the entry cleaner around GMS is GMS sacred specs, so jean submitted that pierre quite a while back it's been under sorry review. I believe yesterday, you jus probably took a look and assigned it to Lantau.

C

So since it's the entry clean up kia, the pr is entry. We we kinda need to get it in before the code. Freeze the earlier the better and the code. Freeze is kind of creeping up next week, I believe so.

C

If anyone has any feedback, please take a look. Otherwise I think I will continue to ping you to our plant out that she has assigned to yeah.

F

Sorry so I was looking at that I just assigned to lontel in case I. He has been was to look at it first, okay, it's a cleanup here and I think makes sense. I, don't think there'll be any problem getting it by Aiko. Freeze, awesome.

C

Thanks a lot appreciate it, the other main thing that's remaining for the beta is getting the test. Infra up so Jeremy and Sean has been working on the Windows testing PR. The idea. The idea there is to introduce this other extension that sets up the DC the the domain controller on one of the notes. I think Jeremy is just cleaning it up, but it shouldn't be ready and at Lena, if you have any cycles, please take a look and please post any comments. If you have any concerns, yeah.

H

We we actually tested it. You know in a job to see if the deploys property and and if it works with with our inference. Oh that's, that's great I've looked, but only while you know fugitive Lee over over the script. In the extension, nothing popped, my eye as a common, but I I will promise tomorrow. I'll look thoroughly and I'll leave any comments.

H

If there is one and short as to the tests we said last week, then we're going to discuss exactly the form in which we will enable this either a new job or something else. I, don't know. If anything anything else, materialized Patrick is not here. So I guess maybe I'll trust you to take this on slap. Okay,.

C

So the basic thinking there that we were planning on is to introduce a test that just looks at whether the tag is set on the nodes like if the gym and if they are set, then the test will continue if they're not set, then it will just skip the test. So that makes sure that things don't fear. If.

E

C

The feature is not enable if the GMs a if the DC is not set up. Basically yeah, that's so that's perfectly reasonable should be somewhat safe. I think that SPR is outstanding, so that'll come in pretty short and with those three in place GMS, I should be really quick.

A

Question? How does the test figure out I'm, assuming part of your test, is also setting up the domain controller right and.

E

A

Some of them knows knows in the domain. So what? How does he know not to do that if you're running an environment that may or may not support it like, for example, if you're running in another kubernetes, prowl environment right or would.

C

A

To do that, yeah.

C

So the way Jeremy set it up is he did it as a aks extension. Okay,.

A

C

It's part of a Gleaners template that referred to the extension and only trigger in the case of a sure aks for now, and if you want to anybody, don't say GC, then we have to figure out how to set up the DC. But then, as far as that, III test is concerned that I just look at the tags and continuous.

A

Yeah yeah no I understood that test. I was just wondering about the intra part. Okay got it. Thank you. Okay,.

C

So yeah that was about GMS a and the other item I had was I, wanted to ask. If people already have experience, building docker images based on Windows through brow, mainly for again around the CSI proxy one feedback from Michelle was from six storage was, and we figure out a way to get.

C

You know, building of Windows docker images in prowl, somehow I asked in the brow channel as well. They said one way around. This could be to set up this federated windows build cluster I, don't know the details are on this. I was curious. If someone has any what.

A

Kind of Windows images are these: is it like? Just images? Are you using for testing in kubernetes or some something else.

C

So this is going to be one of the production images. It's around a component called the CSI, no driver registrar, it's an image that needs to be there as part of it. Okay, so.

A

These are these an image that will build and ship as part of the master components for for the important see as I work.

C

It's time to make the pause image I thought, but wasn't sure how the positive in this you know.

B

Yeah I, don't think that's being made in prowl right now.

H

From what I know is made in an Asian pipeline, I guess part on that side is managing as to the question of building images in four windows in brow.

H

There's there's definitely not nothing as such. However, brow is flexible enough, so we can write whatever workload we need in it. Basically, what prowl does is just look at whatever repo you. You tell it to look or look for whatever command. Whatever comment you want, and it will do.

