[CRUNCH-539] Use of TupleWritable.setConf fails in mapper/reducer

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 0.12.0
• Fix Version/s: 0.13.0
• Assignee: Gabriel Reid
• Reporter: Gabriel Reid
• Created: 07/Jul/15 16:43
• Updated: 06/Aug/15 04:39
• Resolved: 09/Jul/15 18:32

Description

In (at least) more recent versions of Hadoop 2, the implicit call to TupleWritable.setConf that happens when using TupleWritables fails with a ClassNotFoundException for (ironically) the TupleWritable class.

This appears to be due to the way that ObjectInputStream resolves classes in its resolveClass method, together with the way that the context classloader is set within a hadoop mapper or reducer.

This is similar to PIG-2532.

This can be reproduced in the local job tracker (at least) in Hadoop 2.7.0, but it can't be reproduced in Crunch integration tests (due to classloading setup). It appears that this issue is only present in Crunch 0.12.

The following code within a simple pipeline will cause this issue to occur:

PTable<String, Integer> yearTemperatures = ... /* Writable-based PTable */
PTable<String, Integer> maxTemps = yearTemperatures
                .groupByKey()
                .combineValues(Aggregators.MAX_INTS())
                .top(1);   //LINE THAT CAUSES THE ERROR