Hadoop expects your keys and values to implement Writable and serialize themselves (a lot of these are actually hand-written expecting the instance to get reuse

Hadoop expects your keys and values to implement Writable and serialize themselves (a lot of these are actually hand-written expecting the instance to get reused for each input tuple). There's optional and fairly clumsy glue that makes Avro work in a key or value.