Kubernetes SIG Cluster Lifecycle, 8 Jan 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: 2020-01-08 - Cluster API Office Hours

Description

2020-01-08 - Cluster API Office Hours

A

Hello and welcome today is Wednesday January, 8th 2020. This is the first meeting of close the coaster API office hours for 2020. This is a sub-project of a sick coaster, life cycle. We also have meeting etiquette and this meeting is being recorded.

A

So if you do want to to say something, please use the raise hand feature in zoom and I will make sure to pay attention and, if you haven't already done so, please add your name to the attending list in the agenda document and if you have anything that you would like to discuss, feel free to add it into the agenda and first up something we started doing recently is: if there's anybody who is new.

B

A

Would like to introduce themselves, please feel free, there's no obligation. So if you aren't comfortable doing that, that's perfectly fine, but if there is anybody feel free to say, hi.

A

Okay, moving on straight through to the demos and POC Sandra, you've got a demo for us today. Yeah thank.

B

You so I, don't know how much of a dent I mean. I should have practiced before I started this, so everyone knows about the Machine load, balancer kept that people have been working on, I, assume right, the or that Moshe has been working on the idea that it would be great to have a generic machine load balancer for deploying load balancers of different types, whether it's you know, nsx on vmware, f5, netscaler, whatever and so I'm sure in the issues in Kathy. You all have seen me crop several times relating to H.

B

A and cap V like VMware, does not have a built in solution for providing a an external control plane in point to clusters, so I've been heavily involved in or pushing for having this load balancer model.

B

So recently, a before word a little near Christmas I started working on this and and we actually merged an initial implementation of this into cat V last week and or I guess, yeah Sunday would Monday so Cathy now has a chai and I thought I'd kind of review. How it was implemented in cat be to hopefully show the way this load balancer model could work on at a broader scale and Cathy fort for other providers as well.

B

So what I, I guess I put the easiest thing to do, and just kind of be able to go over the CR DS real, quick. All I did. All we did was introduce a new field into our infrastructure. Cluster called load. Balancer rest all right, the this load, balancer ref or the load balancer ref points to at this time and H a proxy load balancer, the H, a proxy load, balancer CRD and the vSphere are the infrastructure.

B

Cr DS, that's V, star, cluster and vSphere machine have no relationship to one another other than this load, balancer, ref and I'm about to make myself a liar. Just so you're aware, if you do go, look at the code, the vSphere cluster controller, sorry, the vsphere cluster controller does currently watch H, a proxy load balancer, mainly because I didn't feel like implementing the generic external Watchers. That cappy did using the infrastructure ref, but there's there's a single o type right now and we're be planning to introduce NSX soon and then we'll introduce the external Watchers.

B

The way Cathy does but other than that, there's there's no relationship and so well. How does that actually work? Well, um everyone knows and Cathy. When you do a cluster reconcile right or in it from an infrastructure provider, you don't mark yourself as ready until your infrastructure, ready I always have thought, that's a little confusing because from a infrastructure provider a point of view, ready just means you ready, doesn't necessarily mean it for structure rated.

B

It's not called infrastructure ready, but an upstream copied the field they copy is called infrastructure ready up until I introduced this. This was always true like it. We added the finalized or true, because we're always rated. We had no infrastructure. However, now, if we have a load, balancer ref using an unstructured reader, we do to that. We do a couple of things. The first thing we do- and this will be something like a generic load- balancer controller- will do in the future.

B

Hopefully is it sets the owner of the load balancer back to the copy cluster and again, that should probably be something a generic load.

B

Balancer controller does that that lives on copy or maybe I'll, move it to a another controller here, but it just does it here, because it was easy to do because I have to read the data anyway, so it sets the owner and that's to ensure that the load balancer gets deleted after the copy cluster gets garbage collected after the Kathy cluster, okay and then the next thing it does is just again uses unstructured reader to check to see if the load balancer object is status, ready and has a status, and so that's sort of the only runtime requirement for that load.

B

Balancer. If you need to have a status ready, that's a pool and a status address, that's a string and that status address is used as our control plane in point, and if somebody has set a port on our spec, we will use it. Otherwise we have a default API endpoint at 6 443 and we law, ok, control, plate, endpoint discovered to be a load balancer. So again, we're no longer ready.

