[ 
https://issues.apache.org/jira/browse/HBASE-6050?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13279029#comment-13279029
 ] 

stack commented on HBASE-6050:
------------------------------

Good one Ram.

So, we are talking about the parent region?

It does seem wrong that we would recreate a parent region dir in the 
distributed log splitter.

How about we remove that dir creation code?  I can see our making the 
recovered.edits dir because it won't always be there but creating all of its 
parent dirs is not right.  My guess is that the mkdirs was done because it was 
just easier than verifying parent dir present.

If parent dir not present, log the fact that there is no target region into 
which to put the edits and move on I'd say.
                
> HLogSplitter renaming recovered.edits and CJ removing the parent directory 
> races, making the HBCK to think cluster is inconsistent.
> -----------------------------------------------------------------------------------------------------------------------------------
>
>                 Key: HBASE-6050
>                 URL: https://issues.apache.org/jira/browse/HBASE-6050
>             Project: HBase
>          Issue Type: Bug
>            Reporter: ramkrishna.s.vasudevan
>
> The scenario is like this
> -> A region is getting splitted.
> -> The master is still not processed the split .
> -> Region server goes down.
> -> Split log manager starts splitting the logs and creates the 
> recovered.edits in the splitlog path.
> -> CJ starts and deletes the entry from META and also just completes the 
> deletion of the region dir.
> -> in hlogSplitter on final step we rename the recovered.edits to come under 
> the regiondir.
> There if the regiondir doesnot exist we tend to create and then add the 
> recovered.edits.
> Because of this HBCK thinks it to be an orphan region because we have the 
> regiondir but with no regioninfo.
> Ideally cluster is fine but we it is misleading.
> {code}
>         } else {
>           Path dstdir = dst.getParent();
>           if (!fs.exists(dstdir)) {
>             if (!fs.mkdirs(dstdir)) LOG.warn("mkdir failed on " + dstdir);
>           }
>         }
>         fs.rename(src, dst);
>         LOG.debug(" moved " + src + " => " + dst);
>       } else {
>         LOG.debug("Could not move recovered edits from " + src +
>             " as it doesn't exist");
>       }
>     }
>     archiveLogs(null, corruptedLogs, processedLogs,
>         oldLogDir, fs, conf);
> {code}

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: 
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira

        

Reply via email to