[HADOOP-7611] SequenceFile.Sorter creates local temp files on HDFS

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 0.20.2
• Component/s: io

Environment

CentOS 5.6 64-bit, Oracle JDK 1.6.0_26 64-bit

• Assignee: Unassigned
• Reporter: Bryan Keller
• Created: 05/Sep/11 17:44
• Updated: 15/Jan/15 03:58

Description

When using SequenceFile.Sorter to sort or merge sequence files that exist in HDFS, it attempts to create temp files in a directory structure specified by mapred.local.dir but on HDFS, not in the local file system. The problem code is in MergeQueue.merge(). Starting at line 2953:

            Path outputFile =  lDirAlloc.getLocalPathForWrite(
                                                tmpFilename.toString(),
                                                approxOutputSize, conf);
            LOG.debug("writing intermediate results to " + outputFile);
            Writer writer = cloneFileAttributes(
                                                fs.makeQualified(segmentsToMerge.get(0).segmentPathName), 
                                                fs.makeQualified(outputFile), null);

The outputFile here is a local path without a scheme, e.g. "/mnt/mnt1/mapred/local", specified by the mapred.local.dir property. If we are sorting files on HDFS, the fs object is a DistributedFileSystem. The call to fs.makeQualified(outputFile) appends the fs object's scheme to the local temp path returned by lDirAlloc, e.g. hdfs://mnt/mnt1/mapred/local. This directory is then created (if the proper permissions are available) on HDFS. If the HDFS permissions are not available, the sort/merge fails even though the directories exist locally.

The code should instead always use the local file system if retrieving a path from the mapred.local.dir property. The unit tests do not test this condition, they only test using the local file system for sort and merge.