B

If there's a load, balancer ref set we're no longer ready until this load balancer has provided us an address, and so how does that with something like H, a proxy load balancer? How does it actually do the load balancing? Well, it's a pretty basic type, now I say: there's no relationship. Now the first thing I said: I'm gonna make a liar to myself, and here you think, I'm gonna make a liar to myself again I'm, not in the process of creating this I also created a pcard mcat be called vSphere.

B

Vm has no relationship to Cappy. This is just used to deploy a virtual machine and it was because I didn't feel like rewriting or refactoring a bunch of the code. I had that deployed VMs that was related to Cappy, like it would literally look into a Kaffee machine for things like the name.

B

I didn't feel like refactoring, all of that or copying all of that doing, a copypasta, so I was like okay, well, I'll just turn that into a generic provisioner, so it has a clone spec, because the H, a proxy load balancer, is deployed using an OVA on your V Center, and this is this is Chuck. This should look familiar to you. This is kind of like the cloud init data, the bootstrap data and the cube ATM config.

B

This is, if you want an optional ssh user to ssh into the load balancer you can otherwise whatever and again our status, and you can see that I mark it as this field. Both of these fields are required as part of the portable load, balancer, multiple model and it's inspected. So people know that it's optional, but it's really required at runtime. Okay. So what what actually happens? How does this work? Well, all this should look very familiar. You know we reconcile it.

B

So what is it watch so it watches vSphere VM, because that's how it gets deployed again, not related to Kathy. It watches a copy machine using this control, plane machine to H, a proxy load, balancer function and I documented what it does here. Basically, it says: hey, anytime, a copy machine is reconciled if it's a control, plane machine and an address and I can get the cluster and that cluster has an infrastructure s and that infrastructure F has a Bres and that load balancer ref points to me or something of my type.

B

I should reconcile myself that should be reconciled. So, in effect, it's saying reconcile me whenever any of the copy machines that belonged to the same cluster that owns me get reconciled any of the control plane machines, okay,.

B

Get owner cluster deletes.

B

I'll go over delete last, so all basic stuff, so the H, a proxy load balancer itself, has a cast a couple of Secrets as a secret for its certificate. Bootstrap data also its config.

B

This just stands up the OVA and then finally I mean this is all just H, a proxy specific stuff right create config secret, but it's all a load, balancer config. Here's the thing that the other load balancers might also need to do right. Reconcile back-end servers, get all the copy machine resources for this cluster. So this is. This is the part that may be able to bake me made generic no get the control playing machines, and then this is H. A proxy specific.

B

You know, add to the backend load balancer, and so finally, when all that's done right, it marks itself is ready and it has an address and it's like cool. Okay, now the VCO cluster can say alright well, I've got a load. Balancer I've got the address and I'm gonna set that as my control plane in point and guess what hey everybody I'm so copy can go, do its thing, and this gives us the ability to deploy a chai clusters with cat be in the way that doesn't tie the H, a proxy implementation to cat V.

B

It ties it to vSphere in a way because we're deploying this as an OVA right, but you could easily see how the NetScaler or f5 implementation doesn't actually need to deploy a VM to get a load balancer. Those would already exist, so their own controller would be draft simpler. It would just be well where's. My config, so I know how to talk in the NetScaler and then it would be reconciling my back-end servers, but what's true what's generic?

B

Is that reconcile back-end servers as well as I need to watch all of the control plane machines and I mean I could run through what it looks like, but it's just going to take a couple of minutes. I think it's kind of more important to review how it was implemented here, because, again, aside from that watch, everything is done with an unstructured reader and the H a proxy load. Balancer CRD has zero relationship to VC or cluster of Eastern machine right.

B

The only relationship it has is this load balancer rest, so you can easily see how anything else could drop in there and again you seen and I talked about it like the vSphere cluster controller is currently responsible for setting the owner on the load balancer, but again that would that would be a generic load balancer and the load balancer controller in the future and I only ate it here because it was I had to I had to get that object anyway, because I had to read the status, but I hope this is kind of evidence to how this could be made.

B

Generic and it's implemented in a generic and portable way so that other people could easily slot load balancers in that again only know about Kaffee machines right, there's, there's nothing in the H, a proxy load balancer that knows anything about cat V, so yeah any questions.

B

um If there are any questions, I guess I'll show this real quick. What does actually looks like in a test that infrastructure that the cluster spec load, balancer ref I mean it's literally like just like? It would be from the kappa cluster, with an infrastructure breath and again like this. This would be like that that load, balancer control or maybe a generic load balancer type. You could see how that might sit inside of the copy contributor.

