Kubernetes SIG Apps, 24 Feb 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Kubernetes SIG Apps 20200224

Description

Led by: Janet Kuo
Co-hosts(s): Matt Farina

Discussion:
- implement job allocation completions index enhancement https://github.com/kubernetes/enhancements/pull/1072
- Previous effort https://github.com/kubernetes/kubernetes/issues/14188
- Pod Disruption Budget [Morten]
- https://github.com/kubernetes/enhancements/pull/1501
- https://github.com/kubernetes/enhancements/pull/1502
- StatefulSet feature - volume resizing [Ken]
- Reconcile PVC
- Previously https://github.com/kubernetes/enhancements/pull/660 and https://groups.google.com/forum/?utm_medium=email&utm_source=footer#!msg/kubernetes-sig-apps/ffClkRzptdA/DH36i9J2GwAJ - KEP needs some cleanup
- CronJob GA
- Barni: https://github.com/kubernetes/enhancements/pull/978 Addressed all review comments. Waiting for LGTM approval.

Demo:
- GKE Application Manager https://cloud.google.com/kubernetes-engine/docs/how-to/add-on/application-delivery [Janet] - 10 mins

A

A

Okay, the meeting is recorded, um welcome everyone to the February 24th C caps, meeting I'm Jana quo and, together with me, is Matt Matt Farina he's going to co-host with me. So today we have a few discussion topics and then finally, a demo and I'm gonna start with this kubernetes cap, you, oh so ken is also on with us. Okay, now you're there.

A

Anyways yeah, oh hey cane and the first topic is this job: completion index enhancement.

A

um From this tab it looks like you also want to implement some indexing function that allows you to shard all the jobs, Iran and I I found that there is a previous discussion about this index. Job effort by Eric toon.

A

So does anyone have some some thoughts about this effort and how we should move forward.

B

When we originally wrote the two jobs Eric put together, a list of nice-to-have features two jobs. The index job was one of those, but we never got actually two out to implementing one of those but I. Remember that the index job was something reasonably simple to implement. So unless anyone objects, I think that would be a reasonable thing to add.

B

But I haven't read the cap yet so I need to have a look at it. How closely it matches what Eric envisioned having but I. Think it's a if there is a reasonable, reasonable thing to have.

C

Yeah I basically agree with the sentiment in terms of the idea that, like so like, if you're running MPI style jobs being able to identify which job you are via an index, that's passed through allows you to, like you know, modify the behavior of the job based on ordinal right. So that does make sense.

C

Yep I think needs a lot of work, though.

A

Yeah the current the way. How is written today and didn't have enough details for how.

B

Yeah yeah quickly scrolling through the proposal it it's just a rough sketch and I haven't seen any discussion in a while. So the question is where the author is still interested in working on this one: okay,.

A

So I can start with asking if the Auster is still going to work on ders yeah.

B

Definitely that's a good. That's a good approach.

A

mmm Does anyone have anything to talk about for this item if not we're moving to the next one? It's about the pod destruction budget Morten is on.

D

Yeah I, don't think, that's that much to say other than still need reveal those two updates to the cap. Okay, those are two issues that we need to address somehow before we will be able to take it. I think construction about just the GA.

C

Okay, that's basically what I wanted to know. What was what what are you blocked on and what can we do to help so got it.

A

Okay and the next one is the staple set feature that can we like to present well.

C

I don't want to present, but I just wanted well. I mean I want to write a cap to go, do to implement biome resizing for staple set, but just want to see if anybody thinks we should not do that now. I think it's I think it's time like we, they have volume resizing for most of the CSI implementations and available on all major cloud providers, so I think being reactive to PVC modifications, at least for the size, makes sense. But we have to talk this storage about the expectations with respect to the pond lifecycle.

C

If the storage is resizing, they also expect us to restart the pod, for instance. That would be kind of in that. That's one thing: it's not I think as an initial start sinking PVCs, because we don't really sink with PVCs and like reconcile them inside a staple set control right now. We could probably do that so I like to put a cap forward and like this week, so they say by the time we meet again.

C

We should have a camp out and just have people take a look unless there's anybody thinks we should not do this now.

C

I'll, take that as let's do it, okay,.

A

Tap at some point for this feature: yeah.

E

I also think I saw the six storage folks asking about Bulmer sizing and I'm, not sure where they ended up. They might have been a PR or an announcement already. There.

