Kubernetes WG Apply, 3 Apr 2018

Previous Meeting

⏯

youtube image

►

From YouTube: Kubernetes WG Apply 20180403

Description

More info on this WG can be found here:
https://github.com/kubernetes/community/tree/master/wg-apply

A

Hi everyone. This is where this is. Where working group reply, it's a little early, the one it's an April 3rd. This is the second will count the last two as a single one. It happened not two days. This is the second like sprint planning meeting I have a bit of an agenda that I'd like to go through so I think I want to spend the first I, don't know two or three to five minutes.

A

Looking at what we did last week, then Antoine and I have been putting together a document yeah like a road map of what we can be doing so I wanted to spend five or ten minutes. Looking at that, and then I want to make a list of things that we want to do next week by next Tuesday sounds good, sounds good. Okay, I'm gonna present my screen, oh yeah, yeah,.

A

Okay, I have my screen going to.

A

B

And hangouts I'm, presenting to the screen internally, okay,.

B

I'm gonna make.

C

This window, smaller.

A

Okay, so let's just go real quick and what did we do so Antoine its? Are we a real working.

C

A

And we determined we didn't yet cap, so I didn't.

C

Have a done in column AI way to imagine the notes, yeah.

A

Insert a column whoa right there.

C

A

I'm just gonna copy, whatever gigolo does, which is which sprint was done in okay, yeah I'm, not very ideological about how how we track stuff, like this I'm, just doing whatever make a feature branch nope. No, we didn't do that and we.

B

D

A

To do it, how I at least not for yeah I'd say most of it. We, the conversions.

D

It sure that we can, at.

A

Least, make a little bit of progress without making a feature branch, so we will go over that when we look at our map that we started finished collecting test vectors. I'd say we made progress, but we definitely didn't finish that yeah Jenny's working at Jenny's, working on that and I know I saw a thingy from her. So the blue and I said yes, yeah pull request, so, okay, yeah, okay, describe open API inspection tags. I did see.

B

A

Two of you edited my document, so we got some description. We.

C

Have some not all of.

A

Them yeah, not all yeah yeah, but we have at least one. We have enough to get make progress. Yeah, yeah, okay, do.

C

We want to finish that you.

A

Know? Okay: this seems like it's not done not done done, of complete yeah. Okay, finish different route, hahahaha! No, it's not done make a branch in the cube, open, API repo, we I! Guess we decided, we don't need a branch there either. No in fact, I've already sent out PRS and should include that we're just gonna be merging into I.

D

Bet Maddie.

A

Didn't know about this a little earlier morning permitting any.

A

Many was on vacation. Just got back, I had one extension tag to the open, API refunded branch, so you've started on this, but it's not quite done yet. um Actually, the other.

D

A

I could yeah I've gotten all the way through to JSON swagger, with list attribute equals atomic I.

E

A

Later so, I mean if GM owners produce any pennies.

C

A

She's a little bad thing, though, it's good to have a single value, both yeah. We definitely want everybody everything to have at least two people that are following along at home. Two.

E

People that are out that are understanding what we're doing I just mean at some point, probably cleaner. A sure that people in this room, but to be honest, I, mean II, has been very sponsor. So fellow another.

A

Poem yeah and many sits right next to you too yeah. So oh Jenny says she added a lot of test vectors to the merged into doc, but it's not done okay, yeah, so progress, but but not done I think we signed up for two weeks worth of worth of work. So this is like sorta half done this extension tag and won't fit to the open reaction it.

A

So it's not merged okay! Well, we'll leave it open and that.

C

There's a pol implement one go.

A

Idl tag, tag I, don't know how these two items I'll do front, yeah, they're, actually kind of the same. No I was thinking, one is about the generator and one is about the opening yeah repo. So the one list attribute in in in pods containers all the way through to swagger she'll, get.

C

A

C

So I should not have listened to you yeah. It's good.

A

Hold on a second fan, I see your chat.

A

And file I kept and done it's done. I found it I, don't know. What's supposed to happen next, I.

C

Can think of ways Joe I know he wants more things to be done. Definitely, but I don't.

A

Know really ones busy, probably take the stuff that we worked out in this document. That I'm gonna show. In a moment the man yeah yeah, yeah and and like copy paste that into right to their.

A

Fan is asking a question in chat and I wanna postpone that until we talk about what we're gonna do next print, that's not okay! Okay,.

E

A

Because next I want to oh it says the meeting log over supposed to take a.

A

Second well, I find the document which I don't apparently I, don't have it open in my window, let's was David open.

