Cloud Native Computing Foundation KubeCon + CloudNativeCon North America 2018 (Seattle), 16 Dec 2018

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Deep Dive: Virtual Kubelet - Jeremy Rickard & Brian Goff, Microsoft

Description

Join us for Kubernetes Forums Seoul, Sydney, Bengaluru and Delhi - learn more at kubecon.io

Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects

Deep Dive: Virtual Kubelet - Jeremy Rickard & Brian Goff, Microsoft

This deep dive session will take you through what it means to be a provider or developer within the virtual kubelet project. We will go deeper into the virtual kubelet interface and discuss design decisions that were made in the last few months. We will also discuss the proposed roadmap to getting virtual kubelet to a 1.0 release. 

To learn more: https://sched.co/J8IB

A

Thanks for coming.

B

Everybody I know.

A

We're across the street and it's raining so thank you for coming. We're super excited to talk about virtual people, but today I have a bunch of stickers too. So if anybody wants a cool virtual cubelet sticker afterwards come see me and we'll pass them out so I'm Jeremy, and this is Brian hi. We work at Microsoft on virtual cubelet, amongst other things, and we're gonna go into a deep dive today. So why are we here? Why are you here? Subversion people is a pretty new project in the sense of the scope of CNC f.

A

It recently got accepted as a sandbox project. It's been around for about a year, though, and we've had a small community kind of coalescing behind it. So today we want to talk about what virtual couplet is.

A

So everybody has some context about why you would want to use it what it does and then we'll also go into a little bit of context on how its structured, how the code is laid out, where you might find things if you're looking to either implement new capabilities or features for virtual cubelet or if you want to fix some bugs or if you're, just curious about how it works, and you want to use it before we get into what VK is, though, let's talk about what a cubelet is, so we're going to talk about a virtual cubelet.

A

What's a cubelet, so cubelet is a software component that runs on your nodes and it's real responsibility is to make containers run on that node. So you met you make a pod in kubernetes the cubelets, what's responsible for actually making that thing wrong. It does some other things too, like exposed synchronous and points for getting logs and exacting into the container and we're gonna go through mostly the the creation of the pod and running of the pod today.

A

So, let's think about how the cubelet works when a node comes up and gets registered in kubernetes, the coop was installed on it and it's configured and it has a name and then it says: hey, kubernetes, api server. My name is whatevs I'm a node. So so at that point, you've got to note registered with kubernetes and then you, as the awesome, kubernetes user, can say: hey I need to make a pod spec here's my pod, please run it for me. It may or may not have some constraints attached to it.

A

You create that into via the API and then the scheduler says: hey, there's a new pot. I need to figure out where to sign that whatevs has some space, let's assign it there, so it updates the pots back and says no name is whatevs.

A

Meanwhile, the cubelet is running and it's doing a watch because it's just a quebradas client. It listens on the API and says, give me all the pods that have spec node name equal to my node name, which is whatevs. So you've created your pod. It goes through the API server. The scheduler says this thing can go to whatevs, so this you know in this flow the cubelet picks that up and then says what I do with this thing it needs to reconcile the state, is it running? Is it not running?

A

Do I need to make it running it figures out what to do, and then it uses CRI to start to container so that flow happens and whatever node you have in the system and the qubit is really kind of abstracting that whatever the runtime is, it could be, docker can be rocket and it's distracting that concept of a node in the cluster. So we know what a cubelet is now. What is a virtual cubelet?

A

C

Attested order Brian I apologize for putting whatevs in there like I, didn't know, you're gonna, say.

A

This is so nicely whatevs. Every.

C

Time so virtual cubelet, it acts basically just like a node in the cluster for the most part, kubernetes schedules pods to it. It makes sure the pods run at set. We're done see you so I can just replace the cupola with virtual cubelet. For some reason, no, not really so.

C

The coop looks really good at running workloads on a node on a hill star on a VM. Whatever you want to call it physical machine, it doesn't matter, it manages the resources on a node.

