Josh Berkus <josh@agliodbs.com> writes:
>> I don't think this one is that far away either. I've been holding Bryce
>> and Ramon's feet to the fire on the issue of possible downside, but so
>> far there's not really much evidence of any *actual* as opposed to
>> theoretical downside.
> What sorts of operations would we test which could potentially show a
> performance downside? I have to admit I don't really understand what
> use-cases this patch is meant to improve.
The patch is meant to improve performance in cases where the outer
relation's key distribution is heavily skewed, by introducing a fast
path for keys matching the outer's most common values (MCVs). But it
does that by potentially sacrificing performance for non MCV keys.
So the case that's of concern is where the distribution is just skewed
enough to trigger the patch's behavioral change, but you don't actually
get a win because there are too many non-MCV keys.
Note that as it's coded, the outer relation's skew is what triggers the
behavioral change. It's not real clear to me how skew in the inner
relation's distribution affects things.
regards, tom lane