Kubernetes SIG Testing, 2 Mar 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: sigs.k8s.io/kind 2020-03-02

Description

https://docs.google.com/document/d/1b9Ppm7ZT_tMWRs5Ph1zGJJKb5nF9c3ZHzMwg1olJIrc/edit#

A

B

Hi I also am time down by my laptop Charles I'm time bound by my laptop battery Oh.

C

Has been interesting, No happy, Monday.

B

Like, oh so I guess, cube con is cancelled, but also maybe we're trying to rework the contributors. I, don't know what.

A

For amia, oh my.

B

A

B

For you, for you, Amsterdam yeah, the officially is not canceled, but like my employer, Google, just like canceled all travel and the same here. It's done so also I was discussing folding this meeting into sync testing, which actually hasn't had agenda items in a while, and this one's been kind of like infrequent ish, like as things have gotten more stable.

A

B

On your shoes and stuff so well.

A

Yeah and I'm, seeing I'm like well I'm, the only one in here and there's no ongoing calendar of invites and like things, must be stable. There like stable.

B

Enough and the remaining things that people want are usually like they're much more than like a zoom call. They're like I'd, like clusters to you, know just magically pop back when I reboot and the doing that is going to be like involved change, not just a like Oh, we'll hash. This out in, like five minutes in a video call, gotcha.

A

Well, I appreciate you making the time here anyway, even if it's just to you and I, we can talk about the issue. I filed yeah I'm sure you had some background context over it with Duffy yeah.

B

We'd actually been discussing the I get intermittently for a while I, definitely think it's a cool cool idea to approach I think I mentioned most issue like that I mean so we won't be able to like switch everything to see all right sure, like a bunch of things, don't and going in fact it pretty much is Padma and crow. Oh sorry, oh and container D they implement it, and it's also actually alpha, which is fun and until very recently, devs and I had fixed this.

B

It was a you could only get the API by like forking kubernetes, it's at least staged now for, like the API.

C

Spec, but so it seems like it makes a gigantic bundle at the end of the day, yeah.

B

So I think what we I think. What we should probably try to do here is leverage Franco Craig could also has some nice things like you. It has its own config, that's well known so, like which end point do I use and stuff. So you don't have to like reinvent the wheel there. It should look pretty similar to like the docker apartment, one except we'll have to write like creo, spec or CRI spec like pod JSON or something for each node, because you know Craig home run is not a nice like tons of flags.

B

You can do this too, that you, you have to like given a spec, but that should actually make things mostly easier for us and we shouldn't actually need the. We shouldn't actually even need their library for that, because, despite being alpha that, like they're, pretty wary of doing any major changes to the types and if they do we're, gonna need to dispatch over versions anyhow, so got business in the API.

A

B

Libraries are all the crew Bernays libraries are like, so you import it at a version. Oh you wanted to.

B

Make two binaries or something so we can make in binaries by having incredible, binaries and I think that should work. Fine Duffy was also mentioning the network. Stuff I, don't think, there's gonna be a blocker, so much as with darker we're, probably gonna have to start leveraging some of its network quirks to fix like Coast, reboot and stuff. So it's just gonna have to be like darker and then everything else is just gonna, be like you make a good container. It gets a random IP.

B

We find out what the IP is and then before we like bring up kubernetes. We coordinate that everything knows. Ok, these IDs, the we're gonna, probably move away from that in the darker back in specifically, um because there's no way with any of these to not have the IPS get randomized again, you like to restart the machine or something and then that just breaks everything, because everyone's, like oh I'm, I'll, give you this in point.

B

If it's not there, whereas what docker it looks like with some hacks, we should be able to actually use some magic DNS than those which container is what and I looked at a couple ways we could like implement this ourselves, but it's pretty pretty messy so like we can, like you know, we could try to like run our own DNS. That's like aware of the containers or something but getting all the network right is pretty polska I think it's fine I!