C

Does things like mount stuff from the hosts into a pod that you asked for all this kind of stuff that the virtual could actually will not do because we're specifically not targeting a like real node that you would that you would think about, but some things that you might use it for is bursting to some cloud containers as a service which Microsoft happens, they have called a CI Azure container instances. Aws has far gate, there's, probably someone leaving out terribly sorry for everybody. It's not on purpose.

C

There's other service platforms as well. I've seen people using it to bridge orchestrators. So, like me, sews I've, seen working on a connector there to allow you to schedule from kubernetes over to me so sauce on something else for joints, Tritan, cool stuff. You can schedule stuff. These are just ideas. You can schedule stuff like an IOT control plane. Maybe these things have a persistent connection. Maybe they don't when they come online.

C

You want the stuff that kubernetes said the run to run on it and like what flip an LED from green to red or something like that and I. Don't know it's pretty cool you can or there's somebody in Microsoft made something thing to order pizza. That is I. Think that was a real thing. Wasn't it yeah? It was real and I'm sorry.

C

So how does it work? That's pretty much it like it's magic, but no, not really it's not magic at all. So this might look familiar virtual Kubler comes up. It says: hi I'm, a node, a node user creates a pod cuz we're in a scheduled the pod virtual couplet.

C

We have a watch in virtual couplet, saying: hey I, want to see everything that that matches this list of things that I should be looking for, specifically for the node name. It runs a state record reconciliation to see what it needs to do. Don't need to create it when you delete it whatever and then you know makes the determination. Oh I need to create this pod, so I'll go ahead and create it. So what is what does virtual kugel?

C

It actually look like there's three main components: we have the virtual coop kulit core framework that is kind of responsible for the kubernetes side. We have providers as kind of responsible for for other stuff he'll get into in a minute, and we have this the daemon that's kind of this virtual couplet distribution that has all this stuff kind of pulled together. So, okay.

A

So, on the previous picture, we saw the daemon it contains the framework and whatever provider you're going to use. So what does a provider? So in the grand scheme of things? What we're trying to use virtual qubit for is to allow you to schedule a pod anyway, like in the normal way you would with kubernetes, but have it run in something that doesn't really represent a node, so the provider itself is really responsible for doing that.

A

Translation between what is a pod and what is that thing that needs to run so in the case of a CI, we're gonna make a resaurant ainur group and put any number of containers that match whatever containers you have in your pod, it doesn't necessarily have direct access to the kubernetes api.

A

It's a separate thing and what it does instead is conform to the api that we provide, which is called the provider interface and that provider interface I will see in a second first like what providers do we have a CI we mentioned a second ago. Atos Fargate vmware has one for vSphere integrated containers. So if you're running vSphere on premise- and you want to use Vic to you- know abstract away resources for running containers, you can use that there's one for Alli cloud, just one for y way. Brian mentioned the iot edge thing.

A

We had a demo at our booth. That was pretty cool. Somebody had this little robot from unki that you can program and they have a little edge device next to it and they use virtual cubelet to schedule workloads to it, and it was doing computer vision and reading signs and printing the text out, and he was able to update that by deleting pod, which ended up deleting the thing from the outer edge of device and deploying a new workload to it, which contained the new logic.

A

Ok, so let's jump in and see what the provider interface looks like if code opening, yeah, ok, we're gonna make this bigger how's.

D

C

A

Okay, so let's collapse this down for a second and close this. So this is the virtual pubertal cubelet codebase it didn't get them and we'll give you the github repo in a little bit, it's structured very similarly to how other kubernetes code kind of looks. We have a CMD package that has like the root thing, other code that goes along with it. So that's like the entry point where the daemon code lives like the root thing. Yes, the root thing.

A

We have some examples: providers which is all gonna talk about right now, so in the providers package, you'll find all the implementations that we currently have. All those things live within the virtual cube codebase.

A

So if you want to add a new provider for something you're welcome to submit a PR to add one we're going to look at the azure provider in just a second, but if we go down to provider go, this is the interface that you would have to implement if you're going to be a provider for virtual cubelet and as we walk through this, it's gonna talk about mostly lifecycle of pods, so create pod.

