Apache Mesos Containerization Working Group, 24 Aug 2017

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Mesos Containerization WG 08242017

Description

Agenda and notes:
https://docs.google.com/document/d/1z55a7tLZFoRWVuUxz1FZwgxkHeugtc2nHR89skFXSpU/edit?usp=sharing

A

B

Hello funding, guys hello.

C

Hey yeah that's way for a few minutes for folks to join.

B

C

um Joseph, do you have anything to to present like if you want to present anything like you can't share the screen, see if that works, I.

B

Don't have any slides or anything, but the doctors there yeah.

C

Correct I should see if you can't share the doc.

C

Yes, I can see that.

C

Yeah, okay, I think: let's get started, we have a pretty tight schedule today, Joseph, you want to start.

B

So getting used to this zoom overlaping.

C

B

All right, snow standalone containers are going to be a new concept that we're going to introduce into the container I, tzer's and specifically the agents.

B

So the concept behind Stanley chairs is that we want to be able to launch containers but doing so directly through an agent API, rather than going through launching a framework which connects to the master, which then has has to talk to the agent, and then the agent has to launch container that contains it executor and then give that executive. Your task to then launch them like command line thing underneath there's Palen containers bypass most of this and just basically provision a container image and then launch command line test underneath and the logic that's backing.

B

This is the same as other mesas containers, just without the or whereas the entry point is not an executor. It's just a command line task. So.

D

Sorry, just before you get into some of the details, can you explain the motivation? Yes,.

B

So the well I guess the direct motivation is to be able to launch what we would consider admin tasks on the agent and these admin tests would have I guess more complicated requirements than just some dependency. That's already installed on a machine like we want to launch an app.

B

Specific dependencies and then launch these on the agent before it connects to the master or without going through and off recycled, so launched these containers. So it's basically like treating the Macy's agent like the docker, runtime or command.

E

B

And our first use case will be to launch a CSI plugin, a container storage interface plug-in that will talk to the agents and provide oriT resources.

D

So, where do these demo containers get their resources? What resources? Maybe you get into that later on and they should just wait till you explain it. Yeah I mean.

B

So the resources right now are this isn't actually implemented yet. But the goal for our Minimum Viable Product is to have these resources just not be tracked anywhere. We would assume that when you start the agent, you will reserve some amount of resources for admin tests and the resources that you use. The sandwich containers will just come from that reserve set.

B

Basically the number resources on the on the physical machine or the virtual machine. The ones that are advertised to the maysa master are the ones that are reserved for other things for the agent. So.

D

It's kind of a pool of resources that all standalone containers would pull from and sort of share as they came and went. Yeah.

B

The goal for later on is to use the same interface that we use to provide or we're going to add four resource providers, so the agent itself will be a resource provider of CPUs and memory and GPUs and storage and it'll to have that interface. It'll need to update the number of resources that the agent is advertising that master Bruce, often and stanlon.

B

Containers can use roughly the same code path to say, I'm using like once a few and some non memory, so that that amount of CPU memory is no longer available for other tasks in races in the Mises cluster.

D

Data, so one last question before I: keep blocking you from continuing, since these aren't tasks, will we be able to do things like run the equivalent of tasks exec to get inside them? If we want to debug, what's going on yeah.

B

So these aren't tasks, but they are still containers, so you will be able to launch yes with containers in them and nested container sessions. As long as we support, or as long as there is enough fiddling in the validation within the existing API is to do so. I I think that's I think there should be yeah. The only thing you really needed.

D

A container ID yeah.

B

I think there will be minimal changes to support this or minimal changes required. No we're just support, nested containers and mr. sessions. They should basically just work once we've implemented Sandlin containers in your API.

D

B

Alright, so, as mentioned, Sam containers have no executor ZnO tasks, and this also means that they don't have health checks, non-kill policies and status updates, so I mean these. These three things are implemented in the executor and because we don't have an executor for sale in containers, fees won't be sorted right away. We might consider adding them later on, because you know they might be needed, but for MVP, it's not part of. um So besides that certain containers share the same exact semantics, other maysa containers. They use the maysa container hazard.

B

They will support you CR. They will support their images and if you have some amount of Isolators enabled it will go through the same isolation, mechanics and the same isolation code path as other containers, we'll have to add a little bits or we'll have to add a.

