Kubernetes SIG Network, 24 Sep 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Gateway API Meeting for 20200924 (APAC Friendly Time)

Description

Service APIs Meeting for 20200924 (APAC Friendly Time)

A

All right this is september 24th. uh It feels like uh we have made some huge progress today. Thank you to everyone uh for helping get the last few prs in um I just updated. Just as this meeting was started, I updated the release checklist.

A

uh We moved a bunch of stuff down at in the past 24 hours, uh so we're now just waiting on status conditions and the very large documentation section, um and I really I should add. To be honest, I should add the api cleanup, because I think that is something we want to have as part of this.

A

um But we are. We are making great progress. I think um so. Yeah thanks to everyone for uh contributing here, and I I had to say I think harry and I must have been thinking the same thing independently, but I looked at the list of pull requests and we had four open, which is fantastic.

A

um It's lower than it's been a long time and I looked at the issues and we have you know when I looked there was 80. I think harry's already closed some. So thank you, uh but today seemed like a great day to spend some time focusing on issues.

A

uh So with that said, uh it looks like harry has added uh a lot here. um It looks like uh these are all ones you think might belong in v1 alpha one. Is that correct? Gary.

B

Yeah, so I went to the list of the issues and I've closed some. I have pinged the authors on others.

A

Sorry, I I'm I'm sure everyone can hear you but me, but I screwed up my audio somehow so one second.

B

Might be me, can somebody else confirm if you can.

A

Hear me sorry, I missed whatever you were saying: okay,.

B

A

Harry, I think, you're fine.

B

Okay, so yeah, I went through all these almost all the issues. At this point uh I closed some, where I was confident that things were settled. I've pinged, uh multiple people on some issues where I think they settled, but we probably could use some consensus and then I, while I was going through this, I thought, like some issues, need some discussion. Probably you know like we should at least discuss like. Is this part of even alpha two or this is something part of even alpha one?

B

And if yes, then you know, if somebody can, you know, volunteer on that. So just do some charging of these issues that that would help us.

A

That makes a lot of sense. uh Thank you for getting this list together. This is great.

A

Danian yeah, it looks like you're on this call gateway. Endpoint publishing. uh Do you remember what this one was about.

C

Yeah, if you scroll down to the bottom, I think maybe it needs to be closed, or maybe we decided to leave it open. But I believe, when we discussed this, that it was, we discussed it and decided that it was implementation. Specific.

B

B

Okay, so I think that's okay, so then the controller can figure out how they want to like uh populate the state. Okay,.

C

Right, this is uh yeah, so the gateway uh decides. Okay. Should I create a gateway instance and expose it via node, port uh via service load, balancer type, so on, and so.

B

Forth that makes sense.

B

A

We will close that for now then um see what's next.

A

The universe of resources that can be routes- uh this is a great one. My guy do you happen to remember the background here.

D

More or less um it's something that as an implementer, you need to know what resources do you need to watch uh what things can that label selector match.

A

And I think we have already solved this now right, because uh our route selector includes a kind or a resource. Now I forget which field we use, but you can now only select one kind of route per listener.

A

If I'm remembering that correctly.

A

That would solve this issue right.

D

A

A

Okay, next one yeah okay, uh this feels like a documentation, request more than anything right, because we clearly support uh both at this point, um but we have maybe not documented that. Well enough um is, is there any specific thing? We have need to add to documentation to answer this question better.

B

I think this was opened at a time when we didn't have anything else other than dc, like significant other than http route. So maybe, if we can bring jay back and ask like, do you think the.

D

B

E

I think we can just close it. uh He can just put a comment saying reopen because it's pretty clear when we go to calculate the docs that is in.

A

So that works for me.

A

Yeah, okay: this is something uh that I think we can is that does anyone know if there's a coup builder or crd validation that can enforce immutability?

A

Oh yes, I answered that right here and it is still not a thing. So we just we, we probably need another label or something like to indicate things that will be will need to be solved with a web hook. uh I know we have several things. um Surely there's nothing already yeah? Okay, uh maybe all.

A

Anyone mind if I just create a validating webhook label or maybe webhook.

B

Yeah, maybe just area webhook yeah.

