Re: Fix BRIN numeric_minmax_multi_ops out of range value range merging

On 11/2/26 23:49, Tomas Vondra wrote:
> On 10/31/26 05:11, shihao zhong wrote:
>> Hi Tomas,
>>
>> Thanks for your patch.. 
>>
>> I think your code is right, the subtraction only fails when the result
>> is too
>> big for numeric.  
>>
>> Two small things.
>>
>> 1. A NaN boundary gives a NaN distance. The
>> brin_minmax_multi_distance_numeric
>> return infinity there.
>>
>> 2. In 0004, GiST can call numrange_subdiff with v1 < v2. Overflow
>> there should give -infinity, not +infinity.
>>
>
> Thanks. I've pushed the three BRIN fixes. Here's an updated fix for the
> multirange issue, with the -infinity for (v1 < v2).
>

I've pushed the remaining part, with some minor tweaks (mostly wording,
stale comments, and so on). I realized this may actually affect more
places, including inserts into GiST indexes, selectivity estimation,
etc. Hence the minor tweaks in the commit message.

I have tried to construct a test exercising the -infinity case, but that
turned out a bit ugly. It either depended on platform (formatting of
"infinity"), or triggered it through GiST page splits etc. I decided
it's not worth it, so we test just the +infinity case.

Thanks for the investigation and patches!

--
Tomas Vondra
may the buildfarm be merciful