B

Think for probably the cases where you're gonna do something like like Duffy was talking about like container dapi implemented against, like like these forums or something like that for that sort of thing, I, don't think you're going to be doing like machine reboot or changing IPS or whatnot. So it kind is still mostly targeting ephemeral test clusters and we just want to like be able to upgrade the case a little bit for all the people are using docker locally.

B

We should be able to set restart on the container and use like DNS, and we don't have any weird environmental stuff there, actually that, because it's us, it's actually fixed in docker and then that also solves some of the awkward issues with like if you're, using like docker for Mac, then like.

B

If you have like a loopback proxy on the host, then how does the container actually get it? Oh? Well, it all just goes through this magic when the docker sets up itself, and then we not worry about it. There's gonna be a little bit of disjointedness, but I don't think there should be too bad and we do intend to try to keep the augment functional for now, and we are looking at like ignite, may be pretty much anything that lets us. Take it docker image and run it we're looking at as like.

B

Would this be useful and ignite is interesting, because it's Siam's.

A

What's ignite I'm, not familiar with that, it's.

B

A project we've, um it's kind of weave.

A

Ignite uh yeah.

B

Yeah, so it's changed a lot since I last had time to play with it, um so I'm not sure how much it fits today. But the idea is like okay, but actually we get quick beams is pretty nice, so we may may do something like that as well. I think that also may or may not actually PC right, but as long as we're like you know, it doesn't really add any binary weight and like it won't be fully supported, probably at first but it's there.

B

A

Well, that makes sense- and that's a very similar sort of a use case to what what we're trying to do here so, in particular on the product one of the product managers for fusion and workstation right and what we did was we actually ported container D to the Mac and we're sort of making sure that you know we adhere to the spec and we play nice there.

A

So, from our perspective, it's like, if we can get the C air I plug in to work and then kind can leverage the CR I plug in then there's like a you know, a happy. You know fitting up together there, but it one of the questions that we had was: do we focus more on making that CRI plugin compatible or should we focus more on like getting a cry, cuddle story working better um and we don't think we get an answer on that. Yet, but yeah yeah.

B

And so, like anything that I mean from kinds perspective, what we'd like to do is invoke a client binary that has all the details of like how do I talk to the API and stuff, and just has a you know like fairly simple set of commands for like a list and whatnot, and so it's both CTR and Craig cuddle offer that we'd be pretty fine, adding like a CTR one I mean it wouldn't really be any different from like we have a special department, one as much as PubMed wants to claim it's docker compatible.

B

It's not quite so like we already today have like.

B

There is code for like invoking apartment and handling that properly, and those interfaces may change a little bit but they're fairly straight for me to be able to label things as belonging to a cluster by name and the list for them, and we need to do a few things like find out the IP address of the container or like run a command in it, and then there's one pretty nebulous function that takes the spec for a cluster and actually makes the containers exist that one's a bit gnarly, but it at least means that, like whatever you do to make that happen, do not care just like make it happen.

B

Yeah just accept this file and then, when it's done, giving an OK return or not great, and then after that, other code can be like. Okay, I want to list the nodes for the cluster it get handles to live in like executive, stuff, okay, okay, it's pretty simple. Let's.

A

Go we've done a lot of the plumbing, so I feel like we're. You know we have some options here. Go ahead. I'm gonna run for my church.

A

Now, if we, the zoombini, doesn't and by the time he gets back.

A

B

A

So in thinking about it, then maybe it sort of makes sense for us to focus on either CTR or cry cuddle, and we can get that out sort of on our side pretty quickly, and then we put it together a longer-term plan. Since how do we sort of use a lower-level provider like the CRM plugin, or does that not make sense? I'm.

B

Not sure what is lower level I think, like I, think those should actually work. Pretty fine I think the only thing that might get weird is like, if it's like, say we actually target container D directly. I could imagine someone having something like regular container supported and then also like magic containers using runtime class or something well. It's so like I could see like if I were building something like this I might say.