H

Can be done is not that complicated. We do need a build node. That brow should connect to some of the work. There is not necessarily this. The same word cloudy. Has this work where he's building the test images for the Windows testing and it's the workflows pretty similar in the city? This is I, don't think it's triggered by prowl at the moment, but the thing is that you're, using a remote Windows nodes to.

H

So there is a bunch of words that may relate to this in one way another so.

A

So I guess one thing: I should know this deep one to talk with Patrick when he's back and seafood can so the way we've done it right to have the seat to sit windows tools, rapport if you can put there all the artifacts are necessary to build that image. Like your docker files, everything else early needs go ahead and do that there. So at least we have a centralized report that in the kubernetes umbrella that has all the artifacts and then he Patrick can incorporate that as part of the hopup line.

A

So he can build it for you and even hosted in MCR, then that would think of our workflows. That way, you have a single report as all our all the things that connects who need for Windows.

C

Okay sounds good. This will be slightly different, I think. Oh, this is a CSI image so but yeah we can work that.

C

Thanks and I'll also follow up with Adelina for some pointers to setting up console.

A

Great thanks next item is their honest, username Claudio great work on that. So it looks like we're almost ready to release, or at least proceeding yeah.

I

And the future merged a few weeks ago and the condition actually merged this week or late last week, so the only part remaining is the entry tests which everything there's one small note which I am dressed addressed to ban and you Joe in order for this to, as you are aiming to add feature gates with Windows Ronis username I have left the comment on the pull request there, which relations. This is well because Susan, a miss, you know, features. So that means it's disabled by default.

I

So without that is the person gonna fail, but we should run those run. Those tests for 116.

A

Okay, make sense, who's the ultimate okay, you just automate really approval for their for the for this and Venice walk all right, cool.

A

Right then, we skip the earlier two items on our agenda. One of it is a caching behavior for writing files in Windows, who added that or so.

D

Yeah, this is thing I added. So basically the problem I have is about testing our into a TCP D driver. We have some tests very simple tests, so if we started a one-pot and then that part will write greatest file and the writes some data into the file and then kill this part, and after that we create a second apart, which will use the same volume, the gcpd volume and read the content from that pod.

D

We see this test flicky, sometimes and when I rescued it I see. Even though the first part finished writing. We also can verify after it writes. The content is correct and then here is part and better. When the second part starts, the content of a file seems sometimes still empty. The file exists, but the content it just like empty. So that's why it failed to verify the content in the second file in a second apart and I.

D

Think the the reason is seems like the first part when he write and somehow is buffered and it doesn't not. It does not flush the buffer onto the disk, yet so, when a second of how stars? So it doesn't see any content, the leanings. We have the same, exactly tests so far. It does not flaky I. Think the writing. The file way I try to approach his poly for Linux. They just use echo to direct the content to the file for Windows I used the same thing first and then later I use the windows API.

D

Like the said content, the PowerShell commands said content, but I think they all flaky and I. Don't know suppose when the file finish. Writing it. Should the file closed should flash the content onto disk I, don't know where, like we can have a command to add, to make sure it's flush and for Linux you know even something wrong like it's not flash yet, but when the part killed we and mounts the volume a month.

D

I think it's Auto also automatically flush the buffer to the disk, but Windows we seems don't have such thing, because we knows we just create a singling and remove singling. There's no guarantee the buffer is flushed. So.

C

Even in the Windows case, I think it's still a GCPD volume right and Men. That gets like that. This small Cisco I know the file system flushes for sure. So that is pretty interesting.

A

Even if it didn't flush deep I mean if it was, cash was cached at the file system. So when you go back to read it, you should read it back from the same spot. It's not like the file system changes from one port, but.

D

When the first part killed it will the volume will detach and when the second part starts it, my legs start on different node and then it will attach and to a different node. It's not the same, and we don't use this mount. So if you see the mount windows file, we just create a singling. There is no dismount function called. Oh, maybe.