A

Go ahead, yeah.

A

Cool, I will do that sure.

A

And this probably does not need to be a bug, but it should be frozen all right.

B

Well, there's kind.

A

Of a bug, okay,.

B

Yeah, but this this can be uh opposed to even alpha one or yeah.

A

B

Yep awesome yeah.

A

So it's not part of a milestone right now and I think it's fine to leave it that way. um If maybe we need a v1 alpha, 2 mile yeah, I don't yeah.

B

I already have created it.

A

A

Surely it's reasonable to have this by v1 outfit too.

B

Okay, this one is where we do. We want to drop like the selector, like from the.

D

Oh right, suffix.

B

Right so that seems to be like. We need to settle that on, given that we have we love selectors, and we have like enough individual number of selectors.

A

If we ever want to rename these apis, it could just be selector.

B

A

Because uh yeah, uh that's a great question. um Does anyone this feels like something where it'd be good to do a quick survey of what we've seen in upstream? uh You know. Obviously the thing I think of right away is service deployment, whatever that all very clearly have selector in their name, um and I think, even network, I think selector inside the name makes sense.

A

I don't know um what do you think harry or anyone.

B

A

Should decide if this.

B

Is v even alpha one or not, and then, like you said it seems to even alphabet right. It.

E

Seems yeah this this kind of stuff seems to be like not super critical, just pick one, because this is it right now. If we change it, then we change it.

A

I I think my preference is if and- and I know I have gone against this in the past, but I think for the sake of consistency having selector in the name probably makes sense when it's a selector field.

A

Does anyone feel strongly any other way.

E

Okay, I'll run with it.

B

um Let's give, can you assign that to me just get it done.

A

A

A

Yeah, okay, um I think I think this may be a dupe. Well, no, I don't have an issue for this, but this seems highly related to the conversation. The you know, kind of broader conversation of if we want to even have defaults at all, uh for these object references um this. This was something that, uh in in the initial api review or api feedback, I got from tim. He pointed out as being strange and maybe not as helpful as we thought it might be.

A

He didn't explicitly say that we needed to remove it, um but it it did seem. You know, like the absence of config maps, you know we've already removed config maps as things config maps default, but the absence of those resource names could just be confusing. Potentially um was that logic?

A

um I don't know I, it seems like there's not an easy way to fix this beyond that, though, um beyond just either or was there. I think bowie has suggested in the past that we should be using aliases instead of though, that may not that may not solve this.

E

Oh, is this the documentation thing yeah we have to go through and instead of using go, aliases just define them structurally to create a new type.

E

Because I think the use of aliases is wrong, but, as I recall, the use of aliases it's just for like temporary type migrations or something, and they don't behave properly. Instead, we should say: typhoo is a bar. Instead of typhoo is an alias of bar, which is the equal sign.

A

I think I thought we were doing. I think we are doing the equal sign right now.

E

No, we shouldn't use the equal sign.

A

Okay, so we okay, we should not, and we should use aliasing instead.

A

Or I know I I get it I get it, I get it yeah. I get it okay. So, instead of this you're saying that should become, I should become the same thing just without uh equals.

A

C

Her the equal sign.

E

Okay, I think if you do that everything works and you don't get any of the weirdness okay.

E

So if someone wants to take the stop, just go through the code base and just get rid of the equal sign, and I think that fixes all the dot generation issues. All the strange comment generation issues.

A

Okay, that would be amazing, if that's all, it is.

B

Should this go in scope or can be out.

E

I, like this seems pretty easy, like just get rid of equal signs. I think we have to do this anyways, because the talks right now get generated badly. Sure.

A

A

Optional listeners.

A

I think we've decided against this right.

B

Yeah we decided against it and made it more strict. So but there's.

D

To be some conversation.

B

Here that we need to refer back.

A

I think I so this I think this has generally been solved by the idea of tcp route and udp route. Like this. This conversation started before those routes were even a thing. uh I know, I believe that there was still some desire to avoid using a route entirely, um but I I don't know I. I don't think this is something that we need to solve and be one alpha one at least maybe the thing is we don't need to close this, but we also don't need to solve it in v1, alpha 1.

B

That sounds ok.

