Fix: ETW span ownership lifetime

[fmutlu68]

Fixes #3849

Changes

• Hold shared ownership of the ETW tracer in Span to prevent dangling references
• Update span creation to pass shared ownership
• Testing: ./ci/do_ci.sh cmake.c++17.test

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 82.10%. Comparing base (3057616) to head (1f76b52).
⚠️ Report is 4 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #4070   +/-   ##
=======================================
  Coverage   82.10%   82.10%           
=======================================
  Files         385      385           
  Lines       15989    15989           
=======================================
  Hits        13126    13126           
  Misses       2863     2863           

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[fmutlu68]

Could you please have a look here ? also @ThomsonTan It passes all tests

[lalitb]

ETW exporter is a bit different from other exporters: the tracer name maps to the ETW provider name/GUID, so creating multiple tracers for the same name is not the intended hot-path pattern. Applications should normally create one tracer per ETW provider name and reuse it for the process/component lifetime.

Because of that, I’d prefer not to fix this by adding shared_from_this()/shared_ptr ownership work to every span creation. StartSpan() is the high-throughput path for ETW. Could we instead document the intended ETW tracer reuse model and/or have TracerProvider retain created tracers so the lifetime protection is handled at setup/GetTracer() time rather than per span?

Also @ThomsonTan suggestions would be helpful, as he has more experience on this.

[fmutlu68]

ETW exporter is a bit different from other exporters: the tracer name maps to the ETW provider name/GUID, so creating multiple tracers for the same name is not the intended hot-path pattern. Applications should normally create one tracer per ETW provider name and reuse it for the process/component lifetime.

Because of that, I’d prefer not to fix this by adding shared_from_this()/shared_ptr ownership work to every span creation. StartSpan() is the high-throughput path for ETW. Could we instead document the intended ETW tracer reuse model and/or have TracerProvider retain created tracers so the lifetime protection is handled at setup/GetTracer() time rather than per span?

Also @ThomsonTan suggestions would be helpful, as he has more experience on this.

Thanks for the thorough explanation, I completely agree regarding the hot-path overhead; atomic reference counting in StartSpan() is not ideal for high-throughput ETW logging. The alternative approach makes perfect sense. I will revert the shared_from_this()/shared_ptr changes on the Tracer/Span and instead have the TracerProvider retain ownership of the created Tracers (e.g., via an internal map) to ensure lifetime protection at setup/GetTracer() time. I'll push the updated implementation shortly.