ISTM there is no right or wrong answer. There is just a question of what is most useful. AFAICT, the code for dictionaries (and therefore the code for sets) has always had identity-implies-equality logic. It makes dicts blindingly fast for common cases. It also confers some nice properties like making it possible to retrieve a NaN that has been stored as a key; otherwise, you could store it but not look it up, pop it, or delete it (because the equality test would always fail). The logic also confers other nice-to-have properties such as: * d[k] = v; assert k in d # assignment-implies-contains * assert all(k in d for k in d) # all-members-are-members These aren't essential invariants but they do provide a pleasant programming environment and make it easier to reason about programs. Another place where identity-implies-equality logic is explicit is in Py_RichCompareBool(). That lets methods in many other functions and methods work like dicts and sets. It speeds them up and confers some nice-to-haves like: * mylist.append(obj) implies mylist.count(obj) > 0 * x = obj implies x == obj # assignment really works There may be lots of other code that implicitly makes similar assumptions. I don't know how you could reliably find those and rip them out. If identity-implies-equality does get ripped out, I don't know what we would win. It would make it possible to do some cute NaN tricks, but I don't think you can defend against the general problem of funky objects being able to muck-up code that looks correct. You get oddities when an object lies about its length. You get oddities when an object has a hash that doesn't match its equality function. The situation with NaNs and sorts is a prime example:
sorted([1.2, 3.4, float('Nan'), -1.2, float('Inf'), float('Nan')]) [1.2, 3.4, nan, -1.2, inf, nan]
Personally, I think the status quo is fine and that practicality is beating purity. High quality programs are written every day. Numeric programmers seem to have no problem using NaNs as-is. AFAICT, the only actual problem in front us is the OP's post where he was able to surprise himself with some NaN experiments at the interactive prompt. Raymond