Hi, On 2020-06-13 01:06:25 +0200, Tomas Vondra wrote: > I agree, we should revert 4cad2534da and only project tuples when we > actually need to spill them.
There are cases where projecting helps for non-spilling aggregates too, but only for the representative tuple. It doesn't help in the case at hand, because there's just 5 hashtable entries but millions of rows. So we're unnecessarily projecting all-5 rows. But when there are many different groups, it'd be different, because then the size of the representative tuple can matter substantially. Do you think we should tackle this for 13? To me 4cad2534da seems like a somewhat independent improvement to spillable hashaggs. Greetings, Andres Freund