Save namespace can cause NN to be unable to come up on restart
--------------------------------------------------------------
Key: HDFS-1921
URL: https://issues.apache.org/jira/browse/HDFS-1921
Project: Hadoop HDFS
Issue Type: Bug
Affects Versions: 0.22.0, 0.23.0
Reporter: Aaron T. Myers
Priority: Critical
Fix For: 0.22.0, 0.23.0
I discovered this in the course of trying to implement a fix for HDFS-1505.
Per the comment for {{FSImage.saveNamespace(...)}}, the algorithm for save
namespace proceeds in the following order:
# rename current to lastcheckpoint.tmp for all of them,
# save image and recreate edits for all of them,
# rename lastcheckpoint.tmp to previous.checkpoint.
The problem is that step 3 occurs regardless of whether or not an error occurs
for all storage directories in step 2. Upon restart, the NN will see
non-existent or corrupt {{current}} directories, and no {{lastcheckpoint.tmp}}
directories, and so will conclude that the storage directories are not
formatted.
This issue appears to be present on both 0.22 and 0.23. This should arguably be
a 0.22/0.23 blocker.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