B

Oh you have to like opt into this runtime class, so you can have like hybrid nodes, in which case somehow we need to tell kind that we want this I I, don't know what that is supposed to look like, but otherwise, like I mean kind, doesn't care too much about it at that detail. It just wants to create containers running this image, and it has a bunch of parameters for that. But almost all the parameters are up to the provider so like.

B

If you know your providers, the VM, you may not even need to do all of these mounts and things and the other thing that a provider can do is say. Oh well, since I'm doing VMs off in vSphere I can't, you know, bind him out to the hosts.

B

So any extra host mounts if I see that config I'm just gonna, throw an error provision, I'm kind of expecting that like if we do cata firecracker at some point, that will be the case even locally, but you just can't do buying bounce so we'll just reject the config with an error and it's fine and it's just procedure ray and then all of all of the rest of it all it cares about is I need to know like where this thing is like.

B

How do I talk to it and I need to know I want to run a command inside this machine right now, it's assuming as route but like sure, because it's some throwaway nodes somewhere and it's gonna run things like keep Adam so and some other straightforward stuff, gonna run a couple of cube, cuddle commands, it'll probe some stuff might grab logs. All of those things are actually set up in a way that, as long as you can like pipe input and output to running a command, it doesn't really care how it happens.

B

So we were even kidding about like we very likely won't do this, but we totally could just implement like GCE support, because if the life the interfaces are written, it doesn't really care about that. It just needs to be an exact thing so like we could use SSH I.

A

Searched Shane's, it's abstracted such that you know, we've got a lot of options. It.

B

Pretty much just needs to be able to pipe things in and out of commands running on particular nodes. It needs to be able to list nodes, delete nodes and it hands the light closer config and says make it, and we have some utils to like help with that, but it all it's doing in that is making the notes exist in their own bootstrap state. And then all the bootstrapping logic just takes the list of nodes and, like runs commands on them.

B

So we also currently expect the hostname etc to match up, but we don't actually depend on. We don't currently depend on doing something like asking a machine from inside the machine like what is your hostname, but we do sort of expect that, like they're, like the containers are named, and we currently make the kubernetes name and the container name and the host name the same because easier for users- and that makes sense a docker I, don't think much would break we.

B

We care about the name that the provider tells us but like that, could be a mapping or something something to keep it track of and you're in. That's just gonna be something like a Craig, Cole LS and some filtering on the output to only return like the ones matching that cluster and that's a label so like it should map pretty well because we're not so far we're not leveraging much darker specific, and we only have one or two Dockers specific work.

B

Arounds going around like again this that you can tell kind I want this node to have these extra bind mounts, which is useful.

B

Docker only accepts absolute path so, like we have some logic to try to like relative path resolves like, but all of all the things that are like that are pretty much all pushed down into that like provision function and it's up to the provider to say I, don't support this or try to support any of the config fields and all it's expected to do is succeed or not, and then you're expected to be able to like list the nodes, life and then exec stuff.

B

So it should be like they should be pretty doable against Craig huddle or CTR, or even some bespoke binary. Like doesn't really matter, hmm I mean it would also he. It would be probably not a great idea of a doable to go like the pod member out and write a fake, docker, CLI or something I think it would be much saner to just wrap that with a little bit of coding kind.

B

What we do want to avoid, though, is something like importing the actual CRI for now, because it's stable and because we are in general, avoiding picking up most dependencies kind is really easy to embed and also for the purposes of testing kubernetes. We don't have like dependency cycles this that we only depend on like one library, they actually comes from the kubernetes repo and that's just the version parser and it's it's staged so we're not depending on Kate's directly and while that does get periodically published by them. It's we're not super slick.

B

We don't expect changes to that code and we wouldn't be super sensitive to them anyhow, and it would break everyone if versions changed, so that one seems fine, but otherwise we have wait depending on core kubernetes things in the main library so that there isn't this.