C

What did I even call that document, but it's not that you want the link. I can't give me the ring and I'll add the document Ginga in the drive. Oh yeah, if you just move it into the other, is okay. You found.

D

C

A

A

Realizing where we realized well Sean.

C

A

Antoine and I were and Jenny we're talking that we needed a more detailed map of what we're supposed to be doing here. So I broke down. I've got two sections here: one is a what we're doing and the other is. Why we're doing it, though? That's my computer vision, we.

D

Need an eight laptop.

A

Well, I'm just predicting to two different sharing.

D

Systems, at the same time,.

A

Yeah and and I have like you know, I could join me from tabs open. So.

A

Briefly, go over what and then we can go over. Why that's not obvious so first first on the list is a tool. Change changes, as vision of our our API system is like type. Echo has got an ideal. It's basically an ideal, definitely right out with ghost trucks and tags in the comments next to the fields and types and I think we need to build that out eventually into something that is a API compiler. Basically, so the input is this x that go on this directory structure.

A

The output is like the serialization deserialization, the the code, to put it in a converter, serializer, the like stubs for registry. You know put your put your read strategy here that sort of thing your validation goes here and I think also. It needs to put out the open, API spec and it needs to put out a CRT definition and eventually that somehow needs to generate a client too.

A

So it's like there should be horn system and making a kubernetes api that either installed a C or D or if you out, throw that and install some other way in service of that. Like that, that's great, but we need that right now. What we do need is a way of describing the topology of our types so that they can be merged correctly, and we have like some merge tags in there.

A

We're gonna leave those into place, deprecated them, but not remove them for, like ever, probably but and added some more and tags that are directed at like declaratively, describing our API rather than giving instructions or hints, were or imperative rich operations.

A

So next handler stack changes. If you look at like the output of the the thing that Phil's team has is some topic merging, it requires the open API schema, so we have to make the open API schema available to the handlers from which will be calling this code and there's two sources of open API aspects. There's the stuff. That's the output of the open, API library that many put together and there's also stuff that comes through the validation field and series. So both of those need to get plunge through to some default implementation of apply.

A

The thing we're doing to apply to to implement applies a new content type or the patch handler I believe that, like we can decide this when we are closer to actually implementing it, which should be who you like in a week or two but I, think we're gonna have one apply, that's targeted at users and one apply this targeted deck controllers. The other option is to give kind of a grab bag features that any caller of apply can turn on or off I. Think it's better to provide targeted flows where we determine the features. I.

A

Think controllers generally want to say, like I, want these fields to look like this I, don't care what I said last time. This is what they should be now if somebody said something else like ignore them. I'm the source of truth here and I. Think users generally want a three-way merge where they're like this is what I wanted to be now and the system looks at like the last time. You said this, you said the same thing again this time. Clearly you don't want to change it, so I'm gonna leave it question.

E

About that you're late merchant controllers- yes, does that mean that controllers aren't deleting fields or that they do so like imperative, Lee.

E

A

A great question: let's talk about some other.

E

A

A great question was talking about earlier, so.

A

Another large change to patch is that now it may need to call post like right now. If you control apply and create objects, the control playing version of it also be able to create objects. I know, there's a section in the update strategy about whether it's allowed to create things or not, and it's on some resources and often others I think you need to allow it for apply. Even if it's not a lot of a regular foot, so I don't know if that means changing all resources or if that means making apply, ignore it.

D

Give me patches is doing HTTP amended, you mean like they could. They may need to create, instead of just not correct.

A

D

A

Mean patch for things and then all three deployed doesn't know it calls a patch, okay, yeah, three things now I, don't think I can create things now make.

D

No, it depends on the resource, I think one or two resources allow it to, but that.

A

They must be reusing those the field from the point strategy, yeah yeah, so we're gonna, you're not doing or that or or change it everywhere. Yeah.

D

I did have a question about apply, creating things so and we can file this away as a separate conversation, but is the assumption that everyone who uses apply is providing enough of the object that it would pass validation if the object didn't already exist.

D

A

Ensure yeah, you have to give enough the. If you don't supply required fields and and the fields that you omit don't have defaults, then it wouldn't pass validation and you wouldn't be able to create it. Yeah.

D

Okay, so isn't that going to lead to people putting fields in there apply so I passes, validation and then, like claiming ownership of it, even though they didn't necessarily care I like.

E

A

Let's look out for examples of that when we.

