[HDFS-2525] Race between BlockPoolSliceScanner and append

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 0.23.0, 2.0.0-alpha
• Fix Version/s: 0.23.2
• Component/s: datanode
• Target Version/s: 0.23.0
• Hadoop Flags: Reviewed
• Assignee: Brandon Li
• Reporter: Todd Lipcon
• Created: 01/Nov/11 04:47
• Updated: 16/Mar/15 18:14
• Resolved: 15/Feb/12 18:56

Description

I wrote a test which runs append() in a loop on a single file with a single replica, appending 0~100 bytes each time. If this races with the BlockPoolSliceScanner, I observe the BlockPoolSliceScanner getting FNFE, then reporting the block as bad to the NN. This causes the writer thread to loop forever on completeFile() since it doesn't see a valid replica.

Issue Links

• is related to: HDFS-3100 failed to append data