There are big difference between Hive UDF and UDAF. I found that when I was developing UDAF. Normal UDF usually process one row into one value. And Hive jobs are executed as MapReduce job (of course other can do run such as Tez or Spark). So in the case of ordinal UDF, it is only necessary to run mapper. However UDAF is different.

UDAF must collect the data which are processed by mapper finally. In terms of MapReduce, this task can be done by reducer. Generally mapper and reducer do not receive the same input. How can UDAF distinguish this type of difference?

Mode

The initializer of UDAF can be called in each mapper and reducer. So we can distinguish the input difference by using mode passed at the initializer.

init(GenericUDAFEvaluator.Mode m, ObjectInspector[] parameters)
          Initialize the evaluator.

This is passed to GenericUDAFEvaluator init method. Although I’d like to omit the detail of how to write UDAF here, we can figure out the difference of each mode passed init method.

mode description
COMPLETE from original data directly to full aggregation: iterate() and terminate() will be called.
FINAL from partial aggregation to full aggregation: merge() and terminate() will be called.
PARTIAL1 from original data to partial aggregation data: iterate() and terminatePartial() will be called.
PARTIAL2 from partial aggregation data to partial aggregation data: merge() and terminatePartial() will be called.

iterate is called at the input of mapper and merge is called at the input of reducer. So you can select by seeing this mode. When you find COMPLETE or PARTIAL1, that means a corresponding task receives input from table records. When you find FINAL or PARTIAL2 that means a task receives mapper output. The code can be like

if (m == GenericUDAFEvaluator.Mode.COMPLETE || m == GenericUDAFEvaluator.Mode.PARTIAL1) {
  // Configure for mapper input. (e.g. check argument length)
} else if (m == GenericUDAFEvaluator.Mode.FINAL || m == GenericUDAFEvaluator.Mode.PARTIAL2) {
  // Configure for reducer input.
}

And also you can select output type by using this mode in the same way. Thank you.