B

Okay, thanks Tim, but I really I mean Moshe has done the lion's share of the design here, I I, pretty much stole everything from him and and honestly from a great design of Kaffee. The whole notion of just the rest, the the control plane machine to the load. Balancer function was fun anyway, but yeah.

B

So hopefully this will allow others to kind of see how this is done, an actual action you know and and and maybe start developing their own, but it would be great to kind of move some of the stuff into copy itself, at least like the some of the more generic actions like the load about setting the owner on the load balancer and the one other interesting thing, I suppose like for Andy, we assert that all H, a proxy load balancer resources to this are removed.

B

But of course that happens after we delete the cluster and I part. As part of this, you saw the vSphere VM CRD. One of the other things I did was I refactored the deletion logic to be more like copy. So we don't do controller owner references anymore. We just do owner references and then, when we delete anything I said I would show the delete. It's gonna look very familiar to you because we do very.

B

We do something very similar where we say okay, can we get it as of a timestamp delete, but again after the cluster is deleted, our tests still take like another 20 seconds, because now we have to wait for the load balancer to be garbage collected. If this, the load balancer ref for the notion of it, exists in an upstream copy that could be part or the last thing that you all do and we wouldn't have to wait for garbage collection anyway. That's it and the questions see.

A

Michael, you got your hand up.

C

Yeah, it looks like the slowed balancer. A reference on the Koster object. Is that correct? Yes,.

B

Sir, it's on the it's on the VCR cluster, the infrastructure cluster.

C

So it's provider-specific Andre, no.

B

Sir, the H, a proxy load balancer, doesn't actually know about the V, the V Stern machine um it when it does anything, it works with copy machines, but remember, like III, didn't modify copy, you know, I, don't have a local, they do have a local fork, a copy, but I didn't want to go, modify the copy cluster. We actually needed this, so I couldn't necessarily wait for Kathy to have something like this upstream. So instead I put the ref on the infrastructure cluster, but yeah you could have a reference like this on the cap. Equations.

C

Oh yes, yeah! That's what I'm getting I see it's on the vSphere cluster as opposed to just a whatever we're calling it neutral cluster. Well,.

B

The Cappy cluster right, but I I, didn't want to just build cat V with a outstanding fork of Kathy. With this you know, maybe this goes upstream or maybe it never does, but you.

A

Know p1 alpha 3 is is big enough, and so what we asked Andrew- and you seen in motion to look into- was doing this for the vSphere provider for right now and that the long-term intent would be learn how things went with with this vSphere specific implementation and then ultimately, ideally adapt it and move it into cluster API, proper and.

C

What is the mechanism for assigning particular machines to a particular load, balancer.

B

So the H a proxy load balancer anyway, it gets deployed every time you deploy a new vSphere, a new cluster with cat B, it's a 200 ish Meg OVA and they use its link mode. So it deploys in seconds so for every VC or cluster. You would have a very small VM out. There provides a load balancer, but regardless the actual implementation of how, with the question you just asked, it looks to say: okay well, I know, I know the copy cluster owns me.

B

So it fact, if you look up here, it will under its reconcile, it won't go any further until yeah an owner F has been set on it. That points back to a copy cluster, okay, so it's owned by a copy cluster and when it wants to discover what it should assign to its virtual server, its load balancer right, it says, give me all of the copy machine resources for this cluster and then filter those into the control plane machines.

B

So this would be I mean you could consider this implementation specific, but I I think this is fairly generic right. No.

C

I, don't think so, like our set up, you know we have the routers on workers and those are the things that need to be added to load balancers. Your.

B

Control plane machines, aren't your control, plane machines, no.

C

We have you know, routers, you know for ingress, know what ingress this.

B

Is just this is just the control plane load balancer, this isn't for any other type of load, balancing yeah.

C

That's kind of what my take away was in I was hoping for something more generic where you could just prescribe I want some kind of load, balancer and then I want to add whatever machine, she'll load balancer, and this seems to be very much more tightly coupled into various layers of the stack, rather than being like it's on a portable thing. Well,.

B

It's a machine load, balancer cut I mean it's for load, balancing the. What did I do did I introduce a bug. This always happens. huh There's a net, a GCR issue. That's not my issue! There's an GCR issue, that's interesting!

B

Yeah! So I, I and I encourage you to pick that up with the cap. Yeah.

