Kubernetes API Expression Working Group, 29 Sep 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: WG API Expression Bi-Weekly Meeting for 20200929

Description

WG API Expression Bi-Weekly Meeting for 20200929

A

Let's get started all right: welcome to working, welcome to working group, api expression, it's the 29th of september 2020.. um I put something on the agenda. I want to talk about very quickly because I I've been fiddling around with the last few days and uh um feel a bit stuck.

A

um I prepared a few cases, so the problem I I have right now is that force when, when we do status, wiping and things like that, we have a lot of stuff that we take from the object to actually decide like if we want to wipe for a feature- um and we already decided that we do it as good as we can like we, we can't look at the object, so we do what we can. If a feature is disabled, uh we just drop to.

A

We just ignore the field, but we have a few things that I'm not sure how we can define right now. The path specification we have in smd, mainly like we have things like this, where we say we wanna, we would want to say, um ignore every field in a volume, every csi in any volume- and I haven't seen a way we can do that right now um and yeah.

A

It gets even worse with multiple of those, and we have things like this, where we we're currently on based on an annotation prefix one annotation gets removed um that I haven't seen before.

A

um That's even harder to do like we probably would just ignore that case and in general, like I had the question like how good do we want to handle that uh we talked about it a bit before like? Do we just want to make it a bit better and live with a few things we can't solve right now, which I think would be fine for now. But what are our expectations for ga and what are there any features we know of that? We should definitely cover good.

B

Yeah, can you I don't know if you could like it has done. That should be, let's focus maybe on the first one, the uh the volumes yet sorry do we does everybody understand what's going on here.

C

No, I'm still catching up.

A

All right so um before we put an object to storage um this status strategy, resets some fields that the user is not meant to set, and that's done in code like this, like there's uh um yeah it, the the storage layer, looks at at the previous object and it feature gates and decides. um This field should not be persisted if the user specified it, um and so what we try to do with the uh the reset fields, as we call them right so far is we we tried find uh wait. I have it. I have.

A

I have some open here.

A

um We define uh field sets that we know get wiped or reset at that level to be ignored by by our apply. So you won't get ownership if you send them um because we they will get reset and for some fields there is some some logic to it to when they get reset that it's based on the object that we can't figure out in our current process, uh because we can't look at the object like that's that the the stuff here is internal api objects.

A

um It doesn't translate to anything we can map to and say: okay just changed, so we we have only the static way right now and we can look at feature gates, but only at them like here. In this case, that's a very common one. We check that a feature is uh not enabled and that it's not currently used by the old object by the previous object and if it would be used before we wouldn't reset it and that's the logic we can't already do so. We just right now.

A

The the current approach says: if a feature is disabled, just do not assign ownership to it.

A

That's as good as we can do so far, but there's some some parts like this, where we don't have a way to specify that we want to reset every field every csi field in every volume, at least not that I know of so far.

A

We can only specify a path like potspec expect volumes and then the index one and then csi.

C

Right, we can't say wild card there, yeah.

B

A

Yeah and like this uh this case, I I I don't think we can handle right now, uh how many yeah and there's a lot of those like um I've. I've been I've been digging on on the on the pod util that gets um well yeah um yeah. The drop disable feels it's like it's like a quite a few that get reset based on some conditions.

A

Some are easy to do like some. I already did, but some are some some ranges, some conditions that we can't handle and yeah.

A

I mean I I wouldn't mind. I also wouldn't mind just pairing with somebody to go over the entire pull requests, find a meeting, take an hour or half an hour and just go everything, because it's probably easier than everybody trying to review it, but that's still an issue we should think about. How could we want to solve it and.

C

Yeah this is uh yeah. I think you understand the problem.

B

So one part of this program also is that delete setting is done on a on the internal object rather than a version object, yeah and because it simply has the version and object. Obviously, we can't just call this function. The reset function to get it reset and then apply the manage fields.

B

We need to somewhat be able to know without having to convert.

A

Yeah yeah, and that's also why, in like, um I think, I think, custom customer resource uh uh has like.

