I wrote:
> This might be a bug in the LIKE estimator (if so, it's in 8.2.3 as
> well). I don't have time to look closer right now, but can you show us
> the pg_stats row for orders_smaller.order_number?
Oh, never mind that ... on inspection, the NOT LIKE selectivity
estimator is obviously broken: patternsel() doesn't realize it needs
to negate the results of the operator when applying same to histogram
entries.
regards, tom lane