[
https://issues.apache.org/jira/browse/HADOOP-5473?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Amareshwari Sriramadasu updated HADOOP-5473:
--------------------------------------------
Attachment: patch-5473-2.txt
There is a small window at TT, when KillTaskAction and reportDone are processed
together, the state will be SUCCEEDED and still TT cleanups map outputs.
So, taking up option (2) looks like proper solution.
Patch marks the task KILLED, if TT reports task as SUCCEEDED, but there was
user request to kill the task.
> Race condition in command-line kill for a task
> ----------------------------------------------
>
> Key: HADOOP-5473
> URL: https://issues.apache.org/jira/browse/HADOOP-5473
> Project: Hadoop Core
> Issue Type: Bug
> Components: mapred
> Reporter: Amareshwari Sriramadasu
> Assignee: Amareshwari Sriramadasu
> Fix For: 0.19.2
>
> Attachments: patch-5473-1.txt, patch-5473-2.txt
>
>
> The race condition occurs in following sequence of events:
> 1. User issues a command-line kill for a RUNNING map-task. JT stores the task
> in tasksToKill mapping.
> 2. TT reports the task status as SUCCEEDED.
> 3. JT creates a TaskCompletionEvent as SUCCEEDED. Also sends a killTaskAction.
> 4. Reducers fail fetching the map output.
> 5. finally, the task would fail with Fetch failures. After HADOOP-4759, the
> task is left as FAILED_UNCLEAN task, since the task is present in tasksToKill
> mapping.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
