homepage

Short version: __get__/__set__/__delete__ attributes on descriptor objects
(as opposed to their types) are treated inconsistently as part of the
descriptor
protocol: the documentation and support code includes them; the core
implementation doesn't.
Example:
 class D(object):
 __get__ = lambda self, i, o: 'D'
 class C(object):
 d = D()
 d.__get__ = lambda i, o: 'd'
 d.__set__ = lambda i, v: 1/0
 c = C()
According to pydoc and inspect, and the description in the reference manual
(section 3.4.2.3), d's __get__ and __set__ override D's:
 >>> inspect.isdatadescriptor(C.__dict__['d'])
 True
 >>> help(C)
 class C(__builtin__.object)
 | Data descriptors defined here:
 ...
 | d
 >>> type(c).__dict__['d'].__get__(c, type(c))
 'd'
 >>> type(c).__dict__['d'].__set__(c, 5)
 ZeroDivisionError: integer division or modulo by zero
According to CPython, they have no effect:
 >>> c.d
 'D'
 >>> c.d = 5; c.d
 5
PEP 252 notes: "For speed, the get and set methods are type slots", which
points to the CPython behavior being an intentional concession for
performance.
Should CPython respect descriptor object attributes, if reasonable
performance
can be maintained? Otherwise, should the documentation and support code be
changed to ignore them?