Istio Config Working Group, 28 Jun 2018

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Config Working Group 6 28 18

Description

Meeting notes: https://docs.google.com/document/d/1P3p7zOpX66hPoZBi_CiC36JW7JmoaLWqE2sgHvdq5tY/edit#heading=h.p7dex81gbrn7

A

Who's the people package change this limo, okay,.

B

A

B

Hey Scott, hey.

C

B

B

Okay, imma be able to get Gardner much crowd, I guess I.

C

Think you're very interesting. Thank.

A

B

B

Just to go quickly through some status so on my sides, Gary propper.

B

Done so, it is this, the server side can serve in secure traffic.

B

B

So I appear out that allows us to drive geometry purely metal data model listening to like describing the things that we need to listen to on in a metadata format, so that is out so that should cover all the types that's currently known, but it is fun to getting out of segments as we make changes to the API is a series and then I have a follow-up er. A cleanup here to tighten the metadata model.

C

B

That you don't accidentally create like injects wrong coding in in the metadata model and then going forward. I am going to work on the mixer integration for an MCP.

A

Seems to be server, rotations and client rotation is to begin today.

A

And then I'm planning to get integrated into make sure we can see make sure they want those implemented on the pilot Park. But with our timeline is for that those who can get and then the corresponding I think we wanted to have a custom client GP as well just to debug some other stuff.

B

Yeah actually I start Gary. I was waiting on till your care to lend it's a very simple frame. You can just expand them. Okay,.

A

Which we can decide if we want to be that in the gallery itself, or is he still under immigration? Okay, short term? What we want to do, but then we want to test pilot without starving all Akali due to communities I.

B

Think we can go either way for me number one I think: what's your one of the hair is.

A

You guys think if we want to hook it up to my guess profile based. We want up it up to a launcher. Yeah.

B

So I mean we can be MC people that we think really the bucket file based implementation, I think and anybody can just create a wrapper around them, pay the service out at the time. It sounds like the most experienced thing to do. Yeah.

A

Mc P, so it has not my boy or generic to the yeah resources. Some.

B

A

Many dances, texas, beginneth, okay, so to that and then on the validation side, cleaning up loose ends for so there's some telemetry we're adding some additional configuration options for the validation and.

B

B

Yeah so folks, I actually.

B

I've been working on the documents for some time, just trying to kind of rationalize something like relation related stuff that I think we both we are doing and we need to do so. I captured the document. I'm not gonna, go over the document in detail, so I know I signed it right before the meeting. So there's not a lot of time to actually have an in-depth discussion about the document.

B

I just want to quickly go over it, and if you have any comments on it, please feel free to leave it on the document and then, if there is enough contention- and we can discuss further in the next week- maybe but the idea is okay- can we actually rationalize the copic validation strategy in history right? So we actually had a lot of organically grown code, protein mixer and pilot for doing validation. You know it actually tried to do if it's.

B

You know things that in a very interesting way, which I, don't think is is is healthy right. One example is we try to do referential integrity checks, but we kind of did it have Hazard Lee? So it was not easy. We eventually decided to accession that, to that so I decided to write this document, capturing the principles and then kind of discuss what we can do in this space.

B

So the main thing is like there's a lot of discussion about relation whenever we start talking about validation. It starts to you know, going to the you know. All I can really need to validation at this level at this little like okay, let's put a stake in the ground, validation that we are talking about at the ingestion time through the API server. That is the validation that we are interested in. So there might be other points in which rotation can be extremely useful, which is great. We should do that.

B

It's just that this is not the scope court, so it needs to be a static operation. We shouldn't, like the old nation, should be based on this stuffing information that's available, rather than looking at the some runtime states. That might be, you know changing over time. So if you don't do that, essentially you end up in a place where you can admit something.

B

It is because of some healthy state now which can become unhealthy in the future, and you know your relation is effectively broken, so just to keep things simple, the one of the principles this to keep the validation, it's cooked, a single resource. We have a lot of code actually in currently available in the validation where it tries to look at other resources, especially its referential integrity right. So there's a description of what's wrong with referential integrity and those kinds of things here so essentially the gist of it is look.

B

We should just look at a single resource at a time to make a decision about it and that thing is valid or not so, and then the rest of the document is about how to violate that principle in a principled way. So the idea is okay: let's do single resource validations. First right, then, let's start with whatever is available, so communities actually have a city, ID malediction mechanism standing in 1.9 based on open, API, schemas.

B

So I think this is a great like the first line, World Vision thing and if we can actually satisfy the validation needs using this on me great, we should just do that, but if it's not something that we can do it it's, there are certain things that actually we can do the just purely based on this right. So we already have an admission controller based model community submission controller based model. Here, that's currently implemented in Gally, it's like online. Now it's right.

B

You know security, validating configuration, but it has some custom code that is pulling from pilot and mixer. So the idea is okay. Can we kind of generalize this thing right, so the the things that we can do it is.

B

We can look at things in a case by case basis, one case as the mixer adapter configuration and mixer instances. Template definitions right so for those we are mixing is already moving towards the model where there is a set of method, a finish. The resources for these things.

B

Come in all right, so essentially the idea is: maybe we can put additional information in these resources so that the admission controller can actually read the constraints from those resources and then apply them. So the canonical example is the the hinder configuration right, so mixer actually supports complications for handlers. These are third party resources and code, so we don't know the the constraints a priori right.

