Cloud Native Computing Foundation Envoy Community, 10 Mar 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: CNCF Envoy Community Meeting 2020-03-10

Description

CNCF Envoy Community Meeting-2020-03-10

A

Not sure which other maintainer czar joining I see we have snow and folks in Cambridge are there's Matt great, hey, hey.

A

These are items to the agenda that you'd like to discuss: I have the color API fronts, but sure there's more.

A

Okay, well, let's get going then oops can add stuff as we go. So one thing I'd like to just bring up is something an issue that Google's faced recently. Yes, we actually added in a PvP annotation, which actually you realize later may be interpreted as a breaking change. So one of the things we've become pretty strict about in the API is is not breaking api's at the proto level or semantic layer. That is, you know like removing features which have no way of being really expressed and I. Think that's great.

A

One area we haven't been particularly strict up on historically has been adding in additional PGV annotations, which you know, because there are annotations, they were never really considered to be part of the core proto ABI, but in reality they actually are because you know we start rejecting configurations when we fail these PGP checks. So I guess I'll just like to ask if we're just generally comfortable as a p.m.

A

particularly the API Shepherds, who are here, which I think is just mad at myself, being much more careful about pushing back on pgb changes and making sure of these. They, the only safe thing to do, is to make a PG V change, weaker, I, believe within a major version and having in a stricter PG B check is actually a breaking change unless there was already an implied contract and I and I.

A

Think that's where things become a bit murk here and you can see in this particular case that there really was an implied contract that to you, for example, you no characters in the middle of an had a name that would be bad homeboy rejected that anyway, it would crash. But if you're talking about like fool, RFC compliance for how to name his values and seeking and being very strict about the character stats there that didn't I, don't think that was actually necessarily an implied contract.

A

But that's debatable I said they say any sort of RFC compliance proxy should enforce these things. I.

B

Think that makes sense one one thing that I wonder if we should track from a feature perspective is, if we could add some type of validation hook to PGV.

B

It would allow us, for example, to do something like change like change, one of the validations, but put it behind a runtime hook and then actually go through a deprecation period and I I mean it would be some work like I'm, not saying that it's easy, but it seems doable, and that would allow us to potentially do things like this header change, which are probably a good thing to do. But, but we probably can't do by that definition until we cut v4 yeah.

A

I mean the question is for other uses of the API. Is that an acceptable thing to do as well? Like you know, we move to the major versioning and the clocks suckhole around the API is to sort of you know, provide folks with these no stability guarantees and now we're saying well. The API is allow this yesterday, but in three months time and they stopped allowing this that might make other API uses unhappy. I think.

B

That's probably true, I I think there's a bigger question here, though, which is that I mean we don't guarantee that PGV is supported in every language in which is API might actually be used. So III guess it's like. Do we even have an official policy on do it like? Do we assume that the PGV validations are a critical portion of the API like I? Think the answer is yes, but it's something that we should probably document so.

C

This PGI donations not actually a exercise in every single my bill. They.

B

Are in envoi, but we don't know, for example, that the people that are building control planes like we don't. We don't know that they're using those validations potentially in their tests, or something like that right, because they could be building a control plane in any language. That protobuf supports right.

A

And I'm not even sure what the status of G RPC is I mean G RPC is the other big consumer of the api's. Today, yeah.

B

And and since their core is written in C, I highly doubt that they're doing any validation but yeah not sure well,.

A

They have that three languages in many of the api's go Java and ready I I.

B

Guess my personal feeling is that I I think the answer is yes like we. We should probably assume that pgv is a core portion of the API and we have to treat it with the same deprecation stuff that we do elsewhere.

B

It does seem, though, that there should be some judgment here which is like in in the case of this header. One I mean this seems like the kind of change that we probably should be able to make and I'm just suggesting that there may be a technical gray area where at least an envoy if we could also annotate a change with like a runtime feature flag that would allow us to go through the normal deprecation period. Well,.

A

I feel, like you think that the annotations are a convenience for developers. You can always implement some, that's in a PGP annotation in code in envoy rights, and we could do that are the rejection there. So we could always have envoy just using normal mechanisms adding to the bootstrap or something like that. Hey, hey, I'm gonna, be really strict about enforcing header, RFC values and just do that there versus doing it in the PGP annotation right.

B

C

Many more time flex would definitely help a lot wandering from right now. Is that, like we have no way to turn this off yep yeah.

B

And and again like just like quickly brainstorming, the technical solution, I could imagine that you could add a flag to the PGP annotation, which is something like a feature flag name, and then we would just have a hook where we could plum that into envoy runtime system. Americans.

C

Make it more difficult for me, especially since the dislikes nice, to be specific to the core implementation, not like other users, currently rely on yeah right.

