[
https://issues.apache.org/jira/browse/HDFS-5988?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Andrew Wang updated HDFS-5988:
------------------------------
Attachment: hdfs-5988-1.patch
{{FSImageFormat#Loader}} was incorrectly basing the decision to populate the
{{FSDirectory#inodeMap}} based on if the old fsimage layout version supported
inodes. We only see the error with the new PB-based image, since it iterates
through {{inodeMap}}, while the old fsimage saver would traverse the directory
structure instead.
I also added a bunch of trace/debug logging to OIV, which was helpful in
tracking down this issue. Trust me, lot of effort for a one-line fix :)
> Bad fsimage generated after upgrade
> -----------------------------------
>
> Key: HDFS-5988
> URL: https://issues.apache.org/jira/browse/HDFS-5988
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: namenode
> Affects Versions: 2.4.0
> Reporter: Andrew Wang
> Assignee: Andrew Wang
> Priority: Blocker
> Attachments: hdfs-5988-1.patch
>
>
> Internal testing revealed an issue where, after upgrading from an earlier
> release, we always fail to save a correct PB-based fsimage (namely, missing
> inodes leading to an inconsistent namespace). This results in substantial
> data loss, since the upgraded fsimage is broken, as well as the fsimages
> generated by saveNamespace and checkpointing.
> This ended up being a bug in the old fsimage loading code, patch coming.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
