Re: slow query, different plans

From: Tom Lane

To: Midge Brown

Cc: pgsql-performance@postgresql.org

Sent: Friday, August 03, 2012 11:26 PM

Subject: Re: [PERFORM] slow query, different plans

"Midge Brown" <midgems@sbcglobal.net> writes:
> I'm having a problem with a query on our production server, but not on a laptop running a similar postgres version with a recent backup copy of the same table. I tried reindexing the table on the production server, but it didn't make any difference. Other queries on the same table are plenty fast.

Reindexing won't help that.  The problem is a bad statistical estimate;
it thinks there are about 700 rows with applies2 = 256, when there's
really only one.  That means the "fast" plan is a lot faster than the
planner gives it credit for, and conversely the "slow" plan is a lot
slower than the planner is expecting.  Their estimated costs end up
nearly the same, which makes it a bit of a chance matter which one is
picked --- but the true costs are a lot different.  So you need to fix
that rowcount estimate.  Raising the stats target for the table might
help.

regards, tom lane

--

I added "and ts is not null" to the query and the planner came back with .102 ms. The problem area in production went from a 10 second response to < 1 second.

 

Thanks for the responses.

 

-Midge