D

Yeah, it's fine to keep that up. As a conversation, I just have seen a fair number of issues where people said well. I need to make the object so that apply works, whether the object exists or not. Yeah.

A

So I think if you make it and then somebody else changes the field, then by the three way, and you never touch that field again by the three-way merge rules.

A

Somebody else owns it and and and you're not gonna accidentally mess with it.

A

Maybe we should survey objects, make sure that there aren't I think actually what you're describing is an API bug like if the user plausibly doesn't want to set that field, and maybe we maybe I shouldn't, be required.

D

Well, I mean it's skills that are required to exist owned by somebody in the object.

E

Uninitialized almost.

D

Well, yeah people were surprised by how apply we behave differently like apply, would work if they omitted some fields. If those deals were already set on an existing object, but it wouldn't work to create the object in the first place. So when it was a single actor thing that was easier than easier to understand. I, guess, but now that we're talking about multiple actors, I'm just trying to understand what people will likely do and how that will, how that will play with each other yeah.

A

I wonder if we could build out like a confused, submit a partial object when he goes into an uninitialized state until the rest of the epoch shows up I. Don't.

D

A

We should do that in the our first take on this, but I actually I can imagine that being kind of useful. So maybe we can think about that later anyway, we're not going to go through the kids through this document, if I don't hurry up. Okay, so that's the patch Handler fix, update, improve emerging code, make sure it is library ID, so it can be called control plane without dragging into world our dependencies.

A

Take all the test vectors that we've collected and run them through this code.

A

A

We're still debating a little bit whether we get everything we need with just three-way merge that can also implement to merge so but we'll try this code before before writing a bunch, although yeah, but we can talk about that once we once we get this stuff invoking the right place, then it's easier to get a read on it: yeah, okay, metadata, API change.

A

Obviously we're gonna have to feature gate this, but we're going to move the last applied annotation from an annotation to a personalized citizen under metadata.

A

Conversion defaulting spec changes I need to write this up. My last few days have been a disaster and I haven't had a chance, but this is. This is basically David sent an example of house.

A

Basically, there's there there's groups of fields that need to go together like some fields, don't make sense of other fields aren't set. So even if you don't default the entire object in order for a conversion to work and actually accurately capture the users intent on the other side, we need default one or two fields.

E

Is that true the like, if you think of the input, oh I,.

A

Can send you David's example right: it's not Sarah, but I didn't write that out. I think the best result I think the best solution to this is to give the defaulting not to like turn it on and off, but they have. It have two modes, one where it just globally defaults, everything and the other where it looks at the fields which are set and, if necessary, defaults additional fields.

A

The other option I think that's probably the best option. Let's not design it here.

A

Validation changes since we're going to have the open, API schema available to the handlers anyway, pesum into tor, like we can call a validate thing on them. For example, the thing that I really want to validate once and for all is that associative lists don't have duplicate keys in a hilarious bug where, if you added an environment variable twice and then tried to you, remove remove it with apply it, remove both of them, which is love, I'm sure it amused me okay, admission chain: this is this is probably a future.

A

This doesn't have to be in like the first alpha, but I think. Eventually, we want to take yeah. Eventually, we want to take what mediating a mission webhook or mutating plugin does put it in the form of apply and drop it in the last applied state map, so that people can see what the mutating admission plugin did. I think, eventually, we want like anything that mutates the advocate its its own list there. So it's very visible who said what? Because they're going to be like a design proposal for each one of these things, I think.

E

A

Least at least we may combine some of these things, but I think this is all the things that I think the project involves, and then I've got a separate timeline section which gives like a reporter, because some of these are more important than others.

A

The putting handler just like same same idea, if you call put, we should compute what you did and put it in the last applied state edition of dry run dry run, is critical to know what you to be able to predict what the system is going to do. When you add something. I think this is important. It's also kind of hard. We have to add it to every you taining verb and we have to thread it through. The web hook api the admission webbook api, because web books cannot have side effects on the.

E

Dry run, but you can I see yeah I'm.

C

E

In the system, but they still needed me to have Gothic right.

C

A

E

C

Side effect, the sewing think he's doing something understand even a validated a.

A

Side effect, like deductive cortas,.

A

Version of skews gate safety, like I, think honestly I think we should get it working in like alpha and not worry about version skew for a good and then and then once once we have that, then we can build up a test we and make sure everything works. I think the tricky case is when you haven't old, acute control and a new API server. I think we might.

E

Get this for free, that's a lobster's I think they pay it simple enough. We're just sending like one.

A

One raw file well, but an old acute control is trying to write the annotation.