A

Ok, should I add it to v1 alpha 2 or just leave it undefined.

B

We can leave it undefined or we can ask people to take another look just so. This is not going stale forever. I don't know like, should we move this forward or just let it be.

E

Yeah rob can we create a milestone, called like backlog or something because I it would be good if, like no milestone means, we still have to treat it and then milestone. I.

B

Don't know how we would want to do it.

E

Yeah there we go.

A

Discuss domain match versus wild card match for listener host names.

B

Yes, this is, I think I and james were discussing this and it came out of that domain. So we have a hostname match type on the listener inside the gateway, and we just want to like what should be the name like right now, it's named domain. Is that good enough, or should it be wildcard?

B

That's I think, that's the, and so if it's wildcard it should be star.example.com right so which one is better. The.

D

First, one is what we have.

B

Right now, that's I think, the gist of this.

A

A

The requirement of a wild card feels, I don't know I I have not. I did not pay attention to this. Initially uh boy, it looks like you, you'd left some comments. Did you have any preference.

E

Here domain match versus wild card match for listener host names. Let.

B

Me take a look at this again.

E

A

E

Of load it back up yeah, can you assign yeah just assign it to me? Okay,.

A

A

I I do not think that I personally don't think this needs to be in v1 alpha. One.

A

They seem similar enough to me that we're not losing anything significant.

E

Okay, let's put in the backlog since we kind of tried.

E

A

huh Yeah, this is an old one february, 10.

A

Is anyone? Does anyone remember the discussion around this this? It feels like.

B

I couldn't get.

A

It wrong, but I think.

B

This is about like the sni in the tls handshake has food.com, but the http request that you get you get on those on that connection or session. I.

E

Think we saw the stingray because we pushed the host name up to the listener block and you can't get a mismatch.

B

A

I'll I'll just.

B

A

No ping james to.

B

See but then you'll again have the any match on the listener and you can still.

E

B

This right with any match.

E

It's not clear like is it that you absolutely because it's somewhat in like it can't be guaranteed by the config on some level.

B

Yeah I mean yeah and like in a single listener, even without any match. You can actually get into this right because you can still configure a listener which has the same port, two different host name, two different certificates, right, food.com.com and then the request path. So this is something you cannot enforce in the api. You have to document it and you have to rely.

C

D

B

To enforce it in their proxy path,.

E

Yeah because it's like what, if they do, yeah it's okay anyway, let's follow with james and then I think this is not possible unless we just state that you shouldn't do it.

B

Yeah, I think it's it's mostly about just documenting.

E

B

Yeah, don't do that.

B

A

Gateway and gateway class. Yes, this was a fun discussion.

A

A

A

Okay, I thought we had already resolved this. There was definitely a lot of confusion between what a gateway and gateway clasp were.

A

B

I forget which one of these I'm using.

E

Yeah yeah once we document it, I think we can just double check and then this one will be closed.

A

I feel like this is something that will naturally be solved when we update documentation and if it isn't, it should be. um So I'm I'm fine with adding this to the v1 alpha and milestone if that works for everyone.

A

Okay, one last one on this list. Thank you harry for uh getting this list together. This is helpful to highlight some of these um when a gateway class is deleted. The spec does not address what happens to any gateways that refer to that fact.

B

This is now a bigger problem because it.

A

B

Controllers resign gateway class that status provisioned gateways.

A

Yeah and well it yes.

A

I think the the ideal option is option two uh there. There is no perfect way to do that, but it the the idea of preventing the deletion of gateway class seems reasonable. Here uh to me.

B

Is that like putting a finalizer on the gate, I.

A

A

Which, obviously you can you can bypass that, um but that seems like a reasonable initial attempt. We still this. This is indicative of a much larger issue. That's tougher to solve of the you know when gateway classes, change or when parameters change- or you know any any of these things change.

A

How does that affect originally provisioned gateways that have used that class.

A

C

B

What kind of permissions do you need to put a finalizer? Do you need like update permission to the gateway class, or do you like a is status.

D

B

Sub resource like right.

D

B

Like gateway class dot status enough, because if you need right on the gateway class, I think.

A

That's not ideal yeah.

