[HBASE-17206] FSHLog may roll a new writer successfully with unflushed entries

• Type: Bug
• Status: Closed
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 1.4.0, 1.1.7, 1.2.4, 2.0.0
• Fix Version/s: 1.4.0, 1.2.5, 1.1.8, 2.0.0
• Component/s: wal
• Hadoop Flags: Reviewed
• Assignee: Duo Zhang Assign to me
• Reporter: Duo Zhang
• Created: 30/Nov/16 13:35
• Updated: 06/Apr/17 00:02
• Resolved: 02/Dec/16 01:22

Description

Found it when debugging the flakey TestFailedAppendAndSync.

The problem is in waitSafePoint.

      while (true) {
        if (this.safePointAttainedLatch.await(1, TimeUnit.MILLISECONDS)) {
          break;
        }
        if (syncFuture.isThrowable()) {
          throw new FailedSyncBeforeLogCloseException(syncFuture.getThrowable());
        }
      }
      return syncFuture;

If we attach the safe point quick enough then we will bypass the syncFuture.isThrowable check and will not throw FailedSyncBeforeLogCloseException.

This may cause incosistency between memstore and wal.