Details
Blocker Description
Periodically, reduce tasks hang. When the log for the task is consulted, you see a stacktrace that looks like this:
2007-10-18 17:02:04,227 WARN org.apache.hadoop.mapred.ReduceTask: java.io.IOException: Insufficient space
at org.apache.hadoop.fs.InMemoryFileSystem$RawInMemoryFileSystem$InMemoryOutputStream.write(InMemoryFileSystem.java:174)
at org.apache.hadoop.fs.FSDataOutputStream$PositionCache.write(FSDataOutputStream.java:39)
at java.io.DataOutputStream.write(DataOutputStream.java:90)
at java.io.FilterOutputStream.write(FilterOutputStream.java:80)
at org.apache.hadoop.fs.ChecksumFileSystem$ChecksumFSOutputSummer.writeChunk(ChecksumFileSystem.java:326)
at org.apache.hadoop.fs.FSOutputSummer.writeChecksumChunk(FSOutputSummer.java:140)
at org.apache.hadoop.fs.FSOutputSummer.flushBuffer(FSOutputSummer.java:122)
at org.apache.hadoop.fs.ChecksumFileSystem$ChecksumFSOutputSummer.close(ChecksumFileSystem.java:310)
at org.apache.hadoop.fs.FSDataOutputStream$PositionCache.close(FSDataOutputStream.java:49)
at org.apache.hadoop.fs.FSDataOutputStream.close(FSDataOutputStream.java:64)
at org.apache.hadoop.mapred.MapOutputLocation.getFile(MapOutputLocation.java:253)
at org.apache.hadoop.mapred.ReduceTask$ReduceCopier$MapOutputCopier.copyOutput(ReduceTask.java:685)
at org.apache.hadoop.mapred.ReduceTask$ReduceCopier$MapOutputCopier.run(ReduceTask.java:637)
The problem stems from a miscalculation of the checksum file created in the InMemoryFileSystem associated with the data being copied from a completed mapper task to the reducer task.
The method used for calculating checksum file size is the following (ChecksumFileSystem:318):
((long)(Math.ceil((float)size/bytesPerSum)) + 1) * 4 + CHECKSUM_VERSION.length;
The issue here is the cast to float. Floating point numbers have only 24 bits of precision, thus will return short values on any size over 0x1000000. The fix is to replace this calculation with something that doesn't cast to float.
(((size+1)/bytesPerSum) + 2) * 4 + CHECKSUM_VERSION.length