[HADOOP-17584] s3a magic committer may commit more data

• Type: Sub-task
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.2.0
• Fix Version/s: 3.3.5
• Component/s: fs/s3
• Assignee: Steve Loughran
• Reporter: yinan zhan
• Created: 12/Mar/21 10:07
• Updated: 24/Feb/23 15:51
• Resolved: 24/Feb/23 15:51

Description

s3a magic committer isRecoverySupported() is false, so will restart all task after application master restart for am jvm crashed, leaving pendingset in magic path not to clear. pendingset name format is jobAttemptPath + taskAttemptID.getTaskID() + ".pendingset", and jobAttemptPath is actually jobIdPath not JobAttemptIdPath in s3a magic committer. These pendingset files are overwritted by new task commit.

But if in new am attempt, a speculative task overcomes origin task, so pendingset file in last attempt may be hold for job commit, the data for commit is wrong

Issue Links

• relates to: HADOOP-17833 Improve Magic Committer Performance