B

Like dependency cycle of like how do we test the changes, if we're also going to use them later, we do pull it in in things that are not in the main binary, though, like we have a network runtime for like really simple notatum networking that we run by default, that that does actually use kubernetes communities proper, but it spilt them frequently and it's a client. So it's not doing anything more than like listing nodes and finding out their IP m so again, like basically can't be broken, and it's totally.

B

It's like in a separate, go module and everything in Sandy that may I think so far. That might that probably will even work. But if not worst case will just say, like this provider uses different images, or maybe you there's like a field where you say, like don't turn on the default CNI, you can just say like with this provider. You need to like we require that that is set to true and we're going to like. Have you install like whatever you want, calculus um thing or.

A

Like you know, we could ship and bundle our own with it and then refer to that yeah.

B

Just then it will have to use a different docker image, sherry, edits, yeah, but I think.

A

B

Think it's actually even reasonably likely that the the current won't work. It's not super exciting it pretty much just uses net link routes to say like this node has this outsider. So if you want to reach any of these IPS just go by that code, IP and otherwise it's doing very standard, plug-in stuff host local I Pam and like p2p, for the local net, so like those should work anywhere.

B

A

Well, this is encouraging I feel like there's a you know, like I said, there's a short term like you know, we can wire it up if we pick I, gotta, lower or CTR, and then work on a longer sort of a road battle on the provider together. Yeah.

B

I I actually don't expect a whole lot. Turning to changing the provider, I think it's going to wind up having to change when kind changes. What functionality we offer like. If we change like how we support the images is something we might want to like have more details about a provider but I get a pretty high level. The main thing that's gonna change is the is the configuration that's being passed to the tool? Okay, the, but like the high level of like here's, a configuration. Give me some like there nodes that I could finish.

B

Bootstrapping has worked pretty well and then everything after that is like okay, I want to handle to a node, so I can exact stuff to do the bootstrapping order. Okay, it's time to clean up, I'm gonna delete the notes now yeah yeah, here's the list of nodes I want to believe and all of those require, like being given an instance from the provider, so you can hide extra information inside the go code there if you need, but currently for the actual providers.

B

We just have the name because that's sufficient but like if you wanted, since everything will actually go to your provider and make a list call before it does anything with a note you can, you know, do whatever you need there, yeah that makes sense so like I, actually, don't expect that to need to change too much and I. Think if we can't accomplish this with like Greg coddle, then we have a real problem with, like the eCos sure yeah.

B

That makes sense, and if not also, if you're specifically using container need the container D API is really nice to work with and they have their own client binary, which we actually used inside the notes and we've been pretty happy with that. Yeah.

A

We so we should tech preview of fusion, and we have this thing called Project Nautilus, which is you know? How do we get containers to all work and whatnot and part of that effort was you know we have our own binary for container D it's running inside of you know the package and whatnot we create a nine PFS, storage and wire it all up, and so yeah I feel like the work that we've done to sort of get that support and that working you know said.

A

Hopefully that sets us up so that we can take the next step here and then I guess the end. The question is like: how can we give out and contribute something? That's not just gonna be for our own interests, I mean I, you know yeah. We want to get our confusion in a workstation and that's cool I kind of don't care about vSphere, but I. Don't want to just do this for our own sake. I think that there's a meta value here I mean.

B

So if it's supporting like a standard API like container to your crank at all um I think Craig Cole is somewhat more generic, but I would be pretty fun with continuty. If we get container you that also opens us up to options and CI of considering like running against that as a lighter thing that starts faster and stuff like, especially when we don't need to build the images and we don't or like we don't expect to need docker at some point of.

A

B

Like the build steps, then we can offer it like, say other teams, or even some of like urbanised own issues, to like try using some of these other ones. Another one that I want to do at some point is: is some vm type options for local developers to say, like okay? Do that? Obviously we want to avoid entirely being like a product but having some validation thereof like what does this take will also be really useful to know so that we can say.

