hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Sergey Shelukhin (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-3643) Close the filesystem handle when HRS is aborting
Date Tue, 16 Apr 2013 20:29:16 GMT

    [ https://issues.apache.org/jira/browse/HBASE-3643?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13633304#comment-13633304

Sergey Shelukhin commented on HBASE-3643:

WAL solution should work (see last comments in HBASE-2231) unless I'm missing something.
If we wait for WAL IO fencing properly, than old server cannot write to WAL after new server
sees it.
Which means if old server completes the compaction before fencing, it will indicate it in
WAL and new server will not read/will attempt to delete the old files.
If it finishes compaction after fencing it will fail to write to WAL.
So ZK doesn't need to be involved.
> Close the filesystem handle when HRS is aborting
> ------------------------------------------------
>                 Key: HBASE-3643
>                 URL: https://issues.apache.org/jira/browse/HBASE-3643
>             Project: HBase
>          Issue Type: Improvement
>    Affects Versions: 0.90.1
>            Reporter: Jean-Daniel Cryans
>            Priority: Critical
>             Fix For: 0.95.1
> I thought of a way to fix HBASE-3515 that has a very broad impact, so I'm creating this
jira to *raise awareness* and gather comments.
> Currently when we call HRS.abort, it's still possible to do HDFS operations like rolling
logs and flushing files. It also has the impact that some threads cannot write to ZK (like
the situation described in HBASE-3515) but then can still write to HDFS. Since that call is
so central, I think we should {color:red} add fs.close() inside the abort method{color}.
> The impact of this is that everything else that happens after the close call, like closing
files or appending, will fail in the most horrible ways. On the bright side, this means less
disruptive changes on HDFS.
> Todd pointed at HBASE-2231 as related, but I think my solution is still too sloppy as
we could still finish a compaction and immediately close the filesystem after that (damage's

This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira

View raw message