A

So the the proposal is for a generic machine based load balancer that uses a label selector to select machines, I.

B

Didn't implement the label? Selector I cut a hard coded that to be just control playing machines, I skip that it and, in this case, Andy yeah, but you're right that that is what the cap says. Yeah.

C

I would be concerned yeah the tight coupling the control plane, we're talking about you know if this load balancer exists, the thing's not going ready till a load, balancer and building all this other logic into these other pieces. It just seems like I'm, not sure to you, know, seems kind of intertwined and I. Think like it just seems.

C

Like you just say: hey I want a load balancer, it gives you a load balancer and then, when a load balancers done it goes on the next step and later does those label selectors to add machines or whatever it really don't see like the need for all that intertwine right there yeah.

A

So this is a proof of concept and it will evolve over time. We do have a couple solutely motion and then you seen.

D

So thanks really good start and I. Think it's very well done so. I do have some comments, but I think we should rather move those or front. So the question is: should we create an issue in CAD fee for some of these comments or put them in a separate cap? That's on the one part, and then the second part is they were.

D

There was some discussion about a contractor e in Cappy Parappa, so I'm going to start working, the inner 60 load, balancer and I'm, not sure if you mentioned something about in the 60s I like the way you guys are on that yeah.

B

Yeah you seen has been looking into that as well. So maybe you two could sync offline. We have a need for that. So yeah, let's sync offline, on that, it's I want to circle back to the comments. I would say: let's focus on the cap that we have you know.

B

Vmware we have product needs around the load balancer, and so this will be what it is for now for cap V, but I think everything should be driven through the cap and then we could update this as we go along, and maybe this stays a POC for your cap. Maybe something else becomes the POC, but we could chat offline I mean for all warning, like I didn't check with Moshe before I implemented. This I just called it a POC of his cap because it was based on it, but he didn't agree to that.

B

So I mean this shouldn't be representative of anything Moshe decided, and so we could talk in the future about whether we want this to actually be that motion cool.

D

So and then the next part was: what are we going to kind of do with the contra bin Cappy proper? We still planning on putting something there, or should we move efforts to to disappear? See I'm happy to build upon this POC and and and move it upstream to cap, be prepared a later date, I just yeah.

A

Yeah I mean that's not on me. Yeah I was gonna, say I'll talk to Tim, but I, see Tim, saying, let's evaluate for alpha 4, so I would say you know for alpha 3 over the next couple of months. Whatever code y'all are working on, can either live in Kat V or we can create a contrarian in cluster API and you can go wild in there and then we can evaluate the proposal and official implementation, at least of the Machine load. Balancer generic controller for alpha 4.

B

Just FYI we I did create a contributor rectory inside of cat V. If we want to put anything there, because I had to create open API bindings for h8 proxy stay to play, I actually didn't want to use h8 proxy as of a month ago, but I found out, they had I thought it was jank, because I thought it could have to do everything via SSH, but they have this brand new data playing API server and I've been working with their engineers on improving it.

B

We just finished a PR for mutual cert off that doesn't require basic as well so um they're doing a lot of work to help as well give them credit.

A

Well, thank you so much Andrew and everybody else who's been involved. Thank.

B

You and by the way I, if you see it, is deploying a two node cluster. So cat V has a che. It's it's a long time coming all right. Thanks, I run.

A

All right give.

B

Me just a minute to get my.

A

Setup back so I can see when people are raising hands all right. We have a discussion topic about tilt cluster API performance, pre loading images and don't know who wrote this so Maya is that you.

A

Thought I heard somebody chime in all.

E

Right, um hello, there.

A

We go can hear you now.

E

Who knows who can say all right.

A

E

Oh shoot they can hear you. We.

A

Can hear you yes.

E

F

G

E

I think my audio is trying to come out through a microphone. So let me just fix that sure: hey hello, can you hear me now? Yes, or can you still hear me I guess so about that books hey? So this is Maya from tilt the Deb tool that at least some of you have been using to develop cluster API, which is super cool and I've got Nick and Dan here.

E

Also, who you can't see but I promise they're here, and we wanted to ask you all some questions about weather and how you folks have been using tilt to develop on cluster API and where we can help with performance and then experience that.

A

Sounds great, please go ahead so.

E

Is there just in case, anyone in this chat doesn't know who we are and what we do tilt is a micro service, dev tool. The idea is that we run all of your all of your Cooper neni, stuff and all of your services.

