AvroParquetInputFormat should use a parameterized type

[asfimport]

The AvroParquetInputFormat currently extends ParquetInputFormat<IndexedRecord>, which works for regular MR cases. But Spark's hadoopRDD and [newAPIHadoopRDD](https://people.apache.org/~pwendell/spark-1.1.0-rc3-docs/api/java/org/apache/spark/SparkContext.html#newAPIHadoopRDD(org.apache.hadoop.conf.Configuration, java.lang.Class, java.lang.Class, java.lang.Class)) methods (correctly) create a RDD with the types from the InputFormat. This means that the RDD always uses IndexedRecord rather than the correct type.

The AvroParquetInputFormat should be AvroParquetInputFormat<T extends IndexedRecord> extends ParquetInputFormat<T>

Reporter: Ryan Blue / @rdblue

_{Note: This issue was originally created as PARQUET-132. Please see the migration documentation for further details.}

[asfimport]

Ryan Blue / @rdblue:
I've added the change in PR #84.

[asfimport]

Thomas White / @tomwhite:
We should change it to AvroParquetInputFormat<T> so that we can support Avro reflect in the future (see https://github.com/Parquet/parquet-mr/pull/308). If we changed it to AvroParquetInputFormat<T extends IndexedRecord> then we can't at a later point change it to AvroParquetInputFormat<T>, because adding a type parameter is compatible, while changing the bounds of one is not (see https://wiki.eclipse.org/Evolving_Java-based_APIs_2#Evolving_API_Classes).

If we don't do the reflect work at this point, we could throw a runtime exception if T is not an IndexedRecord.

[asfimport]

Ryan Blue / @rdblue:
Thanks for taking a look, Tom. I've updated the PR to add AvroParquetInputFormat<T> as you suggested. We don't actually get a reference to the type at runtime to inspect and throw an exception, so we will have to do that elsewhere. That update can wait until we add support for reflect.

[asfimport]

Ryan Blue / @rdblue:
Issue resolved by pull request 84
https://github.com/apache/incubator-parquet-mr/pull/84