[
https://issues.apache.org/jira/browse/HADOOP-5394?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12682658#action_12682658
]
Devaraj Das commented on HADOOP-5394:
-------------------------------------
Some comments:
1) The file jobtracker.info (the restart count file) must always exist in the
system directory, and the value in the file should be 0 (indicating the JT has
started fresh). RecoveryManager.getRestartCount should be changed accordingly.
The update to the info file should be like:
{code}
if (infoFile.exists()) {
delete (infoFile.recovery);
} else {
rename (infoFile.recovery, infoFile);
}
count = readInfoFile();
write (count + 1) to infoFile.recover;
delete infoFile;
rename (infoFile.recover infoFile);
{code}
2) Add checks for info file in the testcase.
3) The restart count need not be logged in the JobHistory file.
> JobTracker might schedule 2 attempts of the same task with the same attempt
> id across restarts
> ----------------------------------------------------------------------------------------------
>
> Key: HADOOP-5394
> URL: https://issues.apache.org/jira/browse/HADOOP-5394
> Project: Hadoop Core
> Issue Type: Bug
> Components: mapred
> Reporter: Amar Kamat
> Assignee: Amar Kamat
> Priority: Critical
> Attachments: HADOOP-5394-v1.2.patch
>
>
> This can happen when the jobtracker gets restarted more than once. In such
> cases, the jobtracker depends on the jobhistory file for the next restart
> count. If the new restart-count is not flushed to the file then there is a
> fair chance that upon next restart, the jobtracker might schedule a new
> attempt with an existing id. This can cause problems not only with the
> side-effect files but also can cause the jobtracker to be in an inconsistent
> state.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
