Details
-
Sub-task
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
None
-
None
-
None
Description
KafkaMetadataLog's appendAsLeader expects the base offset to match the LEO. This is enforced when KafkaRaftClient uses the BatchAccumulator to create batches. When creating the control batch for the LeaderChangeMessage the KafkaRaftClient doesn't use the BatchAccumulator and instead creates the batch with the default base offset of 0.
This causes the validation in KafkaMetadataLog to fail with the following exception:
kafka.common.UnexpectedAppendOffsetException: Unexpected offset in append to @metadata-0. First offset 0 is less than the next offset 5. First 10 offsets in append: ArrayBuffer(0), last offset in append: 0. Log start offset = 0 at kafka.log.Log.append(Log.scala:1217) at kafka.log.Log.appendAsLeader(Log.scala:1092) at kafka.raft.KafkaMetadataLog.appendAsLeader(KafkaMetadataLog.scala:92) at org.apache.kafka.raft.KafkaRaftClient.appendAsLeader(KafkaRaftClient.java:1158) at org.apache.kafka.raft.KafkaRaftClient.appendLeaderChangeMessage(KafkaRaftClient.java:449) at org.apache.kafka.raft.KafkaRaftClient.onBecomeLeader(KafkaRaftClient.java:409) at org.apache.kafka.raft.KafkaRaftClient.maybeTransitionToLeader(KafkaRaftClient.java:463) at org.apache.kafka.raft.KafkaRaftClient.handleVoteResponse(KafkaRaftClient.java:663) at org.apache.kafka.raft.KafkaRaftClient.handleResponse(KafkaRaftClient.java:1530) at org.apache.kafka.raft.KafkaRaftClient.handleInboundMessage(KafkaRaftClient.java:1652) at org.apache.kafka.raft.KafkaRaftClient.poll(KafkaRaftClient.java:2183) at kafka.raft.KafkaRaftManager$RaftIoThread.doWork(RaftManager.scala:52) at kafka.utils.ShutdownableThread.run(ShutdownableThread.scala:96)
We should make the following changes:
- Fix MockLog to perform similar validation as KafkaMetadataLog::appendAsLeader
- Use the LEO when creating the control batch for the LeaderChangedMessage
Attachments
Issue Links
- links to