A

uh On our custom resources, but we we sometimes we specify we have to specify for every version like the mapping we put into the strategies versions, but the strategy itself is not or it's on the internal version, but like we, we had that we had that's already the third solution. We had some approaches where we converted back and forth that didn't turn out too well um it. It works right now, uh the the tests we have. I I added this is working.

A

um The problem is just that: it's not complete and I'm mainly lacking those those fields to reset, and I don't know how we want to solve that. I also spent like last day looking at how fuzzing works. If somebody has more information or not, I want to add more tests to it using fuzzing. We talked about it a while ago.

A

I haven't gone that far yet.

C

It's funny you should mention that I was. I was wondering how you were testing it. I was hoping that there was a fussing solution, because we have to basically test all the types yeah.

A

C

Not yet there there are some examples. I'm like I can help you pull up, I'm sure we can do it. Yeah it's doable um and there are good examples to pull from.

A

Yeah right now, it's only testing that status is handled properly. It gets uh those cases, don't even get tested yet right now, it's just the test is um applying a test data object. It's the same that we use for for the apply integration test and then it modifies status and applies status with a different field manager and checks that there is no conflict and if there's complicated fails, so it just makes sure that status doesn't get old on when it's applied on or when it's added on a non-status call.

A

That's the most basic thing. I have right now.

B

Okay, so if we look at it, I think the the fields that is never going to be good enough, because we are always going to miss some dynamicity in the fields that we can say it depends on the current state of the object. Yeah, we don't have the battle matching uh the the star, so I and I don't think we can go and should go um the field search into infinities uh and it doesn't fix the dynamicity anyway.

B

So it sounds like the solution we have today is maybe a good workout, but it's never going to be perfect, so we're going to have those. uh The question is: how bad are these bugs? Are we happy with these folks going to ga? uh Will we ever want to fix these bugs?

B

What's the solution.

D

B

So can you describe the verb.

A

B

Happen if we keep going this way.

A

um Yeah, so in with the with all the fields I saw in in this pod thingy at least um the box we would encounter. Is that a feature um it? It always comes around a feature not being enabled or being enabled and when a feature is disabled and somebody applies it that they might own the field, um even though the fear of the field does not exist. Like some some, I I saw I I I made notes for them, but I have to look them up.

A

I saw one or two where the feel where there was also comments like the field will get reset, because some other mechanism takes care of it. It's with some feature enabler disabled, but in general it's just. The user will own a field that they will not set, and it's mostly like very.

A

Very specific to that feature for some parts that were more concerning it's like um I think it was, was uh something like multiple eyepiece for for one part or something like that where we might cause conflicts uh with with uh internal, I mean with operators, but those should always force, so it shouldn't break the use. It will just cause conflicts that we, that that are not real conflicts.

A

The worst case.

B

Yeah the fact that we changed the managed fields, even if nothing changes, will cause uh a diff to be produced and the objective returned to ncd. Even if there's no way he's done true.

A

B

A

C

Yeah, the only good news about that is that, like, if they're applying the same thing over and over it will be consistent right, it'll be a different, so it won't cause like right. It won't cause that won't necessarily cause churn.

C

um It will just you know in the initial time when they introduce that field, they'll cause a right.

A

Yeah and if if a controller goes ahead, and are we sure about that retakes the field it will, it will make them reapply a change like we go back and forth.

C

Oh yeah, so if you had two, if you had two things fighting over it, then one would be adding it. The other would be removing it.

C

That's true of a lot of things, though,.

A

I'm also not sure if it's okay.

A

I just want to say I'm I'm not sure if at that point the decision to persist hasn't already been made like I I don't. I don't know the the chain exactly like when when when we decide not to write to ncd, um but it might already be too late at that point,.

B

Okay, just before we save to entity, we compare the object, but the problem is that if we change the managed fields, then the objects are going to be different and starting yeah.

A

B

If we make any change to the managed fields, it's too late.

B

um Can I propose a so I know we've proposed many different ways of doing that, so maybe we already published that. But what about you control the object?

