[jira] [Commented] (MAPREDUCE-489) Force the task tracker to exit when the task is complete, prevents nodes from dying due to resource starvation from impropertly written map/reduce tasks

["Allen Wittenauer (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/MAPREDUCE-489?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14065462#comment-14065462 ] 

Allen Wittenauer commented on MAPREDUCE-489:
--------------------------------------------

Doing a quick pass through the source seems to indicate that we don't always call system.exit on the way out.  We probably should.

> Force the task tracker to exit when the task is complete, prevents nodes from dying due to resource starvation from impropertly written map/reduce tasks
> --------------------------------------------------------------------------------------------------------------------------------------------------------
>
>                 Key: MAPREDUCE-489
>                 URL: https://issues.apache.org/jira/browse/MAPREDUCE-489
>             Project: Hadoop Map/Reduce
>          Issue Type: Improvement
>            Reporter: Jason
>            Priority: Minor
>              Labels: newbie
>   Original Estimate: 1h
>  Remaining Estimate: 1h
>
> We have map/reduce jobs that sometimes run additional threads that are not at daemon priority, and these threads prevent the Task from properly exiting. When enough of these accumulate, the node falls over.
> The included patch forces the Tasks to exit when completed.
> Index: src/java/org/apache/hadoop/mapred/TaskTracker.java
> ===================================================================
> --- src/java/org/apache/hadoop/mapred/TaskTracker.java  (revision 608611)
> +++ src/java/org/apache/hadoop/mapred/TaskTracker.java  (working copy)
> @@ -1801,6 +1801,8 @@
>          // This assumes that on return from Task.run()
>          // there is no more logging done.
>          LogManager.shutdown();
> +
> +        System.exit(0);
>        }
>      }
>    } 



--
This message was sent by Atlassian JIRA
(v6.2#6252)