B

We don't know the constraints of that mesh network, so these can be described in a CIT by itself that describes the adapter I saying that okay, whatever configuration this needs to be supplied, needs to comply with this open, API, spec right and then our admission controller can actually look into that resource. Saying that, like you know, whenever a resource of this particular type comes in I'm gonna apply this open, API spec right. So it's a very generalized mechanism. It just says the source of the.

B

We just need to know the source cid that contains the suspect and the destination cid that we should have quite expected so and then the final one is. We have an expression, validation problem. So in the template definitions, we have expressions that needs to be well dated, so the the hook is currently doing this. But this is about you know, kind of generalizing and rationalizing it a bit more. The IEA is, you know for templates.

B

We are just going to look at the template definition and you can also look at the attributes available to us and then we a somewhat generic infrastructure that actually you know what's over, that girl over that and then ability is that the expressions are of the right type.

B

Essentially, this is the gist of the document and I just want to show a little prototype that I did to kind of enable this work. So one one main problem that we have here is okay, so you know these like. If you want to create, for example, the open API specifications right, there's no good way to palm things from up top at this point right, so I started the whole thing. Some schooling around this and I'm happy to get like your feedback on it, but the main idea is in the API, its repo.

B

We can actually create some, you can we can annotate our resource packs using some options like proto options. This is a one example of it and the example is like. We actually have a bunch of options. You know you say: hey here's a spec, the kind is this there's some metadata. This is the group. This is the version name spaced from the same proto. I can actually create two different sorry through different CS specs right.

B

So this is the second one I can create and when I rounded to I can actually create interesting metadata about this right, so I can actually create a CID that finishes right. So this is the CID definitions that were created from the from from the proto that I showed earlier. Alright, so they're actually two of them, so one of them for the attack not for button one.

B

So there's a this is the regular kata stuff that's created out of it, but in addition to this, we can actually creates enough glue codes so that we can use communities on API machinery to generate client libraries for our resources, which we don't have today. So this has been asked, which is a sense that people want to be able to programmatically access your custom resources and be able to like run on them.

B

It's just that we don't have the API written that we can't generate the API for it, so this allows us to actually create the API for it as well yeah, so I actually have a get another product for that from here.

B

There we go, and so it's the same idea here: I have the photo here. This is an older example that I had so we create this envelope, and then we can point the Trinities code generator to this envelope right. So this is the primitive message. Time is a resource we create an envelope out of it which the kubernetes machinery understands says. We need to generate a client for this and then once we run the client's generator, we can actually create these client sets. You know the communities API like API is targeting our resources.

B

It also allows us to hang a place hand. The open API specifications for validations as well. You know you can't imagine expanding the tool so that the just by looking at the structure of the pro tool, we can actually create an open, API spike. In addition, we can use some proto annotations to further capture some constraints so that they can be placed in this CID definition as well.

B

So all you know we should be able to put all of our metadata into the API zip repo and generate all of the interesting artifacts from there, and you know, validation related stuff being one of them. So.

B

Okay, I'll be happy to field any questions concerns if not feel free to comment on the documents. We are heavily understaffed right now, I'm, not sure how much time we can spend on this. But this is one of the areas that I think. If he spent some time late I mean today, it will actually help with the project helps quite a bit.

C

So this integrates directly with the kubernetes machinery, then the queue control CLI will actually give errors back to the operator. No.

B

So that's a difference. Oh you're. Are you talking about the validation yeah? Yes, so that's the main idea. So in this CIB definitions started with 1.9, you can actually put open. Api, specs and API missionary will directly apply those and if it fails, you're getting you're gonna get an error message. Okay,.

C

That'll be very useful. Yes,.

A

C

Just know that that is something that that I've run into several times where bad config just goes in and things don't work, and then you have to kind of investigate in the logs and figure out. Why? Yes,.

A

One of those that shouldn't be, we should go from nothing to something decent, but.

A

Consistently apply in there's architectural things need to be cleaned up from that right.

C

Yeah, well, you know, part of the problem is I'm I, almost always use cute control and safety of control, so I think sto control does some more validation and keep those you know. Yeah.

A

Actually, right now, it's the same validate if you we'd want to know, but history of control does terms of validations the same thing that the server-side validation will do, except it's not consistently applied. So we're not checking mixture config file in writing. I think there's some other nice things about teaching. Kubernetes are seamless.

B

A

There's some talk about doing, server-side, apply and and what state that's it? But if the server is aware of our schema, then it can do more intelligent merges.

A

So if you wanted to do a patch or apply, we know that doesn't work for C or these right now, but but you in order for that to work you mean it needs to understand humans need understand the structure of our data, complete in addition to the validation we get that as well, which is nice right. It.

C

Makes sense but but all of this is non contextual- it's just purely validating the config in a vacuum. Correct, yes, yeah.

A

I think they're, contextual I.

C

Mean yeah, it doesn't understand the state of the system or anything like that. Just purely right.

B

B

There are certain things like validation of like ten things, for example, we need access to the Sun definition of what that technically should look like right and that's not something that you can study we capture in a series to somehow injected into the system, as we add more temples right. So for that purpose we already have some seers, so you can imagine those years being system state as well as access right.

C

B

Okay, folks I, if you have any comments, feedback any interest in implementing this, which was interesting, please let us know any search the by with yes also, we switch to thank you, so we just meeting two bi-weekly so instead of like canceling every week, I think we can. This will work out better.

B

C

Said why alright thanks good, alright, see you guys in.

C