And I'm probably repeating myself too, but: the predictability of (and difficulty of changing of) hashing for numeric types is why I'm strongly opposed to hash collision / slot collision limits:  they'd end up disallowing reasonably natural looking Python numeric sets (e.g. {2**k for k in range(n)} for smallish n).  I don't think core Python should be solving this issue at all---I think that's a job for the web frameworks.  Christian's idea of providing more suitable types in the std. lib. sounds like the right direction to me.