When to inline definitions of metaclass in Python? - python

Today I have come across a surprising definition of a metaclass in Python here, with the metaclass definition effectively inlined. The relevant part is
class Plugin(object):
class __metaclass__(type):
def __init__(cls, name, bases, dict):
type.__init__(name, bases, dict)
registry.append((name, cls))
When does it make sense to use such an inline definition?
Further Arguments:
An argument one way would be that the created metaclass is not reusable elsewhere using this technique. A counter argument is that a common pattern in using metaclasses is defining a metaclass and using it in one class and then inhertiting from that. For example, in a conservative metaclass the definition
class DeclarativeMeta(type):
def __new__(meta, class_name, bases, new_attrs):
cls = type.__new__(meta, class_name, bases, new_attrs)
cls.__classinit__.im_func(cls, new_attrs)
return cls
class Declarative(object):
__metaclass__ = DeclarativeMeta
def __classinit__(cls, new_attrs): pass
could have been written as
class Declarative(object): #code not tested!
class __metaclass__(type):
def __new__(meta, class_name, bases, new_attrs):
cls = type.__new__(meta, class_name, bases, new_attrs)
cls.__classinit__.im_func(cls, new_attrs)
return cls
def __classinit__(cls, new_attrs): pass
Any other considerations?

Like every other form of nested class definition, a nested metaclass may be more "compact and convenient" (as long as you're OK with not reusing that metaclass except by inheritance) for many kinds of "production use", but can be somewhat inconvenient for debugging and introspection.
Basically, instead of giving the metaclass a proper, top-level name, you're going to end up with all custom metaclasses defined in a module being undistiguishable from each other based on their __module__ and __name__ attributes (which is what Python uses to form their repr if needed). Consider:
>>> class Mcl(type): pass
...
>>> class A: __metaclass__ = Mcl
...
>>> class B:
... class __metaclass__(type): pass
...
>>> type(A)
<class '__main__.Mcl'>
>>> type(B)
<class '__main__.__metaclass__'>
IOW, if you want to examine "which type is class A" (a metaclass is the class's type, remember), you get a clear and useful answer -- it's Mcl in the main module. However, if you want to examine "which type is class B", the answer is not all that useful: it says it's __metaclass__ in the main module, but that's not even true:
>>> import __main__
>>> __main__.__metaclass__
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
AttributeError: 'module' object has no attribute '__metaclass__'
>>>
...there is no such thing, actually; that repr is misleading and not very helpful;-).
A class's repr is essentially '%s.%s' % (c.__module__, c.__name__) -- a simple, useful, and consistent rule -- but in many cases such as, the class statement not being unique at module scope, or not being at module scope at all (but rather within a function or class body), or not even existing (classes can of course be built without a class statement, by explicitly calling their metaclass), this can be somewhat misleading (and the best solution is to avoid, in as far as possible, those peculiar cases, except when substantial advantage can be obtained by using them). For example, consider:
>>> class A(object):
... def foo(self): print('first')
...
>>> x = A()
>>> class A(object):
... def foo(self): print('second')
...
>>> y = A()
>>> x.foo()
first
>>> y.foo()
second
>>> x.__class__
<class '__main__.A'>
>>> y.__class__
<class '__main__.A'>
>>> x.__class__ is y.__class__
False
with two class statement at the same scope, the second one rebinds the name (here, A), but existing instances refer to the first binding of the name by object, not by name -- so both class objects remain, one accessible only through the type (or __class__ attribute) of its instances (if any -- if none, that first object disappears) -- the two classes have the same name and module (and therefore the same representation), but they're distinct objects. Classes nested within class or function bodies, or created by directly calling the metaclass (including type), may cause similar confusion if debugging or introspection is ever called for.
So, nesting the metaclass is OK if you'll never need to debug or otherwise introspect that code, and can be lived with if whoever is so doing understand this quirks (though it will never be as convenient as using a nice, real name, of course -- just like debugging a function coded with lambda cannot possibly ever be so convenient as debugging one coded with def). By analogy with lambda vs def you can reasonably claim that anonymous, "nested" definition is OK for metaclasses which are so utterly simple, such no-brainers, that no debugging or introspection will ever conceivably be required.
In Python 3, the "nested definition" just doesn't work -- there, a metaclass must be passed as a keyword argument to the class, as in class A(metaclass=Mcl):, so defining __metaclass__ in the body has no effect. I believe this also suggests that a nested metaclass definition in Python 2 code is probably appropriate only if you know for sure that code will never need to be ported to Python 3 (since you're making that port so much harder, and will need to de-nest the metaclass definition for the purpose) -- "throwaway" code, in other words, which won't be around in a few years when some version of Python 3 acquires huge, compelling advantages of speed, functionality, or third-party support, over Python 2.7 (the last ever version of Python 2).
Code that you expect to be throwaway, as the history of computing shows us, has an endearing habit of surprising you utterly, and being still around 20 years later (while perhaps the code you wrote around the same time "for the ages" is utterly forgotten;-). This would certainly seem to suggest avoiding nested definition of metaclasses.

