[jira] [Commented] (TEZ-1029) Fetcher can fail to report input failed event upon connection error

["Hitesh Shah (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/TEZ-1029?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13963637#comment-13963637 ] 

Hitesh Shah commented on TEZ-1029:
----------------------------------

Looks good. 

Question on:

{code}
+    if ((reportReadErrorImmediately && readError) 
+        || connectError 
{code}

Should the above be "(reportReadErrorImmediately && (readError || connectError)) ? or do connect errors get reported immediately always? 

> Fetcher can fail to report input failed event upon connection error
> -------------------------------------------------------------------
>
>                 Key: TEZ-1029
>                 URL: https://issues.apache.org/jira/browse/TEZ-1029
>             Project: Apache Tez
>          Issue Type: Bug
>            Reporter: Bikas Saha
>            Assignee: Bikas Saha
>         Attachments: TEZ-1029.1.patch
>
>
> When there is a connection error then the fetcher fails all the maps but does not report input read error to the AM. So the AM does not get any information to rerun the source (if many fetchers report read errors). A connection failed also results in the shuffle-scheduler deciding that too many different map sources have failed and so it should fail itself. Multiple attempts of the reducer fail and the job fails. Looking at MR Fetcher code, a connection issue is also reported as a read error to the MR AM. Planning on making the same fix in the Tez Fetcher as was made in MAPREDUCE-4772.



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