[
https://issues.apache.org/jira/browse/HBASE-8653?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Sergey Shelukhin updated HBASE-8653:
------------------------------------
Description:
Putting it in .1, feel free to move to .2.
We have observed some compaction errors where the code was creating a new HDFS
block, and the file would not exist. Upon investigation, we found the .tmp
directory delete request on namenode from master IP shortly before that. There
are no specific logs on master, but one thing running at that time was
CatalogJanitor. CatalogJanitor calls
HRegionFileSystem::openRegionFromFileSystem with readOnly == true (in fact,
everyone does); if readOnly is true, HRegionFileSystem nukes the .tmp directory.
We didn't go thru details on how it arrived there (or if there may have been
other culprit), but it appears that deleting stuff if (readOnly) is not the
intended behavior and it should be if (!readOnly). Given that readOnly is not
really used (or rather is always true except some inconsequential usage in
test) perhaps entire cleanup should be removed.
was:
Putting it in .1, feel free to move to .2.
We have observed some compaction errors where creating HDFS block, and the file
would not exist. Upon investigation, we found the .tmp directory delete request
on namenode from master IP. There are no specific logs on master, but one thing
running at that time was CatalogJanitor. CatalogJanitor calls
HRegionFileSystem::openRegionFromFileSystem with readOnly == true (in fact,
everyone does); if readOnly is true, HRegionFileSystem nukes the .tmp directory.
We didn't go thru details on how it arrived there (or if there may have been
other culprit), but it appears that deleting stuff if (readOnly) is not the
intended behavior and it should be if (!readOnly). Given that readOnly is not
really used (or rather is always true except some inconsequential usage in
test) perhaps entire cleanup should be removed.
> master seems to be deleting region tmp directory from under compaction
> ----------------------------------------------------------------------
>
> Key: HBASE-8653
> URL: https://issues.apache.org/jira/browse/HBASE-8653
> Project: HBase
> Issue Type: Bug
> Reporter: Sergey Shelukhin
> Assignee: Sergey Shelukhin
> Priority: Blocker
> Fix For: 0.95.1
>
>
> Putting it in .1, feel free to move to .2.
> We have observed some compaction errors where the code was creating a new
> HDFS block, and the file would not exist. Upon investigation, we found the
> .tmp directory delete request on namenode from master IP shortly before that.
> There are no specific logs on master, but one thing running at that time was
> CatalogJanitor. CatalogJanitor calls
> HRegionFileSystem::openRegionFromFileSystem with readOnly == true (in fact,
> everyone does); if readOnly is true, HRegionFileSystem nukes the .tmp
> directory.
> We didn't go thru details on how it arrived there (or if there may have been
> other culprit), but it appears that deleting stuff if (readOnly) is not the
> intended behavior and it should be if (!readOnly). Given that readOnly is not
> really used (or rather is always true except some inconsequential usage in
> test) perhaps entire cleanup should be removed.
--
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