F

Are cecilia kept for this? I think it stalled out like last year, though I'll see if I can find it there's a there's. I didn't see a cap there's a issue for it. I.

C

Didn't see, any I haven't seen an actual design for how this should work that address the concerns that you would have with integrate with staple set. But there is definitely an issue.

C

Or maybe it was a late cap I, don't.

E

Know yeah my just were to sing with six storage, but they ended up with it. Yeah.

C

It looks priests all about looking at the issue, there's open for it.

C

Michelle was looking at it at one point, but yeah.

F

Disrupted a much face, the PR that didn't get merged.

C

Well, was this a PR for, let me take a look. There was one PR that was trying to fix sort of an issue like if you delete the PVC and the staple set, because the staple set is insensitive to it.

C

But it wasn't clear: oh wait here we go okay.

B

This one I haven't seen yeah. There was a threat. I just found the same link paste it in her chat. That was like almost a year ago. It was from helmet from a hat, so there's actually two wings and it. Let me link the original thread from from say, gaps.

A

A

A

A

So, what's the status of this one? Oh it still active.

A

A

Can do want to take a look at this cap instead.

A

A

C

Yeah actually looks pretty good. We can actually work from this.

A

Okay, then, the next step will be continuing this on this old cap yeah.

C

I mean there's only a couple of things: that kind of need to be cleaned up a little bit, yeah.

A

Okay, does anyone have anything to talk about for this one?

A

If not, we can move on to the next topic, which is the cron job on GA plan, um so.

A

Currently, there seems no motion for this effort, so does anyone have interest to take this forward? I.

B

Have lots of interest a little bit less time.

B

I'll see what I can do, I can't promise anything. I would love to see this progress forward. So what the time constraints is probably the biggest issue. Barney was interested in working on it at one point and actually kind of put a cap forward. I was reading it um if I remember correctly, Jordan and I left there a bunch of comments there, I'm not sure I haven't looked if the kind of comments were addressed. I, don't.

C

Think I think it kind of stalled out there. So I was kind of interested if he was interested in picking it up or.

C

But if you can pick it up, that'd be great too.

B

I'm well aware of that, one.

A

Okay, we can maybe sing a flying for who is going to help for cron job ga.

A

So before we move on to the demo, does anyone have anything that they want to talk about.

A

You you I would take the silence of us. No, so um today, I'm going to do a demo for the new feature in GAE, it's called application manager, but for this demo I'm going to showcase what how you can integrate the sea caps, sub project application, CRE and use that to group your applications, and in this feature we also use another kubernetes project called customized.

A

um So let me start with this: I have a slide actually um so.

A

Let me start with tuck: you talk to everyone about this structure, how we do this application manager, so this GK application manager is on a controller, that's running in your clusters. So what the controller does is that it grabs the manifests from somewhere, for example, a github repo and then it pulls them and install them as application, and it uses application CRT controller to report the status of those applications and group them and then show the version and some other metadata about this application.

A

So we have some COI tool, called app control that creates and repositories using and customized format, and so for this get up space deployment. We allow you to create different environments for your applications. For example, you can create staging and production in this graph and then in the first repository you can. We call it application configure Pastore.

A

You can separate your manifests for each environment by folder and then in this another deployment repository and all the manifests will be rendered using customize and the manifest for different environments will be isolated and separated by using git branches. So you can imagine that, and the human is modifying the customized format manifests. You can have some robots does during those manifests, and maybe it's some scripts or something running in your pipeline. You will render it for different environments and then some other reviewers can review it.

A

So that's basically it um so for this demo I'm going to show you a very simple example app using this structure. So let me go to this so I'm here now in my app so I before this demo, I already started scaffolded some the repository structure using the COI tool. So I have this. This application repository here. It's called first app and you can see that I put the customized Yama file in the in the staging environment and if you click inside, you can see it's pointing to my base configuration.

A

So if you know customize, you don't know customize, it's just a you have a base layer for all your as the base of your app and then in this base layer. You I have nginx service and deployment and then, in my staging layer, I just put some patches, for example, I change. The number of because of my deployment from three to one and then in a second in the secondary, pass three. The deployment pass.

A

Three I have this first SAP deployment, and here right now it doesn't have anything because I haven't rendered and I haven't built those role manifest for my staging environment, but you can see here. I have a staging branch for my staging manifests.