B

Okay, if you want to go your own, you can like figure out how to get like cata working or something and we're enough. I expect that's gonna, be a lot more involved for users and it won't be available back, but I would like to be able to point to like okay, you want to run a local one with like, like you like kind, but you kind of need the isolation of VMs I think we can still do it pretty fast, but we need, like you.

C

B

Integration, it looks like most of those things are going to be inevitably through container d, since it's just a like nice, extensible container runtime, it's premature and well-maintained yeah so like it looks like even if we say okay, we want to do with like cata as an auction or or I. Think even ignite has rebased onto it.

B

Having something that's able to target container D one way or another is going to unlock that already at first, like I, think most of the work there is just gonna be: can we target container D and then we'll just probably need to do kind of like we did with the pot man and doctor one and take that provider and fork it a little bit and say: okay, they share some utils but like when we're using fusion.

B

We can support these options and when we're using data, we have to support only these options, but like a huge amount of how does this work is going to be the same code and probably we can actually factor it into the same code, nice, so I I, actually just think that, like implementing this will be useful as long as it's like I'd be willing to accept it. But it will be nice if it is like. You know, container D, API or something instead of like a fusion API yeah.

A

Like and that's that's what I don't want right like I, definitely want to make sure that a there's, a dependency that isn't inherent to a commercial product and B. You know like I, said I.

A

Just don't I would put this I mean yeah I manage this product and these product lines and whatnot, but I've been an open source guy since, like 99 man and like I, remember back in the day when I know, I needed help with like open BSD and the only way to do that was to email, Theo to rat-like, and he is awesome. He was very helpful.

A

Oh yeah, that's a total bug in PF, your take mice, you know I fixed it, you get an advance copy of it whatever and- and you know this so for me- I have this I've always had this altruistic. Like you know, we I ran a data center in trial for a little while, and it was all on Gentoo Linux. So you know I like doing things the hard way and it's just been a community that I've always wanted to give back to.

A

But me not, being you know a strong engineer, you know I, you know we wrote this v cuddle thing. I kind of hacked together, like the original command structure, then I'm like okay, guys make these commands work now and off. They go right. Yeah.

B

Yeah well, I I think this is gonna, be pretty useful. I'm gonna need to Kim coming for the moment to focus on Tucker it's where most like. We have a pretty good existing user base that needs helping and right now, I'm. Actually, specifically looking at like increasing the tests we can run, but at some point I expect that, like it, there are loads of testing or development of kubernetes itself, where it's really gonna make sense to use one of these alternative.

B

Runtimes, no I'm, just keeping our options on the table, even though we kind of went in with the name I've dropped most of the acronyms anywhere like we're. Definitely not trying to be comfortable that, but we are trying to do, is avoid being coupled to any of them too bad. We're going to allow like a tiny bit here.

B

I think we're gonna have to leave an option to specify which darker network you want once we start using one, and that will probably just be a field that basically every other provider is like nope boy or handle.

A

It on your own, if you're.

B

Careful yeah, but like some but like some users are probably gonna need to be able to set that so we'll, probably wind up having that one field. That's somewhat darker specific, like for most of the other providers, they're gonna use, C&I or something themselves, and it's not really gonna make like there isn't a network to pick like it's the network, so I'm expecting to have to go into like say an apartment provider and just be like it.

B

Doesn't it doesn't support this or something or it behaves differently, but that's fine through an error like it I think we already expect that and other than that I really like that one's a bit of sticking point cuz. It's gonna really improve the behavior for most of our users, but if it, if not for something like that I think we can keep everything pretty generic and we've even modeled the config, where we can after like see right types like for the mounts or port forwards or something they mapped pretty closely to them.

B

They're, not actually the upstream types but they're like you, a.

A

B

Is pretty much just like typecast to go from one to the other, so.

A

That's cool so I guess the other question that I would have is like. How can we? How do we help? You know like I've, got a bunch of guys or like where do I start writing code yeah.

