Diabolical bug involving code that behaves differently in a class decorator vs. a metaclass

bug.py

# bug.py

def decorate(func):
    print('Decorating', func.__name__)
    return func

def wrap_methods(cls):
    for name in vars(cls):
        if name.startswith('f_'):
            setattr(cls, name, decorate(getattr(cls, name)))
    return cls

class Meta(type):
    def __new__(meta, clsname, bases, methods):
        cls = super(Meta, meta).__new__(meta, clsname, bases, methods)
        wrap_methods(cls)
        return cls

print('---- Class Decorator ----')

@wrap_methods
class Spam1(object):
    def f_1(self): pass
    def f_2(self): pass
    def f_3(self): pass
    def f_4(self): pass
    def f_5(self): pass
    def f_6(self): pass
    def f_7(self): pass
    def f_8(self): pass

print('---- Metaclass ----')

class Spam2(metaclass=Meta):
    def f_1(self): pass
    def f_2(self): pass
    def f_3(self): pass
    def f_4(self): pass
    def f_5(self): pass
    def f_6(self): pass
    def f_7(self): pass
    def f_8(self): pass

output.txt

Runing the program, you get very different results for the two classes. The class decorator
version always does what's expected.  Not so much for the metaclass.

For example:

bash % python3 bug.py

---- Class Decorator ----
Decorating f_4
Decorating f_8
Decorating f_7
Decorating f_3
Decorating f_6
Decorating f_2
Decorating f_5
Decorating f_1
---- Metaclass ----    
Decorating f_4
Decorating f_3
Decorating f_5
Decorating f_2          # Notice: Where is f_7?  It's gone!
Decorating f_1
Decorating f_6
Decorating f_8

Try running it again:

---- Class Decorator ----
Decorating f_3
Decorating f_7
Decorating f_5
Decorating f_1
Decorating f_8
Decorating f_6
Decorating f_4
Decorating f_2
---- Metaclass ----
Decorating f_3       # <<<<<<
Decorating f_7
Decorating f_1
Decorating f_3       # WHAT?!?! This was just wrapped above
Decorating f_5
Decorating f_8
Decorating f_6
Decorating f_4
Decorating f_2

So yes, you're iterating over the class dictionary, changing the value assigned to some of the keys,
and effectively seeing the dictionary "change size", dropping one or more keys or iterating over
the same key more than once.  It randomly changes run-to-run.  Code behaves differently depending
if it's invoked from a metaclass or not.  

Enjoy!

bisect pins the blame on the patch for http://bugs.python.org/issue20637 at https://hg.python.org/cpython/rev/16229573e73e

One potential consequence of that patch: it means that instances of subclasses of type would have started to attempt to share keys with instances of type, whereas previously key sharing would have been implicitly disabled for all classes with a custom metaclass.