[HDFS-3914] QJM: acceptRecovery should abort current segment

• Type: Sub-task
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: QuorumJournalManager (HDFS-3077)
• Fix Version/s: QuorumJournalManager (HDFS-3077)
• Target Version/s: QuorumJournalManager (HDFS-3077)
• Hadoop Flags: Reviewed
• Assignee: Todd Lipcon
• Reporter: Todd Lipcon
• Created: 10/Sep/12 21:55
• Updated: 10/Sep/12 23:21
• Resolved: 10/Sep/12 23:21

Description

Found this bug with randomized testing. The following sequence causes a problem:

• JN writing segment starting at txid 1, and successfully wrote txid 1, but no more
• JN becomes partitioned from NN, and a new NN takes over
• new NN is also partitioned for the "prepareRecovery" phase of recovery, but properly connects for the "acceptRecovery" call
• acceptRecovery copies over a longer log segment (eg txns 1-3) from a good logger
• new NN calls finalizeLogSegment(), but gets the following error: JournalOutOfSyncException: Trying to finalize in-progress log segment 1 to end at txid 3 but only written up to txid 1

This is because the "syncLog" call (which copies the new segment) isn't properly aborting the old segment before replacing it.