B

Way of preventing old Isolators from being the best nominators, because there might be some third-party Isolators that depend on the executor field or the task fields and those will not necessarily work off the paths right off the box.

C

With standard containers, yes Joseph, what's the plan on that like are going to introduce a like isolator capability or just like a virtual function? Like return, a bool say if it supports standalone container or not yeah,.

B

I mean both are valid, viable approaches, I guess. The way we wanted to go in the long term is through some sort of enumerated capability. Okay,.

C

Yeah I think that'll be great, otherwise I think it might have not too many virtual functions that that does not scale I. Think.

B

So far, we'll just have two capabilities which is launch message container and launch standard container yeah and the I guess the semantics behind each shares. A nested container has container ID you with the parent and then a standard container, we'll just have more blank fields in the arguments that have passed into the isolator. So.

D

Sorry, what's the maybe you just said this: what I missed it, but what's the motivation behind having to API is instead of just consolidating on one well, like we've got launched, nested container? Why why now a new launch standalone container? Why not just have a launch container.

B

So we won't super. Well, we don't want you have once you have a large container because well when you, when you launch a container directly on the agent, it is always going to be a Stanley container. It's a I'm, not talking containers here, so we're going to have a call for launch sample containers and then one for launch nest of commuters. Yeah.

D

But I guess what I was saying is that if we had one you know I just called a lunch container. If you had a parent ID on your container ID, then it would automatically become a mystic container. It was being like yeah.

C

I think that's the interesting point, but I think we already have actually launched the nasty container and I. Don't think we can change the name of that API so that that API has to be there for backwards compatibility. So, like I, don't think it makes sense to hijack that API to support I launch a standalone container. We don't have a choice but adding a new API for that. Okay,.

D

But there's no like difference in the actors know that the FBI would actually be doing yeah. There's.

F

B

Like but that's basically all there is the the actual plan for this is fairly similar. Although there are some differences in the return semantics for scaling containers versus nest containers.

B

Today, online containers are meant, as are there I guess they are designed for the operator to use rather than for frameworks and tools, and that so the launched and container will be more of an idempotent launch, whereas launched nest. Entertainers like if you can give it the same, ID it'll fail press. When you launch the same ID and a standard container, it will succeed.

C

So so, just one comment on the item: potency you're, saying like the launch scan own container is idempotent based on the ID or based on all the fields like, for example, if I have the same ID, but next call I specify different command different containing. What's the semantics there? Yes, it's Colonel.

B

Adjust based on the ID and because.

C

I don't expect.

B

That we'll be saving most of the other fields in the agent.

C

Okay, I see it'll, be nice to craford I like the idempotency. It's only on the ID yeah.

B

It's this little comment: if we container with the same ideas already running okay,.

C

D

C

D

You just you, allow it to keep running and go or.

B

Yeah, so the idea is: if the operator knows that they are changing a field and then they launch for the same ID. Then if they get a Tio's here they know. Do you first kill and then relaunch.

B

This is primarily just support things like you know. If you were to launch an asteroid potater through a supervisor service like a AB starter system, D, they repeatedly launch, we would expect those things to succeed.

D

F

D

A little strange to me, but okay to go just seems like a weird thing to get back. If your. If your next plan is to kill the container but.

B

E

If the next sorry go ahead, a little confused about why you'd want it? What if.

G

You, if your use case is lamenting these things under our supervisor, then an HTTP API doesn't seem like a natural interface. For that know, most things that launch processes are gonna, be the security at command.

E

And why she appeared.

G

E

Of thing, so why did you end up with a HTTP API.

B

And so we're gonna have an HTTP API, regardless of the I guess. The final interface like there may be a CLI component after we've implemented these API. So the CLI could then be a masive CIA.

B

It could just look like a normal command line.

D

But I just I, just read your your comments on the side. Then that answers my question about why they throw too so.

C

So so so James I think like if you look at doctor so doctor hadn't daemon, and it provides HTV, API and doctor client. It's like, for example, dr. run into doctor trying to talk to the HD, API and insistently unions. You can specify like darker and in fact guys, I think it's the same model essentially yeah.

G

I kind of get it, I maybe there's some background which I'm not really kind of present in here there's a lot of things that are similar so like is this supposed to be analogous to the kubernetes daemon sets, or was it supposed to be for developing I?

G

Guess I'm, not really is the use case, but this isn't all that obvious to.

