Open Telemetry Uncategorized, 7 Apr 2023

Previous Meeting

Next Meeting

⏯

youtube image

►

From YouTube: 2023-04-07 meeting

Description

HTTP Sem Conv Stability WG

A

All right folks.

B

Good morning,.

C

Hello good morning.

D

A

Prepared for the blog post, I haven't click the button, because I I want the ECS folks to confirm.

B

Little typo here.

A

Yeah I'll fix that. Thank you.

B

Yeah I read over it last night, um uh their their blog. Also, um it's good it's good. It's a good uh convergence of efforts.

B

Yeah, uh let's uh I've, put some to kind of quickly triage um one is on this issue. um uh Christian had asked about keeping this one open as a tracking issue.

B

um I wasn't really sure like I mean we're tracking it in the projects, um I don't mind having a tracking issue, um but maybe there's maybe a tracking issue for all ECS related stud alignment.

B

C

A

Supportive, my only question is I I feel we probably should've got more ecis folks, yeah like in whatever media like we should have a regular meeting with them. Instead of us thinking about something. This is a little bit differs from their side.

B

And what do you do? You have any thoughts on a tracking issue versus a project say a dedicated project for ECS alignment, I.

A

I feel a dedicated project would be advisor. The minimum thing is, we have a regular meeting with ECS folks and we can use that to identify home from ecis would be a good fit for this collaboration. So far, Alex seems to be the only one active I know recently, there's a another one joined and she's working on the elastic client semantic combination, so it's maybe like they're trying to ramp up and which was really connect with them. Frequently.

B

Okay, yeah I will I'll set up a project um and yeah. Then maybe we can use that to drive. Do you think it's worth a working group or merging it into the existing General semcon yeah? So so I.

A

I feel it's probably better to merge with the translate like I I feel like charge. Storage, probably is the the proper guy, in my mind, was who should be driving this.

A

But the problem is here: we have a group here and then we have a different goal and then we have some work with UCS folks and it seems we have different channels.

A

So my suggestion is that we probably should consolidate the channels.

B

So we already have project for this for the general. Does it make sense to have a new project for ECS I.

A

Would suggest not or I have too many I want us to have less.

B

Okay, so uh uh just a label in here probably yeah, okay.

A

Yeah and I I know Josh has some thinking. The problem is I like it's hard for for us to be in the same meeting like Josh is not here. So it's hard for me to speak for him. Yeah.

C

uh Yeah, what I was thinking that it comes up in a specific work groups like fast I'm, sure messaging. So if we want to track those work, maybe we should have some prerequisite that before the semantic conventions become stable, they should be aligned, but essentially the I think the in terms of work. This particular.

E

Seeds should uh on board and to ACS at their own pace before they are stable.

F

D

B

Okay, so we can kind of try to manage it at at a high level from the semcon working group, um but yeah. It makes sense for the specific sigs to drive that.

B

um For this one Riley um uh was there any discussion in the TC of when or how to make this change?

B

Should we I know there was some consideration of.

A

My understand is, metrics is pretty new, so the concern is less and Jack has a PR uh which added the hint or the device. So we believe that that should make it smooth.

B

Got it so no no need for transition period once we have the hints implemented so.

A

So far we haven't heard any strong escalation. This sounds like it's an okay thing.

B

um So this was this topic came up in the spec meeting this week.

B

um Some a couple folks when reading the Otep, the ECS alignment Otep assumed that in cases where it was sort of a tie of like picking one or the other, for you know, but there was no Advantage, um no improvement uh that we wouldn't see any improvement from the ECS modeling that we would pick the hotel side.

B

um So I think that is an open, um I thought we were going to align in that case with ECS, since they are stable already and we are not.

B

Daniel asked Daniel had requested in the spec meeting for the TC to make you know to say what we were doing here.

D

A

So so my understanding is, the teaching me happened on Wednesday and I mean is going to respond back and the general idea is there's no uh there's no simple thing like we'll. Just start from UCS or Roja starts from open time to cement information. uh More likely, it'll be a scenario based, for example. If it comes to container kubernetes, you see I started too early and later a lot of things change, so it might make sense to better, along with the resource model which Josh is leading for HTTP.

A

Since the ECS one is more mature and we see a lot of good reason to switch to that for next one. As long as we can have a list of the pros and cons, then we should make the decision here instead of based on Uber like just ecis or start from hotel.

