Major Description
In a process of making Nutch2 run on Hadoop 2.3.0 + HBase 0.98.1 we encountered java.io.EOFException's like ones described in this mail thread: http://www.mail-archive.com/user%40nutch.apache.org/msg12644.html
We applied a patch mentioned there and got our setup running but being very unstable: it would fail with an ArrayIndexOutOfBounds exception whenever we try to generate a batch of some 50 or more pages to fetch.
We investigated the problem and discovered that in working setup of Nutch2 + Hadoop 1.2.0 + HBase 0.94.14, PersistentDeserializer is used for deserialization during reduce phase, and not AvroSerialization.AvroDeserializer. The reason for this sudden swap of deserializers lies in GoraMapReduceUtils#setIOSerializations method. It uses StringUtils.joinStringArrays and this method uses HashSet under the hood. Two more serializations were added to io.serializations property in Hadoop 2.3.0 compared to Hadoop 1.2.0 and this results in AvroSpecificSerialization being placed on top of serializations list.
After we have patched GoraMapReduceUtils#setIOSerializations, having explicitly set PersistentSerialization to be the top of the list, we have fixed the problem with instability. Moreover, we don't even need to patch Avro now, just one simple change in Gora and everything works like a charm!
So we propose to move PersistentSerialization to the top of serializations list.