[KUDU-2081] LBM - possible data directory sync race

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 1.4.0
• Component/s: tablet
• Assignee: Unassigned
• Reporter: Dan Burkert
• Created: 27/Jul/17 19:26
• Updated: 27/Jul/17 19:26

Description

Inspecting the LBM code, it became apparent that there is a potential race when multiple threads are sync-ing the same dirty data directory in LogBlockManager::SyncContainer. The issue is that when two threads race to sync the same dirty directory, one of them will 'win', and start syncing the directory. The other will 'lose', and immediately return early, and assume the directory sync is complete (which may or may not be the case). After discussing with haohao and adar we think it's mostly benign do to the way the LBM is used, but still probably worth fixing.