B

What about we call this function, so we have to convert the object to internal.

B

We call this function that is going to clear the object and then we convert back to the version and then we remove the things that have.

A

Really moved uh we we already tried that um the problems were that the conversion um lost intent on not defined fields. I think.

B

Yeah, but what if we only consider all the things that have been removed, not the things that have been added.

A

B

Right so you you take the field set of the object. um You then convert the object to internal apply. The applied setting then convert back to the version.

B

Take the field set and look for the things that have been removed.

B

Not the things that I've changed, but only oh, I didn't, but only the things that I've been removed. I I suspect we've tried that before yeah that wasn't.

A

B

Reason why would you.

A

Yeah, I I don't remember the exact reason. I think we had some problems with the crs.

A

I'll I'll, I can look up the pr discussions again.

C

A

Want to do that, like I, I know I've I've we at least once built that.

B

Yeah, I I can't believe we would have a primary class for that, because we don't care about zeros. Do we? Oh, they can have status, though.

A

B

A

The current approach handles crs probably like if, if status is defined as a sub resource and a custom resource, it will get reset and not owned. That works. Luckily, that's the first time crs work with this.

A

B

I I suspect, as long as we're trying to have a different mechanism, this is going to be hard to have the two map. So this is what I write a long time ago.

B

Yeah so, first of all, they're going to be different mechanisms are not going to have the same semantics and plus they might default in time anyway. One's gonna be updated and the other is not um so really.

C

B

If we could have the same thing that be perfect,.

A

Yeah, okay, that's um so.

B

Maybe spend very little time looking into it, why we uh discuss this option and maybe document it somewhere. So we don't try again in two months.

C

I mean it seems sound right, like you can get the field set from any object, and so, if you computed one before and then computed one after and dipped them, it should work.

A

Yeah I there there were some some problems with conversion and um I don't remember what all of them were.

C

Yeah and if there's a bug in that, we probably need to figure out what the bug is and fix it. If we can, because it sounds like a really useful capability.

B

Yeah but I I'm sure, there's something that makes it fundamentally impossible.

A

Yeah, um I think um if the user didn't like, we don't get a a full object, all the time with apply. So if we go to versions we get those fields set to some default. I yeah, I think conversion did defaulting or what was some so at to some extent. How was it? Oh.

C

It's defaulting in a mess with this anything.

E

B

Basically, you have to default, but it's only going to add fields that remove some so that shouldn't matter what we killed out here are the fields that get removed.

B

Don't they get reset to the value of the of.

A

The previous subject, but after we reset the end version and then we go back to a version, object and the I think there was defaulting, or at least something on that path like because we work with an unstructured. I.

C

Feel like this is still trackable like, but yeah there's a lot of moving parts. So maybe I'm missing something, but.

B

Okay, I suggest we uh maybe kevin, can you look into it and then maybe we yeah.

A

B

A

I'll I'll pull out the last three people requests and.

A

D

I can find out more, but I I think we need yeah. We need to set up a meeting, yeah and figure this out.

A

All right, that's that that already helps me. I feel uh every time I look at it, I I I feel a bit stuck sometime recently, all right um yeah. I think that's that's it for for this one.

A

uh Anybody else has something to bring up here.

C

All right, I'm back.

A

And let's see uh declarative defaults.

B

A

The normalization.

B

Thing because I merged it into the defaults, um but okay, so the cap is uh the gap between yesterday. If you want to.

C

B

Please do uh if.

C

It's nothing clear enough somewhere.

B

Please you can tell me and uh if you think it doesn't solve the phone, tell me please, and hopefully we can do that before code announcement freeze, which is going to be next week. Also.

A

C

Yep yeah, I think we need to get it to um our approver reviewers sometime this week, yeah absolutely yeah. um I think I was pretty happy with it. I'll do another pass today, just to see if there's anything else. Thank you.

B

I'll look at it after meeting and then yeah and then we'll ask anyone uh these just the defaults that is mostly implemented by jenny, which must need the uh the default piece to be present.

