I wrote:
> Jeff Janes <jeff.janes@gmail.com> writes:
>> Hash index is not concurrency safe, starting in REL8_4_0 and up to HEAD.
> Ouch. This used to be okay, because adding new entries to a hash page
> always added them at the end. The 8.4 changes to keep individual hash
> pages sorted by hashcode broke it :-(.
Actually, now that I am looking at it, that patch COMPLETELY destroyed
hash indexes. The search logic requires that index entries within a
page are ordered by hash value. Although the insertion code preserves
that property, neither _hash_splitbucket nor _hash_squeezebucket make
any attempt to do so. So it's not just a transient concurrency issue,
you can easily get corruption of a hash index leading to permanent
search failures.
Ugh. Mea culpa for letting this one through.
regards, tom lane