A

Update pod delete pod get pod and then it starts to deal with some of the synchronous operations like get container logs, exec and container get pod status get pods, and then it deals with things that are more node like so the cubelet really deals with a node right, but we're not really a node. So we need to provide some of these things like what conditions are available? Can we schedule to a CI ID be a condition? What capacity does my abstracted thing have like?

A

If we have an endless bucket of space, can we schedule forever or do we have a constrained bucket of space? Can we not get scheduled forever so.

E

Again, like all these things represent.

A

Like the the virtual node so operating system, is it a Windows or Linux node, and then we also have an optional interface for doing stats, so you want to collect stats from your providers, one of the things the framework allows us to do. This is an optional interface that you can implement. Okay, so really the provider abstract away your end thing and it's responsible for taking this interface and turning it into reality. So, let's take a look at one.

A

What one looks like: let's go, just look at the azure implementation, so here is AC I go.

A

This is the Azure container instances provider and we'll just walk through it and see what looks like so in here we've got stuff, that's related to initializing, an ACI client, so you know we're importing a bunch of ACI stuff, making a new provider here things that matter for Azure are like what region is it going to run in what resource group do you want to use things specifically matter for ACI is what operating system do you want to create the containers as what kind of a memory and CPU defaults?

A

Do you want to give to it? Do you want to run it in the V net? So all these things that are really out of your specific, not really kubernetes specific. You know this thing is encapsulating and hiding from you and really only implementing that interface. So we can scroll down and we see it builds up stuff. We can see what regions are available, but then we we get into the implementation of the interface. So more scroll down, Smarr miss it sorry, yeah, don't pass a provider.

A

C

A

I broke the bills already we're gonna compile this right now. So here's here's where we start implementing that interface right. So this is gonna. Take a kubernetes pod object and you know virtual people is not abstracting away. What a pot is it's dealing with pods and it's expecting you as the provider to deal with pods but you're, going to turn that pod into whatever the rep like the real thing is that you need to deploy. So in this case we're gonna make an a CI container group.

A

That's what the representation of a pod is an AC I can have one or more containers. I can have ports exposed. It's really similar to what a pod looks like in kubernetes. So we're just going to walk through that pod definition and turn it into the right thing for a CI. Then we scroll down, and eventually we use the a CI client to make that thing. Okay, so you do your pod creation, the scheduler signs it to virtual qubit.

A

The framework in a second we'll talk about says: oh I, need to make this thing and it uses a provider interface to actually invoke that. Okay, so that's what they provide. Our interface looks like and if we scroll through here again, you'll see more like, let's see what else we have delete pod update pod, which we don't implement, you get pod.

A

So all the things you would need for those lifecycle operations and then further down on this, we have some implementation of like resource constraints and statuses for a CI, but we're gonna come back to the PowerPoint now and Brian's going to talk about the framework where's, your PowerPoint Safari Oh,.

A

We're back to full screen good.

F

A

Should make it fullscreen sorry.

C

Don't need it, it's, okay, all right, so the framework, it's kind of it's this thing, that's basically a reconciliation loop. If you know what a cooper nineties controller is that's essentially what virtual couplet is, or at least the core framework we work from a shared Informer cache. That's basically, this local cache that kubernetes client provides of of stuff that you're watching, and you might share this with other things.

C

If you happen to want to build your own personal, couplet style daemon, we have an HTTP server for handling things like requests for logs and and like pot exec, and these are very all the synchronous actions that that the hub, API server will call in for and of course, we also add things like tracing and logging. That's just you kind of get for free by using the core framework. You can get some pretty awesome details in terms of how long it took to create this pod.

C

How long it took to come, go through the work queue and all that kind of fun stuff so take through some of the the.

C

Code for that framework, so we will go through that route thing.

C

So this is just the the main CEO daemon implementation. We go through a whole bunch of set up stuff like setting up the kubernetes client setting up the provider that the user asked for.

C

These are all pretty basic things. I, don't think I need to show them, so we create this via Kubla thing and then we go ahead and run in this is where the magic happens or but not magic.

