[jira] [Created] (ACCUMULO-2172) Concurrent compactions before machine failure may cause uneeded recovery

["Keith Turner (JIRA)" <ji...@apache.org>]

Keith Turner created ACCUMULO-2172:
--------------------------------------

             Summary: Concurrent compactions before machine failure may cause uneeded recovery
                 Key: ACCUMULO-2172
                 URL: https://issues.apache.org/jira/browse/ACCUMULO-2172
             Project: Accumulo
          Issue Type: Bug
            Reporter: Keith Turner
             Fix For: 1.6.0


[~vines] noticed that minc finish events are not synced to the walog.  I was looking into this issue to see if it was a problem and I noticed a possible issue w/ this.   If the following sequence of events occurs, its possible that duplicate recovery may occur.

 # Tablet T1 writes File F1 to metadata table after minor compaction
 # T1 write minc finish event to walog (no sync is done and info not in walog)
 # T1 adds F1 to set of files in tablets memory
 # T1 major compacts F1 into F2
 # Tablet server serving T1 fails (and nothing synced walog)
 # T1 recovers data in F1 because there is no finish event or file in metadata table

Step 2 must sync before step 3, because step 3 makes the file eligible for major compaction.  Its seems like the sync could be done using group commit for efficiency.



--
This message was sent by Atlassian JIRA
(v6.1.5#6160)