A

The also will be like that, and Armin probably will will respond back from like referencing, that he says voice. Okay,.

B

A

So for for this, one I think for this group. What we can do is we elaborate? What's the motivation like I remember that mentioned: there's some some nice thing they want. We want to change. So if, if there is strong motivation and the actual cost of breaking something is less than the motivation, then it seems like it did fit.

B

Yeah so, and although this is these, are the only ones at this point that I think we're not um we don't have any reason to believe uh one is better than the other I mean they're literally just you know it's pretty much. Just ours is abbreviated versus the other side. Yeah, um so I think we're not really sure how to recommend one or the other.

A

Me it's also fine like if we have a strong preference, we make it clear, so we educate the others for just come here and try to help. If we're saying it's a tie, then we talk to Easy as folks and see. What's the potential turn for open times, what's the potential term for them, and then we should act as reasonable people to make a make a call where to go and the worst case is it's going to either damage this side or left side and the Damage seems to be equal.

A

Then we'll have a voting or something right.

B

Okay, that sounds good yeah.

A

So so, essentially the position from TCS. We don't want to make a uber call back, saying: hey if we don't know we'll just starts from ECS or Roger starts from hotel, because that doesn't seem to be rational. We want to be rational with the understanding that, if you change anything whether it's ECS or open times, someone will have to suffer and we need to understand what's the actual cost benefits that makes conscious decision instead of some Uber like Direction, Just, ecis or hotels management, information.

B

B

Do you have any thoughts on next steps, or this.

C

uh I, do not think that what I suggest is super helpful anyway and I I. Don't think this is a critical problem, that's blocking the stability of HTTP semantic conventions, so we can keep it open. We can postpone it.

C

um I mean the the this particular issue does not resolve the problem that they have.

E

B

Okay, so I'm just going to yeah I, will remove that as a blocker. Then.

B

B

This one I assume we can as we're as it's sounds like we're going to at least remove the schema translations as something in the stable spec or there will be a uh adult think this should be a semcon blocker anymore. I mean uh stability, blocker.

B

So if no objections I will remove that.

B

um So I on this one, so I'd sent the pr to essentially remove plymetry stability to uh schema transformations.

B

um It seems like there's still um some strong feelings about um schema translations um being something that we want to do in the future, um and the concern is then: if we want to do it in the future, um adding it in the future could be considered a breaking change versus.

B

Adding it now to stability, but not using it, not allowing it to be used until some future time undetermined future time is maybe less technically breaking, although I mean it's sort of potato potato I think um any thoughts here.

C

I I, like the idea of moratorium, um I'm, not sure how to express it, but I think we're caught. In the same conversation, the stability for what for for instrumentation or for consumers, and if we.

C

Like if we can say that we require reinforce it for instrumentations right and we promised some period for consumers, but we don't.

C

We are not. We don't guarantee that that attribute names remain the same forever, that in future it might become a valid change. If we have tooling think some clothes like this may make sense to me. What do you think trust in everyone here.

B

Far as wording I can definitely try to take a stab at that in the in a PR and see what it looks like.

C

B

Yeah: uh let's try that.

B

I feel like that's the best way, to see what it uh if it makes sense when we make sense, and yet.

C

Thank you for being so diplomatic and facilitating this.

A

I have a question regarding the like stabilities, because multiple things next there's a the spec itself, there's also how instrumentation libraries or how the maintainers should do, and also how how the user should expect this I feel like. Even if we do a great job on the specification part, the instrumentation Library owners could still try to follow the spec in a way that they break the user like. If the spec is saying, there's a new major version. We have the freedom to make some breaking changes. The instrumentation Library might decide.

A

Okay, I'll have flag or I'll have some like. Maybe parallel packages supporting both for the transition or they'll just say how to shift and delete the previous package and my God finished we're trying to use the spike to handle those things like how maintenance should do everything or how the users really expect that it's probably it's probably too big scope.

A

It's like the English dictionary, we'll say like this. Word is no longer used, it's considered as attention to English, but we don't force people to to either use it or or stop using it. It's individual's decision.

A

The spike is just falling on the specs. It's regular law. We Define the law, whether we want to follow the law or not or not, whether it's illegal, it's unlawful. It's your problem! The law is not going to say anything about. The law is going to define it.