C

So this is in the V Kubla package, if your happen to be following along in that regard, so you can see some things automatically where we've got this this logger here that that we pull that we actually are attracting from or sorry extracting from the past. In context, we start the HTTP server. We register the node, we start a reconciliation loop that checks the pod statuses from the provider and sync that back to kubernetes, and then we have this this other thing here that runs to sync to kubernetes into virtual Kubla.

C

So we can figure out what needs to happen. So we, you know, start up a work. Queue go down to the workers. Some of this is probably, if you've written a kubernetes controller is probably pretty standard. It's pretty good, relatively simple I, don't want to say it's simple, because it may not be pop something off the queue you can see. We got tracing in here and we go and figure out what this thing is and something happened.

C

Here we go there, we go, we run the thing that says: go ahead and sink it, and if it fails, we added back to the queue we have a certain number of Macs retries, which is nice, because things fail, not that you know not all that uh Knopf tainha specially well like AC, isin and preview. So.

C

Just we we get this info from the from the queue. We look up the pod that that, from that key, that we got from the queue and we determined we need to delete it or do we need to create it. It's a relatively simple thing here: we're just syncing pod we're getting notifications from kubernetes and we're determining oh. Does this pod have a deletion timestamp or does it not have a deletion timestamp? Do we need to create something or not, and we pass that on to along to the provider to deal with it.

C

Fix this I want to fix it. There we go cool. So what do we have coming up? Basically, we want to hit targeting a 1.0 release for probably March, so we have a whole bunch of stuff and github labeled VK 1.0, some of the things that that came out top of my head. While is writing this port forwarding?

C

We want to have metrics, so we have tracing already, but we don't have metrics for VK, so you can kind of track that stuff and we just have to kind of figure out what actually needs to be measured. Support for external volumes, there's some more work to do to split the core of rostral couplets framework, the reconciliation stuff from the actual daemon setup.

C

There's been some some kind of mixed-up things in there. We're trying to split some of that out now things like supporting pod updates, which we don't currently support, even though there's not that much to update, but it's still important, some stuff that we're I mean all that stuff we need help with. But there's more, we want to make sure a provider. Interface is is good to go like it works for most people. We really don't have to change it after 1.0.

C

So if you are thinking about writing a provider, please do give feedback on the actual interface Mauri to e tests. We have some it's not enough. Never enough. We want to pass performance tests. Some things don't make sense like we can't do host mounts. We can't like we don't support mount propagation and anything dealing with the host at all. So there will be some things that we expect to fail, but for the most part we should be able to support things like right.

C

Now we don't have the downward api support that we need to get in and some other things and again more work. Making the core virtual could let more of a framework that you can consume and potentially build your own thing. That's not the the daemon that you saw in the code base here and bugs and missing features and all that fun stuff. So we want you to help us.

A

This is a pretty new project in terms of CN CF adoption, but it is a sandbox project. Now we have a lot of work to do still on a bunch of things and we think it would be great to have a lot of community support. If you have use cases, you can give us those if we don't meet them already, if you can think of other providers, if you want to work on any of those things that we just talked about, that was really the context for for doing this deep dive.

A

We wanted to give everybody a sense of how it works, how things structured, how you could find things if you want to dive into the code yourself, we do have a that. Obviously, there's github repo. We are in the kubernetes slack in virtual cubelet. You can join us, ask questions in there if you want to dive in and find a good issue. That's a good thing to start with. You can ask in there we've got a couple: people that are really encouraging. You could be new contributors. We have community meetings every other week on Tuesdays.

A

Those are at 11:00 Pacific time we do the Bunn zoom, so you can join us and have good good, chats and again first people to CNCs sandbox project. So we left a bunch of time at the end here to answer questions. If people have questions, we can also do a demo if people want to see a demo, so it's kind of choose-your-own-adventure. Now, if people have questions, go ahead and ask them now and there are no questions and we'll do a demo and you can, you can see. Oh we're.

A

Gonna have Nick run the microphone around so that your questions can be recorded. I thought.

C

You were gonna, do the any question.

G

Yeah any questions.

C

Yeah, that's right.

