On 2016-12-16 09:34:31 -0800, Andres Freund wrote:
> > To fix his issue, we need something like your 0001. Are you going to
> > polish that up soon here?
>
> Yes.
I've two versions of a fix for this. One of them basically increases the
"spread" of buckets when the density goes up too much. It does so by
basically shifting the bucket number to the left (e.g. only every second
bucket can be the "primary" bucket for a hash value). The other
basically just replaces the magic constants in my previous POC patch
with slightly better documented constants. For me the latter works just
as well as the former, even though aesthetically/theoretically the
former sounds better. I'm inclined to commit the latter, at least for
now.
Andres