E

B

So this can be used for development, but it's primarily aimed for running additional components. Alongside the agent such as I mean there's like existing demand for things like admin tasks in basis where, at the start of an agent, you would launch one or two extra things that would be like, maybe a proxy you too agent endpoints or like some extra security features or database of long sided.

B

This would be roughly the same thing where we would launch CSI plugin that is actually offering the storage resources of the agent and we can't have, or in our newer model of resource providers, with agents will need to have the CSI plugin launched alongside the agent outside of the resource offer cycle. So.

G

That, logically, that's as if the agent being a framework for a task that runs from the agent, you can it's one logical way of thinking about.

A

C

I think so, like I mean just like I won't call you a framework because there's some like no like all for psychos. It's like a very dumb, like Indian system, he'll, say like just monitoring the house of the process and relaunch. If you fail since I, it's very simple way of doing like monitoring of a supervisor for a given container did.

D

You guys consider doing something like that, instead of having standalone container having sort of a framework that is launched by the agent when it first starts, that gets the first set of offers to choose. So I can grab a bunch of resources to act as the resources.

B

Are not necessarily or they're, not necessarily tied to a master on it.

B

The the end goal is that you would launch a Mesa agent and then launch a Salem container without having a master in the loop, so the master might not even be up when you're doing this got it. So.

D

Then you have to have an agent.

B

D

Have to have a you know if you don't have to have a master, so then it makes it really look like what you were saying before, like a dr., D or something or you can launch container the.

C

Reason I mean we can't like, for example, like to put the one. The use case for CSI I mean the other. The other alternative is like we just directly talking to dr. Deema for those CSI plugins, but we don't want to do that because we don't want to like strictly tied to dr. demon and all the api's I, don't want to straight I, to doc a demon and and then providing an abstraction layer. Its agent, like makes more sense to me.

G

So then, is this identical to just launching a container with our office, because, again in terms of the isolation, is the isolation still done by the agents.

C

B

Things going through the exact same code path as when you launch a container. Normally oh I,.

G

See, okay, so all right! Okay! This is it's a little bit more obvious to me now. Thank.

A

B

We recently refactored the maysa container answers so that it takes the same arguments for launching a command task or default executor task or nest container or in this case assume. So all those go to the secret valve.

G

So in terms of operator, api's kind of thing standalone containers are visible through the same upload API as any other container. Is that right.

B

Yeah there there will be some fields that are going to be missing, of course, but if you go through with links agent, / containers, it should still be there I'm, not sure what the what the output would look like an agent state, but we'll have to add that as well. Yeah.

C

/ containers we have to maintain backwards, compatibility I, don't think we UN showed I can ask the container in / containers, so we only show, like top of our a secure containers, to be backwards compatible. So even either. We ask some parameter to the request. Allow operator to specify hey, I, wanna, see all containers, including as to container and schedule container. Otherwise we have to use a different point for that. Just so sick of like backwards compatibility.

A

G

E

Does that make sense future.

H

Yeah, that makes sense Thanks.

B

Okay, um where were we.

B

So the is continuing off on the API itself, and so you've mentioned that the launch command for standard containers is idempotent primarily so that you can make the same call twice and expect both this is you, and, as mentioned, we have resources that won't be accounted for in the master. At this moment, I'm gonna leave a giant to do in the code when this stuff is a purview.

B

But besides that, the actual fields here are identical to launch nested container you've still got the container ID and we're not going to enforce any uu IDs in the container ID this moment.

B

Maybe it's because we want some like.

B

It could be a human-readable ID instead of a UUID yeah, since I've got the command info and the resource of sending for Joseph.

G

It sounds like you can't tell the difference between a REIT a launch, a repeated.

E

Launch or just a ID collision well.

B

They're going to be different status, code returns and the first, if you've launched an container the first time, you'll get a 200 and if you do the same call again, you've got a to it soon, but.

G

If you launch a different container whose ID happens to collide, you would get it to it. Yes,.

A

G

Verify that the other fields are identical, yeah.

B

So we're gonna do that, because when you restart the agent containers will still be recovered normally, but we at this moment we're not saving the other fields, we're not checkpointing them the disk at all. So that point we just have the container ID for a stand container and you try to relaunch it. Then we had to be consistent right, okay, um yeah, so we're not checkpointing these things at the moments, because it's extra code and we're not sure if it's necessary yet, but it's certainly doable to checkpoint all these other fields.

