[
https://issues.apache.org/jira/browse/HBASE-16788?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15562850#comment-15562850
]
Gary Helmling commented on HBASE-16788:
---------------------------------------
[~tedyu], adding the archiveLock does not change the semantics of the existing
read/write lock. Multiple concurrent readers can still obtain the read lock.
However, only one thread would be able to read the list of compacted files and
archive them under the archiveLock.
I also looked at your v2 patch. It seems problematic to modify the contents of
the compactedfiles list (by setting archiving flags on individual files) in
getCompactedfiles(), since this code path is only called with the readlock
where you don't typically expect mutations to happen. I don't think this
necessarily matters in this particular case, but I think it would introduce
confusing semantics going forward. It also adds a lot more complexity to the
code that the simple lock I introduced around archiving.
Do you have any objections to my patch?
> Race in compacted file deletion between HStore close() and
> closeAndArchiveCompactedFiles()
> ------------------------------------------------------------------------------------------
>
> Key: HBASE-16788
> URL: https://issues.apache.org/jira/browse/HBASE-16788
> Project: HBase
> Issue Type: Bug
> Components: regionserver
> Affects Versions: 1.3.0
> Reporter: Gary Helmling
> Assignee: Gary Helmling
> Priority: Blocker
> Attachments: 16788-suggest.v2, HBASE-16788.001.patch,
> HBASE-16788.002.patch, HBASE-16788_1.patch
>
>
> HBASE-13082 changed the way that compacted files are archived from being done
> inline on compaction completion to an async cleanup by the
> CompactedHFilesDischarger chore. It looks like the changes to HStore to
> support this introduced a race condition in the compacted HFile archiving.
> In the following sequence, we can wind up with two separate threads trying to
> archive the same HFiles, causing a regionserver abort:
> # compaction completes normally and the compacted files are added to
> {{compactedfiles}} in HStore's DefaultStoreFileManager
> # *threadA*: CompactedHFilesDischargeHandler runs in a RS executor service,
> calling closeAndArchiveCompactedFiles()
> ## obtains HStore readlock
> ## gets a copy of compactedfiles
> ## releases readlock
> # *threadB*: calls HStore.close() as part of region close
> ## obtains HStore writelock
> ## calls DefaultStoreFileManager.clearCompactedfiles(), getting a copy of
> same compactedfiles
> # *threadA*: calls HStore.removeCompactedfiles(compactedfiles)
> ## archives files in {compactedfiles} in HRegionFileSystem.removeStoreFiles()
> ## call HStore.clearCompactedFiles()
> ## waits on write lock
> # *threadB*: continues with close()
> ## calls removeCompactedfiles(compactedfiles)
> ## calls HRegionFIleSystem.removeStoreFiles() ->
> HFileArchiver.archiveStoreFiles()
> ## receives FileNotFoundException because the files have already been
> archived by threadA
> ## throws IOException
> # RS aborts
> I think the combination of fetching the compactedfiles list and removing the
> files needs to be covered by locking. Options I see are:
> * Modify HStore.closeAndArchiveCompactedFiles(): use writelock instead of
> readlock and move the call to removeCompactedfiles() inside the lock. This
> means the read operations will be blocked while the files are being archived,
> which is bad.
> * Synchronize closeAndArchiveCompactedFiles() and modify close() to call it
> instead of calling removeCompactedfiles() directly
> * Add a separate lock for compacted files removal and use in
> closeAndArchiveCompactedFiles() and close()
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