B

So are you is what you're saying that the stability is an instrument more of an instrumentation concern, investment.

A

So I'll give one example uh like we will talk about hey like uh the users depending asks why they cannot change that. We have to be super careful, so my argument would be each individual 6. They know how important that is like if I'm the rust, safe and I just shift a an alpha version based on the http.trl and now the spec is asking me to change to Euro. Adult School I would be super happy, just change it, because I just released the Alpha version eight hours ago like why.

A

How I need to do all the side by side and switch? But if you're, the like trust, here's a Java guy. You know that your package is used by one million customers. Then you should be careful and that's your decision.

B

So this is or specifically around like transition periods transitioning or is it more General? Are you making a kind of a more General argument about things that that we should limit stability requirements in the spec.

A

Well, I I feel a lot of how folks they're imagining me, despite would be very prescriptive about how maintenance can just act as a robot, and my answer is: no maintenance should have their judgment and the specs shouldn't be too frustrated about that.

A

Like we're saying this is the this is the stake in the ground and we understand there's Legacy things and the transition will require some effort, but here's our general thinking and each maintenance. You should use your brand and decide what was the best thing.

B

Okay, so the idea is, we would still provide some guidance, but we would also say you know: maintainers yeah,.

A

Okay, one example like like like there are folks asking hey: the TC should make a decision whether we should just start from ecis and start from old health. So during the TC discussion, I think most folks have the idea, like we don't try to answer that, because if you're saying he said like it seems like if your father and mother both got in trouble- and you can only say one or you're- going to save your father or mother- that's a stupid question.

A

We're not going to sponsor that directly because anyways that's going to be a stupid answer. So we're going to give a smart answer by explaining the rationality like we want to be reasonable and we need to understand the situation and once we understand that we make functions decision. This is why we never said we'll start from ecis over stars from Old House, Memphis convention and I feel here we're we're seeing the similar trap.

A

If we just try to address the direct question, most likely we're going to make people unhappy, anyways different, manage the petition expectation by telling them hey. That question doesn't seem to be very reasonable and we we should take a step back and I want you to be accountable. Yes, the maintainer should make decisions.

B

E

B

About this, just briefly um because part of the I understood, tigrin's, concern about transition period being or the transition period is more about applying to Consumers vendors, specifically yeah and so I think he's wanting. He doesn't want their vendor. He doesn't want people to pick up the rest even like the rest, Alpha, instrumentation, I think and start sending the new Telemetry.

A

But my argument is how he stopped that if the rust it means it starts some like device instrumentation if they start an awful version of device and you haven't prepared for that, how can you stop them? What's the difference here.

B

I think the difference is how critical HTTP is for basic user experience in pack-in vendor products.

A

I understand my argument: has the the right system is just very new? Nobody has the expectation I think it just work. It's Alpha version, they'll, take feedback and based on the feedback, maintainer will decide. What's what makes sense for them. Maybe they'll just be said: okay, it's a mess. I'll just wait for the specs to be stable uh before that I'm not going to shape anything or maybe they'll just realize. Oh, there are a lot of users actually doing doing this like something they have to exist and stuff that doesn't understand the new schema.

A

So we have to do side by side. It's their decision.

B

Yeah I will um I will put together sort of an updated proposal here and try to factor all of these. All of that in and I'll probably use the spec meeting on Tuesday to try to get more to go through that and get more eyeballs, but I I understand what your point is and I think that it I would like. I would like to give maintainers that judge that discretion also um I just want to confirm with folks other folks.

A

Yeah I understand yeah, so my general comment on this is, it seems. A lot of people have very high hope that in English dictionary will just tell people how to write a document. Maybe we should set a different expectation.

A

I understand like like this is just my personal opinion.

B

um All right anything before we go on to um discussion for um uh Mateo, primarily for materials, lidmilla and I and I'm, hoping both of you can stay a little longer.

C

I can I just uh a quick comment: I updated the ECS, a PR to include HTTP only so just calling out if people want to review it, nothing to discuss nothing new.

D

B

B

All right: let's talk about client spans.

B

B

B

G

Actually made sure that it is looking as it should be and that it's working and so on. So this is like, hopefully, the final version or the final before comments version and uh let's take a real implementation of the low level way of instrumenting an HTTP client.

