Losing Constant Folding for UDFs is not really a case, right? Because you can imitate the same effect: instead of typing “dataFrame.select(myUdf + 1)”, you move the +1 into the body of myUdf — which may even make more sense logically

As for predicate pushdown — that can be a really beneficial boost, but applies only to filters. So not at all for UDAFs or UDFs which just compute or construct something. For those, Ser/De may be a big issue, though