refactor(profiling): replace bytecode wrapping with sys.monitoring (PROF-14200)

[vlad-scherbich]

← Prev PR: #17624

3.14 vs 3.15 (table view)

3.14 vs 3.15 (flame graph): python

3.14 vs 3.15 (flame graph): native

Description

Refactors ddtrace/profiling/_asyncio.py to replace the functools.partial + wrap() monkey-patching mechanism with direct attribute replacement throughout, and introduces a sys.monitoring PY_RETURN hook layer on Python 3.15+ for task-creation hooks.

Context: The existing wrap() approach intercepts calls by replacing a function in its owning namespace with a partial-applied closure. Direct attribute replacement (_original = f; def _patched(...): ...; module.f = _patched) achieves identical runtime behavior with less indirection and no dependency on get_argument_value.

sys.monitoring for task-creation hooks (Python 3.15+):
asyncio.create_task and TaskGroup.create_task return the new Task object, which we need in order to call stack.link_tasks(parent, task). sys.monitoring CALL/PY_START events don't expose callee arguments, but PY_RETURN gives us the return value directly — making it a natural fit for these two sites. On older Python the same sites fall back to standard attribute replacement.

Changes

• set_event_loop, _GatheringFuture.__init__, _wait, as_completed (asyncio) and new_event_loop, set_event_loop (uvloop): converted from wrap() closures to direct _original_X / _patched_X attribute replacement
• create_task, TaskGroup.create_task: sys.monitoring PY_RETURN hook on Python 3.15+; direct attribute replacement on older versions
• New helpers: _register_return_hook(), _unregister_all_return_hooks(), _py_return_dispatch(), _py_return_handlers dict, _monitoring_tool_id
• Drops functools.partial, get_argument_value, and wrap imports
• sys.monitoring path is gated at >= (3, 15) — not rolled out to 3.12–3.14 yet (see TODO in code for follow-up)

Testing

CI: tests/profiling/collector/test_asyncio.py, tests/profiling/test_scheduler.py

Manual A/B validation

py3.14 (baseline) vs py3.15 on movies-ab service — covers experiment window from branch start to present:

• py314 vs py315 comparison view

Risks

Low. The behavioral contract is unchanged — the same events trigger the same link_tasks / weak_link_tasks / track_asyncio_loop calls. The main risk is a subtle argument-passing error in the direct-replacement wrappers; these are covered by the asyncio profiling tests.

Additional Notes

• sys.monitoring tool IDs 4–5 are used (free custom slots; 0–3 are reserved for debugger, coverage, profiler, optimizer). Tool 5 is tried first to minimize conflicts.
• One-line fix to _memalloc_tb.cpp is bundled in this branch.

[datadog-datadog-prod-us1]

 

✨ Fix all issues with BitsAI

⚠️ Warnings

🚦 16 Pipeline jobs failed

Changelog | Validate changelog     

DataDog/apm-reliability/dd-trace-py | build linux cp315: [amd64, cp315-cp315, v85383325-751efc0-musllinux_1_2_x86_64]     

DataDog/apm-reliability/dd-trace-py | build linux cp315: [amd64, cp315-cp315, v85383392-751efc0-manylinux2014_x86_64]     

View all 16 failed jobs.

ℹ️ Info

No other issues found (see more)

🧪 All tests passed
❄️ No new flaky tests detected

Useful? React with 👍 / 👎

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: 670d7db | Docs | Datadog PR Page | Give us feedback!}

[pr-commenter]

Benchmarks

Benchmark execution time: 2026-06-01 21:06:16

Comparing candidate commit 0af646f in PR branch vlad/ddtracepy-315-profiling-asyncio-monitoring with baseline commit 653e008 in branch main.

Found 0 performance improvements and 3 performance regressions! Performance is the same for 597 metrics, 9 unstable metrics.

scenario:iastaspects-stringio_aspect

• 🟥 execution_time [+569.224µs; +613.534µs] or [+14.520%; +15.651%]

scenario:iastaspectsospath-ospathbasename_aspect

• 🟥 execution_time [+107.851µs; +112.955µs] or [+24.894%; +26.072%]

scenario:span-start

• 🟥 execution_time [+1.358ms; +1.569ms] or [+8.564%; +9.893%]

[cit-pr-commenter-54b7da]

Codeowners resolved as

ddtrace/profiling/_asyncio.py                                           @DataDog/profiling-python
ddtrace/profiling/collector/_memalloc_tb.cpp                            @DataDog/profiling-python

[github-actions]

This pull request has been automatically closed after a period of inactivity.
After this much time, it will likely be easier to open a new pull request with the
same changes than to update this one from the base branch. Please comment or reopen
if you think this pull request was closed in error.