hadoop-mapreduce-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Harsh J (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (MAPREDUCE-6238) MR2 can't run local jobs with -libjars command options which is a regression from MR1
Date Sat, 31 Jan 2015 18:19:34 GMT

    [ https://issues.apache.org/jira/browse/MAPREDUCE-6238?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14299919#comment-14299919
] 

Harsh J commented on MAPREDUCE-6238:
------------------------------------

Many thanks for working on this - it is a rather often observed issue. In following your analysis,
the fix looks good to me (essentially we reset the paths after the remote upload to be of
the correct lookup FS).

Would it be possible to have a test case also track this for preventing future regression
on this? It is surprising that the default tests that do run local job runners have never
caught this before - but thats probably cause they did not make use of HDFS as the default
FS.

> MR2 can't run local jobs with -libjars command options which is a regression from MR1
> -------------------------------------------------------------------------------------
>
>                 Key: MAPREDUCE-6238
>                 URL: https://issues.apache.org/jira/browse/MAPREDUCE-6238
>             Project: Hadoop Map/Reduce
>          Issue Type: Bug
>          Components: mrv2
>            Reporter: zhihai xu
>            Assignee: zhihai xu
>            Priority: Critical
>         Attachments: MAPREDUCE-6238.000.patch
>
>
> MR2 can't run local jobs with -libjars command options which is a regression from MR1.

> When run MR2 job with -jt local and -libjars, the job fails with java.io.FileNotFoundException:
File does not exist: hdfs://XXXXXXXXXXXXXXX.jar.
> But the same command is working in MR1.
> I find the problem is because when MR2 run local job using  LocalJobRunner
> from JobSubmitter, the JobSubmitter#jtFs is local filesystem,
> So copyRemoteFiles will return from [the middle of the function|https://github.com/apache/hadoop/blob/trunk/hadoop-mapreduce-project/hadoop-mapreduce-client/hadoop-mapreduce-client-core/src/main/java/org/apache/hadoop/mapreduce/JobSubmitter.java#L138]
> because source and destination file system are same.
> {code}
>     if (compareFs(remoteFs, jtFs)) {
>       return originalPath;
>     }
> {code}
> The following code at [JobSubmitter.java|https://github.com/apache/hadoop/blob/trunk/hadoop-mapreduce-project/hadoop-mapreduce-client/hadoop-mapreduce-client-core/src/main/java/org/apache/hadoop/mapreduce/JobSubmitter.java#L219]
> try to add the destination file to DistributedCache which introduce a bug for local job.
> {code}
>         Path newPath = copyRemoteFiles(libjarsDir, tmp, conf, replication);
>         DistributedCache.addFileToClassPath(
>             new Path(newPath.toUri().getPath()), conf);
> {code}
> Because new Path(newPath.toUri().getPath()) will lose the filesystem information from
newPath, the file added to DistributedCache will use the default Uri filesystem hdfs based
on the following code. This causes the 
>  FileNotFoundException when we access the file later at 
>  [determineTimestampsAndCacheVisibilities|https://github.com/apache/hadoop/blob/trunk/hadoop-mapreduce-project/hadoop-mapreduce-client/hadoop-mapreduce-client-core/src/main/java/org/apache/hadoop/mapreduce/JobSubmitter.java#L270]
> {code}
>   public static void addFileToClassPath(Path file, Configuration conf)
>     throws IOException {
> 	  addFileToClassPath(file, conf, file.getFileSystem(conf));
>   }
>   public static void addFileToClassPath
>            (Path file, Configuration conf, FileSystem fs)
>         throws IOException {
>     String classpath = conf.get(MRJobConfig.CLASSPATH_FILES);
>     conf.set(MRJobConfig.CLASSPATH_FILES, classpath == null ? file.toString()
>              : classpath + "," + file.toString());
>     URI uri = fs.makeQualified(file).toUri();
>     addCacheFile(uri, conf);
>   }
> {code}
> Compare to the following [MR1 code|https://github.com/apache/hadoop/blob/branch-1/src/mapred/org/apache/hadoop/mapred/JobClient.java#L811]:
> {code}
>         Path newPath = copyRemoteFiles(fs, libjarsDir, tmp, job, replication);
>         DistributedCache.addFileToClassPath(
>           new Path(newPath.toUri().getPath()), job, fs);
> {code}
> You will see why MR1 doesn't have this issue.
> because it passes the local filesystem into  DistributedCache#addFileToClassPath instead
of using the default Uri filesystem hdfs.
> We should do the same in MR2.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message