B

I, don't know I think I had this problem a lot. Everyone like wants to on this project, because it's popular but like it's not a massive code base and like I, like I'm, active enough on it that like when some little bug comes up, I, usually just squish it same day ciao. So that has left like some kind of gnarly stuff that doesn't really need code so much as deciding how we're gonna do it hmm like the host reboot thing is one of the biggest sticking ones for people. I.

B

Think I've actually got that now, but it's been entirely like which terrible trade-off we make here in terms of like we need to muck around with the network and stuff. It's gonna be weird you're, definitely not supposed to be using these things. This way, how do we make this work? The actual changes are probably not gonna, be much good like the whole repo, like maybe like 10 K lines or something like it's, not a it's. It has never been much code and we intend to live with it.

B

It's been like oh no you're, trying to like run darker and darker. What like awkward specific hacks, do you need to employ to do that? I'm gonna, try to come up with some more I've. Actually, I've got a teammate here it popped in and on the call quietly that is in a similar boat like wants to contribute more and I've, been like here's an issue, but keeping the keeping them up enough has been hard even with our like existing ones. So I'll.

B

Definitely you know if I, if anything like obvious, comes up like oh here's, something I'm working on weight or that we want and then is outlined, but we're not working on, but mostly a lot of the issues we get wind up, not being kind issues. It's like people. Misunderstanding like we're using container tea as a rent, I'm expecting darker or like they don't yet understand, kubernetes or things that are infeasible like people want load. Balancer support, it's like sure.

B

First go to docker and get them to like allow us to reach containers by there, peas and that will do all kinds of cool things and until then, like you get port for its because that's what docker has so like. Even that one like we've looked at like could we build some crazy hacky thing, but it's so far, they're all so happy that like falls.

A

A little bit outside of the scope right, yeah.

B

So like I, there are, there are things like that, like they're definitely features that we still want that we are not working on, but that's because they seem somewhat infeasible, at least with the current backings, like we like darker on linux shirt. But as soon as you get on Mac or Windows, you didn't the networking restrictions really block us from building those last couple of things that people want. It.

A

Could definitely be an area where we could really help. We have our own networking stack and it's the way that we're kind of delivering container support right. Now, it's it's a little different than how docker does it, but I could see where we can really help with that. But anyway, yeah.

B

I think the problem has been that, like, unfortunately, like a lot of people are, you know how doctors open source, but then they're, using like the darker for Mac or darker for.

A

Windows and those are the.

B

Stuff they're doing with their network stack is restrictive and we haven't found a way that we think is good to work around. That really was one it's like I want to find. Who owns that and say: can you like add a tap adapter to your VM and everything will be great and I like we could try to ship something like that in kind, but all the ones I've seen so far require something like modifying the docker app and that's not a patch at one. Oh.

A

Yeah and then you're, you know dealing with supporting it on all the different platforms. Are you going to implement that very differently and yeah.

B

So there are things like that that, like yeah, this would be great and if someone comes up with a fantastic but so far, I'm fairly convinced that they're just not feasible, but you can also have some fairly big incoming code that I'm discussing back for now around like people want dual stack support. It's like yeah, but that's.

A

In like a to two versions of Kate's running in two different clusters,.

B

Sometime ipv4 ipv6.

A

B

Started it like in kubernetes, this is fairly a thing and I don't carry all of the mess while like no one else is willing to.

A

B

Don't do that see if they can figure it out a little bit, I get it in like one real cluster somewhere at all, and then maybe this in. So if anything I've been like holding back something once because it's like oh, this is going to make massive changes and be a mess and, like guys, aren't settled yet and like no real clusters have this.

A

Yeah, so how much do you need to test something? That's not in a real cluster right.

B

It's like yeah, this code never runs in production. Is it real code, see.

A

If a tree falls in the forest, that's nice.

