Kubernetes SIG Cluster Lifecycle, 4 Aug 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Cluster Addons meeting: 2020-08-04

Description

Meeting notes: https://docs.google.com/document/d/10_tl_SXcFGb-2109QpcFVrdrfnVEuQ05MBrXtasB0vk/edit#heading=h.q9sl7lw9laa3

A

Hi everyone: this is the cluster add-ons, so project meeting for august 4th and everyone's doing pretty well. Today, we've got an empty agenda for the most part, but have been getting a lot of work done behind the scenes, but particularly with google summer code.

A

If you have something that you want to chat about, please do add it to the agenda or just speak up about something and um yeah justin. You wanted to start off.

B

uh Sure I can't happen. I uh I think when we were debating what's put on the agenda, I think I was just joking that, uh like we seem to be, I certainly seem to spend a lot of time just trying to keep up with the excellent work that our google summer of code person samtoshi is doing. um I think she has some pr. She has a bunch of pr's up that I think, are uh getting into a state where they can merge.

B

There's a very, I think, there's an interesting one, which is uh perhaps worthy of feedback more on the tool itself, which is around uh a coupe cuddle plug-in to show owner refs. um I think that's that's an interesting one. The the implementation is limited by what api server offers. So this is one where we're saying like if such a tool existed- and it was a little bit faster, would this be useful to people?

B

um I think I mean tend to start playing with it, but if other people play with it as well, I think that would be very valuable. It's not just in add-ons, but also generally, but in the add-ons case, um uh an operator, the operator cr.

B

Typically will uh have an owner ref, either directly to all its objects or indirectly towards objects. So it's a nice grouping mechanism there. That uh would be good. One there's also, I think, some great she's doing some work to create a generic operator which is essentially based on the idea that a lot of the.

B

A lot of the time you generate a an operator with the pattern with the add-on pattern uh using cube builder, and then you don't change it, and if you don't change it, uh you probably don't even we can sort of bypass the generate stage and just share a common implementation with a couple of hooks, and so satoshi is looking at doing that, and I think it's working pretty well- and I think also uh where she's headed is uh generating, manifests generating manifests that are perhaps better suited to to what uh we we use.

B

The cube builder manifests the the the manifest the cubiller generates are uh very extensible and a little bit verbose at times and most of the time, I personally, for example, normally just deleted them and start again, and I can uh can we do something there and that sort of ties into a lot of the work she's been doing elsewhere on our back generation and on um this. This generic operator idea.

B

So yeah a lot of a lot of great pr's in progress from uh the google summer of code project. There is less than a month remaining, I think the it seems to be aligned with months.

B

So I assume that august 31st is the end and um I think the priorities are essentially to get everything uh squared away all those pr's merged, like those sort of work, threads finished um and working on a blog post to describe you know, google, summer of code, what it's like to participate in a google summer of code project in kubernetes and um uh the work that that that she did and so hopefully we'll be able to get that onto the kubernetes blog. That is our hope.

B

um We are lee, I think, is he sent an email previously talking to the people that would approve such thing.

B

That's all I got.

C

uh The owner reference work in particular looks really interesting. um Is there any sort of discussion of a server-side component to provide that information.

B

I think what we're arguing is: uh let's first, let's first demonstrate that it's valuable and then let's like make the strong case for a server-side component, because you know this. This information exists already on the server because it's used by the garbage collector. um So I agree with you that it should be something that we have server-side.

B

uh The intent is basically like make it work and show that people like it, even if it is slow and horrifically expensive to evaluate um and then uh and then like say, hey, hey, api server like we are totally hosing your api servers. Do you want to make this a little more efficient.

C

I think um the openshift console does something similar, but to get around performance issues uh you have to specify the kinds that you want to watch up front so say, like my, the cr-type watches like manages pods and services, so watch for pods and services in this name, space that have owner references and then that's not terrible but yeah. Exactly exactly what you're saying, there's a performance problem. If you open that up to everything on the cluster.

B

um It might not be yeah it'll, be interesting to see like uh right now. It's the current implementation like loops through all the kinds uh in a serial loop. It will be interesting to see how bad the end user performance is. If we loop through those do those in parallel. um I think that'll be interesting. It's obviously very very expensive. It might not actually be expensive in terms of human wall clock time, but I I agree with you.

B

We want to optimize that, but first like that, I think the meta strategy here is to to make it um demonstrate that it's useful, like we don't want to. We don't do all this optimization. If people don't people don't want it.

C

Yeah, absolutely, it seems really useful.

B

Cool and there's it what's the what's, the is: what's the open shift command, is it an oc, command or.

C

It's not a command, it's just part of the console, so the ui does it in the browser.

B

As part of the console, you could do a watch and you could have your own external server-side component right or I don't know how your.

C

B

Works but anyway,.

C

uh You mean you could do a watch of the information in the console and expose that again.

B

I'm more like you could. If you have a if you have a long-running server, you could watch every object in the system and sort of like keep a lookup table. um I don't know whether you would like how granular would make that look a table. Would it be like just the object, just the kind to kind references, or would you try to keep something more accurate than that, but um you could do such a thing anyway. I I yeah.

C

Yeah, that's kind of how I imagine, if you were to do this, like as part of the garbage collector or something like the garbage collector, just kept the cache and threw it through the data somewhere in the server um so that you could query it later.

B

Yeah- and I think my understanding is the garbage collector- has that, but uh we should, I don't want to speculate. This word.

C

Yeah it does, it does keep it cache. But if, if I remember correctly, it looks at namespace at a time and then clears everything out so just like walks through the cluster to try and limit how much space it uses.

C

That makes sense.

B

I'm sort of wondering like what happens when you have a bajillion objects.

A

That's interesting to think about how the garbage collection behavior could change, depending on the number of namespaces in your.

A

Cluster or just become more latent, basically, but.

C

Someone probably has numbers.

C

A

Yeah owner refs are really like kind of purely interes interesting feature like inside kubernetes, because the data store, but it doesn't really have that feature natively built in right. So.

A

The fact that they're there and that they're needed- but they have these interesting constraints.

A

um Yeah, I don't know if it was discussed before, because I I was out the last time we had this meeting, uh but we did work on adding some preliminary get uh support for the packages in the channels of the declarative pattern based operators with the idea just that we would like to be building the ecosystem, that's more extensible for the users and or the people deploying these operators and um yeah. That might be of some interest.

A

It's baked into some of some sochi's um declarative pattern, pull requests, and that would be the natural place as well to extend for the oci package, support or docker image based.

A

Just heads up there and we are at the end of our agenda, so great conversation on the unref stuff. Does anyone else have things that they wanted to bring.

A

A

Okay, well, that's probably the end of our meeting. It's awesome as well always to hear from all of you if you're staying healthy and taking care of each other, and we will keep chatting in the slack channel and see each other in two weeks.

B