E

So you can configure everything in one place. You can run everything with a single command. You get all of your logs in one place, so you can tell when things are going wrong. You and any changes you make to your files, get really quickly propagated into your running containers. You don't need to wait a million years for things to build that sort of thing and later, since that, like I've, seen that cluster API already has a tilt file, which is super cool, so who?

E

Who here uses tilt for developing on Oster api.

A

I'd say a good number of us do so. Chuck did the first bit of work few months ago to get us started with tilt and then I recently got the tilt file put in place and did the support so that we have a couple federated repository. So we have a core cluster API github repository, and then we have individual provider repositories and I wanted to have a way so that we could just have a single tilt file that spun everything up together.

A

Even though not everything was in the same github repository and so I think I was chatting.

E

With you about.

A

That my on slack the other day, how.

E

Do you feel about the the the multi repo single tilt file structure? Is that feeling? Okay? Is it do you wish it were different? Is there are there pain points there personally.

A

I think it's great and I don't have any problems with it, but Chuck or anybody else. Who's been working with tilt over the past couple weeks, I'm into please yeah.

F

I think the way that Andy set it up is pretty cool. I haven't had time to play with it. Yet because I've been doing a bunch of non tilt related work, but I will I'm looking forward to using it more regularly to have better feedback I just added that chance to play with it too much yet cool.

E

Yeah um so I noticed your top file is pretty kind, specific and I'm wondering how what that thought process was like how opinionated you folks are about everyone who uses tilt to develop copy, should use kind or, if you, if there were folks, were interested in using other clusters to develop just because there's a lot of kind related overhead in your till file and one solution. Some of that would be just use a cluster, that's less of a pain so case about your thoughts there and.

A

Er I saw your hand went up, was that for oh I,.

B

Was just she went down a different topic, so I just decided to lower it and let this topic go first.

A

E

About that later, you're welcome to shoot us an email or a slack. If you don't get back to it,.

A

There's not I, wouldn't say: there's actually all that much that's kind specific, like there's some stuff about getting it preload getting the certain manager, I'm, just pre-loaded and I, know nadir, I'm gonna, let you speak because I know you're, not using kind. So yeah well I'd like to hear your thoughts. Yeah.

G

I I actually have a local clustered, I use and maybe because I have some other picked pink bits and pieces in there like lucky for logs and so I can search on them. Inca fauna, so we've actually made the kind specific stuff additional. So you can opt out of it opted into kind by default and then you can set. You can say, I, don't want the image pre loading stuff and that more or less worse than my cluster I'm add any issues with it. Oh.

E

I see, there's a pre. Let me just for kind boolean nicely done it I mean. What's that cool, that's very exciting and I'm glad that has been working mostly working for you.

A

Chuck, you've got your hand up yeah.

F

I guess I'm curious. Are there other kubernetes clusters that you would recommend or suggest using in place of kind.

E

I mean normally like, like we often recommend docker for desktop, but I just tried to build Cappy on docker for desktop and I felt, like my laptop, was gonna blast off so um but maybe a remote cluster like pke it'll, be like the initial pushes will be a bit of overhead because you need to like push stuff to a docker registry. But after that the the live updating is pretty fast, and so it shouldn't be too much of a big deal. If your cluster is remote and.

F

You're thinking of like a long live cluster, yes,.

E

F

Cut it yeah. My development workflow has definitely been more short-lived clusters. So if I'm developing something I'll make a kind cluster, do my work and then just get rid of everything. It's.

E

Just for the sake of having like a clean slate when you start developing next yeah, that's pretty and.

F

I, don't want my clique, my kind kind clusters, don't last across computer, restarts right.

E

Right right, it makes no sense um yeah, so I think our our team's next bit of homework certainly is to figure out how to speed up the the kind load process or, at the very least, give you some visibility into like yes, we're still working, we tilt hasn't just frozen an abandoned. You I also wanted to know as far as speed that I noticed, y'all have a live, updates, enabled which is awesome. How do you feel about the speed of those? Is it appropriate? Is it still too slow?

E

How often does tilt say, like whoops can't do a live, update, I'm gonna do a full docker build. What have you been? What have been your experiences with that feature.

A

Once I got through making sure that I had all the paths set correctly and the live update was getting triggered. I would say it's great. We were accidentally using the a flag for go, build to rebuild all packages, so live update was taking about a minute and then the deer got rid of that. So now it's like eight 10 seconds on my laptop. If I need to change something so I would say it's great.