B

Yeah but other than that I think it's mostly just like feel. Like the most helpful thing. We've got like a couple of vmware people actually like Duffy, has been hanging around and just sharing their expertise with all the users and stuff I'm, like fixing a bug here and there, but in terms of features, there's like most of them are related to testing kubernetes alike.

B

They're, like I, mostly just been settling on the designs of like code, is pretty small yeah one of the changes we wind up needing to make her actually in another project like I'm trying to get some changes to move along and container to, you know, fix a few things or like kubernetes itself. We we've kept.

A

It pretty late well that make sense like and I get that and that kind of aligns with how I kind of assumed and figured this this was gonna go, which is, you know, hey well, it was an idea.

A

You know, there's this sort of abstraction on the CRI layer that we can all take advantage of, and then you know you guys and architects and and whoever else can decide how do we actually want to do this and then you'll come with to plan like okay, here's we think the way to do it and as a result of that, a body of work will emerge. Yeah.

B

Yeah I mean I, think we're gonna find there's a few more like open questions like how does this relate to mini cube, which is also now implementing some of the stuff and even borrowing some things and like whatnot, so we're mostly still focusing on like how do you test kubernetes with this, and that mostly involves changes like in the kubernetes tests? But I we do weird like I. Do the local dev is useful and regardless of how that settles, people are gonna, be doing it and I sparked reams of it.

B

So no there's a few things to show up there, but honestly, one of the nice ones will be giving them some options of a docker, so like I've just recently had a teammate work on like the pod, Med 1, but I'm pretty happy to like pass off working on container D or something it looks like we're.

B

Gonna have him work on some other kubernetes testing stuff entirely like again with real clusters, so yeah I kept looking forward to carries here and I hope to see people around like one of the place of one of those helpful things has just been. We get tons of people coming in and they're like. Oh, how do I X and often it's just some small under misunderstanding, but like Asian aliy, it is some like deep, deep early bug and.

A

When they're coming in, are they coming in through slack or github, issues are both mostly.

B

Slack but also github issues anywhere else. We haven't taken to monitoring any other channels. Okay, like most of the people on this project, myself included like part time, do it we, you know, there's all their stuff so like or not like quite ready to use staff like a stack overflow, occasionally come file initiative like here's, the slack or also there's like we do mention like there's the mailing lists for like sick testing and stuff, but it's pretty rare that we see a question through there. Most of them are get up issues. It's like.

A

Yeah, that makes sense. Actually it's interesting trying to find, even as an example of you know, ways that could help just trying to find this meeting was tough. I had to get it from Duffy cuz I went to the kubernetes community page and I'm like is there.

B

A calendar here is there anything here, there's something but yeah most of them. The easiest thing is in the. If they have one in there slack in the description they have it button, yeah I I said I may very well fold this down, it's been it engines, ik testing, I'm, trying to do more there again and like we haven't, had super high attendance to this meaning we're doing most discussions through, like other, like async means yeah like yes well, like one of our one for larger contributors.

B

Antonio is like an engineering manager at SUSE and Spain. So, like time zones, it's just easier to use black email. What is broad.

C

B

A

Gotcha well, it's been super helpful I. Think that answers a lot of what I wanted to discuss an ant and get asked you know I think we have a pretty good idea of some of the options ahead of us, we'll wait to hear or some of the discussion I've got a meeting a couple hours with Duffy later today, so we can go over so.

B

Now this biggest question is whether you want to do CPR or I'm, not really sure but happy to receive so yeah.

A

And that's that's what I'm gonna be talking about Duffy with I'm, not sure the implications of picking one of the other, so I want to make sure that we're sort of thinking about it and I think.

B

Probably the biggest obstructed, Craig Codel is that it's supposed to be in coniferous, whereas, like CTR, is more like the container to you guys saying like okay, here's the tool like we have an API that you should use like here's the thing but I think in practice, they're in a pretty similar space and depending on what all you want to be able to do. The like CRI is a much more restrictive API like we've. Had we actually like again, we don't do everything with CRI.

