Should stable HTTP instrumentations published under the OpenTelemetry org implement all Recommendeds and SHOULDs?

[trask]

I personally would expect stable HTTP instrumentations published under the OpenTelemetry org to implement all Recommends and SHOULDs.

Part of the issue is that Recommendeds and SHOULDs can often not be added later to stable instrumentations, e.g. since adding new metric attributes is a breaking change.

If there is a reason why a particular instrumentation chooses not to implement a Recommended or a SHOULD I think that this should be documented along with the rationale (e.g. it is prohibitively expensive for some reason to capture a specific attribute from a specific library, or a specific library provides no way to capture a specific attribute).

[trask]

Discussed in Semantic Convention WG today.

It looks like we are fairly well covered already by RFC wording https://datatracker.ietf.org/doc/html/rfc2119:

SHOULD This word, or the adjective "RECOMMENDED", mean that there may exist valid reasons in particular circumstances to ignore a particular item, but the full implications must be understood and carefully weighed before choosing a different course.

I would say that not implementing a SHOULD is probably worth opening an issue each time (in the lang-specific repo), both to track the reasoning, and to provide visibility and an opportunity to comment from the (lang-specific) community in case there are any differing opinions.

But I'm not sure that this requires any special wording / guidance in this repository.