Kubernetes SIG Windows, 28 Apr 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Kubernetes SIG Windows 20200428

Description

Kubernetes SIG Windows 20200428

A

Hello, everybody and welcome to another sick windows meeting it's the 28th of April and as always, this is a recorded meeting. So please adhere to the CN CF code of conduct be aware that our meetings and just like everybody else's, are not getting posted on YouTube, not for the last month. We're trying to fix it. It's not up to me. The six contributor experiences tries to fix that. So you know I'm just waiting for them, I, basically just upload them on YouTube right now, so you can keep them there.

A

If anybody's interested in a past meeting, just ping me and I'll I'll, send you a link but go ahead. Craig I was.

B

Gonna say if it's about adding it to the playlist I figured that out last week you.

A

Figured it out so so yeah I used to be able to go back to the old Eddy door and do it, but if there's a better way, I will pink, you and I want to know they hit it. George I didn't know. So if you know how much money one.

A

Okay: okay, cool thanks n I.

A

Even wrote your notes, so I don't forget about it: okay, cool! Let's take a look at our agenda here: Moe's and Adelina. You have the first item.

C

I'm saddened as I have microphone problems and.

A

No problem we can go, we can go to the next item if you want and I don't see, Mars on and he might join us late. If you want a way to can do that, yeah.

C

Please, because my gosh then he's everything blocked everything on my computer.

A

I'm not gonna, say it's because they're running an Oracle cloud, but maybe it's because of running an Oracle cloud and you have what I sure and GCC here. So you.

B

A

Hear that the zoom made a deal with Oracle yeah, they did so it's big thing all right. Let's go to the next item. Jing. Are you online I.

A

Think you are the item that we need a proper on month, vol 2. We need to probably unmount the volume during volume teardown when a pod is deleted. Can you talk a little bit about that? Please.

A

All right, I guess we're more than one person is having microphone problems today.

A

Adaline, are you ready to go back to you, since it doesn't look like okay go ahead.

C

C

Necessarily an easy way to to do it. I.

A

Can you also give a brief overview for the thing so because people that just feel the video might not know? Okay.

C

Okay, so basically, the idea is that we need to run different OS versions on in the round robin fashion on the same job. So basically, let's say one day: one Windows version the other day, another and so on. Now, because of the way a guys engine runs, so they basically because we feed it the template and it.

C

Puts the image version and all of the data in that tend to be a parameters passed on by cube test and in cutest by the job itself, there's no way other than in the job image I mean in the in the wrapper script for cutest to to actually implement this round-robin thing now, because they they change the whole approach to something called pod, utils, they're, basically deprecating that wrapper that I'm talking about and that's basically bootstrap, it's a Python script and it's called the bootstrap script.

C

So, basically, now all the jobs are decorated with something called particles. Then those handles most of what bootstrap bed, and that is.

C

Sown in the kubernetes building, whatever all that stuff now, the problem is that it's highly firm, I, not only try from the punch, would provide a specific logic and those decorators, but I don't even know if there is necessarily a way to do it from what I saw. So, basically, it wouldn't work that way just having the same job and it's switching between between the images and I would also add that it would not be clear from a test grid standpoint. At any point. What we're running the alternative to that is I mean.

A

I mean they're what we're running that that's an easy problem to solve right, you could say like Monday, Wednesday Friday, we run these like Tuesday Thursdays, who run this and Saturday Sunday I mean we could yeah. You can make that.

B

A

So know what we're running, but if it's hard for us to be able to do that dynamic configuration. That's that that really.

B

A

Of the problem right, that's that's there. What do you run when can be deterministic right? Can we make that dynamic? Reconfiguration is a problem that we're looking to solve yeah.

C

We can add another layer, we can say in cube this, for example, so I look into this. So, for example, at this moment in cube tests, we do have a randomizer for say the location of a particular cluster that we create. So we don't crowd up anyone location, so it just it just selects locations at random. If you don't specify one, we can potentially do that with the images in the sense that we can add a flag to instruct this round-robin thing in queue test.

C

Now the downside there is that you would have to I, don't know how, if we can specify the images between it needs to choose from. Let's say another I mean another flag, another parameter or something like that. I, don't necessarily am clear on the mechanics on how we would do that, but it is possible I, don't think it's the most extraordinary best approach in the sense that the way it's just a feature and it just belong in there is it's a use case. It doesn't belong in the intention in the queue test.