G

So it actually implements like the full HTTP recent spec that we have it creates span for every, like actual HTTP request sent over a wire and what we discussed later six weeks ago, maybe for a connection errors like everything that happens before the first attempt, uh and in case it like fails the operation that there is sort of like a fake or mortgage, it is, is pen created like around the whole operation. So there is some Telemetry about the like.

G

A let's say, virtual, like logical, logical, HTTP, operation and uh error that gets thrown by the line is captured. Yeah.

B

G

Yeah so I think this more or less matches the spec ER that uh yeah that one that I also created a long time ago.

C

Oh, does it work? How do you inject? Is it like this fake span? Is it conditional.

G

Yeah, so uh trust if you can open, yeah and open uh connection a response of Interceptor yeah, so uh I just first of all capture the start. Time of the whole operation in case I like need to create this one later, I have to know where the whole thing started and in line 45. If there were any reasons, if nothing was sent from any sort of regent, I agree this band basically and it's uh yeah.

C

And if there were retries and it will become a sibling of the individual, tries.

G

If there were any retries the condition and the you know, the number of reasons is zero will be false, so this one won't get created.

C

Oh, so it's either or right.

G

It only is created in this situation when there is like an error connecting to the like the remote endpoint I, see.

E

C

And then we had some conversations on the pr as far as I understand this this retry, so it cannot capture at least 4K HTTP or exactly the connection level things it's not HTTP but like it does not retry. When response is received right.

G

So like when the response I'm understood for, can you like explain it a little bit more.

C

Yeah so I think for okay HTTP. If I remember correctly, they have the network Interceptor right and uh the Interceptor that higher level right so and the retries they apply uh are for exactly for connection problems and uh things like the DNS resolution.

C

Basically, when there is no response that was even received and that Network Interceptor reacts on to the different HTTP codes and then it's basically in the user and application responsibility to write retry logic when response is received uh and I wonder like would we what would happen if we get that on our 500 response from this server? What will we create here.

G

In general, it depends on the HTTP client.

D

G

Both okay, HTTP and reactor energy, like I, have a separate PR for reactivity, which is uh a little bit more messy than this one, and it's still not not ready, but uh in these two and I'm pretty sure that at least a few other uh clients like like which code I read. uh They actually retry things like resend things like um not retrace, but.

E

G

E

G

Directs or authentication attempts internally so in this PR, uh if you like scroll down to the abstract like HTTP test I've added several tests that actually exercised this and that redirect repeatedly and there's a a little bit Busters that kind of tries to do the authentication part I'm glad that you've been over there but yeah. That's uh that's another story, but yeah in case of Oak hdp like uh I'm, pretty sure that, like I'm served absolutely certain retries get captured and I am pretty sure that, uh like authorization, resends also will get captured.

B

And redirect redirects and.

G

Yeah redirects, sorry, my mind is just you know: it's evil, uh yeah redirects get captured and I'm pretty sure that the also application, any sort of like authentication um redirects recent will also get captured. I haven't tested it yet I don't want this VR to get too big, but yeah.

B

Yeah yeah so Lynn Miller's lid Miller. Your question was specifically around retry, like retry on failure, kind of a scenario yeah.

C

My question was mostly to understand how behaves in different cases yeah, but uh and if we have something similar between retry inside HTTP, client and retry, with some additional user Logic for 500 or something.

G

Yeah so like, if you like, you have a spring application. You set up like a rest, template I, believe and there's like a retry utility, then Spring Class that allows you to retry on, like whatever you want right.

C

Not necessarily so spring is too complicated right, so what I mean is uh yeah when I, if I use, okay, HTTP or native uh I might add the Handler of some sort? That will do the retries and uh yes, it really depends on HTTP client, but when we Trace in the video twice for a k, HTTP I assume it's a network Interceptor and it's the last one. So it's if there is another Interceptor that does the pre-trial Logic for for 500, then we will Trace multiple 500s.

C

G

Yeah I think I understand now so uh I think we can.

G

You can write instrumentation in the way that secures for this, so we can make sure that we at least provokage to be I'm like the same like for every HTTP client, but for okcp, I'm I think we are. We should be able to implement instrumentation in a way that always injects like the hour uh Interceptor like as the very last or like almost last Interceptor executed, so anything that the user creates himself and puts into the Interceptor chain it will be executed before.

G

So, if any sort of like user driven retries happen, they still will be counted properly.

