Spark's UDAFs appear to be serializing and de-serializing to/from the MutableAggregationBuffer for each row. This gist shows a small reproducing UDAF and a spark shell session:
The UDAF and its compantion UDT are designed to count the number of times that ser/de is invoked for the aggregator. The spark shell session demonstrates that it is executing ser/de on every row of the data frame.
Note, Spark's pre-defined aggregators do not have this problem, as they are based on an internal aggregating trait that does the correct thing and only calls ser/de at points such as partition boundaries, presenting final results, etc.
This is a major problem for UDAFs, as it means that every UDAF is doing a massive amount of unnecessary work per row, including but not limited to Row object allocations. For a more realistic UDAF having its own non trivial internal structure it is obviously that much worse.