Home | About | Sematext search-lucene.com search-hadoop.com
 Search Hadoop and all its subprojects:

Switch to Threaded View
Avro >> mail # user >> Avro new mapreduce API

Copy link to this message
Avro new mapreduce API
Hi there,

Can u give me some examples or explaination about programming with
pure org.apache.avro.mapreduce interfaces??

-------- Save your time, continue if you know how --------

All of my programs are writing with hadoop's new MR1 interfaces
(org.apache.hadoop.mapreduce), so I want to use new
org.apache.avro.mapreduce of avro too. But it doesn't work for me.

The program takes input of avro data and output the same.
The main idea behind my program is subclassing hadoop's Mapper
and Reducer against avro wrapped key/value.

Here is a block of my job driver :

        AvroJob.setInputKeySchema(job, NetflowRecord.getClassSchema());
        AvroJob.setOutputKeySchema(job, NetflowRecord.getClassSchema());



The definitions of MyAvroMap and MyAvroReduce subclasses respectivly are

    public static class MyAvroMap extends Mapper<AvroKey<NetflowRecord>, NullWritable,
     AvroKey<CharSequence>, AvroValue<NetflowRecord>>{ ... }

    public static class MyAvroReduce extends Reducer<AvroKey<CharSequence>, AvroValue<NetflowRecord>,
            AvroKey<NetflowRecord>, NullWritable>{ ... }

The methioned NetflowRecord is my avro record class. And I got running exception

    java.lang.ClassCastException: class org.apache.avro.hadoop.io.AvroKey

By reading hadoop's and avro's source code,
I found that the exception was thrown by JobConf to make sure
the map key is a subclass of WritableComparable, like this (hadoop1.2.1, line759)


But the avro shows that AvroKey and AvroValue are just a simple wrapper
**without** subclassing Writable* interfaces of hadoop.
I believe that, even without testing, I can get through that using old mapred interfaces,
but its not what I want.