Re: BRIN minmax multi - incorrect distance for infinite timestamp/date

On 10/17/23 22:25, Tomas Vondra wrote:
> Hi,
>
> Here's a couple cleaned-up patches fixing the various discussed here.
> I've tried to always add a regression test demonstrating the issue
> first, and then fix it in the next patch.
>
> In particular, this deals with these issues:
>
> 1) overflows in distance calculation for large timestamp values (0002)
>
> 2) incorrect subtraction in distance for date values (0003)
>
> 3) incorrect distance for infinite date/timestamp values (0005)
>
> 4) failing distance for extreme interval values (0007)
>
> All the problems except "2" have been discussed earlier, but this seems
> a bit more serious than the other issues, as it's easier to hit. It
> subtracts the values in the opposite order (smaller - larger), so the
> distances are negated. Which means we actually merge the values from the
> most distant ones, and thus are "guaranteed" to build very a very
> inefficient summary. People with multi-minmax indexes on "date" columns
> probably will need to reindex.
>

BTW when adding the tests with extreme values, I noticed this:

test=# select '5874897-01-01'::date;
date
---------------
5874897-01-01
(1 row)

test=# select '5874897-01-01'::date + '1 second'::interval;
ERROR: date out of range for timestamp

IIUC this happens because the first thing date_pl_interval does is
date2timestamp, ignoring the fact that the ranges of those data types
are different - dates allow values up to '5874897 AD' while timestamps
only allows values up to '294276 AD'.

This seems to be a long-standing behavior, added by a9e08392dd6f in
2004. Not sure how serious it is, I just noticed when I tried to do
arithmetics on the extreme values in tests.

regards

--
Tomas Vondra
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company