Request for adoption in OTel SDKs of the W3C Trace Context Level 2 spec (which is now in Candidate Recommendation status)

[kalyanaj]

What are you trying to achieve?
As part of the W3C DT group, today we published the Level 2 (aka version 2) of the W3C Trace Context spec in "Candidate Recommendation" (CR) status: https://www.w3.org/TR/trace-context-2/.

Please see the above link for full details, to summarize here a main change from the Level 1 version (which OpenTelemetry has already adopted) is that the Level 2 spec includes a new trace flag in the traceparent header called the "random trace id" flag. When this flag is set, it conveys that at least the right-most 7 bytes of the trace ID have been generated in a random (or pseudo-random) manner. This can be helpful for samplers / sharding logic etc. as with this they can get stronger guarantees that traceID has been generated in a random/pseudo-random manner.

What did you expect to see?
Based on the above, want to update the OpenTelemetry community on this spec for your evaluation and to consider prototyping/adoption in OTel SDKs. Assuming the traceID is already being generated in a pseudo-random manner, this would involve setting the above new trace flag to reflect it.

Note: The exit criteria for this standard moving from the current "Candidate Recommendation" status to "Recommendation" status is to have at least two implementations.

Additional context.
https://www.w3.org/TR/trace-context-2/ has the details.

[dyladan]

The concrete changes here that are relevant for OTel:

1. if trace id is randomly generated (most if not all SDKs) set the random flag
2. propagate the random flag if a traceparent is received

[Oberon00]

2. propagate the random flag if a traceparent is received

Isn't propagation of unknown flags already required by the previous W3C spec? Would still be good to verify I guess.

Late EDIT: This is wrong. See comment below #3411 (comment)

[carlosalberto]

if trace id is randomly generated (most if not all SDKs) set the random flag

I wonder if we should strongly suggest all SIGs do generate random ids (if still there any SIG that doesn't).

[dyladan]

3. propagate the random flag if a traceparent is received

Isn't propagation of unknown flags already required by the previous W3C spec? Would still be good to verify I guess.

no. unknown flags SHOULD be set to 0

[jmacd]

I believe we should add flags to the protocol for Span context and links. Otherwise, a tail sampler and downstream consumer will not be able to know whether the TraceID has the random flag.

Same issue is written at least once: open-telemetry/opentelemetry-proto#382 (I'm pretty sure there's another copy of this written somewhere else).

[marcalff]

Now that the following is done:

• Add W3C-specified trace flags to v1 Span proto opentelemetry-proto#503
• Prepare release 1.1.0 opentelemetry-proto#523

can this item, W3C Trace Context Level 2, be adopted in the specifications ?

Looking to implement it for opentelemetry-cpp.

[dyladan]

The Trace Context level 2 is in candidate recommendation phase and the W3C is actively looking for adopters. It would be very helpful to us (w3c working group) if OpenTelemetry (including opentelemetry-cpp) can implement this at least experimentally to ensure it works as intended.

[marcalff]

The Trace Context level 2 is in candidate recommendation phase and the W3C is actively looking for adopters. It would be very helpful to us (w3c working group) if OpenTelemetry (including opentelemetry-cpp) can implement this at least experimentally to ensure it works as intended.

There is a chicken and eggs problem with the process then.

Preparation work has been done already:

https://github.com/open-telemetry/opentelemetry-cpp/blob/497eaf43e5676ae7982f7119a2ac70d09211a6f5/sdk/src/trace/tracer.cc#L83-L93

I am waiting for the opentelemetry-specifications to change, to mention that supporting level 2 is ok (experimental maybe), to uncomment that code and actually support level 2 in the implementation.

[tsloughter]

Currently implementations use either the right most 63 or 64 bits for sampling. We should all adopt 56 bit but only once this is adopted?