It doesn't make sense that cmp and hash are set independently, does it? (original) (raw)
Right now, cmp= and hash= are treated as independent options, both in attr.s() and attr.ib(). This does't make much sense to me.
If you do attr.s(cmp=True, hash=False) then you get a broken class that violates Python's invariant that objects which compare equal must hash equal (if they are hashable at all).
If you do attr.s(cmp=False, hash=True) then you get a class with a really weird hash that has lots of collisions (Foo(x=1) and Foo(x=1) are different objects so they compare non-equal, but their hashes are the same.) Also, it is broken if the object is mutable, because the hash can change over time. ...and now that I think about it, objects that have cmp=True also have the same problem with mutability. Actually I guess all of my attrs classes are currently illegal, because the defaults are to generate illegal classes. I guess I should go fix that.
Anyway, I think the only plausible configurations are:
cmp=True,frozen=True,hash=Truecmp=True,frozen=True, hash undefined (attrs doesn't have a nice way to do this right now I think?)cmp=False,hash=False(useobject.__hash__)cmp=False, hash undefined
And then for attr.ib, why are cmp and hash separate arguments? Wouldn't it make more sense to drop the hash argument, and just use the logic that if attr.ib(cmp=True) and we're generating a hash method at all, then we should include this attrib?