[jira] Commented: (HADOOP-1087) Reducer hangs pulling from incorrect file.out.index path. (when one of the mapred.local.dir is not accessible but becomes available later at reduce time)

["Devaraj Das (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HADOOP-1087?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12488364 ] 

Devaraj Das commented on HADOOP-1087:
-------------------------------------

This is strange actually. If the file.out.index does not exist in the directory, then we should get an IOException and the handler for IOException in the doGet method sends a SC_GONE status in the HTTP response. This should be seen as 410 on the client and not 500 (in the exception reported by the client: 'Server returned HTTP response code: 500'). '500' means that there was an internal error in the servlet which could have been caused due to some other problem like a NPE (could have been due to something like HADOOP-1123). Another thing pointing to this direction is that the map output will be declared as 'lost' in the same exception handler code in the doGet method, and the JobTracker will reexecute the map. So the job should not hang. 

> Reducer hangs pulling from incorrect file.out.index path. (when one of the mapred.local.dir is not accessible but becomes available later at reduce time)
> ---------------------------------------------------------------------------------------------------------------------------------------------------------
>
>                 Key: HADOOP-1087
>                 URL: https://issues.apache.org/jira/browse/HADOOP-1087
>             Project: Hadoop
>          Issue Type: Bug
>          Components: mapred
>    Affects Versions: 0.10.1
>            Reporter: Koji Noguchi
>
> 2007-03-07 23:14:23,431 WARN org.apache.hadoop.mapred.TaskRunner: java.io.IOException: Server returned HTTP response code: 500 for URL: http://____:____/mapOutput?map=task_7810_m_000897_0&reduce=397
>   at sun.net.www.protocol.http.HttpURLConnection.getInputStream(HttpURLConnection.java:1149)
>   at org.apache.hadoop.mapred.MapOutputLocation.getFile(MapOutputLocation.java:121)
>   at org.apache.hadoop.mapred.ReduceTaskRunner$MapOutputCopier.copyOutput(ReduceTaskRunner.java:236)
>   at org.apache.hadoop.mapred.ReduceTaskRunner$MapOutputCopier.run(ReduceTaskRunner.java:199)
> 2007-03-07 23:14:23,431 WARN org.apache.hadoop.mapred.TaskRunner: task_7810_r_000397_0 adding host ____.com to penalty box, next contact in 279 seconds
> This happened when one of the drives was full and not accessible at map time.
> and one mapper
>     public void mergeParts() throws IOException {
>       ...
>       Path finalIndexFile = mapOutputFile.getOutputIndexFile(getTaskId());
> failed on the first hash entry in mapred.local.dir and used the second entry
> Afterwards, first dir entry became available and when reducer tried to pull through,
>     public static class MapOutputServlet extends HttpServlet {
>       ...
>       Path indexFileName = conf.getLocalPath(mapId+"/file.out.index");
> it used the first entry.
> As a result, directory was empty and reducer kept on trying to pull from the incorrect path and hang.
> (wasn't sure if this is a duplicate of HADOOP-895 since it is not reproducible unless I get disk failure.)

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.