Tanuj Khurana created PHOENIX-7984:
--------------------------------------
Summary: Fence writer on sync failure to prevent false-success RPO
loss
Key: PHOENIX-7984
URL: https://issues.apache.org/jira/browse/PHOENIX-7984
Project: Phoenix
Issue Type: Sub-task
Reporter: Tanuj Khurana
Assignee: Tanuj Khurana
A block-full or explicit sync() that fails after closeBlock() has already
advanced block state leaves the LogFileWriter in a torn-down state. A
subsequent retry sync() then finds nothing left to flush and returns success —
a false ACK. That false ACK clears currentBatch, silently dropping the unsynced
records (RPO loss). The retry succeeds against a writer whose durability
barrier was never actually crossed.
Fix (mirroring HDFS DFSOutputStream single-shot semantics)
- LogFileWriter: latch the first append/sync failure. Every subsequent
append/sync fails fast rather than re-driving the already torn-down stream, so
a fenced writer can never report false success.
- ReplicationLog.apply(): a retry is only meaningful on a fresh writer. On
failure, request rotation; if no fresh writer can be staged, surface the
original failure instead of burning a retry attempt on the fenced writer.
Recovery remains the higher layer's responsibility: rotate to a fresh writer
and replay the unsynced batch. The failing event's own record is recovered via
replayFailedEvent.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
