An Avro Key/Value file written via raw MapReduce contains records that follow the schema generated by the org.apache.avro.hadoop.io.AvroKeyValue class.
If these files contain specific or reflection-based records, there is currently no easy way to read them in as specific or reflection records. Using the basic public Crunch APIs, they can only be read as generic records (that also contain generic records).
A method should be added to the Avros class which allows specifying specific PTypes to be used for reading the underlying data types within a raw MR output file.
Link to related discussion that inspired this ticket on the user list: http://s.apache.org/es