I tried to google something about it. Why do non-data descriptors work with old-style classes?

Docs say that they should not:
"Note that descriptors are only invoked for new style objects or classes (ones that subclass object() or type()).".

class Descriptor(object):def __init__(self):self.x = 1def __get__(self, obj, cls=None):return self.xclass A:x = Descriptor()a = A()
a.x>>> 1

Thanks.

You are right to question the documentation. I've tried looking through CPython sources to find an explanation, but be warned: I'm no expert.

From my understanding, attribute lookup and descriptor __get__ invocation occurs in instance_getattr2 (chosen extracts):

v = class_lookup(inst->in_class, name, &klass);
if (v != NULL) {f = TP_DESCR_GET(v->ob_type);if (f != NULL) {PyObject *w = f(v, (PyObject *)inst, (PyObject *)(inst->in_class));}
}

So either I am missing something, or nothing in the implementation requires a new-style object (which contradicts the documentation).

For the record, I tried recompiling Python to restrict descriptor invocation to new style classes objects, but it actually brought up a gigantic mess. I learned in the process that class methods themselves are implemented as descriptors: this is the mechanism used to return bound or unbound method objects depending on the usage. For example:

>>> class A:
...     def foo():
...         pass
...
>>> A.foo.__get__(None, A)
<unbound method A.foo>
>>> A.foo.__get__(A(), A)
<bound method A.foo of <__main__.A instance at 0x000000000229CC48>>

As a result, it seems that preventing descriptor invocation for attributes of old-style objects or classes would also prevent method calls on them, at least with CPython implementation.

Once again, I'm no expert and this is the first time I dive into Python implementation, so I could very well be wrong. I've filed an issue to try to clarify this.