Tomas Vondra wrote:
> Yes, that's clearly the culprit here. In both cases we estimate here are
> only ~4000 tuples in the hash, and 9.3 sizes the hash table to have at
> most ~10 tuples per bucket (in a linked list).
>
> However we actually get ~3M rows, so there will be ~3000 tuples per
> bucket, and that's extremely expensive to walk. The reason why 100MB is
> faster is that it's using 2 batches, thus making the lists "just" ~1500
> tuples long.
>
> This is pretty much exactly the reason why I reworked hash joins in 9.5.
> I'd bet it's going to be ~20x faster on that version.
Thank you for the explanation!
Yours,
Laurenz Albe
