Comment by life_is_short

> for aggregates with variable-length per-group state, we don't have a good size estimate

Can you say a bit more why?