Neil Conway <neilc@samurai.com> writes:
> I was looking at this a bit in IRC, and I was more concerned by the fact
> that 7.3 was 20% than 7.2 on the same hardware, when they both used the
> same query plan (consider the data at the end of the URL above, after
> the execution of 'SET enable_mergejoin = off;').
Hm. Are we sure that both versions were built with the same
optimization level, etc? (My private bet is that Wade's 7.2 didn't
have multibyte or locale support --- but that's a long shot when we
don't know the datatypes of the columns being joined on...)
> Also, is it expected that the cardinality estimates for join steps won't
> be very accurate, right? (estimated: 19 rows, actual: 765 rows)
Well, it'd be nice to do better --- I was hoping Wade would look into
why the row estimates were off so much.
regards, tom lane