H

C

Obviously, why would I say something like that.

F

Thanks for an interesting talk, and my question is about mock provider, if you could elaborate on how can we use it and what use cases of using it and question? Basically, do you see if it's feasible to use it as the kind of integration for it to do this integration tests for operators.

C

So yes, so the mock provider, we actually actually just recently added II, do we test and we actually use the mock provider specifically for that, but yeah, it's it's totally just for you know, testing out features. Maybe we need to add something to the tour and we just want to run that against something without having to like fire up real infrastructure yeah.

C

Does that answer your question?

C

Not quite yeah.

F

I guess not quite okay, for example, if I, for example, if I create an operator right, what I want to be sure does, for example, if certain event happened right, then certain I don't know. Resources like deployment has been skilled or I. Don't know. Persistent volume has been something and so on, and the status and the sub resource of my energy has been updated right. So I want I, want to address right. Such kind of scenarios and and right now, I feel that I can do. I cannot do this in in communities.

A

F

H

A

Test operators like in sense of operator framework things that you're building I'm necessarily the operator framework but like.

F

Yeah a theory Australia's well or if I wonder if I wanna, I don't know too long any kind of information, for example, I bribed my possible playbook I will eventually need something similar yeah. It really doesn't matter the operator or in anything else, yeah I,.

C

Think that's a little tricky to answer so first yeah you could totally use it to do that, so you fire, purser, Kubla and essentially schedule it to that with a mock provider, and that should work but I unless you're, specifically looking the target vertical I would recommend. Looking at the couplet I, don't know what they might have available for in terms of like dummy dummy interfaces for CRI, but that would to me I would I would try to look at a CRI integration to do that so you're running on our real couplet.

C

If that's, where you're expecting to run your stuff.

E

So I assume virtual cubelet is calling into the cubelet api in receiving api calls from from the server api. It so is the cubelets api, a stable, kubernetes api or so.

A

The cubelet api is pretty thin. It's really scoped to those synchronous operation like exacting into a container and getting logs. The other parts of that are really driven is just being a Cabrini's client. It's not really a cubelet api, it's just watching for pods or watching for services. Those sort of things so.

E

The virtual cubelet is monitoring, what's happening and yeah.

A

It's it's operating just so, if you go through, if you dig through the cubelet code, you'll eventually find the spot where it sets up a watch using just a kubernetes client with client. Go virtual Cuba. Does the exact same thing so.

E

I assume it's watching for certain rests operations, and how do how do you guarantee that, if I, if I, throw a virtual cubelet binary into my cluster, is it's actually gonna watch for the right things, depending on the kubernetes version, I'm running on my cluster yeah? So you want that.

A

C

So the basically the only thing the couplet does is so really anything you make changes on the API server. Then everything else is just reacting to those changes so like the couplet is looking for stuff.

C

That has been scheduled to the node name of that node that the coolest been registered with, and likewise with referred to, okay, that we're just watching for for things that are scheduled to the node name, that we've specified when you started virtual public so like when it's time to create a pod, the there's nothing telling the crew blade or the virtual couplet go ahead and create that pod. It's just this is now in the API server. Then the scheduler even come and say Oh someone's added a pod.

C

We need to schedule that it's not hasn't been scheduled yet so it goes and updates the pod with with the node name that's gonna be scheduled to, and then we look at that and say: oh there's, something that's been assigned to.

E

Us for us, okay, yeah, so that just another question is when you, when you start a virtual cubelet and bind that to some providers somewhere in.

B

E

Or or asier, what kind of resources do you advertise I mean how do I make sure that I only schedule specific workloads on through this virtual cubelet yeah? So what.

A

uh There's a couple a couple things there there's two parts of your question: I think right. Now, it's up to the provider to report what resources are available in a CCI. We don't do a very good job of that. We need it's one of the things we should do better. Well, we say that it's like pretty much an endless bucket and a CI does have quotas exactly.

E

Right into those, it could look very attractive to two server API saying: oh, this is yep. This is a use.