C

That needs to be added yeah.

D

I'm not sure like for mount Windows. If you see the mountain Linux right.

C

D

We have a month and month, but there's no corresponding like mean start stinking windows. So I don't know what we knows. We should maybe check this yeah.

G

This is the nation: can you file a github issue and paste the link in the slack I will pull in the storage folks from windows and I?

G

Looked at the set content, there's no flush tricks like explicit parameter, so I, don't know exactly what's happening here, so I will be able to loop in the right folks. If you have a get up issue yeah, you can write an email as well and I'll paste. My email id I can look under.

D

Yes, so my Slyke is Jing ji, gxu 97. So can you pick me so I can try yeah. Thank you. Thank you very much.

D

Yes, that's it! So that's my kind of issue I want to bring it up.

A

Thank You G, alright, and the last issue here is maintaining kubernetes end-to-end test images. Adelina did that come from you or claudio.

F

No we've been talking about trying to merge the windows test images with other than this ones, merging in the sense of managing them together to remove a lot pinpoints that we have had liked has been broken because people change only the entry, linux images and test cases. So because the way we are doing it now, we basically have very ad hoc a building and pushing process. So I talked to the next, so I talked to some of the folks on sick testing and also team.

F

Talking about this and basically I wrote down a summary of what the different options, what we could do and the options, including like maybe building a using GCE. We can some scripts like. While we talk about similar a little bit similar to CSI discussion about building an image that we can just create a VM and execute and build the image remotely and push it to GCR. So there are different options and different pros and cons.

F

I just wanted everyone to take a look if you're interested, and maybe we can move forward and decide what to do.

A

All right, thank you. Yeah read this and I'll make sure the pad to go so there's a chance to check this out as well, and hopefully everybody from our community gets know. If you have any comments or suggestions, add them to this dog. Thank you. I.

H

Just have a question here: how does this align with the work you've been doing in regards to images.

I

I'm not sure does this include the four requests from from me and I've done regarding the windows test images, synchronous, yeah,.

F

I, don't think it comes with yours, it's just who is going to actually how to ask how to push and build? Who is going to do that and your quiet environments? Should we do this because, right now, even for Linux, you have to basically poke someone to build and manually push and I know there has been. We have been discussions about automating the process and there are also some technical issues about building this outside of GCE. So this is just something that I wanted. I shouldn't want to input as well.

I

I also had that free to write a document regarding the windows of the image building process. Recently, I've pasted another link in the chat, but I haven't thread this document, yet so I'm not entirely sure. What's in it, but the way I've been doing it basically, two nodes, one in US and one windows and the windows. Node is basically configure Windows notice configured with dock remote management, which basically means I, can build the docker images from the Linux nodes.

I

For example, this sources has been included in the image material script from the kubernetes repo, so basically just basically make we're also build for Windows as well.

H

Just hear you Joe I'm, assuming this is more not on infra discussion and actual mechanics discussion, as in.

H

What I want to say is if there's any any discussion about automating this process, and you know in a pro job or something like that, and if there's no available bandwidth for it, we can lend a hand in making more testing that route. Job put this it's basically in the interests of this sig to have this as fast as possible, and that's.

F

Good to know, I don't know if it's actually on the roadmap or anything I've just heard about some people may have their psycho whoa Thomas. We're not including this I just wanted to make sure that whatever we decide to do that would actually work for Windows as well. And yes, the talk is more about how to I guess how to merge the two processes of using the next time with images, because you can build it with any Sony credentials to push to GCR, basically use the doctor manifest yeah yeah.

F

So take a look when you have time and we can discuss on sack or we can set up a separate meeting. If you want okay.

H

Great excellent, that's that's I just want to reiterate our our this position to do something in this regard. If there's need okay. Thank you. Thank.

F

A

All right, I think that was the last item on our agenda and were pretty much out of time. Well, thank you all for for attending appreciate your your work on this as you're pushing for the 1/16 release and we'll see you guys next week. Thank you thanks.

C