A

Yeah, that's! That's! That's! That's blocked on this one yeah yeah.

C

Yeah but it's good.

B

Almost stops, though, that's good all right, that's.

C

Great, so we can just pick it up.

B

I could actually maybe get the code back into shape while the camp is being reviewed so that we have that ready. When we, uh when we get to get ready.

C

Yeah that makes perfect sense.

A

All right, then, we we just talked about this one um scale. Sub resource uh is uh andrea. Looking at this- or I I just saw some discussions on slack.

B

F

Yeah I haven't started yet. I just started to read the various issues and prs around. I'm probably gonna ask questions uh as soon as I start working on it.

A

Yeah, I I think I looked at it before. I might still remember some things: cool yeah. Oh, look! It's like yeah.

E

Please ask questions.

B

A

Okay, how about dangling.

A

Fields, wait is that new in the list yeah.

B

We had a bug, we solved it in. uh We had a bargain.

B

We fixed the dangling fields in 119, but we've never uh plotted these. That's the different thing. Oh yeah, that's the different thing. Did we fix this one? Yet? No, no! No, but we still need to back up that um the day we fix it.

C

Yeah, that's the only.

B

Thing, that's that phase.

C

Yeah, thank you so.

B

I've set two backboard lists, one for the actual change to the dangling fields and one for the pack that we haven't fixed and I don't know if you want a backpack to 116 or not. But I suspect we want to backpart. As far as we can.

C

Yeah, I think I think, we'll just request them all and see what the um the release team thinks.

A

Is it something that's um that's a apply only or does it also update? This is apply only, I believe, no.

B

A

Yeah I mean, I think we started having proper versioning with the 161 like when we fixed the the smd version. So sometimes back porting was was a bit tedious, but we should be able to backport it.

A

D

um Students again.

A

I think that's it.

A

B

Okay, yeah: we haven't done much on that. But again I'm not super worried about that plus, it's probably something we can do uh before the next release, not 120 but yeah, which is the one yeah before 121 yeah. So we have some time.

C

Yeah, I agree it just needs to be done, but it's not like a long problem.

A

Yeah, maybe maybe it doesn't con, considering it um before, because if we want to change flags, I think there was some some limits on when we allowed to do that or how? How long before we have to do it yeah.

A

I think we had it with what we wanted to remove for renamer flag a while ago.

A

Okay, yeah, um I don't think something else happened here. I've never heard anything else.

A

C

We have something: yeah that'll, be okay,.

A

Okay, convertings to smg schema structural schematic, smd schema.

B

Yeah we need someone to look at that.

A

Okay, should we go okay, yeah? That's sounds good.

A

Okay, um api service is still. We still need somebody that knows more.

B

Yeah- and we don't know if it's super important yeah.

C

Yeah, I think that's the thing we need to figure out is if this is actually a priority.

B

But yeah but as you said last time joe, I think again it's going to be hard to argue that this is not needed.

C

That's what makes it tough if you just ask people if it's needed they're going to say yes, so we kind of have to like yes, the right way yeah we need to. We need to think about it ahead of time. Think about what it would really mean.

C

Why why something like this would couldn't be done, post gaa or could it um and then kind of like champion an outcome that we think is saying rather than just asking people yeah, okay, uh empty politics? You can put my name next to that for now um I can. I can start to think about that and get ready to talk about that.

A

Yeah, if he could.

A

Type: okay, um the empty port, selector.

C

Yeah this is the um the fix for um being able to support, um adding or removing atomic from types is um out. The pr is a bit out there. I think antoine was trying to look at it. There's a link to it already.

C

I think it's, I think it's.

B

C

Yeah I've left a few questions, but I think it's ready if we want to go with that yeah that bottom, that very bottom item there um yeah the support.

E

A

Oh okay, that's it there's a discussion going on. Okay.

C

uh Yeah, so that's going uh the other thing. Well, I'll wait until we get to we need when we get to kind of go. I have something to say too.

A

Yeah did we. We still have to investigate that.

