spark: What is the difference between Aggregator and UDAF？

Question

In Spark's documentation, Aggregator:

abstract class Aggregator[-IN, BUF, OUT] extends Serializable

A base class for user-defined aggregations, which can be used in Dataset operations to take all of the elements of a group and reduce them to a single value.

UserDefinedAggregateFunction is:

abstract class UserDefinedAggregateFunction extends Serializable

The base class for implementing user-defined aggregate functions (UDAF).

According to Dataset Aggregator - Databricks, “an Aggregator is similar to a UDAF, but the interface is expressed in terms of JVM objects instead of as a Row .”

It seems these two classes are very similar, what are other differences apart from the types in the interface?

A similar question is: Performance of UDAF versus Aggregator in Spark

Answer 1

A fundamental difference, apart from types, is external interface:

• Aggregator takes a complete Row (it is intended for "strongly" typed API).
• UserDefinedAggregationFunction takes a set of Columns.

This makes Aggregator less flexible, although overall API is far more user friendly.

There is also a difference with handling state:

• Aggregator is stateful. Depends on mutable internal state of its buffer field.
• UserDefinedAggregateFunction is stateless. State of the buffer is external.