On Mon, 2018-06-04 at 11:52 -0700, Andres Freund wrote:
> I wonder whether, at least for aggregates, the better fix wouldn't be
> to
> switch to feeding the tuples into tuplesort upon memory exhaustion
> and
> doing a sort based aggregate. We have most of the infrastructure to
> do
That's an interesting idea, but it seems simpler to stick to hashing
rather than using a combination strategy. It also seems like it would
take less CPU effort.
What advantages do you have in mind? My patch partitions the spilled
data, so it should have similar disk costs as a sort approach.
Regards,
Jeff Davis