[DO NOT MERGE] Rename db|messaging|gen_ai.system to *.provider.name, rpc.system to rpc.protocol.name

[lmolkova]

Fixes #1581

Warning

THIS IS A VERY-VERY-VERY BREAKING CHANGE (even though it affects experimental attributes only).
Backends use presence of these attributes as an indication that corresponding spans follow certain conventions.

• *.provider.name is
  • Leaves extension point to provide *.provider.version and similar in the future, especially when defining server-side conventions
  • it is more precise for messaging/db/gen_ai and feature_flag areas where it represents client perception about backend provider (project/product/service/vendor).
  • There used to be feature_flag.protocol_name, so we can't change it to feature_flag.protocol.name yet due to const name collision Implement code-generation hints to drop/rename attributes in case of a collision #1462 moved to Rename feature_flag.system back to feature_flag.provider_name #1614
• rpc.protocol.name is more precise for RPC where it captures different application-or-higher-level protocol that may work on top of another application protocol (grpc/thrift over http/2)

Despite being problematic, we believe this change is necessary to accommodate future extension of semantic conventions for *.system.* and this is the last chance to make this attribute name right before we declare any of these attributes stable.

Merge requirement checklist

• CONTRIBUTING.md guidelines followed.
• Change log entry added, according to the guidelines in When to add a changelog entry.
  • If your PR does not need a change log, start the PR title with [chore]
• schema-next.yaml updated with changes to existing conventions.

---

Given potentially high impact of this change and also upcoming holiday season, we'll need to keep it open for a while to collect the feedback

[lmolkova]

I'm going to extract it to another PR that we can merge right now to reduce churn in feature-flag area

[lmolkova]

I don't believe it's used correctly here

[Flarna]

I think it is, AWS SDK is modeled as remote call. There is #1622 to add aws_api to the well known enums.

Anyhow, removing rpc.system but keep rpc.method and rpc.service doesn't seem right.

[lmolkova]

it's not a remote call though - it's a logical call that uses HTTP underneath. OTel RPC semconv represent actual remote calls using remote calling protocols.

[Flarna]

Well, that's more or less the same for GRPC which uses HTTP/2 as transport.
Anyhow, likely a topic for a different issue.

But still I think you should either keep or remove all rpc attributes here.

[codefromthecrypt]

I have some opinions on this, which I don't expect to be accepted, but would like on the record even if dismissed.

As genai is typically accessed from web service apis which are commodity and easier than the typical case of database protocols, I would prefer changes to be discussed independently.

For the case of genai, I do strongly prefer whatever word we use for api (e.g. openai which is often emulated) to be decoupled from the provider of that. However, I would like this to advise to use the same constant even if in some cases the client might get it wrong. For example, don't say something like "azure inference client" as the system or provider name, for when it is being used to access azure openai service with the openai api. Especially a proxy client (litellm also has a proxy client library), should attempt to do more than say its name.

The field will be used for both sides, and it seems a step back to end up with client SDK names in either of these fields, vs a web service api, possibly private provider of that. The edge case of local llms, e.g. running llama.cpp embedded, we can discuss, too, but most of the time this is about web services.

more below if you like

---

I think there is ongoing tension treating genai like database, when in practice they are typically accessed via web services apis and so more like a normal cloud service than a database. This becomes even more the case as more and more are instrumented server side, or as proxies. For example, while there are indeed some L7 proxies for database protocols, it is a lot more common and easier to do on genai as most if not all well known services are web services. So, there are backends who instrument the http server side like a normal private cloud product, as well as proxies who instrument both ways today.

So, I would like to decouple the topic of database from genai, especially this sense of provider being implied as a client identifier. I'm not sure "provider" or "system" for that matter intuitively mean "client perceived name" especially when the attribute isn't namespaced as client and also used for server data (in practice even if only in genai metrics are the only server thing so far).

I cofounded a project back in 2009 called jclouds which has very much overlap on the problem of api vs provider of that. When the community got together, probably the single best feature we did was formalizing "api" independent of "provider" of that. So, while I do like the idea of using provider, I would like it to be more strong, so it at least suggests to use an authoritative name for the actual service as opposed to a client SDK name, especially in SDKs that target multiple services.

https://jclouds.apache.org/start/concepts/#:~:text=Providers%20implement%20APIs%2C%20and%20describe,those%20offered%20by%20the%20API.

[lmolkova]

@codefromthecrypt can you suggest what'd you like to be changed/added ?

[github-actions]

This PR was marked stale due to lack of activity. It will be closed in 7 days.