Re: Sum aggregate calculation for single precsion real

On Wed, Feb 15, 2017 at 9:52 AM, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
> On Tue, Feb 14, 2017 at 11:45 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> You could perhaps make an argument that sum(float4) would have less risk
>> of overflow if it accumulated in and returned float8, but frankly that
>> seems a bit thin.
>
> I think that's more or less the argument Konstantin is in fact making.
> Whether it's a good argument or a thin one is a value judgement.
> Personally, I find it somewhere in the middle: I think the way it
> works now is reasonable, and I think what he wants would have been
> reasonable as well. However, I find it hard to believe it would be
> worth changing now on backward compatibility grounds. He doesn't like
> the way it works currently, but we have no way of knowing how many
> people who are happy with the way it works today would become unhappy
> if we changed it. We need a fairly compelling reason to risk breaking
> somebody's SQL, and I don't think this rises to that level.

I know this is said from time to time in this list, but a third option
that wouldn't break anybody's SQL would be using compensated summation
in the input type.

AFAIK, that can only increase precision, but it will cost cycles. The
impact could however fall below the noise and be worth a try.