Follow-up from PR #942 (review item P3, #942 (review)).
The escape logger installed by ccdproc/conftest.py uses a single process-global _ReentrancyGuard shared by all three wrapped functions (np.asarray, np.asanyarray, np.ma.asanyarray). Under a threaded scheduler (e.g. dask's default threaded scheduler):
- While thread A holds the guard (
guard.active = True), a genuine escape occurring concurrently on thread B is silently skipped — the wrapper treats it as a re-entrant call.
- The tally update in
record_escape_log() (_ESCAPE_LOG_COUNTS[key] += 1 on a defaultdict) is not atomic on first insert.
Both failure modes drop real escapes, so the baseline ratchet (CCDPROC_ENFORCE_ESCAPE_BASELINE=1) can produce false negatives — never false alarms — meaning a new escape site reached only from worker threads could slip past CI.
Fix sketch:
- Replace the shared guard with
threading.local() so re-entrancy is tracked per thread.
- Take a
threading.Lock around the tally increment (or use a thread-safe counter).
Follow-up from PR #942 (review item P3, #942 (review)).
The escape logger installed by
ccdproc/conftest.pyuses a single process-global_ReentrancyGuardshared by all three wrapped functions (np.asarray,np.asanyarray,np.ma.asanyarray). Under a threaded scheduler (e.g. dask's default threaded scheduler):guard.active = True), a genuine escape occurring concurrently on thread B is silently skipped — the wrapper treats it as a re-entrant call.record_escape_log()(_ESCAPE_LOG_COUNTS[key] += 1on adefaultdict) is not atomic on first insert.Both failure modes drop real escapes, so the baseline ratchet (
CCDPROC_ENFORCE_ESCAPE_BASELINE=1) can produce false negatives — never false alarms — meaning a new escape site reached only from worker threads could slip past CI.Fix sketch:
threading.local()so re-entrancy is tracked per thread.threading.Lockaround the tally increment (or use a thread-safe counter).