Prototype context scoped attributes

[jack-berg]

Context scoped attributes OTEP was merged: #4931

This gives the ability for for users to add arbitrary attributes to context that will be stamped onto metrics, spans, logs recorded within the scope.

Cool feature which is long overdue. Key use case is to stamp things like tenant id onto all telemetry.

[dyladan]

One question that wasn't made clear in the OTEP: should this only apply to child contexts/spans? In the Java prototype (https://github.com/open-telemetry/opentelemetry-java/compare/main...carlosalberto:context-scoped-attrs-proto?expand=1#diff-0800d66929f0e7caa6527e0eab5f07b4a1c91f111b0ce433a304d8a5a64eb38aR217) it looks like attributes are only pulled from context on span start, meaning they only apply to children (and all downstream descendants).

I have a use case to propose that would require propagating attributes to parents and ancestors as well. When generating span-derived metrics like request duration, request count, and error rates it is sometimes useful to be able to add dimensions to those metrics that are not known at the time the entrypoint span (most commonly HTTP server span or local root span) is created. For example, it may be that you want to split a metric by whether a particular third party service was used, a database was accessed, or a particular feature flag was enabled.

There are a couple other ways this can be achieved, but they have drawbacks:

• Hold a reference to the entrypoint span and set attributes on it in multiple instrumentations. Introduces tight coupling between instrumentations and requires the span to be stored somewhere in context or thread local storage, etc. Cannot be used with OOTB instrumentations which typically don't have a mechanism for this.
• Use a SpanProcessor that copies attributes to ancestors. There is no mechanism in the spec to get the parent of a span (only parent spancontext), so the processor is required to hold a reference to all entrypoint spans. SpanProcessors also only run on span start and end so if an attribute is set while the entrypoint span is still open, that span may be ended by the time its child ends and the window when attributes can be set may already be missed.
• Monkeypatch the SDK to add this behavior. Downsides of this approach are obvious I think.

[jack-berg]

Changed to ready-needs-sponsor and removed myself as the assignee. I agree with this and hope someone else volunteers to sponsor 🙂. I originally opened this issue just to not loose track of the OTEP.