Probability sampling in tracestate specification

[jmacd]

Changes

Specifies how to interpret and use p-value and r-value for consistent probability sampling decisions, including two optional samplers named ConsistentProbabilityBased and ConsistentParentProbabilityBased.

Does not modify existing built-in Sampler definitions.

This is ready for general review.

Implemented in open-telemetry/opentelemetry-go-contrib#1379

OTEPs
https://github.com/open-telemetry/oteps/blob/main/text/trace/0168-sampling-propagation.md
https://github.com/open-telemetry/oteps/blob/main/text/trace/0170-sampling-probability.md

Part of #1414
Part of #1412
Part of #1524
Part of #570
Another take on #1899
Fixes #2224
Part of #2179

[jmacd]

@oertl @dyladan @carlosalberto Please take a look at this draft. I'm expecting at least the two of us vendors to supply implementations of this, while we continue to explore ways of adding sampling support into the traceparent.
@ocelotl Please take a look, I've tried to format this in the manner prescribed by #1210 and implemented in #2003.

[jmacd]

@open-telemetry/specs-trace-approvers the two optional samplers specified here are implemented by this PR in opentelemetry-go-contrib: open-telemetry/opentelemetry-go-contrib#1379

[ocelotl]

Changing to comment instead of requesting changes since these are not critical changes needed.

[jmacd]

@open-telemetry/specs-trace-approvers This meets our minimum criteria for merging. I have been trying to get more stamps of approval, but it's not happening (or not happening quickly). I propose to merge this PR.

[mtwo]

Looks good, a bit long but that's likely necessary given the topic. I couldn't find an example for what I suspect might be a common scenario: when an edge service uses a probability-based sampler but a downstream service decides to sample when it sees a trace with a sampled flag OR some other condition is satisfied (an error, for example).

I think (?) that this is similar to the "interoperability with TraceIDRatioBased sampler" example, though the second sampler in that case uses different criteria.

[jmacd]

@mtwo There isn't an exact example corresponding to your question. The section "Example: Probability and non-probability sampler in a root context" comes close, and there is a related discussion in #2179. As discussed in the issue, a better Delegating sampler API would be nice, but is not critical to the answer here. Sampler API details aside (it's a bit messy, we could use a v2 API proposal to clear things up--see #2179), the child span is expected to have adjusted count equal to the parent when the span is sampled, and adjusted count 0 when the parent was not sampled but there were errors (p=63).

[carlosalberto]

We have more than enough reviews and no standing questions. Let's merge this now and address any improvements in future PRs.