From “Stop Writing Classes” (Jack Diederich, PyCon 2012):

Namespaces are for preventing name collisions, not for creating taxonomies.

He gives the pathological example of

d = MuffinMail.MuffinHash.MuffinHash(foo=3)

(the MuffinHash class, which belongs to the MuffinHash module of the MuffinMail package). At best that should be MuffinMail.Hash(foo=3). In his particular case actually the MuffinHash class should never have existed, and it should have been a plain old dict { 'foo': 3 }.

It’s a good talk, most of its points exactly as applicable to C++ as they are to Python.

Inevitable nit: In C++, our standard library primitives aren’t nearly as fit-for-purpose nor as stable as Python’s. Where Jack prefers Python’s dict over MuffinMail.Hash, I’d say that in C++ it is good practice to introduce muffinmail::hash at least as a typedef, and ideally as a full-blown class, in preference to peppering our muffinmail code with std::map<std::string, int>. In C++, it’s a bad idea to harness yourself to the API of a standard library type — an API you can’t change to suit your needs, but which someone else will change every three years in ways that may make it less fit for your particular purpose. Better to own your own API surface area. See also “Don’t inherit from standard types” (2018-12-11).