D

huh To me, this seems more like another area where we want a web hook.

A

A

You you are right, we could use a web hook to prevent to to enforce two.

D

Yeah, if you used to finalize.

A

D

E

Sorry isn't this basically what a finalizer does.

A

Right but can't you get a get around finalizers sure.

E

But you can't get around.

A

A

Right I mean you can just get rid of the web hook, but if you have permission to get rid of the web hook, you can do whatever.

D

Well, I'm not understanding how a finalizer solves the problem. It seems like that solves half the problem. It says you can't delete this until something finalizes it, but you still need a something that essentially does what we would do with wetbook. That checks whether there are any gateways using the gateway, glass and then removes the finalizer, if not.

E

Wouldn't the gateway class controller always be watching the gateway class and they could tell that the gateway count went to zero and then.

C

B

Would basically say finalized, but then we have another problem. Somebody creates a gateway class as an naughty provision.

E

Oh yeah yeah, it's gonna be hard to solve everything here.

E

uh We just have to see from experience whether or not this is going to be like a very common race condition, or something that, if you like try to trigger it, I guess you could get things confused.

E

Like if you're, if you're, trying to mess up the controller by creating tons and tons of game, I guess it's an attack that you can do. I'm not sure it's like a attack, that's worthwhile, because basically you're just messing yourself up.

B

Yeah, do we know what.

E

Storage classes: do I don't think they have a very good policy, so I know that some implementations don't even handle this, that they just disappeared out from under them. They just kind of like things just go wrong.

B

I think for v 1 alpha 1. We can leave this as undefined and come back if this actually is a problem like this seems like like somebody who really wants to break things right like I, like a user, could very well delete a gateway class by mistake as well, though, but it seems like a problem.

E

Yeah, so we should make sure that you can, you can safely upgrade and delete gateway classes. I think uh someone like maliciously trying to mess with the classes is less less interesting. I guess um you know the unix philosophy of like bathing stuff and then just crashes might be okay here for the extreme cases, if the cost of trying to like completely make it safe. It's not worth the complexity.

B

I agree with that.

E

Because I'm cautious of like adding my books just to guarantee like.

A

A

So yeah, but if we already have a web hook in place, I agree this is not sufficient reason to add a web, but if we already have a validating web book, maybe yeah- I don't know you're right. This is you kind of.

E

Yeah, let's see what the experience is, um the danger of like.

E

Yeah because I don't know like, are there other resources, it's pretty hostile to like prevent deletion in a very permanent way.

E

I feel like most of the kubernetes things has a you can force it if you want- and you know that you're doing something a little bit weird but imagine like there was a bug, your system kind of got into a confused state and now you're just trying to clean it up by just violating all sorts of constraints.

E

um If you absolutely lock people out, then they actually can't repair it in that case, as opposed to sort of 98 of the time you're within bounds, and then something blew up and then now you can just manually try to like force.

E

I think that's why the finalizers have a force is because if you had a bug, if the controller had a bug, you could still recover from that. Whereas if you just completely lock people out, then you know I don't know like they would have to like reboot their cluster like edit. It just seems a little overly harsh.

A

Yeah, I think that's reasonable. uh Does anyone want to volunteer here? I thought I thought there was a good idea here about how finalizers might work, and although we're not really solving this here, I think it would be useful somewhere in this issue, to at least document that suggestion and how it could potentially work.

A

Just so we don't lose this conversation.

B

Yeah just tag me on that rob since I came up with that word perfect.

A

A

All right, I think, that's the end of this list, there's a few well, there's lots more issues to go through. um Thank you. I this feels like something that should also be v1 alpha 1, uh unless uh I think it's reasonably easy to get and- and I would hate to release our v1 alpha 1 that had you know invalid examples or invalid crd. You know whatever.

A

Yeah, um oh yep, okay, yeah this this seems reasonable. John, are you also volunteering to work on this.

A

A

A

If you're talking, I can't hear you.

A

Okay, awesome cool yeah this this would be awesome. I does anyone mind if I add this to the v1 alpha. One milestone.

B

Yeah, I think it should be in there. Okay,.

A

Cool and I will assign john to this- thank you for creating the issue.

