[
https://issues.apache.org/jira/browse/HDFS-2078?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Todd Lipcon updated HDFS-2078:
------------------------------
Attachment: hdfs-2078.txt
Turns out the test broke because it was just changing permissions on currentDir
instead of the rootDir. Since we were no longer calling clearDirectory, it
wasn't triggering a permissions error, and after the saveNamespace it would
succeed in restoring the directory even though the namespace save actually
failed.
Updated patch has TestSaveNamespace set the directory a-w and modify the root
instead of current/ dir to simulate the failure.
> 1073: NN should not clear storage directory when restoring removed storage
> --------------------------------------------------------------------------
>
> Key: HDFS-2078
> URL: https://issues.apache.org/jira/browse/HDFS-2078
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Components: name-node
> Affects Versions: Edit log branch (HDFS-1073)
> Reporter: Todd Lipcon
> Assignee: Todd Lipcon
> Fix For: Edit log branch (HDFS-1073)
>
> Attachments: hdfs-2078.txt, hdfs-2078.txt
>
>
> This JIRA is to address the following scenario/bug:
> - The NN is configured with an edits-only storage dir in /edits and an
> image-only storage dir in /image
> - The image dir fails while it is running. Since the edits dir is still valid
> it does not immediately shut itself down. 2NN continues to try to checkpoint,
> but fails because it can't upload an image anywhere
> - Operator fixes the disk on /image and instructs the NN to restore removed
> storage
> - The 2NN should now be able to download/upload a checkpoint successfully.
> Currently this does not work since the NN clears the storage dir upon
> restoring it. With the 1073 design, out-of-date files aren't a problem, and
> in fact can be used to restore the namespace.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