C

Nice on currently is it an Interceptor or it's a network Interceptor, the one in occasionally that's.

G

What it means it made like a metal shortcut here.

C

Awesome yeah so then we'll have the same, probably behavior, for oh and does logical span. We capture it on the call level on The, Logical level, right, awesome, yeah, so then yeah. Then it aligns with what we have in this fact that we have this logical thing. Plus we have the individual choice.

C

I, don't think we need to resolve it now, I mean we, we don't it's not blocking HTTP semantic conventions or anything. It's just what we should write in this pack and whether it's a feasible thing to implement.

G

Yeah, so it's more of like a general directions for implementers of instrumentation.

C

Yep nice, thank you for trying it out.

F

You're welcome uh can I ask you a question just to clarify my understanding because I'm not that familiar with Java implementation. So we have to uh create this fake span just because uh there is no kind of General stack that intercepts the overall kind of HTTP request flow. So we have to do this kind of in individual components, because basically there is no kind of start uh saying that we can. We can use the greatest the greatest band up front and then, basically uh just.

B

Use http I would say it's less about um is that we want to conditionally capture that outer span.

B

uh We could start it right away, um but we want we in cases where there's we can capture that Network level span. We actually don't want to capture that outer The Logical. One.

F

Got it yeah, basically like I'm I'm, mostly thinking about this from.net perspective, because there we have like this, let's say nested logic, so we just started. uh We start a San and then, if something happens uh we basically just like finish it with uh everything that we can capture there like URL or if there is no response. Basically, it will be something where it might. We might might put zero there or just miss it, but this time still will be there and for retries.

F

If retry happened, then we capture it and create another one and then set this additional attribute like resend account. So what I'm trying to say is basically this Outburst man will be anyway created, so we're going to do this conditional in.net. But since here we have some different Gates Right.

B

But I don't think we want that duplication. We don't want both the outer span and the reach and the network level span, because.

G

It then, even in a simple scenario of like sending out a single, successful, HTTP request, you would get two client spans right, the other one and like the network layer, one.

F

Yeah, that's right! So, basically, it's not about encompassing spam. What I'm trying to say that recent counts will be set only for retracts for the first try. It will be. No recent count.

F

According to the current, according to the current specifications,.

B

Yeah so yeah as long as for like a basic request, you only get one span and I think you're good.

F

Yeah basically well the the at least like a internet. The logic is that if we have just one request- and it was successful- we just created span and that's it. If it's not successful, then we capture everything we can put to the put that put that first uh spam and then we do retrace and then for retries. We put this recent count, one for the next to the reset, Council and so on. All of them will be same links as of now, so there is no kind of parent Childs, uh but yeah.

F

The first one will be created anyway, but since in Java uh you have to do this conditionally right. So it is not enough to just start it uh with the first. uh The first, let's say attempt another reason why, but the first really first I have done so then you can just go like and finish it and capture Warrior again.

G

So Java behaves exactly the same way like Doppler. What you're seeing here on the screen is just a small Lego. It's a like Edge case when uh this, like this band, that it gets created and then like the the last block, is, it is only ever created, even if in case there haven't been like any actual hdb request sent on The Wire. So this is like pretty much only for this scenario when the HTTP client is trying to like, for example, is trying to connect to some address, and this address doesn't respond at all Yeah.

F

So basically we're saying like if, if everything went fine, then this span will be created by underlying underlying library or client or whatnot yeah.

G

There's another Interceptor that actually like creates the HTTP spelling. This is just uh just regard against the connection failure scenario. All.

F

Right so basically, that's.

A

F

Guarantee that span will be creative anyway, regardless yeah yeah got it yeah, that's nice! Thank you.

B

um I would that were what would be the difference here if we did start the span and used this one as the it would always be, the first try and then on the network level, not capture the first try.

G

It would make the code very complicated.

G

That's uh one thing because we would have to like, as the span like downwards to the network Interceptor and then like, depending on what happens in the lower layer like Android and create another one and like.

B

Could you could you always skip the I, mean kind of the way that we're skipping.

D

B

We have a recent counter somewhere. Could we then just always not produce a span for recent count, zero for the first one.

G

um That's your idea, the question.

B

So how about I read the code first and then ask the question.

D

G

B

Make sense because I haven't read the code yet yeah.

G

So one thing that is not shown here and maybe a little bit, not obvious, but uh the the HTTP client resented the counter. That's encapsulated within the class, gets like automatically incremented on every HTTP request.