B

G

Happens when standalone containers use things like persistent volumes, is that possible? Like? Can you do that for here.

B

So non compressible or non shareable resources are going to be a bit more challenging first Eiland container to use mostly because these resources will still be shared to the mess or they'll still be advertised with the master. So there it's really easy to get a collision if something launches a framework or some other test through the normal.

E

Offer cycle and.

B

It lands on the same persistent volume or the same Court, something, and so what we're gonna do. There is, if you're, going to use a volume or a port on a stand container or some other non shareable resource, then you should just exclude those from the agents resources at start.

C

Yeah I think versus ammonia. It's going to be a bit challenging until we should stay where we can't update the agent resources dynamically yeah, otherwise likely it will be a complex. A some firm launches has using the same, purchased ammonia as a standalone container and dances. Only.

C

You have two more to contains just trying to simultaneously acts to the same volume which my breaks on the code so yeah as Joe some legend like compressible resources, is okay for non compressible resources in Fernie to disallow the initial, a or maybe just say, resources here in the SkyDome containers only for CP in the memory, because that may be discs because there's something we care about and the rest for. We don't care about. Just this saga and I start.

B

For MVP, we might just go through that round to.

C

Think about how gee this is a bit mark. um That's a good feedback when should say think about that more especially for fists in volumes.

C

B

B

Okay, so for other API, so we're gonna add um right, have a weight command. So that's, basically a supervisor or an external service thing know when a standard container exits or fails. This is primarily so that the super or the yeah supervisor of the stand container will know to relaunch necessary.

B

We don't expect all standard containers to be long-lived, but if they have to be, then you have to provide a way for them to be notified on exit Joe.

C

So just one question here: so if the container terminate see, if we call wait, it will still return the exit code. Is that correct.

B

It terminates yeah I thought he ran a sitter fail.

C

Access or there okay, but what, if like I, call wait after the container has been terminated? Oh well, I'll still receive like two hundred okay, because that's just an idea, finesse, the container just wanna, make that I thought. We have consistent semantics where.

B

Yeah so I guess we want this to have roughly the same semantics as a weight bid on normal process ID. So we'll keep this container idea around until somebody calls weight on that container ID. So.

C

We're gonna keep the checkpoint directory for, or rather only.

B

Until the first standard container and then after that, the checkpoint and stuff will be added to the normal GC.

D

Yeah sure you'll force it to be there until someone calls wait, so they can get the egg that status and then after it's been waited on exit status has been returned. Then you can add it to the GC list. Yeah.

C

But that's kind of kind of a little different than like the way paired, because I think like that, you have two component involved boy of the first way. You don't know if it's successful or not I'm dating called a second wait. What if the first way to dis succeed, and then we like kind of GC those they're? Actually the second way is gonna fail, yeah I, don't know like. What's the I mean for the from the caller, then it's very kind of like then the caller don't have a way to get the exit code right.

C

So that's gonna unfortunate first, a second call so say.

B

C

The caller ID issue so if I say color, launch the standalone container and then terminates the container terminates and then I call the first wait and somehow aging flashes before the response of that way, return to the kind okay, now agent we restart recover and then things you like that the agent already process to wait so that direction might be G, C and then C. But the client don't know the result.

C

Yet the kind will call wait again and then the second way don't have a result, because the the directory has the the checkpoint Direction has a secure, already being removed or G seed. So how does the client find out the exit code for that container? I mean.

B

In that case, at least in the current design, we return for a for on the second call. If we no longer know what the agent number that that's no idea, all right so I think like, but.

C

I'm not sure if we want to have an G. Oh no.

B

You don't have to.

C

I think the way we handle diagnostic container is we have a local remove, which kind of me like it's kind of like another, call to indicate that it's safe to remove the other artifacts, the socialist a container, including the on the exit code. So if you don't know the result of a way to should not call remove, but if you do know the resultant weight and can safely call remove to to remove all these artifacts there socialist, a container yeah.

D

We added that later it wasn't in the original launch mystic container design, but we added it later for the health trick work that Gaston was doing yeah.

I

Is removed really intended for use by operators, though, is that just meant for the hell checker.

B

Yeah, it sounded like goes mostly for health checks when we have a long-lived top-level container and then infinite child containers. Yeah.

C

But the remote is kind of indication on the fact that the client won't be able to one be able to get to execute anymore. Otherwise, like would have a GC problem, because we don't know where it's safe to GC, that checkpoint directory for the exit code, I.

