-
Type:
Bug
-
Resolution: Unresolved
-
Priority:
Major - P3
-
None
-
Affects Version/s: None
-
Component/s: None
-
None
-
Storage Execution
-
ALL
-
-
Storage Execution 2026-11-23
-
0
-
None
-
None
-
None
-
None
-
None
-
None
-
None
A transaction that commits successfully can still report a writeConcernError(WCE) if a session kill delivered an interrupt to the committing operation before its majority write-concern wait completed. The interrupt is sticky and cannot be retracted, so the client receives ok: 1 with WCE on a transaction that is durably committed. Retrying the whole operation in response is unsafe because the write already committed.
Why the error is not recoverable for clients as-is?
TemporarilyUnavailable (365) carries no error categories therefore has no TransientTransactionError label and clients get no signal that only the commit needs retrying. A client that retries the whole operation after observing WCE 365 on a committed transaction will duplicate the write.
Mechanisms that can trigger this
Any session-kill path that interrupts a checked-out operation can deliver the interrupt inside the commit-to-write-concern-wait window:
- Cache-pressure rollback thread. The periodic thread calls killOldestTransaction() with reason TemporarilyUnavailable (365) when the storage engine reports cache pressure (periodic_runner_cache_pressure_rollback.cpp). This is what produced BF-45709.
- transactionLifetimeLimitSeconds killer. The same killSessionsAction machinery aborts long-running transactions with code 290 (kill_sessions_local.cpp).
- User-issued session kills. killSessions and killAllSessions interrupt operations with reason Interrupted(kill_sessions_local.cpp).
- killOp. Marks the target operation killed through the same killOperation() -> markKilled() path (operation_killer.cpp). This is what the attached reproducer uses.
- related to
-
SERVER-135378 Add logging to killOldestTransaction() to differentiate cases where candidate is identified but not aborted
-
- Open
-