Sampling relevant attributes: MUST vs SHOULD

[trask]

Related to open-telemetry/opentelemetry-dotnet#319

Currently the HTTP semantic convention says:

Following attributes MUST be provided at span creation time (when provided at all), so they can be considered for sampling decisions:

• server.address
• server.port
• url.full

We have an example where this is not possible: open-telemetry/opentelemetry-dotnet-contrib#2012

This may get resolved in .NET 9 (~1 year from now), but I think it's worth discussing MUST vs SHOULD for sampling relevant attributes.

[trask]

Discussed in 10/23 semconv meeting and decided to stay with MUST because without providing these attributes at span start sampling experience is degraded and sampling is important

[trask]

Reconsidering based on #467 (comment)

The current wording:

Following attributes MUST be provided at span creation time (when provided at all)

seems to imply that if a user enhanced an HTTP span (e.g. via some callback) with one of these "sampling relevant" attributes, then they would be violating the HTTP semantic conventions.

Changing MUST to SHOULD seems to be the simplest way to address this (though there are likely other ways).

I also think changing from SHOULD to MUST in the future would be less impactful compared to changing from MUST to SHOULD, so I think SHOULD is the more conservative choice here.

[lmolkova]

I'm okay changing from MUST to SHOULD, but the way I read the full sentence "Following attributes MUST be provided at span creation time (when provided at all), so they can be considered for sampling decisions:" is:

• you have to provide attributes before the span starts to be able to access them at sampling time.
• instrumentation library MUST collect them before the span starts
• instrumentation library that provides a callback to collect sampling relevant attributes MUST do it before starting a span

If these intentions are not clear, let's improve them (and change MUST to SHOULD if it feels better)

[trask]

Discussed with @lmolkova and we agreed that it's good to change sampling relevant attributes from MUST to SHOULD.

• SHOULD is already a strong statement (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.

  and we expect all HTTP instrumentations published under the OpenTelemetry org to follow SHOULDs.

• MUST here implies that users who enhance a span with a sampling relevant attribute (e.g. http.request.header.*) after start are "breaking compliance with HTTP semantic conventions", while SHOULD does not imply "breaking compliance".

[trask]

re-opening until we apply the change in this repo