C

Think just think more on this and I I do believe. Our move is necessary for this case and we can travel around this I think we should keep moving I think we won't have a 28 minutes. We have another presentation, queued up.

G

So I have a quick question: um why aren't there executives in these containers.

C

Sorry, what was the question why.

B

There are not the executives in these containers, that's primarily because the executor would have to read have to register with the agents and be associated with the framework, and we just don't have a framework for the states. Yeah.

D

We could imagine putting some wrapper around it, that did health tricks and some of these other things like a stripped-down executor or something I, think maybe that's what your point is more dreams well,.

G

Yeah because I guess originally I didn't it wasn't clear to me originally that the agent had to be running for these things to launch so that if the agent is running, then you have a full agent stack on the hose. So you have the executives. It seems like you have almost everything you need to make these basically the same as regular containers, in which case, if they are the same as regular containers, then it seems like a good thing.

G

There's only one if there's fewer it's less diversity in terms of how containers will couldn't feel they thought, through the full implications of that I.

C

Think that's the interesting part like. Maybe we can do that. We just need to make the default as secure or like command. Executors like a registration are optional. So if it's yeah.

B

The other would.

D

Also say it would also rationalize why there's two separate calls for the launch container well standalone, container lunch nested container would actually look significantly different at that point, too. Right.

B

And we also know who descent status updates to or where to report these health check results, but.

G

It seems like if you're running I could see us off at something like a CSI service. Maybe you do what that to be health check right, because this thing never come, it never becomes healthy, like you want it to be I. Think the concept of health checks makes as much sense in the deadlines container world as it does in you know. The framework managed container world yeah.

C

I think he'll circuits, not because the client can do how statistically some edge to the endpoint. They.

B

C

Just just a matter of like uh like reusing some, the color in my sauce, yeah I, think that's a good point that we might want to think about how to you. Maybe you could have like a.

G

Designated well known framework ID constant. Where is the framework stuff? We.

B

B

Well, we might add a health check independently of executor z-- in size, standard containers, because there wouldn't be too difficult to like launch an additional process inside a container after we've launched it.

I

Yeah I could imagine the weight, call returning a streaming response that you would get health check. Events on, for example,.

D

You can always just use internally use the mystic contained lunch, that's a container if you had to launch a new process inside there, yeah.

B

Which is exactly what a health check or some health checks will do.

D

Okay, we should probably move on like cheese now.

C

I think we should move on and do you have anything else like a scroll down and you like.

C

What kill okay.

B

It should be relatively straightforward, we're just sending a signal to the containers that we can get rid of it. You know, I haven't exactly decided if we want to make waiting after a kill required.

B

Because I guess, if you, if you specify a signal there, it might actually not be a kill, you could specify any kind of signal to a container and that will not necessarily terminate the container.

B

Okay, besides that, at some point, if we have a service within maces itself that launches stanlon detainers, we're gonna have some sort of wrapper in there, but if we have say an internal components inside basis, let's launch the stand containers. These will still go through the operator API a just like any other call any other caller. So they'll have the same restrictions as everyone else.

B

That's basically all we have at the moment this design.

E

Okay, executives be able to say, executives can launch.

G

Standalone containers yeah.

B

Technically, anyone to launch a standard center if they have the correctly else, authorization, okay,.

G

So an executor as an executive tickets allocated a set amount of resources. In theory, if I'm not satisfied with those resources, I can just launch a bunch of settling containers. Yeah.

C

I think the vendor point is: why do you need to last a standalone consumer versus like a nested container? In that case, yeah.

H

I mean I guess executives are semi trusted anyway, so it's totally fine practice.

D

No, no, no trusted, they are necessarily. You can specify your own custom executor. Never you launch a task, so whatever you want to primer yeah.

B

And we wouldn't- we definitely wouldn't want our default executor in Mesa, so we launched a standard containers all over the place. Those would be harder to manage compared to an s2 container.

C

Okay, so I think we are out of time and thanks for the presentation Joseph, if you, if you guys, have any comments or suggestions, please feel free to talk to Joseph oversight or emails. I think she moved to the next presenter.

C

Okay, that's Moulton, an exploration so JP diamond airport, isolate or that he's being.

H

Okay, I actually have.

H

G

My screen I just have this go through the design, because it's really straightforward.

