[GeorgiaPreda]

RFC to address #8999.

Problem and change

Every deserialized Semaphore started a lease-refresh timer, even when it was only an unused task argument. Task objects can retain these instances, causing active timer counts and idle event-loop CPU to grow with submissions. The bound callback also retained otherwise unreachable semaphore instances.

Start the callback only after successful registration, and use a weak reference inside it. Add regressions covering collection before and after registration, unused arguments not starting timers, and acquisition starting lease renewal.

Verification

• All three new regression cases fail against the original initialization.
• 31 semaphore/lock tests passed.
• Ruff 0.15.16 check and format check passed on both changed files; git diff --check passed.

Bounded CPU comparison

Separate baseline and changed-code runs used four worker processes, ten unused semaphore arguments per task, 1,200 total submissions, 5 ms simulated tasks, and the default lease timeout. Idle CPU was sampled over approximately 6.5 seconds per checkpoint using thread_time on the shared scheduler/client event-loop thread. Both runs used the same /proc compatibility adjustment.

Completed tasks	Original idle CPU	Changed idle CPU	Active callbacks, original / changed
0	3.15%	2.51%	10 / 0
200	4.31%	2.93%	2,010 / 0
600	6.05%	3.36%	6,010 / 0
1,200	7.33%	3.09%	12,010 / 0

This supports removal of timer-driven CPU growth in the bounded workload. It is one comparison, not a statistically controlled benchmark, a scheduler-only process measurement, or proof that the original long-running cluster crash is resolved. Task objects still retained semaphore instances during the measurement; unused instances no longer create active timers.

• Tests added / passed for the selected scope
• Changed-file Ruff checks passed
• Full test suite and all pre-commit hooks

AI tooling: OpenAI GPT-6 via Codex in ChatGPT Work assisted with investigation, implementation, testing and publication.

[github-actions]

Unit Test Results

See test report for an extended history of previous test failures. This is useful for diagnosing flaky tests.

    40 files  ± 0      40 suites  ±0   14h 0m 32s ⏱️ - 37m 3s
 4 163 tests + 3   3 983 ✅ + 4    178 💤 ±0   2 ❌  -  1 
81 019 runs  +58  76 759 ✅ +43  4 239 💤  - 3  21 ❌ +18 

For more details on these failures, see this check.

Results for commit 58ee127. ± Comparison against base commit dc182bd.