[
https://issues.apache.org/jira/browse/HBASE-13651?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14536511#comment-14536511
]
Matteo Bertozzi commented on HBASE-13651:
-----------------------------------------
in the new code this is avoided by writing to the wal before removing files.
if you exit the GC pause and you try to write to the wal, you are not able to
do it because of recoverLease() called by the new RS owning the region.
there is still the possibility that you get in GC pause after writing to the
WAL and you do the remove, but in that case you already wrote that you are
going to remove those files and you should use the compacted one, so the new RS
is loading only the compacted one.
> Handle StoreFileScanner FileNotFoundException
> ---------------------------------------------
>
> Key: HBASE-13651
> URL: https://issues.apache.org/jira/browse/HBASE-13651
> Project: HBase
> Issue Type: Bug
> Affects Versions: 0.94.27, 0.98.10.1
> Reporter: Matteo Bertozzi
> Assignee: Matteo Bertozzi
> Priority: Minor
> Attachments: HBASE-13651-0.94-draft.patch, HBASE-13651-draft.patch
>
>
> Example:
> * Machine-1 is serving Region-X and start compaction
> * Machine-1 goes in GC pause
> * Region-X gets reassigned to Machine-2
> * Machine-1 exit from the GC pause
> * Machine-1 (re)moves the compacted files
> * Machine-1 get the lease and shutdown
> Machine-2 has now tons of FileNotFoundException on scan. If we reassign the
> region everything is ok, because we pickup the files compacted by Machine-1.
> This problem doesn't happen in the new code 1.0+ (i think but I haven't
> checked, it may be 1.1) where we write on the WAL the compaction event before
> (re)moving the files.
> A workaround is handling FileNotFoundException and refresh the store files,
> or shutdown the region and reassign. the first one is easy in 1.0+ the second
> one requires more work because at the moment we don't have the code to notify
> the master that the RS is closing the region, alternatively we can shutdown
> the entire RS (it is not a good solution but the case is rare enough)
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