B

Yeah I mean I, don't I, don't like from a library perspective, I, don't know how PGV could offer something that was general and let's PGV itself had some type of API, where you have to register your feature, flag system or something like that.

D

Yeah doing this would be to deprecate the field and replace it. Yes,.

B

Yeah that that would be the cleanest way of doing this. It would be the most pain to the most users. You know and again, that's why it's like I feel like there's always some judgment here, which is we're. You know we're trying to balance pain to users versus doing the right thing, and it's like this feels like a case in which we could probably just do it, but you know it might break someone, so it feels like we need to at least have some type of feature, flight control, so.

C

There we saw into this lecture led to like the elevation on the path, free right and hold free right is things we charged a lot more likely. Torchic of configuration will be essentially incorrect and why should be rejected and someone sighs depends on what the violations of being and someone's like she will prefer their lies to be fairly strict and they to actually avoid a users. I concur themselves by actually making rewrite that wisely. Black holder requests.

A

I think the more complex is validation becomes the more likely it should just exist in code and envoy, and it shouldn't just be like it's pretty. We don't know. Pgb isn't particularly expressive today, for example like if we had like full seller. You know expressive is you're, not kind of thing that might.

C

I should be as simple as a muster with Slash and that were actually like a brake assist single things. The Mathmos that was that's actually like the extended validation actually be like that, must go be breaking things you.

D

Could also imagine replicating like if, if I, if we had a solution where we deprecated string header name and change it over to a wrapped message, that was like ballad string, header name, you could imagine saying something like when we validate our protobufs, that we we like re-implement the pgb check-in code, because we know what type it is.

C

Nothing that love api turn. It.

D

B

A

B

Just gonna say the one benefit from a from a PG v type. Annotation is that at least theoretically people with control planes or systems that validate config. They could run these things in advance, and that is kind of a nice thing for people that are building.

C

B

To not have to send it all the way to envoy just to figure out that it gets rejected. No.

A

No I agree, I mean you basically have a contract. They're explicit I mean the the concern over API changes. You turn other things as bigger one, because guess what we have a lot of API to in anyway built into the newer major version thing every year, you're throwing away or your old API code and running new stuff.

D

A

That can be, there can be brother for anyone yeah, but that's we're gonna discuss its offline Antonio, but that was basically the community agreed approach to providing stable API is for envoy and other clients such as GRP, see. Now that was a Google request and one we're actually resource, but it's it. Yes, it's nice not pleasant for anyone, but turns out predator. We don't know that actually ants better story is in the predator, what other than never deprecated a feature or basically break. Why compatibility they so on?

A

You know a calendar cycle which again, is it's not suitable for long live projects.

D

Just to get like a closure on this is: is it okay, if I like the next steps that I make would be to like file public issue and make a PR that restricts those validations down to the ones that we already have implemented in code and do that churn of marking the field is deprecated, so the.

C

Code replaces characters as well as through filling out the asserts like no I'd, be okay. So a your feels like change. Someone would be there only.

D

For tests, oh never.

C

D

Yeah those those changes were for tests program.

C

D

Nothing nothing in that PR changed frock-coat. It was just API change at Tesco. Do.

C

The pro code, just like optimal just like do run them things. Yeah.

D

Anyway, does that make sense.

C

B

That's fine, yeah are.

A

You gonna add another like well-known name to PG ve, just gonna go straight seems to make it explicit in there on my code, I like.

D

That's what I'm trying to figure out like it's like adding another field to PG b, unless, like I can get you to monitor that PR like it always takes me like six times as long as I need to just to get people to review.

A

D

But I'm willing to do like a Envoy, valid slash, are slash, slash and like no character check just that we don't have to rewrite it. I mean.

C

Some of the different places yeah.

D

Sorry, if I can just like yeah.

B

If, if you would prefer to do it within PGV, but you feel like your block there because of code reviews, let me fix that so I mean let's: let's talk about.

C

B

Offline like we can, we can get you code, reviews faster or we'll just honestly make you a maintainer of PGV. So, let's, let's just do the right thing like: let's not not do it in PG, because.

D

B

Frustrated about code review, speed.

D

I'll turn that up in quickly, if I can get reviews on that yeah.

B

So if you would prefer to do it in PGV, let's just do it there and then, if you could ping me, if you feel like you're blocked, we will figure out the right way of getting you unblocked.

A

Anyone tried the any of the v3 API stuff and have anything to report I.

B

I kind of doubt that anyone is using it yet and, to be honest, it's like a chicken and egg problem, because I think until we block people from adding things to v2, like people probably aren't gonna move to b3, so I I think we probably just have to move forward with the blocking. So.

A

Is everyone? Okay, with the current plan of record that at the end of q1 cooling, allowing we will stop allowing additions of any new fields to v2 api's and the new features need to be go straight into v3, o v4 alpha api's I.

