Vincent Jiang created KAFKA-14497:
-------------------------------------
Summary: LastStableOffset is advanced prematurely when a log is
reopened.
Key: KAFKA-14497
URL: https://issues.apache.org/jira/browse/KAFKA-14497
Project: Kafka
Issue Type: Bug
Reporter: Vincent Jiang
In below test case, last stable offset of log is advanced prematurely after
reopen:
# producer #1 appends transaction records to leader. offsets = [0, 1, 2, 3]
# producer #2 appends transactional records to leader. offsets = [4, 5, 6, 7]
# all records are replicated to followers and high watermark advanced to 8.
# at this point, lastStableOffset = 0. (first offset of an open transaction)
# producer #1 aborts the transaction by writing an abort marker at offset 8.
ProducerStateManager.unreplicatedTxns contains the aborted transaction
(firstOffset=0, lastOffset=8)
# then the log is closed and reopened.
# after reopen, log.lastStableOffset is initialized to 4. This is because
ProducerStateManager.unreplicatedTxns is empty after reopening log.
We should rebuild ProducerStateManager.unreplicatedTxns when reloading a log,
so that lastStableOffset remains unchanged before and after reopen.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
