[22elix3r]

Fixes #4695

Changes

Remove noexcept from the declarations and definitions of TracerContext, LoggerContext, and MeterContext constructors. These constructors copy resources and, for trace/log contexts, allocate a multi-processor. Failures can now unwind through the already-throwing provider constructors to the caller.

The production diff is only six constructor-signature lines. Runtime GetTracer, GetLogger, and GetMeter remain noexcept.

Add a dedicated allocation-injection executable: runtime tests target resource schema-string copying in all three contexts/providers and internal processor allocation in trace/log contexts. Injection stays out of other tests; CMake excludes the executable for Windows shared-library builds where executable allocation overrides cannot intercept DLL allocations.

Reproduction and verification

On unmodified upstream adebdf0c, each of the six context/provider resource-copy tests aborted with std::bad_alloc instead of reaching its caller's catch block.

With the fix:

• Full configured suite: 1,120/1,120 CTest cases passed.
• All eight runtime allocation tests passed under both ABI v1 and v2.
• ABI-v2 initialization/provider subset: 64/64 passed with metrics exemplar preview enabled.
• C++14 production syntax checks and C++17 test compilation with exceptions disabled passed.
• Allocation regression executable passed with AddressSanitizer and leak detection (linked against the normal SDK build).
• Formatting and diff checks passed; clang-tidy 21 reported only existing warnings in unrelated processor/metrics headers.

CI follow-up

• Add the missing direct sampler/id-generator includes, and retain <cstddef> explicitly for std::size_t.
• Keep the test's replacement new/delete implementations out of line on GCC, eliminating the reproduced optimized -Werror=mismatched-new-delete failures without disabling warnings.
• Set VALGRIND_OPTS only for this Bazel test so its allocation-failure hooks remain active while libc malloc/free are still checked. No tests are skipped.
• Reproduce the Bazel crash locally under GDB: unwinding a failed string-valued attribute copy crashes in std::variant cleanup. Target the Resource's schema-string copy instead, preserving nonempty attributes and the same constructor-propagation check without also testing that separate cleanup path.

The revised six resource-copy regressions were linked against unmodified adebdf0c: every one still aborted with exit 134. They pass with this PR, so changing the allocation target does not hide the original noexcept bug. This follow-up changes only the test source and its Bazel target; the six-line production constructor fix is unchanged.

Follow-up verification:

• Full configured suite: 1,120/1,120 passed; all eight allocation tests also passed under ABI v2.
• GCC 14 -O3 -Werror/maintainer-warning compilation passed for both ABIs; the C++23 GCC 14 Release/maintainer executable passed all eight cases.
• Both ABI allocation executables passed under Valgrind with zero errors and no leaks; the Bazel target also passed with CI's async/preview flags.
• Clang Bazel ASan, UBSan, TSan, and no-RTTI configurations passed the target. GCC sanitizer verification was unavailable locally because its ASan runtime file is missing; Clang used its installed runtimes.
• Local IWYU 0.25 reports correct includes in all three ABI/preview modes; CI uses 0.26. Format, Buildifier, and no-exceptions compilation passed; clang-tidy 21 only reported existing warnings in unchanged headers.

• Changelog updated
• Runtime regression tests added
• Public API reviewed (context constructor exception specifications intentionally relaxed)

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 86.83%. Comparing base (fc137c6) to head (5b77642).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #4703      +/-   ##
==========================================
+ Coverage   86.82%   86.83%   +0.01%     
==========================================
  Files         525      525              
  Lines       20726    20726              
==========================================
+ Hits        17993    17995       +2     
+ Misses       2733     2731       -2     

Files with missing lines	Coverage Δ
...k/include/opentelemetry/sdk/trace/tracer_context.h	100.00% <ø> (ø)
sdk/src/logs/logger_context.cc	93.94% <100.00%> (ø)
sdk/src/metrics/meter_context.cc	93.82% <100.00%> (ø)
sdk/src/trace/tracer_context.cc	82.36% <100.00%> (ø)

... and 1 file with indirect coverage changes

🚀 New features to boost your workflow:

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

[lalitb]

Could you help clarify what caused the crash with the original string-valued attribute? Was it an issue with the test setup or with cleanup after the resource copy failed?

[22elix3r]

It was cleanup during exception unwinding of the partially constructed resource copy, not initial fixture setup. The source Resource was fully constructed before FailAllocation was enabled. The injected std::bad_alloc occurred while copying the owned string attribute; GDB showed the crash in std::variant cleanup (_Variant_storage / _M_reset / std::__do_visit).

I also reproduced this independently with a standalone std::variant containing the same alternative types, without the SDK or GoogleTest, using GCC 14, GCC 15, and Clang with libstdc++. The partially copied variant had index 255 (the valueless sentinel), but cleanup still entered the visit dispatch and segfaulted. That narrows it to the variant-copy cleanup path rather than the fixture setup.

In 71fe088 the tests instead fail the Resource schema-string copy, while retaining a nonempty attribute map. This isolates context-constructor exception propagation; it does not fix the separate variant cleanup issue. I rechecked all six revised resource-copy cases against unmodified adebdf0: each still aborted with exit 134, whereas with this PR they catch std::bad_alloc.

[lalitb]

Thanks, that answers my question. Targeting the schema-string copy makes sense, especially since the revised tests still catch the original bug.

Let's track the remaining attribute-copy cleanup failure in a separate issue and link it from the test comment. I'm happy to approve this fix independently.

[22elix3r]

Tracked separately in #4707, including the SDK-independent reproducer and GCC 14, GCC 15, and Clang-with-libstdc++ results. Linked it from the shared resource-copy test comment in fb55590; the production fix and allocation-injection behavior are unchanged.

[lalitb]

Thanks for the fix. This lets context construction failures reach the caller instead of terminating. I’ve left one question about the changed test before approving.

[lalitb]

Since these are public constructors, removing noexcept changes their exception contract and can affect code relying on nothrow construction. Should we mark this as a breaking change in the changelog?

[22elix3r]

Yes, removing noexcept changes the public exception contract even though this is a bug fix. Done in fb55590: the entry is now under Breaking changes and explicitly notes the impact on code relying on nothrow construction, consistent with #4469.

[marcalff]

LGTM, thanks for the fix and tests.