Re: Improve cube GiST page splits

Hi Rustam,

Thank you for the detailed review and benchmarks!

On September 30, 2026, Rustam ALLAKOV wrote:
> However, I found two cases where data that master accepts now makes
> the page split fail with an ERROR.

Both are regressions in my patch. In v2, coordinate and bounding
interval comparisons use PostgreSQL's float8 ordering. Split-quality
estimates use plain C arithmetic rather than SQL arithmetic that can
throw an error.
Dimensions with a non-finite or zero range are skipped. If no dimension
can provide a split, the existing fallback divides the entries evenly.

I added regression coverage for both incremental insertion and CREATE
INDEX on existing data, including your examples, infinities, and an
underflow case.

PFA v2.

Thank you!

Best regards, Andrey Borodin.