F

And we're doing something a little bit different than I think what the documentation says so we're rebuilding the binary locally and then pushing the binary to the running image versus pushing our diff to the image and having it rebuild the binary on the doctor image. Oh yeah,.

E

That's a great as it keeps your images nice and light. That's right.

F

Yeah because I think there were something like two gigs.

A

F

I was about to being.

A

Able to do a build in the container cool.

E

I'm glad that's working well for you, I didn't I didn't get a chance to look. Are you doing that with a local resource, or how are you doing that.

A

Yeah cool before we do another topic and receive no drew him.this yeah.

B

Thank you. My I just wanted to say that it kind of it goes back to what chuck was saying so I think it's probably gonna be more of a reiteration but I'm part of a work stream team at the VMware, where I've got a couple of engineers and and there's some of their new to cat be and other aspects of kubernetes and I find. That kind is useful because it builds a habit right.

B

They have to learn how to deploy it over and over again, and that actually is useful for making sure they know how to work on the project versus. Oh here's, this long lift thing and then what they're really learning is: okay. Well, how do I interact with this long lift thing and maybe not the end to end workflow so I mean maybe it limitation of resources that people don't like to use kind, but the the other nice thing that we recently did with with cap V and this kind of relates back to crime.

B

So I'm saying this is: until is we used to always have to push images for pre submit tests on prowl, because we would then run them somewhere else to test them, but with kind you know, you don't have to push images at all, which is really nice for doing local tests, and it may not seem like a big deal, but it you know when you have large teams and that are decentralized and to be able to confine you know the test.

B

Workflow, that's repeatable, reproducible and behaves the same everywhere, but also doesn't really need a bunch of remote resources to work. It's actually really useful. So that's why I, like kind I mean to me kind, is like the single most useful thing. That's come out of kubernetes for developers ever tilt, maybe a close second I'm getting used to it.

E

Pretty cool, that's that's, really useful context and makes a lot of sense.

E

Neat did folks have anything else to say either about kind about the speed of live, update.

E

Cool um I noticed that y'all are running nan docker container runtime and are using a start. Sh solution to restart your processes is that working okay just wanted to check in because it's been a point of confusion for some folks in the past.

A

As far as I know, yes, Chuck you're, the one who worked with the tilt folks to get that put in for us right, yeah.

F

Yeah, that's been working. Fine, we haven't had any problems. They're cool.

B

I have a sorry I need to reach me.

B

When you say nan docker is a container D yep. Okay, I wanted to add that maybe the reason that is because the machine images that we deploy with capi are all based on container D, so I think it's probably good that that aligns with what we're actually encouraging people to do when they build their machine images.

B

Maya look, I, don't know if you heard that can.

E

Yeah yeah mute, uh am I unmuted now yeah great cool. um Let me just get back to my notes.

E

Yeah mm I'm, not sure I had anything else, but Nick is there anything you want to get with hardens cool um unless anyone has anything else to tell us or requests about tilt. While you've got us on the line.

E

We're super jazzed that you're liking developing with tilt and we're going to be working in the next couple weeks to speed up developing cluster API with tilt, because we think that addressing some of your pain point its will be a good like way to decide on things to focus on that will help other folks that would like bubble out and help other folks using tilt without us getting totally bogged down and the multitude of things we could be fixing about the product. So I guess anyone have anything else for us. Andrew.

A

B

Yeah, it would.

A

B

To see some integration between y'all and ASCII cinema, if you've ever seen that or work with that.

E

um I'm vaguely familiar say more: what kind of integration you thinking I mean.

B

Just the built-in support for yeah built-in support for, if I, do a demonstration with tilt like after the fact I can say. Oh and I want the the demo from that like I. Just want that part part of tilt could be enable ASCII cinema. So at the end of everything, you've recorded everything.

F

Interested there is a snapshot feature that kind of gives you all the logs of the existing like environment in like a shareable space, which is pretty cool, yeah.

B

I saw that I think what I and thank you and I think what I meant was well ASCII cinema also has the here's how you upload it and here's the shareable feature so well. I could take those logs and then share them. I, don't know I'm being lazy like it wasn't easy, but me till it's an easy button in a lot of ways. So the more easiness you can add to that button, the bigger the bigger you can make that button, the better. It is right, yeah.