B

Think this first transition is probably gonna, be a little painful, but I, don't I, don't see any way around it because I don't think we're gonna flush out the issues until we get people to move. Okay.

A

Well, I think that's the plan then, and uh yeah I'm.

D

Hurry I I just have like one piece of feedback on the API, tooling and I. Don't know if this is just my own problem. Where's everyone else's, but typically when I'm, making dev changes now in envoy, I'd make changes to the v3 API, but the proto format script will only push v2.

A

E

D

Not v3 to v2 yeah.

A

I mean basically, what we will do is again for until end of quarter. Probably you wouldn't make those changes to V 2 and then at the end of quarter. You would. We will flip this around so that v2 is effectively frozen and no one can modify it. Then this really gets to be the canonical source of truth and that will propagate forward to v4 yeah yeah.

B

Like from from an optimal tooling perspective, I think where we need to go from from a North Star perspective is you should only have to edit one proto like everything else should be automatically generated, so it's like all of the you know all of the shadow proto's, like all of the next version proto's like to me. That's a that's a build implementation detail. So if we.

C

Did initially get.

B

To a point where you know you run, basil builds and it just like. Does the right thing: I think that would be a much better developer experience like obviously we're not there now, but it but I think that should be our goal. We.

A

Need to eliminate the explicit shadows that stepheson mention should be done as part of our technical debt burned down, but yeah I mean right now, and you are only supposed to modify a single file. It's just confusing what that file is well.

B

There's that, but there's also I think it's more confusing because people don't generally like it, it used to be where you edited a proto. You do basil builds and you would automatically get all the changes now you have to edit the proto. You have to run the proto formatter.

C

B

You have to do basil build right, so it's just like what I'm suggesting the dev flow should be from an end goal perspective. Is that you, edit the proto you type basil build, and it just does the right thing.

A

That can actually work because you need you, some non hermetic things there, but we can discuss like I, think we're pretty lazy and maybe some other bad life experts in the room, because I know in a bass or Robin you can. You can do some pre non and medic things, but that's different than basil, build I. Think.

B

It's fine again if it doesn't work that way today, but I would hope that we would agree that, like that would be our optimal state. Right like it would be nice if the whole thing was hermetic and like I just make a proto change, and it just does.

A

How do you rewrite the source tree? I mean you need whether there's going to be multiple versions living in perpetuity of each of these api's in the source tree. You know, like always at least one step in which a file gets copied back out of the Basel cache, or something like that to make this work. I would.

B

I, honestly, don't have the answers, but I just feel like that would be where we'd want to get or at the very least, when you type Basel build. There would be type of error. That says like no really like there's a change, but you have to type this other command or something it's just like.

C

B

The the flow today is quite confusing.

F

Recommended way of doing like checking and generated files, you are the check to the bill to verify that it is what it's like. What are they connected to? The other thing that I found is that doing git push now takes quite a while because of the get hooks. So even if I'm not changing any api's I don't have to go, get a coffee while I wait for you.

B

Should do know verify I do.

F

Slammed by the formatter, because I didn't yeah.

B

So what I do like this is horrible, but, like I, do the hook until it starts running basil, then I control see and I dunno verify yeah.

A

Like I feel sufficiently hurt when this should just basically figure out how I mean it's not just with the proto stuff, though we discussed the other day, I. Think as we brought up, it was the Python stuff as well.

A

We running checks which don't need to be run basically during things format or check formats, and you should just basically be able to compute everything, that's going to be pushed, and only if there's an file, that's modified with the proto suffix or a Python suffix that you run the respective checkers on and that's a project for someone who really wants to make this fast. You know you know if it's enough of a pain point for someone, please go ahead and do that. I can create an issue to track yeah.

F

I also think like we're, like format, checking like the generated files during the hook, which, presumably yes, presumably as long as we're as if it's getting checking CI. It's probably not useful for me to check every time I like and they get hooks, but yeah. It's the this that wasn't work that we get on victory like streamline this. What.

B

I would recommend doing is just opening tech debt issues for some of these things, there are people out there that that, like fixing these types of things- and it's like having these issues opened, it'll make it more likely that we can, you know, get people out within the community who might want to work on them. Yeah.

A

Like it's a very quick work around us all just create fast, fixed format which doesn't do this stuff, like it's I, yeah I can make that easier in the preset mid hook, I actually.

B

Almost want to do that just to avoid this bug that is hitting all of us for the rewriting the build files.

D

B

A

It's very weird: what's going on there there's either Sun path issue where it's actually not seeing the file, while some reg X issue, which I'm yet to figure out yeah. Okay,.

A

Do we have anything else on the agenda.

A

Going once twice: okay, okay, three minutes, bye.

C

F