[
https://issues.apache.org/jira/browse/HDFS-6197?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13962090#comment-13962090
]
Chris Nauroth commented on HDFS-6197:
-------------------------------------
Thanks for the reviews, Nicholas and Jing! I'll commit this after Jenkins runs
it.
bq. Just want to make sure I understand the issue correctly, so here multiple
attempts means multiple attempts of rollback?
This would be multiple attempts of beginning the rolling upgrade and rolling
back with no additional edits in between. Probably the most specific
explanation would be to look at the code of
{{TestRollingUpgrade#testRollback}}. This test failed on Windows during the
rename before my patch.
bq. I think we may also want to add some new mechanism to cleanup the trashed
edits for rollback of rolling upgrade (e.g., the first checkpoint after the
rollback maybe).
Yes, this seems to make sense too.
> Rolling upgrade rollback on Windows can fail attempting to rename edit log
> segment files to a destination that already exists.
> ------------------------------------------------------------------------------------------------------------------------------
>
> Key: HDFS-6197
> URL: https://issues.apache.org/jira/browse/HDFS-6197
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: namenode
> Affects Versions: 3.0.0, 2.4.0
> Reporter: Chris Nauroth
> Assignee: Chris Nauroth
> Priority: Minor
> Attachments: HDFS-6197.1.patch
>
>
> As part of a rollback from a rolling upgrade in progress, we discard edit log
> segments by renaming the file with suffix ".trash". If no new edits arrive
> in between multiple attempts, then the ".trash" file may still exist. On
> Windows, {{java.io.File#renameTo}} fails if the destination already exists
> though. This is visible right now as a failure in
> {{TestRollingUpgrade#testRollback}} when running on Windows.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