E

A

Somehow, okay yeah once more find releases yeah. Once we get everything rolled over then we're good but I'm. Talking about that. First slight: transition from.

A

Paint into that, and last thing this can be done in parallel with a lot of other things. Is you change the idea we want to make it more strict. You need to roll that out in a in a piecemeal fashion, like first we put into the thing that makes the idea that gives you. Oh it's like a warnings file or something we check that in we tell everybody this morning.

A

Smile must not get bigger and then one API the time we can go fix all the warnings and hopefully the things that we change our our backwards-compatible like making it feel optional or adding a list tag. Okay, do.

E

We just as a follow-up question for some of the things we may want to do it before we like beta, is this up I.

A

Think if you look at the the timeline I'll share this stock with the group. Everybody need to eat right, okay, great, it's in the it's in the drive, I think Google Docs is either I. Think I still need to share it. You can you put. Another drive, showed oh okay, cool, so it's in the drive so go. Take a look.

A

We are running out of time. See Jenny is a chat message here.

A

Yeah, that was those are previous yeah discussion. Yeah we talked about that fine fan. I saw your request and I couldn't tell. If that was about me. The existing apply code, which were mostly not using or if that is was about filling companies knew we.

E

Should talk about, we need to give us back of what conflict means, unmake, you asleep, and what we're detecting rather than working on the and let's follow up cuz I, think that was part of what you were wearing and to is what, in the conflict me yeah.

C

A

Oh yeah, okay, so we have two minutes left. Is there anything that we're not going to do in sprint to see so first I'm going to move the stuff that we did at one there.

B

I think likely not going to do this. You could probably remove that make a feature: branch, okay,.

A

Your ranch is there anything we are going to do that. I should add.

B

E

This money, not doing things, is better than finishing them. Can they don't have to maintain them.

B

A

I'm gonna add like finish that dock.

C

Take you up on that.

E

Is that your user documentation for the email tags with it in those.

E

These club, open API, didn't have a place for the user documentation to go like a dust from up with one I mean if we just put it in the go rock it gets posted. But if it's, what do you invite user documentation.

C

E

People are gonna use, it was, they begin.

C

The user would be the the API.

E

Author yeah, like as I'm reviewing this thing by having the context of this, is highlight either someone writing an API or someone using an API but needs to understand how things are actually gonna get merged together or in they apply. And you know those are two vastly different categories.

A

Of people- yes, maybe.

E

Our ABI authors need to know how to write.

A

Any of yata API users need to understand how to use the API. The documentation for the API users ought to be generated from the tags.

E

Summit, the conflict stuff not like it day, I mean audition of what these mean has to be. Yes, the API author.

D

E

To like write comments and stuff, no I mean like when the atomic like the the conceptual model. An Associated list needs to be documented, he's different right.

D

E

Pretty single instance, but like we need to conceptually document. This is what I'm associate listings I mean: there's kind of exist for apply now, but no one understands it and we're okay.

A

So that's that's two different kinds of documentation: yeah.

E

D

E

Like I'm, not giving a priority on this here is where this fit into this may, just because I have a feeling, link, I, think reviewing that stuff. Might there might be questions right and reveal ambiguous items.

A

Okay, do we want to do either those things to sprint.

C

If those I don't think so,.

A

So is there currently a place where our tool chain is documented, not.

E

Yes, the there is an apply document. There's couple places, there's a fine.

D

E

No I mean not a place; it should go now. Okay, it should be maybe in the cube generator repos that right good, please a place to start to start and I miss you mean you wanted to be good. Lucks.

A

E

Least some of it I'd expect the one for the AP offers to being go Docs. So it goes.

D

E

Expect the one or users probably not to be good enough, because it's conceptual are you gonna sign up for that? Not this brain okay,.

A

E

Know if I can promise, it doesn't need to be done. This print yeah, okay, are.

A

There any other okay we're over time, so I think maybe I'll cut this off.

A

A fan I'm not sure if your item needs to be track on this sheet or not I,.

E

Don't know what do you think bill and can you and Antoine sync up offline and then decide what the appropriate thing to do is yeah.

C

I was gonna, suggest and then you main and then.

A

We can intelligently, maybe yeah what maybe yeah that's what's gonna they habit of just send stuff like this. As a you know to the group list and yeah that's a good idea, then we can all wear the not guys ever like differently debate, or this is a slack channel.

D

A

C

I could send slack yeah.

A

Awesome: okay, thanks everyone I'm gonna figure out how to stop recording okay, maybe I will.