Greg Stark <gsstark(at)mit(dot)edu> writes:
> One of the things I think has to change with postgres is the default
> selectivity assumptions for inequality operators. They're way to high
> currently.
Maybe so, but 5% is grossly too low. We'd just be throwing ourselves
into a different set of badly misoptimized queries. On what grounds can
you argue that "WHERE x > y" will select only 5% of the rows, if you
have no knowledge about either x or y? (And no, I won't buy the
argument that users are only interested in queries that fetch small
proportions of tables. That argument comes from thinking of only one
class of applications.)
Perhaps more to the point, 5% is still too high to ensure selection of
an indexscan. You'd need something more like 0.5%, which is even more
obviously hopelessly misguided.
regards, tom lane