C

Whatever way we could do it there, however, there's also the third possibility, and that is something that came to what we are already doing. So, basically, if you look at a guy's engine, we have. We are testing right now, 1903, 1909 and, of course, 1809.

C

We can. These are periodic jobs. We can push the triggering period in such a way that they don't, let's say overlap.

A

The idea there was that we I mean if we have resources and you can actually test every single OS and do a full test matrix. That's great, like I, believe when we first brought it up. The idea was that we don't this unlimited cache of resources. So that's why we had recommender or Smith I recommended it to run one every day, a different OS right, but if I have the option to test all of them using.

C

A

C

Great, what I'm saying is that if we reduce the frequency, so we have three jobs. If you reduce the frequency of the jobs to let's say three days, so we know that each of them will trigger every third day right. So it's the same amount of resources for three different jobs because keep in mind a job, creates and destroys a cluster. So there's no resources that keep hanging wrong.

A

Yeah that works, let's.

D

Do that yeah and I think that having a separate job for each OS might give the clearest signal, instead of having to look them up and I believe that each job in the prowl configs? That's where you define, which bucket they report up to in test grid. So.

B

D

Have to keep those separate, and then we can just vary the the frequency of those jobs. I, don't do you know. Is there when I was looking? I only saw a way to specify the frequency in like a total number of hours. Is there a way to kind of stagger that so if we have three jobs that are all set to run at like 72 hours apart, you know that they would run what like 24 hours apart rather than having them all three run every three days, not.

E

D

Think that's a super big deal, but that might be a nice if we can.

A

And you could also configure it to run every week so that way you could duplicate one job price. You can say every.

B

A

Runs every seven days and then you can basically create seven jobs where three of them tells the three OS flavors that you want to test and the other four are duplicates of those.

A

Right, rather than saying around every three days at which point every other week will be non-deterministic. What do you run on Wednesday, for example, you could sell.

B

A

Job to run every seven days and you have for duplicate jobs of the region of three. If three is a number of voices you're using.

C

That may very well be a way to interconnect the job so I know there are parameters.

C

Well, you can specify for each job that so, for example, only one runs one of one instance of the same job runs at one time. I'm, pretty sure you can define.

C

Interconnections between them so dependencies, sorry between them, but that's still on the table to look at you.

D

Okay but I yeah, so I think that we have we have a way forward and then we can just kind of optimize that as if we need to having separate jobs that are run at kind of long frequencies.

D

Okay, so yeah cool, yeah, I think.

B

D

Saying that this should be feasible to support running tests at some kind of frequency, for three different OS versions. Is that correct, yeah.

F

A

All right son sounds good and by the way, thanks to to Craig I was able to toggle the all our playlists, all our videos for the last four meetings to out YouTube as well. So now, I people are interested they'll be able to get all of the old videos, Thank You Craig, alright,.

A

Say Jing you wanna talk about your item. None of you are able to enjoy it much fun meeting. Sure.

G

So basically, I mean kind of investigated, a place relay to you like how to share data between Linux and windows and issue came across. Basically, we noticed you know when a party rights access a volume right there is a volume set hub and the teardown face.

G

So what absurd hub is laying you attached about him to a node and then it will, if it's not format, right, petition format and during the teardown we actually do not do anything. Special just remove some symlink created during the volumes at hub and I noticed they're mighty sure that after volume is detached from the disk from the node that it is in your unclean State, and so, if you attach again and use it in my say, Linux node, so it will fail to month so I think during the volunteered on this.

G

It require some proper like an amount or dismount operation and I tried to use a month. Wow months we were command to place, / p. That means remove the driver directory and the closed file system and it works. So if you attach the linux node with NTFS file system, it can successfully month so I would just want to confirm right. This is a proper like operation. We should do during volume, teardown phase and anyone have some knowledge about it. Right I just want to hear about a penis.

A

G

Reasonable to me, mmm yes, I, think I kind of go some information there saying if you don't do this his mount and detach busy like just.

