Mark Lewis <mark.lewis@mir3.com> writes:
> What you probably want to do is tell PG that you're looking for a
> literal underscore and not for any matching character by escaping the
> underscore, that will allow it to do a much quicker index scan.
The other half of the problem is that the planner is drastically
misestimating the number of matching rows --- it thinks only one
when there are really about a thousand, and this leads it to use
a nestloop that will be very inefficient with so many rows.
Try increasing the statistics target for that column. Also, if
this is a pre-8.2 PG release, consider upgrading; I believe we
improved the LIKE estimator in 8.2.
regards, tom lane