E

We saw since cool thanks for the feedback, all right, I think that's all we got okay.

A

Great well, thank you all for joining and bringing up tilt I. Think it's fantastic and looking forward to seeing what else you can come up with sounds like a plan Thanks all right. Does anybody have any other topics before we move on to backlog, grooming issues that don't have milestones.

A

Okay, let me switch over as usual, I'm going to start on the bottom and work my way up, so we're gonna, look at the oldest ones. First I did add a comment to this one about not using load balancers in alpha 3 and I said I was going to add a comment. I don't know three weeks ago, and it slipped my mind. So my recommendation here, if you're, okay with it andrew, is just to close this because I don't think, there's anything. That's gonna! Stop us from continuing to support single machine. That's.

B

Fine and in fact cat I put this in the notes of the doc, but just another and give a shout out to chuck cat V and, as implemented I think were the first ones I said of cat D, the e to e framework, but Chuck develop along with others. It's part of every presubmit and we are using H a for that, but we are doing a single node without in load balancer. So yeah we'll know real quick. If this stops working, because our e to e will break.

B

But yeah real big shout out to chuck is that that works so well. Cool.

A

Thank you, okay, submit copy providers for given release for scenes. You have conformance verification, sounds like a great idea.

A

I think this is probably something that can happen. Independent of milestone, like it just happens when it happens, party wise, I'm, thinking, important long-term for now, because getting b1 alpha 3 done is more important in my book and I'm gonna just put this in next and we can.

A

If anybody is interested in I, don't know if we're gonna, try and automate doing some of this I know, you have to open up a pull request to one of the CNCs repos, so I, don't know how much we'd automate, but if anybody has any thoughts, please feel free to jump on this issue. The comments.

A

So we haven't awaiting more evidence about cube ATM in it. Not recognizing docker installed, be a snap I think this is a cube ATM issue and the reporter never came back so I'm, gonna, close and Andrew good are.

B

We supporting docker on the machine and we're not building them with docker but I'm curious, and are we supporting docker with the machine images if people replace container D with docker, and is that a question for Cappy or is that some other copies kind of been the upstream discussion forum for the image builder anyway? I.

A

Honestly, don't know the answer to that question. Well,.

B

That I asked obviously because of this question, because this is docker not continued days. I.

A

Would think that we, in theory or I, think support is a slippery word, because we're an open source set of projects and so I think it's like? Does it work with docker versus container D versus something else, and it could work with docker? It probably should work with docker, given that kubernetes works with docker, so I think that it's kind of it's it's hard to say like if you can produce an image that can run kubernetes correctly and it works with docker. Then I don't see why it wouldn't work in a cluster API setup.

A

If you know what I mean.

A

All right is that cool and you yes, yeah, go cyclone lint errors in mango I'm, pretty sure, with the split that Chuck recently did where we now have multiple management processes that this generally went away. I am so basically. This is something that, if you have enough web hooks and enough controllers and you're using queue builder to generate your main go and it keeps adding all the set up with manager calls. You basically end up exceeding the go cyclo limitations for linting and I.

A

Think we we're generally saying well, this is kind of generated code, so we can just go ahead and say: don't rent this for go cyclo, I'm tempted just to close this. Unless anybody feels strongly that we need to fix it. I can.

F

Update the ticket, if you assign it to me, I, can just run this against current made because it did change quite a bit with the manager split. Okay, thanks Chuck, which maybe I should have PSA at the beginning of this meeting, but totally didn't think about it.

A

Yeah so in terms of that take a minute and just say why you do he did yeah.

F

Yeah, so we split apart the main cluster API manager into three different three different managers: there's the cuvette diem, bootstrap manager, the comedian control play manager and then the core manager, which is the same as what you have now or what it was before. It was all bundled into one manager, and now it's split into three managers and we're able to ship three different images with three different sets of VMO, allowing you to not use the coop ATM components. You choose and use your own blue sky provider.

A

Thank You Chuck all right should the cluster API iPod have a cube, barback proxy. Sorry sidecar like the cubed am bootstrapper, does I'm thinking. Probably yes, I, don't remember Chuck when you were doing cat PK. Did you put the sidecar in or do you know where that came from I.

F

Don't know where that came from I've seen it was I. If it's there I'm guessing it's, it was generated into place.

