[
https://issues.apache.org/jira/browse/HBASE-2471?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12859002#action_12859002
]
stack commented on HBASE-2471:
------------------------------
HLog files contain edits for all regions on a particular RegionServer. HLogs
get GC'd only after all edits in a file have been persisted. HLogs may stick
around a good while if any one of a RS's Regions is a laggard flushing. While
the HLog lives, some other Region X may have split and been removed, and then
the RS crashed. On replay, we'll trip over Region X's edits though Region X
had been removed.
> Splitting logs, we'll make an output file though the region no longer exists
> ----------------------------------------------------------------------------
>
> Key: HBASE-2471
> URL: https://issues.apache.org/jira/browse/HBASE-2471
> Project: Hadoop HBase
> Issue Type: Bug
> Reporter: stack
> Fix For: 0.20.5, 0.21.0
>
>
> The "human unit tester" (Kannan) last night wondered what happens splitting
> logs and we come across an edit whose region has since been removed. Taking
> a look, it looks like we'll create the output file and write the edits for
> the no-longer-extant region anyways. This will leave litter in the
> filesystem -- region split files that will never be used nor removed. This
> issue is about verifying that indeed this is whats happening (We do
> SequenceFile.createWriter with the overwrite flag set to true which tracing
> seems to mean create all intermediary directories -- to be verified) and if
> it indeed is happening, fixing split so unless the region dir exists, don't
> write out edits.. just drop them.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