A

Those node right now what we do is when the note is registered, we taint it, so you have to explicitly opt into using virtual cubelet right now. So when you deploy your node, you know the virtual people. It runs as a pod in your container can run other places as well, but you're gonna find it it.

A

So when you register it it starts up, it has a default paint built into it and you can it's over writable and you can change it, but by default, when the node starts up, it comes up with that taint, so things won't get scheduled into it and unless you tend.

E

This to change a dollar agent, you taint your.

A

Your pot yeah, if you specify Toleration on pot, so this is an example of something that will run on virtual cubelet. So you can see it's got toleration.

H

At the bottom, raishin yep.

A

And you get so. The other thing you can do is because that's configurable and this one says a sure you can run multiple virtual cubelets to get multiple notes like if you want to run some pods in asher, and you want to run some pods in a CI or in Fargate, or you want to run some pods on on premise with Vic, or maybe you want to burst from your on-premise stuff out to a cloud provider. You can use multiple toleration there to allow it to run in different things.

I

So I have a question about if some sort of thinking about like it go away: No. Okay, if you're going to do a new provider, it seems like the model you're working on is basically that your provider is an infinite sync and then, which is you know, probably true for essentially all ACI customers right, very large capacity, and that it's I assume you're just expressing the node into the API server as like, very very, very large capacity, yeah right now we are, and that and so I'm sort of curious.

I

Do you actually think virtual cubelets is compatible with something that has a much more finite capacity, or should it always sort of be used with providers that you can? You know more or less tree as an infinite sync, just due to the capacity of the provider versus the relative size of customers, workloads, yeah,.

A

So the provider interface allows you to specify those things. You could give realistic numbers for that VM. Where does that with Vic? It's not I, don't think they represented it as an endless pool of data, it's possible to do, but you could make it be realistic and be more node like, even though it's not a no rates, it's a more elastic capacity or more I'm, not.

I

Sure, I, and so like one brief, follow to that. So how does a CI deal with maintenance of the actual stuff? That's running virtual cubelet workloads? Is it that you, basically just you, know, kill pods and then they'll.

A

Come back up and yeah, so we're running to reconciliation things right. We have one, that's looking at stuff in a CI or your provider and made its query in the provider. So you provide a report status back and then we have one that's looking at kubernetes. So if something's gonna die in I mean a see, I will restart it for you. It's like how easy how it works, but you know the provider interface allows you to do that too.

A

So if it died, didn't capture the underlying system is but for a CI you can do all kinds of things like define, readiness, probes and liveness probes. It looks very similar to like the kubernetes pods back and a CI will, because it's a matter of service will will do a lot of that, for you specifically for a CI yeah.

G

I have two questions. One is that they feel need to call with two providers, but who are more providers? Do you need to run virtual cube, play multiple times, yeah.

A

Is it right now it's configured to have one provider like per instance, so you would want to run multiple pods because it's gonna register it as multiple nodes right like in that case. So you would want to run one per her like variation of virtual couplets. So for us you can run Windows containers where they guess by doing a virtual cubelet to Windows, and you can do Linux containers to AC ivory via virtual node or sorry, bye-bye virtual cubelet, by I specifying Linux as the operating system.

A

For that note, so you're running to V K's in that case and.

G

The second question is that networking, how does netbook ease no that's.

A

A good question: that's that's kind of up to the provider, so in the like I'm gonna use, a CI is another example in. In that case we support a couple of modes. When you start VK up, you can specify the you wanna use a V net, which is like what you'd use to do private networking between like an ApS cluster and a CI. So in that case it kind of looks just like one network. If you don't use that one of the default behaviors for ACI is to create a public IP for things.

A

So when those pots been up, they'll have public IP addresses, they won't necessarily be able to see things back in the cluster. So you got to kind of take that in consideration when you're thinking about what workloads are gonna run there. It's definitely not a generic replacement for cubelet or for a node. It's going to be for a little specific use cases, and you have to kind of consider that with the topology of your network and just how the how VK and the back-end provider are configured like that, though,.

D

For things like exec, when it doesn't really make sense in the particular context, how do you handle that we.

