[
https://issues.apache.org/jira/browse/HADOOP-15183?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16825555#comment-16825555
]
Steve Loughran commented on HADOOP-15183:
-----------------------------------------
Turns out the current code stores all files being renamed anyway, as the list
of files to move is built up incrementally but only pushed to S3Guard at the
end, irrespective of store size. This is a strategy doomed to fail at precisely
the wrong point in a production application.
Also, in some tests, the time to update the DDB table is the bottleneck on
deletion performance.
We need to move to incremental DDB updates along with the deletes. Maybe, even
with partial delete enabled, we should have each thread go:
COPY->Update->DELETE->Update. It'd be moving to one DELETE per file renamed,
but except for small files, that won't be the big expense
> S3Guard store becomes inconsistent after partial failure of rename
> ------------------------------------------------------------------
>
> Key: HADOOP-15183
> URL: https://issues.apache.org/jira/browse/HADOOP-15183
> Project: Hadoop Common
> Issue Type: Sub-task
> Components: fs/s3
> Affects Versions: 3.0.0
> Reporter: Steve Loughran
> Assignee: Steve Loughran
> Priority: Blocker
> Attachments: HADOOP-15183-001.patch, HADOOP-15183-002.patch,
> org.apache.hadoop.fs.s3a.auth.ITestAssumeRole-output.txt
>
>
> If an S3A rename() operation fails partway through, such as when the user
> doesn't have permissions to delete the source files after copying to the
> destination, then the s3guard view of the world ends up inconsistent. In
> particular the sequence
> (assuming src/file* is a list of files file1...file10 and read only to
> caller)
>
> # create file rename src/file1 dest/ ; expect AccessDeniedException in the
> delete, dest/file1 will exist
> # delete file dest/file1
> # rename src/file* dest/ ; expect failure
> # list dest; you will not see dest/file1
> You will not see file1 in the listing, presumably because it will have a
> tombstone marker and the update at the end of the rename() didn't take place:
> the old data is still there.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
