Douglas Alan <darkwater42@gmail.com> writes:
> As I said, in my original post, Postgres's approach would be completely
> reasonable in this case,* if* the rows that it was looking for were
> sprinkled randomly throughout the table. But they're *not* in this case --
> they're all at the end.
There's been some talk of penalizing the seqscan+limit combination
(perhaps by increasing the estimated start cost for the seqscan) if
the WHERE clause involves any variables that have a correlation stat
significantly different from zero. But nobody's done the legwork
to see if this would really be useful or what an appropriate penalty
curve might be.
regards, tom lane