hadoop-common-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Owen O'Malley (JIRA)" <j...@apache.org>
Subject [jira] Updated: (HADOOP-277) Race condition in Configuration.getLocalPath()
Date Tue, 06 Jun 2006 03:12:31 GMT
     [ http://issues.apache.org/jira/browse/HADOOP-277?page=all ]

Owen O'Malley updated HADOOP-277:
---------------------------------

    Attachment: mkdirs.patch

This patch is closer to what we did for the routine above it last week. (Sorry about not fixing
this one too at the same time. It wasn't biting us, but that was no reason not to fix the
obviously parallel code.)  Is there some reason that you need the synchronized block around
the mkdirs? File.mkdirs does a File.exists internally, so you don't need to call it yourself.

> Race condition in Configuration.getLocalPath()
> ----------------------------------------------
>
>          Key: HADOOP-277
>          URL: http://issues.apache.org/jira/browse/HADOOP-277
>      Project: Hadoop
>         Type: Bug

>  Environment: linux, 64 bit, dual core, 4x400GB disk, 4GB RAM
>     Reporter: paul sutter
>  Attachments: hadoop-277.patch, hadoop-task_1_r_9.log, mkdirs.patch
>
> (attached: a patch to fix the problem, and a logfile showing the problem occuring twice)
> There is a race condition in Configuration.java:
>        Path file = new Path(dirs[index], path);
>        Path dir = file.getParent();
>        if (fs.exists(dir) || fs.mkdirs(dir)) {
>          return file;
> If two threads simultaneously process this code with the same target directory, fs.exists()
will return false, but from fs.mkdirs() only one of the two threads will return true. From
the Java documentation:
>  "returns: true if and only if the directory was created, along with all necessary parent
directories; false otherwise"
> That is, if the first thread successfully creates the directory, the second will not,
and therefore return false, even though the directory exists.
> This was really happening. We use four temporary directories, and we had reducers failing
all over the place with  bizarre impossible errors. I modified the ReduceTaskRunner to output
the filename that it creates to find the problem, and the log output is below.
> Here you can see copies initiated for two files that hash to the same temp directory,
simultaneously. map_4.out is created in the correct directory (/data2...), but map_15.out
is created in the next directory (/data3...) becuase of this race condition. Minutes later,
when the appender tries to locate the file, that race condition does not occur (the directory
already exists), and the appender looks for the file map_15.out in the correct directory,
where it does not exist.
> 060605 142414 task_0001_r_000009_1 Copying task_0001_m_000004_0 output from rmr05.
> 060605 142414 task_0001_r_000009_1 Copying task_0001_m_000015_0 output from rmr04.
> ...
> 060605 142416 task_0001_r_000009_1 done copying task_0001_m_000004_0 output from rmr05
into /data2/tmp/mapred/local/task_0001_r_000009_1/map_4.out
> ...
> 060605 142418 task_0001_r_000009_1 done copying task_0001_m_000015_0 output from rmr04
into /data3/tmp/mapred/local/task_0001_r_000009_1/map_15.out
> ...
> 060605 142531 task_0001_r_000009_1 0.31808624% reduce > append > /data2/tmp/mapred/local/task_0001_r_000009_1/map_4.out
> ...
> 060605 142725 task_0001_r_000009_1 java.io.FileNotFoundException: /data2/tmp/mapred/local/task_0001_r_000009_1/map_15.out

-- 
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


Mime
View raw message