G

So if you, if a span, has started like uh if a few like calls Below on the stack, then this will like always be more than zero when an execution arrives back here,.

B

um One question I had uh Lynn Miller on around this idea of General directions for implementers.

B

It felt like maybe like that, would be kind of some of those use cases you were trying to Probe on of like what happens. If this happens, if there's a Handler, a user driven Handler um where like where would it make sense to I mean maybe in the example, maybe in the examples we have.

G

Maybe some like implementation notes and something like that. We don't have this at all in the summative conventions and maybe it would be useful to us something like this.

C

I think, maybe by the time we announce a HTTP stability, we should have some either a blog post or document probably document, because we want to keep it up to date that describes it and I. Think Josh Suarez asked this question. Where should we put things like that? And there were some different opinions? I, don't remember if there was a conclusion, Ted wanted it to be in his pack, and we thought that maybe he opened Telemetry IO would be the right place.

C

I personally think that we can just go ahead and create a document and open Telemetry, IO and capture this best practices or something. How do we? You know the the editorial recommendations on information relationship is facts.

C

And I think it's it's more than than just retries back I think that the semantic conventions are complicated. To now that we want to. We have a reasonably full document that describes it.

G

Yeah so like HTTP has been quite simple. Maybe you accept the recent thing but like, for example, the HTTP server conventions like I. Don't have many like doubts or questions regarding that, but uh I look from time to time at uh in progress messaging spec and this one will meet a lot of like suggestions and uh like best practices for people who will Implement instrumentations, because it is complicated.

C

Yeah I, like I I'm, dipping my toes into a python instrumentations, and they are so far from what's written in this pack or at any time, was written in this fact that I'm scared of this whole life person honestly.

F

Yeah just to throw this idea. um Actually, this topic arises not even uh like I'm on this uh specification six, but also for customers.

F

So, for example, if people want to instrument their services manually, for example, for pops up scenarios or someone uh all these best practices will be really really useful, regardless of implementation right, regardless of uh technology stack, so I'm, I I would maybe even more generalize the suggestion to come up with some set of best practices on telemetry.io, which allow user just to uh you know, give more context how how they should behave or which options they have uh in some particular case when they have when they have some.

F

You know uh situations, for example, pops up or retries or some messaging stuff, uh and then using this this set of like a best practices. We can actually also create some additional uh documents for some specific implementation, let's say 4.net, but always reference them to the like a wide, open, Furniture. That's practice, it's an ability to be really helpful for everyone, not only for my dinners, but also for users.

B

D

B

Suggest, um as far as like the website or here I mean I, think the spec makes a lot of sense, um which gets it's. The spec gets synced to the website anyways, um and it would be people who are they're already coming here. If they're trying to implement this, and maybe even where we have a maybe even the whole examples stuff could move into that kind of a document.

F

Exactly so, I I would say that specifications might be like a or should be as short and like a direct as possible, so people just don't get confused when they do implementations, but all these additional topics, additional nuances, can be explained in examples and best practices or even some samples um like a highlight all descriptions um somewhere else, but yeah we might reference it from from a spec.

B

Well, I'll open an issue.

B

um To propose something like that just to get in case there's Community feedback but I like the idea.

F

Thank you. Thank you, trust it. It's really great. Thank you.

B

um So to the spec issue itself: did we have any open quests that we need.

G

I think the only open question was the one that I didn't understand previously and that we've talked over during this meeting. So probably no, but uh also like, like pretty much like only YouTube trust, because you know I have reviewed this, so uh it doesn't have like any approvals enough of rules.

B

B

C

You just written a profile, probably um I'm. Fine I was not fine with. Some must not think that, let's see if, if I, if did you update the pr after that.

G

uh uh I I last updated it two weeks ago, like with responses to your comments, awesome mostly relaxed some of the conditions, but.

B

So this one is still here: this must not.

G

I will take a look at this next weekend, uh so agreed through the entire conversation again.

B

Let me like: do you briefly Recall? Why, um like, if you're admitting that logical span, what would be? Why would you emit that.

C

Let me very cool, so this is encompassing span right, we're talking, yeah.

G

Yeah, this is a high level weight.

E

C

Wait so now we are we're talking about conditional, so it's either or yeah.

