Bruce Momjian wrote:
> Neil, I have added these item to the TODO list. Do you plan on applying
> this?
No, I don't have any immediate plans to apply it, as unfortunately I
didn't see a performance win :-( It's also possible I'm just not
measuring the right workload, although I don't have time to rerun the
benchmarks at the moment.
The patch does two things: (1) change hash indexes to only store the
key's hash value, not the entire key (2) store index elements within a
hash bucket in order of hash key and search for matches via binary
search. #1 is definitely a win in some in some circumstances (e.g.
indexing large fields or types with expensive equality operators), but
those aren't the common case. I'm surprised that #2 is not a more
noticeable improvement...
One possibility would be to provide an optional implementation of #1,
perhaps via an alternate index operator class. That way people could
select it in those situations in which it is worth using.
-Neil