C

Handle exec actually, so we we do right. Now we do exec and logs we're looking at how to do port forwarding, which but I think that's going to be mostly up to the provider.

C

I think yeah for so for those API requests that come in like that I think we support most things like I come not familiar, but yeah.

A

Like exact is implemented in the provider to a lot of the things are in the provider interface and it's kind of up to the provider. If it makes sense or not yeah for a CI a there is a a Pai operation. You can do. That makes sense in the like the VMware, Vic on-premise scenario, you probably also do that. Other things might not make sense like you, especially that IOT case right, where you're like deploying to some IOT control plane in the cloud and it's doing asynchronous operations to end devices.

A

So it's kind of up to the provider to implement what makes sense, yeah.

C

So we'll wire up the HTTP server, send them and pull them there, io streams and or whatever we need to do and and sent it to the provider. Okay,.

D

So like, in that case, if like it didn't make sense to have an exec function for the particular workload you just like return, an error, I guess.

B

J

One of my questions, gotta answer, cific, that's good! The other question is: are you able to handle secrets and config maps and other objects, like those secrets.

C

And config maps: yes, nothing else.

J

Like volumes and stuff like that, you said did not support it. Correct all right. Thank you.

H

Love are you guys are all on the opposite side, trying.

K

To make you work a little, so you would do it's it's one virtual Googler per provider, which say, for example, I want to have two different configurations for the same type of so I like AC. I can figure this way and that way I would have them they would represent. Is there any way or thought about overriding any of the provider configs via the pod spec, where it's like?

K

Hey we've this configured this one way, but if this comes in via the pot, you know as extra configuration in the pod spec, we can pass that down or I. Think.

C

That would always be up to the provider cuz we're.

K

C

Vk we're not parsing it. Okay.

K

Yeah, so so yeah so it'd be that would be possible because you're passing the whole pods back to the provider. So if they wanted to do that, that would be up to them. Yep cool yeah.

A

I mean you can totally ignore the pots back and order pizza via the Domino's API, to see somebody built that on Monday, we their workshop at the reactor and he built a Domino's ordering provider. So you could schedule a pot and it would make a pizza.

H

J

Do you support static scheduling.

C

huh So what does that mean? Yeah cool.

J

It supports static pod scheduling like their the exists, yeah like how it's used for in manifests and add-ons so I'm wondering if you support that, without relying on the API server I.

A

Don't think we do I'd have to check, because all virtual cubelet is doing is listening for pause being created in passing the spec on to the provider which doesn't have API server access. I, don't think we do.

J

No, not exactly no I was just wondering if they can operate. The virtual couplet I can misuse it such that I run it without the API server, but only to fake pods that they think that they're running to kubernetes but they're, not.

C

You you would have to stand up a control plane to do that. Yeah. We.

J

A

I'll talk to you about it after yeah after yeah. That's interesting up here. Anybody.

L

L

And the question is: what do you guys mean when you say that M BK doesn't support storage because maybe I'm misunderstanding how things work, and that sounds something that a CI is supposed to handle, not BK so I'm confused about that yeah.

C

So, probably probably like volume support would come out of the provider, but.

B

B

A

Yeah not the whole gamut of things.

L

Okay, so that means that every other API or Canaris resource or and we should be supported as well or this was yeah. If not what's the challenge, because that it I imagine Alicia is another Canaries instance running, but with an abstraction of level, a higher abstraction m, so I don't see why anything wouldn't work. Okay,.

C

So the provider gets the a pod spec and if it's in the pods back, they can do whatever they want. We do not give them access to the kubernetes api from virtual couplet. It doesn't stop you from wiring in your own client and you know pulling objects down to do what needs to happen. But VK does not right now provide any way to grab those kind of other objects which will need to figure out and the only the only two things we we give it. Access to. You are fetching secrets and config maps.

C

Is that answer and totally.

L

No, but I'm research, more and I have more questions. I I had this like Janet, yes,.

C

Any other questions, cool.

H

Would any of you, gentlemen, care to drop the mic.

C

A

So we have stickers um come get them. We have a whole bunch of them.