G

Unplug, the disk from node, and if you attach again right in an unclean state for Windows, it will do some automatic repair, but for Linux it won't do it, so it will fail and also if anyone interesting like investigate how to share they have between Venus and Windows. Note, please let me know, or you want to share some knowledge about it. Yeah.

F

G

Yes, that's some other options. I'm trying to be mask into so I just want to worry guy for this type of NTFS right. It's also option. If some people want to use it yeah.

F

G

Mean yeah: that's what is not working on Enix.

F

Examples repairs repair.

G

Your iodized I think there's some limitation on that. So it's just like one of the option we want to explore so.

F

G

F

F

G

That's a good point, so even we do that I say proper dismount during the normal, like a politician right bottom tier dumb, face in case of question crash, so it will still cause problem. Yes,.

G

But I think this property's months doesn't matter. It is for sharing me Windows or Linux I. Think it's good to add because it makes sure all right during the normal operation it should kind of gave a clean State after yes,.

G

Sorry, I kind of loose.

E

Some like this America.

A

Yeah Jing, when you're not talking, please mute that cuz coming from you.

F

So so normally the volume driver would detach the disk.

F

And flush it if necessary, if you want to do so, I, don't know what driver you're using when you share data between globalism. Linux is I scuzzy or some other driver.

G

I would testing our GCPD, so I think detach from our DC PD. That's not to any rat the flush, it's kind of very poor to detach.

G

A

A bits worth investigating this further, maybe a Knick. If you can take that offline with whooshing that'd, be great.

A

A

A

Alright jostling em point slices. Yes,.

H

You have an update on endpoint slices. Do you know that started the work? He actually has the first implementation ready. It's testing it as a Friday. He was basically Friday evening was testing it so I'm hopeful that it will be in the next version of communities, because he was saying that it shouldn't take him too much time to finish testing I'm having it for review. So, depending on how fast the video goes, it might be the code fairly quickly.

A

We're cool and just give everybody some background talk a little about what the endpoint slices would give us.

H

That's a good question. Actually I, don't remember all the details about it. I know that it's a feature that was meted for Linux first and we needed to get it on Windows I, don't know if.

E

So Darren point slices basically um like we all know with the current endpoints API I, think it's in the currents and the current state or in the current design of endpoints there's some perf problems that can occur when there's a lot of endpoints backing a service. Basically, the endpoints object can grow very large and since the whole object needs to be recomputed, whenever any operation is made, you know just a single endpoint. It can lead to a higher strain than.

B

E

On the system and I think endpoint slices is a new, more lightweight resource that is designed to handle. You know on operations on specifically services containing, like many thousands of back-end endpoints, more gracefully and scalable than current endpoints API implementation.

E

And yeah currently Linux Jocelyn's. That already has this feature and they want it to go GA and one 919 it's currently in beta and we're also trying to support endpoint slices. Oh.

A

Thank you for the bike. Ride David.

A

In great progress on this Jocelin team, that's awesome and I'm, assuming we don't need a cap for this right. You can just do you need a cap.

E

Not sure I mean Linux has a cap on this. Maybe we could extend it with a Windows section or something yeah.

A

Let's look into that just to make sure that you know we don't someone doesn't come on day and tell us.

B

A

You're, adding all these things- and you know within probably let them across the program just or or David. If you can take a look at that and make sure that you put a small amendment there, then how it will work on Windows and and and put it out for review, I'm sure you should be fine, but let's, let's, let's partner up: okay, okay, I'll! Take the action item. Thank you.

A

All right, that's all the topics we had for today. Anybody that wants to bring something up any concerns issues.

A

Peter did I hear right: GC, PG, 8 windows, containers on Jackie the other day.

B

Yes, congratulations.

A

That's great achievement. We look forward to a lot of usage of Windows on kubernetes at Google. That's awesome, Congrats on the milestone.

B

Thanks after being complete, we should know a chaos' 88 today.

D

A

That's awesome. Well, congratulations! What is? It is the week of G Aiden, congratulations to aks as well. Maybe I should put that both in the notes here. So are we getting free, unlimited usage accounts for everybody.

B

Thank you. Yes,.

A

All right cool: well, everybody have a good rest of their week and good day. Please stay safe and we'll see you all next week, bye everybody, bye and.

D

Thanks everyone thank.

H