Re: Speeding up aggregates

Hannu Krosing <hannu(at)tm(dot)ee> writes:
> This should also make it easier to implement all kinds of GROUP BY
> ROLLUP|CUBE|GROUPING SETS|() queries.

> Do you have any near-term plans for doing them ?

Not me.

> Is there a variable to set that would disable one or another, like we
> currently have for disabling various join strategies ?

enable_hashagg. I didn't think about one to prevent the old style.

>> Note that even though there's no SORT, the sort_mem setting is used
>> to determine the allowable hashtable size, so a too-small sort_mem
>> might discourage the planner from selecting hashed aggregation.

> Do you mean that hashed aggregation can't overflow to disk, or would it
> just be too slow ?

I didn't write any code to let it overflow to disk --- didn't seem
likely to be useful. (You're probably better off with a sort-based
aggregation if there are too many distinct grouping keys.)

regards, tom lane