[
https://issues.apache.org/jira/browse/HDFS-7646?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14284708#comment-14284708
]
Tsz Wo Nicholas Sze commented on HDFS-7646:
-------------------------------------------
Do we have copy-on-truncate? We discussed this earlier; see [~shv]'s
[commnet|https://issues.apache.org/jira/browse/HDFS-3107?focusedCommentId=14189216&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-14189216].
> HDFS truncate may remove data from the supposedly read-only previous/
> directory during an upgrade
> -------------------------------------------------------------------------------------------------
>
> Key: HDFS-7646
> URL: https://issues.apache.org/jira/browse/HDFS-7646
> Project: Hadoop HDFS
> Issue Type: Bug
> Reporter: Colin Patrick McCabe
> Priority: Blocker
>
> During a DataNode layout version upgrade, HDFS creates hardlinks. These
> hardlinks allow the same block to be accessible from both the current/ and
> previous/ directories. Rollback is possible by deleting the current/
> directory and renaming previous/ to current/.
> However, if the user truncates one of these hardlinked block files, it
> effectively eliminates the ability to roll back to the previous data.
> We probably need to disable truncation-in-place during a DataNode upgrade.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