B

um It's I I mean yeah. The problem is we update the timestamp of the managed fields, even if nothing changed? I need to look into this I'd like to understand why that happens.

B

That shouldn't be too difficult.

A

And quantity fails type checking is that's fine, it's it's related to that. Can.

B

Someone yeah it's it's optional, let's make it yeah I ever be out for that.

B

B

Can someone wait? I just faced you the same thing.

B

Oh wait: can someone look at it and it's it's it's a one night out. So if someone can actually look at it.

A

I'll look at it, um I don't mind looking at something else for once.

B

It's it's gonna. Take you half a second! So okay, don't get too excited.

A

um What was that? What would you what you were talking about on slack that it's it's not uh it shouldn't be blocking. I I remember you, you said something to uh andre yeah.

B

No, this absolutely should be.

A

A

Yeah, I just remember something you mentioned that it's not worth. He looks at because.

B

It's okay: the other aspect of this bomb is not blocking. Okay,.

A

um Yeah, we do not have a document, yet um I don't know I can go ahead and create it or yeah.

A

Oh okay, I'll I'll, create one.

E

Easier awesome.

A

Yes, please: okay, um yeah, that's not ssa scope. We, I guess I'd, say we go with those.

C

um We do need I talk about you both and we'll talk to both the sick chairs. um Oh, are we still not at clango um I'll, wait, one sec when we get down to api clients.

E

um Yeah, uh the bag wasn't good, uncomfortable namespace by focus uh also on script. I don't know if.

A

We're still making progress on that.

F

Yeah, I've created a pr and it's out there: okay, great.

A

uh Yeah the validation code, yeah, okay, clients now go ahead.

C

I think I think we need to move this up for ga. I talked to both of the sig chairs for api machinery, um daniel smith and david eads, and they both felt that um something along the builder pattern lines would be needed for ga.

C

They also felt that we should take only one, but at least that we needed one example of a controller using apply in the main kubernetes coupe code base. Before we went to ga they wanted some evidence that it could actually be used. um So I think we do need to move this one up as required.

C

um I'm I'm hoping to start working on it more seriously in the next week to try and get it going, but uh the idea would be that we would create a feature branch against kate skates and start doing. um I'm prototyping work directly there, instead of um in a separate.

E

Repo sounds good.

A

um Let me let me know those really quick for the document.

C

Yeah, that's blocking yeah.

C

Also um migrating one controller and, like the people were not very strict on that, like david was like you could take the smallest controller. There is and do it, um and that would be sufficient. We might we might set the bar higher for ourselves, but um they just wanted there to be one.

A

All right, I'll I'll put it in the dock cool. uh Where will we.

A

All right, tom's, turning any any updates on.

C

uh We're shelving this, um so you can just mark that as shelved. If you want um okay, we're going to um defer a list of postgame work.

A

Good do we have okay, we already have g required, okay, that it's not the coin as well.

A

That's not in scope, manage fields from isn't that? Oh no! It's done for edit right right. I think I saw pr about that. Yeah.

A

um Okay, that's the that's the status code thing um I I recently closed the cap. uh I think yesterday the cap pr um because we we said we don't wanna, continue with it right now.

A

I I still think from from the client side. It would be an interesting thing to have, but well we don't need right now. Yeah.

B

A

The clarity field manager will probably stay asleep for a long time.

A

All right unions, immutability uh yeah. We still have the equation controllers. We should at least think about if they are ga required, I'll put them on the list.

B

I'm gonna move move it slightly above.

A

Yeah, uh maybe higher now yeah documentation is also.

A

I mean if we, if we say uh a working controller and a client, is part of alpha g of our j requirement, which also make sure that people know how to use them like how to use apply on a controller and not only show it off as it's not as a internally.

A

Okay and we reach yeah great.

A

Awesome anything else we want to discuss.

D

Yeah slowly you can did you have any topic.

B

Of things you wanted to discuss.

D

F

Not at the moment, I just wanted to make sure I was following.

D

Along with things, thank you.

B

Awesome cool, I think we're done.

A

Stop recording.