Regression with 0.20 regarding remote contexts

[valkum]

Bug Report

We see a regression after updating to v0.20 regarding remote context and how traceIds are propagated to children.

Version

tracing v0.1.37
tracing-futures v0.2.5
tracing-subscriber v0.3.17
tracing-core v0.1.31
tracing-log v0.1.3
tracing-attributes v0.1.26
tracing-serde v0.1.3
tracing-opentelemetry v0.20.0

Platform

Description

We have an instrumented function (fn handler) in which we call OpenTelemetrySpanExt::set_parent.
Previously, we used something like this:

let context = Context::default();
let context = TraceContextPropagator::new().extract_with_context(&context, &carrier);
Span::current().set_parent(context);

With v0.20 this creates two traces because all spans created after setting the context will get a different traceId than the
span created for handler.

The example of set_parent uses TextMapPropagator::extract(&carrier). extract just calls Context::currenty() and then calls extract_with_context. If we change our code to use something equivalent:

let context = Span::current().context();
let context = TraceContextPropagator::new().extract_with_context(&context, &carrier);
Span::current().set_parent(context);

we miss a common root span according to Jaeger.

Note that the root span is entered at the time of calling Span::current().context().

I think the regression was caused by #26.

If this is some error at our side, it should be better documented for set_parent. If not, and the PR is also not the cause, I assume we hit a regression in opentelemetry.
We are willing to create a PR once we know which path should be taken.

[valkum]

I read the description of #45 more thoroughly, and it seems this is a duplicate of #45. Feel free to close this or the other one (depending on which one you think contains more context to fix this)

[jtescher]

Resolved in #55