A

C

We can see a screen, I see the dog. Are you to.

D

C

Right screen, yes, I.

D

C

D

All right, cool.

G

Because I'm not.

D

G

D

Showing the whole screen, but we can see the guy. That's.

G

Right, okay, so I guess the motivation for this was we run or I cost us with his Network and we allocate ports and we expect user tasks to use the ports that every now and then we have problems or a user task has problems because it doesn't follow. He doesn't write a contract. As you know, it's not always straightforward to do what people forget, and things mostly seem to work, and so you'll have tasks which take a poor start listening on a port that they haven't been.

G

They don't have resources for, and everything will sort of work and then they'll get some kind of weird problems and health checks. If they don't understand- and someone has to be by that for from a security point of view, if you have tasks which can which for where there's no technical restriction on the two ports, they can listen to, then you know there's lots of possibilities where um tasks about trust if I are able to intercept traffic from other tasks, and lots of them is different.

G

I look at security head-on, so I looked at, but basically the idea is looking and look into how various tools work. The answers are not back to line pretty much. What we want to do is you want to make sure that each container only on orcs or which holds resources, there's no real good, colonel api's. For this.

G

Basically, what you have to do is find all the processes in degree find all the sockets and all the processes in the subaru and then match those suffix up with the subjects instead of loosening suffix crew chief obtained from met men. So this is like an otter. End looks like that, because you end up having four scared, 5 ft or every evidence with them, and so redoing.

G

The same thing that I service, like the disk, the projects discussed later dude we're basically, in the background of periodically collected information, eaten and we've resolved it against the risk of of containers, which we know that there's a.

D

Sorry so so this is more of an enforcement type thing, meaning there they might try and just open random ports or connections on random ports, and if they do we'll will have something running. In the background to.

G

The ideas yep that's correct: once we once we find if we find that a process of your container is listening on a port, it doesn't hold the sources, then both ways that contain the limitation and then your eyes will come back. Okay,.

D

You'll kill the whole the whole task, yeah.

C

So what about you, ephemeral, ports like like, if I, have a process? Listen on a like a port on the Econo for range? Are you gonna enforce that as well? Yes, okay, ports like how do you know this is like a around for stack that fashion anything on so.

G

The contract- sorry, it's actually an extra mark, so they have that there's a configuration knob, which somewhat a person's life in general. My view is that we should be strict and that processes should only be allowed to listen on surface for which they hold their sources. One careful now lid process doesn't break this rule. So anytime we have a built-in resource executor that uses the process live. Photos opens a port with us, listen to the port, so we have a mode which this stuff is called container force.

G

Walkthrough sources only, but has a slightly more obvious name in the contacts series.

G

It's so it's something like a.

G

Check agent for Angel, so if you have, if you enable this option, then what we do is we only. We only. There are five Forks that our theme you listen to within the range of ports, that agent foods hazardous resources. What that means, and how close is that in any container, can listen on any for they want as long as it's not within agent support, page I,.

C

See so you have an additional fact. Is there any use case or I go for this back to me off? Yes,.

G

Then use catch. The use case for this file will be off is if you have a custom executor, but users at h2 uses the HTTP, API and you're concerned I. Think.

C

G

So if you have a say like my executives, HTTP API, so we that doesn't listen only ports- you don't just happen to to the agent. We really don't want music containers listening on arbitrary force. We that's generally a bad idea and we want to give we want to give strong feedback to users when they get it wrong. Okay,.

G

So it's actually so the actual algorithm as I described is fairly inefficient, but there is a kernel patch there's a kind of patch here. So thank you to the Verizon kernel guys, and so this is currently in met dev the next kernel. What this does is. It provides the clap by class ID on the net link API. So that means that we'll be able to make this same mechanism work efficiently and where that work is you apply the net class secret isolator and you apply this network ports isolator.

G

Then, when the in the network, port isolator uses the net link interface to obtain the listening sockets, each listening socket will be charged with its class as a then we can use the class ID to find the C group and the secret watch them every cigarette. Can we we know the resources straight away. So all this I know all this.

C

So James I have a question on this. It's alright saying that the net class ID will be a resources in that case, at a friend or step on each container. There.

G

Won't be in the resources I'm once you, if you're using the net class, then what we would do is do something we're doing something like starting with your container ID look in the sugar to find the net class. Oh, oh.