A

So right now I'm going to start by adding another environment to my first data. So right now, I have the staging environment I'm, going to add this production environment so I add this production.

A

A

Okay, so, as sure does what happens is that it's, it just adds another customisation file under my production folder, and then it points to this same base and I'm going I'm not going. Don't you change anything in production, so the production will use the same manifest. That's the base layer and you also show some information for where this production environment lives and then I'm going to push this change to the kids repository and that change will be pushed here.

A

So if I click here I can see the production setup is here and then that's my staging and production environments for this app. So let's say that I'm now ready to release my app so I'm, giving it a version.

A

Zero point, one point: zero for this first app, so I do that by tagging it and then I push this tag.

A

You so after this tag is pushed I, can just use this command to tell on the tool to to the CLI tool, to render my manifest for staging environment by picking up that tag.

A

A

So the tool is finding that commit for the for this tag and then creating a pull request, so the poor request is created for the second appointment with the fully hydrated manifests. So why do we need a pull request? This is because that we want to have users to be able to audit their manifest before it's being pushed to the cluster, of course, and you can choose to bypass this, but um but on most of the time you would want someone to take a look before this beat. This is being pushed to your environment.

A

So let's say someone not me is okay: with this change to being pushed to the staging environment, they can just approve a merge, or maybe there are some checks here. That's and some robot that's doing the merging thing so after this is merged. um I can just apply this to my staging clustered.

A

And then on in the in the UI web UI there's an integration with application CRD. So if, after this first app is being pushed to the cluster, it should show up here.

A

So you can see that this app is being merged and then push to the cluster and versions and using application CRE. You can see on the status of my app by aggregating of the underlying component and if you are interested in what's in this application, TR is just um the version of this app is here and then all the components under it and the status of this app.

A

So I just deploy my app to the staging environment. The next thing I want to do is to promote it from staging to production.

A

Then so I can run a command to push it to production and again this will create a pour request.

A

So you can review the changes of the.

A

A

A

So in here it's using the.

A

On three replicas for a test equipment, which is the base layer and I'm, not gonna, go ahead and merge this. So if you look at the previous pour request, you can see that the staging is using the one replicas, which is what we specify in customize and then, if I apply this I.

A

This first step running in my production, cluster.

A

A

You know it's here in the production cluster and so um I'm gonna also showcase how you do roll back.

A

So I have the second app here um and also the second app here running in the staging cluster running zero point: one point: zero, so I! So um when you use get ups, usually you don't roll back.

A

In the cluster, you usually wrote back using by rolling forward by adding a new new tag and push make a new push. For example, you do some get revert and then do that roll back, but here I'm going to show you that when there's something happens, this zero point one point: zero version doesn't work and I want to just revert it back to the previous version without creating a new get commit.

A

So to do that, I can see what's previous being deployed, then I can apply to an older version and after that um this second app will be brought back to the previous version in the second.

A

Ok, so that's it um so.

A

Yeah, that's it so um I'm today, I'm, not just gonna Sh, just showing everyone that how you can leverage this urban area is open-source effort to build something that that's easy to use that you can have an overview of something you deploy from your git repo and how you manage your configurations for different environments.

A

Yeah! That's it does anyone have any questions.

A

A

A

See something from the chat.

A

There's a question about what about home.

A

A

As I understand, ahem doesn't support this get-ups flow and also it doesn't um helm, doesn't fully integrate it with application, CRT or customized. Yet two.

G

Things I can I can add to this real quick. This is Matt Farina. Well, as far as get ops goes. If you look so, the term was coined by the folks over at we've works and they created flux and they have a helm operator that can be used with get up a get ops flow. They figured out how to integrate it in that flow.

G

That's one of the reasons we in the home project actually haven't done more with creating controllers that run in the cluster, because they're already tools like that, so it can be used in the get-ups model and as far as the application CRD, that would need to be in a chart. Just like anything else. Home doesn't automatically create it for you, but if it's another package resource it will be deployed along with, like everything else, that's where those two things fit. I'm not sure how it would relate to this tool.

A

Thanks fat Oh does that answer your question.

A

The thing is from Omer: oh okay, good thanks.

A

So if we don't have more topics to discuss, we can and this meeting does anyone have anything that they want to talk about.

A

Okay um thanks everyone for coming today.

E

Thanks Janet bye, thanks bye.