B

Even on the manipulating bearnaise know, levels need to do some quirky stuff, like loading images ahead of time and like CRI, doesn't have a concept of like loading, an image, because that doesn't make sense for a real kubernetes cluster. You can't ask it to pull and you can ask it to run something, but you can't say like here's, a tarball, because that's up to your runtime, it's not something that cubelet cares about like CRI is really the like.

B

What does cubelet need so like you can probably pull it off with crank it'll and then it's broader support and things and like cry cuddle is not something. People should be breaking at all I'm, not sure where container these stances on CT are I, think they probably can't before we're here, but I'm, not sure if they've actually committed to that, but like their overall API is much more expansive and you learn useful than CRI, because it's not just what cubelet means cry cattle. You mean right.

B

Soso, like the craap, I okay, alright cut was gonna be bound by what's here, I can do and see it. Ri is just like what is killing gotcha. Where is like container D has a much more thoroughly like.

A

Gotcha so like, if we were to like, for example, let's say you know, make things work and and make that work against. You know cry cuddle or CTR. What would we need to deliver? Would it be just like our supporting that and then you install kind and then like there's a config where you're pointing it at our implementation of it? Yeah.

B

So, like probably you so usually correct, so cry cuddle reads this like standard file on disk. That says like this is the end point I'm supposed to use and any other details. I, don't think you normally said otherwise, but I think there are other ones.

B

Ctr I think is much more likely to have like a it. Just assumes the default one, but can be told to do other things and I and it normally isn't configured it's just used, and then you might do something like want to specify a namespace or something I'm, not sure how we want to pull that through I'd probably have some discussions with and don't be, depending on which one you go with, whereas it's quite cuddle he's gonna operate indicates.

B

Namespace I have a feeling that CTR make me may make more sense here, but in terms of like maximum neutrality or whatever, that's definitely crack up. Okay, that's what it's like but like. If you can pull it off with either of them, I'm pretty fine with whichever one you want to support. I'm given like I I, think most of what we're gonna wind up supporting with it is gonna, be continue. Debase anyhow, I'm, not sure how much sense it makes two legs stand up, cryo and then run kind against.

B

It, I think for most of the people that want like that ecosystem pod men is probably what they actually want mmm and for most of our users I. It will continue to be darker, but for a few people doing some more interesting stuff, they're, probably going to what like the VM type things that are like okay container T is a mechanism by which we run containers. But then we can do interesting, implementations of containers and for that, just targeting container T is probably sufficient.

B

That's, but it's really down to like, whichever one's going to like make more sense for, like these case, more integration and maintaining this and four kinds perspective. It will just be another thing implementing the writer executive tool and we may have to work out a little bit of like. Oh, we need to pump some extra details, but we can figure that out. Okay,.

A

Digesting a lot of that yeah yeah.

B

In general, I like something like CTR, ideally shouldn't need that kind of conflict. You would expect to be able to like just go to your sea lion, feel like CTR, PS or something like that and it'll do the thing yeah and it should just know where the thing isn't. Do the thing yeah, that's what you mean like that is all of these things are trying to sort of kind of be drop-in docker replacements so like we should be able to treat them as one I'm sure.

B

In practice we might run into a little bit of fun there. We certainly with apartment. Then.

A

We address those as we go right right.

B

And some of those may very well be actual, like in pubmed's case. Most of them were actual bugs, but yeah.

C

A

Okay, this recording will it end up in a meeting notes somewhere, so I can send it to some of my engineers if.

B

It doesn't we poke Aaron cricket burgers with XP, okay, I, don't have I, don't think I have access to that, but it should window post it somewhere. Okay,.

A

I'll follow up on it just want to make sure that you know some of the engineers here. The output of the discussion. That's all yeah.

B

Cool well thanks for coming and I look forward to class appreciate.

A

It then nice to meet you in person. Finally, nice to see you Cheers.