> It's equivalent to our assumption that distributions of values in
> columns in the same table are independent. Making that assumption in
> this case would probably result in occasional dramatic speed
> improvements similar to the ones we've seen in less complex joins,
> offset by just-as-occasional dramatic slowdowns of similar magnitude. In
> other words, it will increase the variance of our results.
Under what circumstances do you think that it would produce a dramatic
slowdown? I'm confused. I thought the penalty for picking a bad set
of values for the in-memory hash table was pretty small.
...Robert