[
https://issues.apache.org/jira/browse/HDFS-8690?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14608866#comment-14608866
]
Colin Patrick McCabe commented on HDFS-8690:
--------------------------------------------
I do think it's odd to abort on {{SocketTimeout}} but not on {{IOException}}.
After all, a client partitioned from the network may experience other
exceptions than {{SocketTimeout}} that have the same practical effect.
In general, the philosophy behind aborting is that the open-for-write files may
be re-assigned to other lease holders at any point once we can no longer renew.
This is why {{DFSClient#abort}} includes a call to
{{DFSClient#closeAllFilesBeingWritten}}. There is probably scope for a larger
cleanup of this code, but it would be difficult and take a very in-depth
review...
> LeaseRenewer should not abort DFSClient when renew fails
> --------------------------------------------------------
>
> Key: HDFS-8690
> URL: https://issues.apache.org/jira/browse/HDFS-8690
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: hdfs-client
> Reporter: Chang Li
> Assignee: Chang Li
> Attachments: HDFS-8690.1.patch
>
>
> The lease renewer special cases SocketTimeoutExceptions to abort the
> DFSClient. Aborting causes the client to be permanently unusable, which
> causes filesystem instances to stop working. All other IOExceptions do not
> abort. The special case should be removed and/or abort should not completely
> shutdown the proxy.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