B

Right um and we don't, but we don't I, don't think we spell that out, though I.

C

Don't think we can I I I'm. This is where I'm hesitant it's it's feasible to do either or everywhere.

G

Yeah so like the occpa example that we've like talked about previously like for the most part, it implements the number one point here, but it does emit like a logical and Compass expand. So it kind of like violates the should not in line 134, but it is like uh because it should not then like it's not absolutely for million.

G

We still sort of can do that for like edge cases, so yeah I I also don't think that there should be like strong either or there probably are situations when in one instrumentations you have to like, depending on the situation like this side, foreign.

C

Essentially, okay, HTTP, or sometimes it's not even the encompassing HTTP clients fan is some extra thing that tells about the error of it.

D

B

Go ahead, it's not uh necessarily an over the wire http span that actually went over the wild wire, but it is an HTTP span in the in the sense that the user code tried to make it.

C

Yeah uh I'm trying to recall my concern: oh maybe we can, if you have time, I, can spend a bit more time on this, but otherwise we can do it offline. What do you prefer.

B

um Yeah, it's late for mitesh.

G

We can talk a couple of minutes more. It's fine, okay,.

C

So then, okay, okay, so we have two modes, then the first one is uh retry mode. This, the second one is logical mode. So the the P2 here, it's the it's impossible to trace, read price.

C

Yes, and your point is that HTTP resent count must not be set.

G

Yeah and perhaps it is like uh unnecessary statement and the way that, because it is a logical, Splendor, natural is no way to like capture recess, so you won't set the recent counter anyway, but uh yeah.

C

So if I'm, a user, okay I think, can you scroll a little bit lower address, come trying to understand what that? What exactly I meant here.

C

C

Okay, so I think yes, so um when I use HTTP client right does things like 503 or 404 or 429.

C

It's definitely my something in my application that retries those because HTTP client can't right shouldn't. Then, essentially, we have two different retries one are redirects and connection failures and DNS. That should be retried anyway and HTTP. Clients might not provide hooks to instrument those, but they like 99 of them, should provide hooks to instrument 404s 503s and things like that.

C

Either they don't it's just every request they make is a try. They don't have a anything that to handle it right or there is some additional hope for users to plug retry logic in so I think my concern is the statement that it's impossible to instrument individual tries I, think we should think about it more because we mean one class of them, but not the other. Do you see what I mean.

G

G

Think I get it so like uh you're saying that we haven't really thought about the sort of user driven. We tries like custom logic like for 29 and 503 yeah I.

G

I also think that it probably mostly depends on the HTTP client library that there I'm pretty sure that there are some that like allow you to capture everything. And then there are some that only like. Allow you to capture example, the the internal like the internal ones like the redirects, and then there are some that only have the user-driven ones, but.

G

B

G

B

This I, like I, liked materials earlier suggestion of just deleting we.

G

Could I think we could I mean it probably shouldn't be like the spec shouldn't be like too restrictive, I think around this? The it should be like I. Think, like this section, should be more about, like a general guidance that there are in very generally too basic and approach this either.

G

We try to attempt like best effort capture, every single issue, because Over The Wire as much as the HTTP client allows you to do this so, like I, think about the user driven retrines versus the like HTTP client, different phrase, in a way that we can only do like what the library allows us to do. So, if, if it does internally some magic and like we tries even more, if we cannot capture, we can capture something well too bad.

B

uh And I mean you shouldn't be using this attribute incorrectly anyways like yeah. If it's not a retry, you shouldn't be stamping that anyway,.

C

Yeah, so if we can remove it, then always.

G

B

Cooler so then, yes, that's what I wanted before pushing I will um push this in the spec meeting next week. I just wanted to get our agreement on it. First.

C

I'll hit the proof.

G

Okay, if you have any more comments, please add them. I'll write review them before this pick meeting I'll also try to join them like next week. It's on Tuesday, right, okay,.

B

And thanks for um thanks for this, this is awesome. I really and I will look through this um I really wanted to get a a fully baked prototype.

B

Or fully baked implementation, not even prototype, it.

G

Would be our first HTTP client that actually implements that.

B

All right, thank you so much for staying long. That was uh long overdue uh discussion.

C

Definitely, thank you I think you know cash convince.

F

Thank you very much. Yeah I'm really really glad that the week really clarified all this stuff. Thank you. Thanks.

D

D

C