[jira] [Updated] (HDFS-7042) Upgrade fails for Windows HA cluster due to file locks held during rename in JournalNode.

["Chris Nauroth (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/HDFS-7042?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Chris Nauroth updated HDFS-7042:
--------------------------------
    Attachment: HDFS-7042.1.patch

The attached patch closes the last committed transaction ID file before pre-upgrade and before rollback.  There is also a small bug fix in {{BestEffortLongFile}} to null the {{FileChannel}} so that subsequent calls to {{lazyOpen}} will reopen it.

I haven't added any new tests.  The problem can be reproduced by running the existing {{TestDFSUpgradeWithHA}} suite on Windows.  The test suite passes after applying this patch.  I've also used the patch successfully in a live cluster.

> Upgrade fails for Windows HA cluster due to file locks held during rename in JournalNode.
> -----------------------------------------------------------------------------------------
>
>                 Key: HDFS-7042
>                 URL: https://issues.apache.org/jira/browse/HDFS-7042
>             Project: Hadoop HDFS
>          Issue Type: Bug
>          Components: journal-node
>    Affects Versions: 2.4.1
>            Reporter: Chris Nauroth
>            Assignee: Chris Nauroth
>            Priority: Blocker
>         Attachments: HDFS-7042.1.patch
>
>
> During upgrade of an HA cluster, the JournalNode attempts to rename the current storage directory to previous.  However, the process still holds the last committed transaction ID file opened while this is happening.  On Windows, this is a locking violation that causes the rename operation to fail.  Ultimately the whole upgrade fails.  There is a similar problem during rollback too.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)