On 2018-01-30 13:57:44 -0500, Todd A. Cook wrote:
> Out of curiosity, I then modified hashint8() as previously described.
> With that change, run time dropped to 11 minutes (also average of 3
> runs).
>
> FWIW, the data values in these sets are sort-of random (where I can't
> explain the "sort-of" in a public forum), but strongly biased towards
> negative infinity. Starting again from scratch, we could probably remove
> the bias, but we have 28-30 billion of these things collected over the
> last 14 years (starting in PostgreSQL 7.4) and it's kinda tough to change
> directions at this point...
FWIW, you could just create a different hash opclass and use it for
those queries...
Greetings,
Andres Freund