Minor Description
When joining two PTables on a key that is a custom writable PType, the shuffler will fail with the following NullPointerException under Hadoop2 if the custom type has been registered using Writables.registerComparable. This happens regardless of whether a specific integer code is provided or the default hashCode()-based value is used.
org.apache.hadoop.mapred.YarnChild: Exception running child : org.apache.hadoop.mapreduce.task.reduce.Shuffle$ShuffleError: Error while doing final merge at org.apache.hadoop.mapreduce.task.reduce.Shuffle.run(Shuffle.java:160) at org.apache.hadoop.mapred.ReduceTask.run(ReduceTask.java:376) at org.apache.hadoop.mapred.YarnChild$2.run(YarnChild.java:168) at java.security.AccessController.doPrivileged(Native Method) at javax.security.auth.Subject.doAs(Subject.java:415) at org.apache.hadoop.security.UserGroupInformation.doAs(UserGroupInformation.java:1614) at org.apache.hadoop.mapred.YarnChild.main(YarnChild.java:163) Caused by: java.lang.NullPointerException at java.lang.Class.isAssignableFrom(Native Method) at org.apache.crunch.types.writable.TupleWritable$Comparator.compareField(TupleWritable.java:317) at org.apache.crunch.types.writable.TupleWritable$Comparator.compare(TupleWritable.java:284) at org.apache.hadoop.mapred.Merger$MergeQueue.lessThan(Merger.java:578) at org.apache.hadoop.util.PriorityQueue.upHeap(PriorityQueue.java:128) at org.apache.hadoop.util.PriorityQueue.put(PriorityQueue.java:55) at org.apache.hadoop.mapred.Merger$MergeQueue.merge(Merger.java:669) at org.apache.hadoop.mapred.Merger.merge(Merger.java:193) at org.apache.hadoop.mapreduce.task.reduce.MergeManagerImpl.finalMerge(MergeManagerImpl.java:804) at org.apache.hadoop.mapreduce.task.reduce.MergeManagerImpl.close(MergeManagerImpl.java:369) at org.apache.hadoop.mapreduce.task.reduce.Shuffle.run(Shuffle.java:158) ... 6 more
It appears that the Writables.WRITABLE_CODES entries are not deserialized from the configuration during the shuffle phase of a join until TupleWritable.setConf() is called. However, because TupleWritable.Comparator is registered as a raw comparator for TupleWritable, the shuffler uses the comparator without instantiating or configuring a TupleWritable instance. As a result, the type codes for the custom types are not available when the comparator starts to run.
HADOOP-10686 made WritableComparator implement Configurable, but this was not released until Hadoop 2.5. If I build Crunch against Hadoop 2.5 and copy TupleWritable's setConf() function to TupleWritable.Comparator, then the shuffle works as expected. However, since Crunch currently targets Hadoop 2.2, this does not work for the current version of Crunch.
As as a workaround, it appears that if the mapreduce.job.output.key.comparator.class property is set in the configuration, then the instance is created in JobConf.getOutputKeyComparator() using ReflectionUtils instead of using the WritableComparator registration. ReflectionUtils will pass the configuration to anything that implements Configurable, so setting mapreduce.job.output.key.comparator.class to TupleWritable.Comparator and implementing Configurable might work for Hadoop versions older than 2.5. I have yet to try this, though, and I have not looked into Hadoop1 to see if this would also work there.
If the shuffle is able to register the type codes via either method above, then there is one small secondary issue that I hit: Writables.registerComparable checks if the type code is already present in the map; if the type code is already in use, then it throws an exception, even if the class being registered is the same as the existing class. With the type codes being initialized during the shuffle phase, any later call to registerComparable for the same type code and class will fail. I currently have my registerComparable call in a static initialization block for my PType, so it is called whenever my writable type is first used under Crunch; in this case, it happens when the reduce phase starts. Checking to see if the class being registered and the existing class are equal inside of registerComparable before throwing an error, similar to the one that is in Guava's AbstractBiMap, prevents this exception from being thrown.
The above was happening using 0.11.0-hadoop2 on Hadoop 2.5.0 (CDH 5.2). The modifications I mention above were made on top of d4f23c4 and also tested on CDH 5.2.