[jira] [Updated] (MAPREDUCE-5060) Fetch failures that time out only count against the first map task

["Robert Joseph Evans (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/MAPREDUCE-5060?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Robert Joseph Evans updated MAPREDUCE-5060:
-------------------------------------------

    Attachment: MR-5060.txt

This is a preliminary patch that removes the extra checks for connection established while setting up the connection to the NM.  Only when the reducer actually starts reading map data does it associate a failure with a particular task.
                
> Fetch failures that time out only count against the first map task
> ------------------------------------------------------------------
>
>                 Key: MAPREDUCE-5060
>                 URL: https://issues.apache.org/jira/browse/MAPREDUCE-5060
>             Project: Hadoop Map/Reduce
>          Issue Type: Bug
>            Reporter: Robert Joseph Evans
>            Assignee: Robert Joseph Evans
>            Priority: Critical
>         Attachments: MR-5060.txt
>
>
> When a fetch failure happens, if the socket has already "connected" it is only counted against the first map task.  But most of the time it is because of an issue with the Node itself, not the individual map task, and as such all failures when trying to initiate the connection should count against all of the tasks.
> This caused a particularly unfortunate job to take an hour an a half longer then it needed to.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira