[jira] Commented: (HADOOP-654) jobs fail with some hardware/system failures on a small number of nodes

["Owen O'Malley (JIRA)" <ji...@apache.org>]

    [ http://issues.apache.org/jira/browse/HADOOP-654?page=comments#action_12454502 ] 
            
Owen O'Malley commented on HADOOP-654:
--------------------------------------

The number of failures on each task tracker is already on the web/ui.

I propose a very simple mechanism where each job can configure the maximum number of task failures that it will tolerate from a single node.

JobConf gets:
   void setMaxTaskFailuresPerHost(int newValue);
   int getMaxTaskFailuresPerHost();

JobInProgress keeps a Map from task tracker id to counts of failures. Each task failure increments the count for the responsible tracker. (Should lost task trackers count as 1 failures or lots of failures?)

When the tracker is asking for new tasks, if it has failed too many times, it is not given a new task. Note that these counts (and penalty boxes) are per a job.

Does that sound like a reasonable starting point?

> jobs fail with some hardware/system failures on a small number of nodes
> -----------------------------------------------------------------------
>
>                 Key: HADOOP-654
>                 URL: http://issues.apache.org/jira/browse/HADOOP-654
>             Project: Hadoop
>          Issue Type: Bug
>          Components: mapred
>    Affects Versions: 0.7.2
>            Reporter: Yoram Arnon
>         Assigned To: Owen O'Malley
>            Priority: Minor
>
> occasionally, such as when the OS is out of some resource, a node fails only partly. The node is up and running, the task tracker is running and sending heartbeats, but every task fails because the tasktracker can't fork tasks or something.
> In these cases, that task tracker keeps getting assigned tasks to execute, and they all fail.
> A couple of nodes like that and jobs start failing badly.
> The job tracker should avoid assigning tasks to tasktrackers that are misbehaving.
> simple approach: avoid tasktrackers that report many more failures than average (say 3X). Simply use the info sent by the TT.
> better but harder: track TT failures over time and:
>  1. avoid those that exhibit a high failure *rate*
>  2. tell them to shut down

-- 
This message is automatically generated by JIRA.
-
If you think it was sent incorrectly contact one of the administrators: http://issues.apache.org/jira/secure/Administrators.jspa
-
For more information on JIRA, see: http://www.atlassian.com/software/jira