Re: hashagg slowdown due to spill changes

Hi,

On 2020-06-12 15:29:08 -0700, Jeff Davis wrote:
> On Fri, 2020-06-12 at 14:37 -0700, Andres Freund wrote:
> > I don't see why it's ok to force an additional projection in the very
> > common case of hashaggs over a few rows. So I think we need to
> > rethink
> > 4cad2534da6.
> 
> One possibility is to project only spilled tuples, more similar to
> Melanie's patch from a while ago:
> 
> 
> https://www.postgresql.org/message-id/CAAKRu_aefEsv+UkQWqu+ioEnoiL2LJu9Diuh9BR8MbyXuZ0j4A@mail.gmail.com
> 
> Which makes sense, but it's also more code.

I'm somewhat inclined to think that we should revert 4cad2534da6 and
then look at how precisely to tackle this in 14.

It'd probably make sense to request small tlists when the number of
estimated groups is large, and not otherwise.

Greetings,

Andres Freund