C

I see it's like directly looking to see, go to find out the class ID yeah.

G

And then one stream and then sorry I'm, sorry, basically when we get containers, we going to define the class ID.

G

Of container ID qualities, okay,.

C

Ice-T no question is like: is there a way to detect this kernel feature so that the isolator can be adapt to your likes? For example, if the kernel features they're using that way, otherwise leaving the hallway yeah.

G

We'll be able to do that. Basically, what will happen is so we need only to I need to do a patch to move in on free to come through the API. But basically what will happen is we'll get they stop it in the flow back from the sub diagonal interface and either they contain right there class how they will be operated or it'll be zero. So if it's zero, then they have to believe the old way.

G

If it's not okay,.

C

Got it sorry, I have another question so uh regarding uh like, without this kernel, patch you're gonna do a polling on the socket diagnosis, information I, think my experience, I kind of like this kind of very slow, especially if you have a container that has like a million emotions like like. What do you have a polling like in troubled I can tune or like? How do you make I mean it's here.

G

C

G

A polling interval container for swatch so just like that container discourse interval. So oh.

C

I see it's pretty.

E

G

Modeled on the.

G

G

So that's kind of it um I've been going through. Ehm he's been doing some very quickly that.

G

One problem that he pointed out last last night is that the resources of the ports in the isolated update method attributed to the I select their executives container, whereas we really need to track them on each mister container. So they need to have nucleus, more thinking and experiment about how to sulfa. You.

C

Know I think that's a yeah I think that's a good.

F

Catch yeah, can you say that again they didn't quit for when, when the isolator gets the.

G

Update call back the containerized wolf Aubry, they all the resources and apply them to the executives container. This is when you have a trio of mr. containers, which means that all the ports for all the NASA containers were active at that point would be then attributed to the executive container. What we want is for each ports to be. We need.

G

We know that force for each nested container kind of discreetly, and- and we can know that when containers get launched but I I just look at this target, but I don't think that's a good way to know that yeah yeah.

C

I think maybe like it's okay, because I think because you don't have like a can change. Yes, you can came down, have a separate hour, so it's all on a host now work like just timing. A check on the top-level container should be sufficient, maybe I'm not that yeah.

G

It's not that satisfying I think it might be the only thing that can work but I think it'd be more satisfying to do the other force that may contain a which asks for 32 actually uses 32 rather than the mr. container B, which are support. 40 yeah.

C

I think the tricky part here is like we right now. We don't support a kanessa container, specifying resources, that's kind of like a limitation at the moment, and the reason we don't do that initially is because for some of the resources it's more more complicated, for example for cpus you, if you do a long as the container, to specify cpus. How do you structure the C group CPU hierarchy to do you need to extract the CPU shares from top-level container or not, and we don't have those answers and also I think so.

D

At the moment they don't they don't even go through an offer cycle, so they don't have the opportunity to request. Oh you.

C

Can owe it I could resource coming for our executor, like all the executor, rewriting requests those resources. I, think that sounds fine. What tasks for that executors container can request resources, but the fact that you want to deal with this nesting nesting structure for C groups like CPU memory, it's gonna be complicated and we don't know the answer, so that's it. It won't be that yet, but you'll be great to see. Kinda do.

D

G

I understand tests in this part series, but I already made to spend some more time to make it more sophisticated. So I can understand these issues.

C

Yeah thanks thanks games.

C

Okay, um James: do you have anything else all.

G

Right, we have anything else: I'm gonna do contain each and I have a design.

G

C

G

C

Yeah I think like: let's do the Pam discussion next time next working group meeting. Does that make sense think ago right six up, we have to do that. Let's do the depend discussion next week. Love sorry next time in two weeks and I encourage folks to to do some like background singing on that pitch.

D

C

Description I didn't know, James has a design dog for pluggable or authentication module support can support, form ASOS and we don't have a discussion on that.

D

C

It's James. It would be great that you can provide some background doing the the same to next time, because I don't think many people understand what that is for, like don't remember, most people don't have the contacts so it'd be great to have some some background inches. Actually, although that is, and what is what things are trying to do, that would be really useful why this okay.

G

I can add some more info to the designed up to just for the yeah.

C

G

C

Okay, so I think we are on time anything else.

C

Okay, so I think we're kind of already and thanks to everyone for joining the meeting, do the presentation thanks and we will see you in two weeks: okay,.