A

All right, the other, uh tiny one that I had an issue of I've noticed that all our license, headers, for whatever reason, don't actually have any contribution. I don't know if this matters, I am the furthest thing from a licensed expert.

A

I just noticed uh it is very you know as an example, here's cluster api, all kubernetes whatever, and you know you look at any one of our types files and we just kind of missed that one line.

A

I don't think that was intentional. I does anyone know any reason we might want to avoid.

A

This, it seems like a really easy thing to fix. I just before we went alpha one, it seemed like we might as well just match up with what the rest of the kubernetes repositories are doing.

B

Yeah, I think, let's do that to usually oso software has that owner yeah.

A

B

A

Cool and then the this one, I think, would be good to discuss here, um harry yeah. You want to give some background.

B

Yeah, so uh we have this route.

D

B

What does yeah.

D

B

Selects the route.

B

uh What does this? Do? I don't know where it is used? No, we moved it from gateway class to the gateway level yeah, and uh this this selects. You know which routes to select. So here uh the way the api is laid out. We have a label selector and we have only same same namespace. So if only same namespace is set to true the label, selector does not kick in now. By default, our behavior is to set the only same name space to true, so that seems a little off.

B

Just with the principle of you know, d like true by default is not a good api design, but uh I couldn't come up with a better solution, while adding.

A

Yeah- and so I agree with the the default here being awkward- I just I- I can't come up with a a good field- name that is the inverse of this.

B

Yeah I mean if you even if you come up with the inverse field name, you still don't want that default right like how will you reconcile that with a with the namespace selector field,.

A

Right right well so.

A

What what you would need is you would need this field would be a field that would enable the use of namespace selector and by default it would be false, and so you can only use namespace selector if field foo is true, um but by default, that field is false. I see that.

B

A

Just I'm struggling with the name that that works here, um yeah or that's similarly,.

B

Clear the other option we have is uh where this is referenced. You know we have like route selector, so we just instead of round name spaces. We have like a route selector and the in the field. If you specify a route selector that kicks in. But if you don't specify it, you select the routes from your namespaces namespace.

A

Got it so I remember we we had done something similar to this before and what we had was a difference between nil and empty and nil meant same namespace and empty meant all namespaces, because yay selectors, um I think the feedback there was that that was confusing, that nil and empty being so close together. But having such different meanings was not great, and so we we kind of split it up, even though it doesn't necessarily need to be.

B

Yeah I mean in our case we'll we'll keep nil and empty selector as the same thing right like it means you're selecting in this in the name space you are in. uh If you do have a namespace selector, then that kicks in.

A

Oh so so well, but the standard is for an empty selector that selects all spaces. That is the kubernetes standard interpretation of that, unless we.

B

Okay, I see what you're saying I don't think we have a way out here so.

A

Yeah I mean yeah I that default like I get why that default is what it is, but it does complicate these kind of differentiations.

B

um Yeah, let's probably just I'll I'll respond here, like the options we have and like I'll, just close, it.

A

Yeah, I don't mind you know. At the very least, I I don't mind a default like an inverse field here, um if there's a a good name, we can come up with that. Has you know kind of an inverse meaning, so we can have a false default instead of a true one. I think that could make sense.

A

I just haven't been able to come up with a name, and I think it's difficult to combine kind of a three-state selector of nil, empty and specified, because no one empty are going to be so close conceptually, but so different in how they're interpreted yeah.

B

The other thing we could do is cross namespace and set that to focus.

A

Yeah yeah, I I think there's something to that of.

A

Something like that, uh multi-name space.

A

Cross namespace yeah cross name space. Maybe that's it. I don't know um yeah.

B

Okay, um I think we have a couple of ideas. Let me see if I can follow up on this one.

A

A

All right, well, I think I think we're at time. Thank you, everyone for the great feedback here. uh I should note I kind of have this massive pr open. I will take some time tomorrow to try and clean it up. I have finally rebased it that was fun, but now there's already an awesome amount of feedback on it. So I'll take some time tomorrow to try and get it up to date and paying anyone who's already looked at it to take another look, um but I think that's all uh we're getting awfully close right now.

A

So, thanks to everyone for the help here, um yeah have a good rest of your week.