B

That's that's true. Different versions of cube builder I've used like what seems to seem to get this weird code that shows up now and then and I could tell it was just generated code and some point can't be at the same thing that or somebody copied a template from another project, and somebody put it there.

A

Yeah, if I recall correctly, this key bought proxy sidecar protects the metrics endpoints from anonymous access all right. So.

A

I'm gonna put help on this one and.

A

I'm gonna go with soon because it is a security related thing. I'm gonna put it in the milestone for now, alrighty close to cuddle will be to support for additional back-end implementations for provider repository, so breech CEO or you familiar with this one.

H

This person is asking to have a different back-end and not only get up, which is something that is somehow, but it is proposing to do these using library. I, don't remember the name so.

A

Yeah all right well, I'm, assuming this was not anything that was in scope, I mean basically what was in scope in the initial proposal was you can pull it from someplace like github or you can override the location of the yamo and whatnot with a config file that says, go get it from some alternative place right, yeah,.

H

In an initial proposal, we are going to support a github and local file system so and- and there are PR for both of them, additional back-end is for my opinions, our facilities, okay,.

A

So I put this in the backlog and it's in the next milestone. Thank you.

A

Cubm bootstrap provider does not take into account the cluster API server port I. Think I saw this over the break or when I got back, and there was some question about taking a look after the holidays. I Andrew go ahead.

A

You're, muted, Andrew.

B

So you say you can assign it to me if Chuck's busy, I literally just did this for Kathy but I, don't know it should be a standard thing. I I'm, using the port as like the vSphere cluster control plan in point port as a possible default. So you could define that, but not the host, and that would specify the port will.

A

Actually think.

B

A

Is specifically talking about cluster spec cluster Network API sir report? Oh that one not the control, plane, input. Okay, sorry.

B

A

Gonna put this in backlog: I, don't think it's anything. We need to change for right now and we can look at it when people have time.

A

A

We need to move the machine deployment annotations to the right API group. We have help on it and cleanup and we need conversion logic. I think this Vince did you move them at all or they're. Still bold, I didn't okay, so I'm gonna do a long term on this and we have helpful on it on there. So that's good.

A

All righty um to find constraints for upgrades crossing cluster API versions. I know you and I talked about this treaty. Oh, is this without getting into too many details here terms of priority and milestone. Is this something that needs to be solved for alpha 3, I.

H

Guess so, because we knew that we went after through the first implementation of cutter upgrade, so we have to choose how to make this. Okay.

A

A

And next one is also you've be to add label to provider, components for easier cluster cuddle adoption, yeah.

H

I would like to cover the use case that an operator that is an existing management cluster I want to start using the cluster cutter and for doing so I'm suggesting that they, each provider, has a label in the next release that that we will be srida from now till the v1 alpha tree gets out, and so basically, when we one alpha, 3 gets out most likely hold the existing cluster and later already a label that makes that option almost automatic.

A

Right, so this is just the provider components like the deployments and our back and whatnot. This is not clusters and machines and everything exactly it is only for the provider components.

A

It seems reasonable me, I mean it mostly just adjusting the customizing amyl that we have right exactly okay, why don't we create individual issues for the various providers that you have listed here and they appropriate give up repositories and make sure that it's clear to do it for the alpha-2 and the master development work, so we'll have two different PRS for it for each one.

H

A

Think you can just link back to this one for all the other ones. Okay, I would take open a day.

A

A

Yes and implement a shared framework for creating test objects also for PTO.

A

B

A

A long term next, yeah.

H

This is a follow-up of a comment on one PR, basically in one PR I'm implementing a test framework that generate cluster of just Custer machines, we fold the relation all the brain and this disk. This is now implemented mastery for testing move across the cattle mood, but JSON of several that disk, and this might be useful for some also for in another place. So this is the they should tracking that the idea. Okay.

A

Yeah and I have some previous experience with other projects, doing like basically creating a builder for individual things so like in bolero, we had a builder for backups and a builder for restores, and you could just say, like I'd, like a new builder with this namespace, this name these characteristics and at the very end you just say build, and then it would give you back the object you wanted. So maybe look at something like that.

A

A

Control plan should scale up serially, not in parallel, and this is control plane. This is needs to come in and it's in the milestone and then bump concurrency of the doctor provider workers to ten Oh.

A

This already has last Oh Thank You Chuck, okay, so we are done I think, there's not really any time for anything new, so Happy, New Year everybody. Thank you for joining today and see you next time.

A