Re: Properly mark NULL returns in numeric aggregates

Hi Andres,

On Fri, Apr 10, 2020 at 12:14 PM Andres Freund <andres(at)anarazel(dot)de> wrote:
>
> Shouldn't these just be marked as strict?
>

Are you suggesting that because none of the corresponding trans
functions (int8_avg_accum, int2_accum, and friends) ever output NULL?
That's what we thought, but then I realized that an input to a comebine
function is not necessarily an output from a trans function invocation:
for example, when there is a "FILTER (WHERE ...)" clause that filters
out every tuple in a group, the partial aggregate might just throw a
NULL state for the final aggregate to combine.

On the other hand, we examined the corresponding final functions
(numeric_stddev_pop and friends), they all seem to carefully treat a
NULL trans value the same as a "zero input" (as in, state.N == 0 &&
state.NaNcount ==0). That does suggest to me that it should be fine to
declare those combine functions as strict (barring the restriction that
they should not be STRICT, anybody remembers why?).

Cheers,
Jesse and Deep