Provide a way for Logs SDK to sample based on span context

[aabmass]

The OpenAI instrumentation currently only records events if the span is recording:

opentelemetry-python-contrib/instrumentation-genai/opentelemetry-instrumentation-openai-v2/src/opentelemetry/instrumentation/openai_v2/patch.py

Lines 59 to 63 in 2756c1e

	if span.is_recording():
	for message in kwargs.get("messages", []):
	event_logger.emit(
	message_to_event(message, capture_content)
	)

This makes sense, since event recording can be quite expensive and we can piggyback on the sampling decision. However it feels like the Logging API should be more integrated and instrumentation shouldn't need to do these checks. Most logging libraries also have an API to check if recording like Logger.isEnabledFor() and/or support lazy evaluation of log fields. So two questions are

1. Should GenAI events be recorded only when a span is recording?
2. Should the instrumentation be in charge of this or should the Logging SDK provide its own API to check for recording or lazily evaluate log entry fields.

[drewby]

We are currently making an implicit assumption that events are recorded as part of a span, but the specification does not explicitly require this—though all the examples at the end show events within spans. It might be worth updating the specification to clarify this expectation either way.

[xrmx]

My thinking is that traces and logs are two different signals and so they should be ~independent.

[aabmass]

I agree it's cleaner to keep them separate when writing instrumentation. Maybe the logging SDK can provide a means to tie span sampling to event sampling. We have some precedent here in Exemplars where the default ExemplarFilter is TraceBased.

[aabmass]

Discussed in SIG and we agreed Logs/events SDK should have way to configure this behavior. Changing to a feature request and I think we may need a spec discussion