"Bryan White" <[EMAIL PROTECTED]> writes:
> I have recoverd the performance lost when I moved to Postgres 7.0 by
> executing SET enable_indexscan = OFF before creating my cursors and
> turning it back on for the inner loop query. It may even be faster
> then before so I am happy.
OK, so it was the indexscans that were hurting. (7.0 has new sorting
code too, so I was a little afraid that the problem might be with the
sorts. Evidently not.)
This suggests that at least on your setup, the default value of 4.0 for
random_page_cost might still be too low. I have not tried to measure
that number on a Linux machine, just on machines with BSD-derived
filesystems. Maybe Linux does a lot worse with random accesses than
BSD? Needs looking into.
> It seems that with index scans the cursors start producing data right away
> (but the overall rate is slower). With sequential scan and sort the report
> gets no data for the first 30 minutes and then runs at about 4 times the
> rate of the index scan.
Right, that's what you'd expect: the sort has to be completed before it
knows which row to deliver first, but an indexscan has no such startup
cost.
regards, tom lane