Related

Why do I get an Error when creating an instanz of a class that has a metaclass

class Meta(type):
def __new__(cls, class_name, bases, attrs):
a={}
for name, val in attrs.items():
if name.startswith("__"):
a[name] = val
else:
a[name.upper()] = val
return(class_name, bases, a)
class D(metaclass=Meta):
x = 5
y = 8
def hello (self):
print("hi")
c = D()
c.hello
Why am I getting this error:
'tuple' object is not callable
line 20, in
c = D()
Cant finde the reason can someone help me out?
(code based on the teachings of the video at https://www.youtube.com/watch?v=NAQEj-c2CI8 )
The last line inside your metaclasse __new__method, which reads
return(class_name, bases, a)
Returns a plain tuple - it should (and indeed, on the video you comment where you got the inspiration for your studies) call either type or type.__new__: there produce a new class that is then ready to be used:
return type(class_name, bases, a)
Although the video is a nice first introduction to playt around with metaclasses, (1) this is an advanced topic, and one actually can make a whole career in Python without ever needing to write a metaclass;
(2) on calling type instead of type.__new__, the video is not technically correct: it works, but it has a different effect in that the resulting class, although being modified upon its creation, will have the regular type as its metaclass - by calling type directly, all the references to the metaclass itself (Meta in your snippet) are discarded. If instead you call super().__new__(cls, class_name, bases, a) , the created class will be an instance of the metaclass (conveyed under the name cls in this call), and all subclasses of the created class (D in the example), will also have the same metaclass and be created through this __new__ method.
One extra detail on your experiment: you are testing for attributes that starts with a double underscore ("__") - The Python compilation process automatically mangle all such names, (but names "sandwiched" between a pair of "__") to include the class name itself as a prefix - so if D would have a __x attribute, it would show up as _D__x in the attrs dict received in the metaclass -
I see in the linked video that is the example there; however this part of the code is never tested or verified: no __ attributes are demonstrated to be preserved. It will preserve reserved "magic" names as __init__, though. As stated earlier, the video is not 100% good for teaching, as the author itself seems to be just showing some of his own experimenting.
And one last comment regarding the specifc error you are getting: whatever the metaclass __new__ method returns is used by Python as the object bound to the name on the class statement (D in this case). So, D becames an "ordinary" variable containing a tuple, instead of a new class. When you execute D(), Python tries to call the object, and you get the error that tuples are not callable.

Why doesn't Python allow referencing a class inside its definition?

Python (3 and 2) doesn't allow you to reference a class inside its body (except in methods):
class A:
static_attribute = A()
This raises a NameError in the second line because 'A' is not defined, while this
class A:
def method(self):
return A('argument')
works fine.
In other languages, for example Java, the former is no problem and it is advantageous in many situations, like implementing singletons.
Why isn't this possible in Python? What are the reasons for this decision?
EDIT:
I edited my other question so it asks only for ways to "circumvent" this restriction, while this questions asks for its motivation / technical details.
Python is a dynamically typed language, and executes statements as you import the module. There is no compiled definition of a class object, the object is created by executing the class statement.
Python essentially executes the class body like a function, taking the resulting local namespace to form the body. Thus the following code:
class Foo(object):
bar = baz
translates roughly to:
def _Foo_body():
bar = baz
return locals()
Foo = type('Foo', (object,), _Foo_body())
As a result, the name for the class is not assigned to until the class statement has completed executing. You can't use the name inside the class statement until that statement has completed, in the same way that you can't use a function until the def statement has completed defining it.
This does mean you can dynamically create classes on the fly:
def class_with_base(base_class):
class Foo(base_class):
pass
return Foo
You can store those classes in a list:
classes = [class_with_base(base) for base in list_of_bases]
Now you have a list of classes with no global names referring to them anywhere. Without a global name, I can't rely on such a name existing in a method either; return Foo won't work as there is no Foo global for that to refer to.
Next, Python supports a concept called a metaclass, which produces classes just like a class produces instances. The type() function above is the default metaclass, but you are free to supply your own for a class. A metaclass is free to produce whatever it likes really, even things that are bit classes! As such Python cannot, up front, know what kind of object a class statement will produce and can't make assumptions about what it'll end up binding the name used to. See What is a metaclass in Python?
All this is not something you can do in a statically typed language like Java.
A class statement is executed just like any other statement. Your first example is (roughly) equivalent to
a = A()
A = type('A', (), {'static_attribute': a})
The first line obviously raises a NameError, because A isn't yet bound to anything.
In your second example, A isn't referenced until method is actually called, by which time A does refer to the class.
Essentially, a class does not exist until its entire definition is compiled in its entirety. This is similar to end blocks that are explicitly written in other languages, and Python utilizes implicit end blocks which are determined by indentation.
The other answers are great at explaining why you can't reference the class by name within the class, but you can use class methods to access the class.
The #classmethod decorator annotes a method that will be passed the class type, instead of the usual class instance (self). This is similar to Java's static method (there's also a #staticmethod decorator, which is a little different).
For a singleton, you can access a class instance to store an object instance (Attributes defined at the class level are the fields defined as static in a Java class):
class A(object):
instance = None
#classmethod
def get_singleton(cls):
if cls.instance is None:
print "Creating new instance"
cls.instance = cls()
return cls.instance
>>> a1 = A.get_singleton()
Creating new instance
>>> a2 = A.get_singleton()
>>> print a1 is a2
True
You can also use class methods to make java-style "static" methods:
class Name(object):
def __init__(self, name):
self.name = name
#classmethod
def make_as_victoria(cls):
return cls("Victoria")
#classmethod
def make_as_stephen(cls):
return cls("Stephen")
>>> victoria = Name.make_as_victoria()
>>> stephen = Name.make_as_stephen()
>>> print victoria.name
Victoria
>>> print stephen.name
Stephen
The answer is "just because".
It has nothing to do with the type system of Python, or it being dynamic. It has to do with the order in which a newly introduced type is initialized.
Some months ago I developed an object system for the language TXR, in which this works:
1> (defstruct foo nil (:static bar (new foo)))
#
2> (new foo)
#S(foo)
3> *2.bar
#S(foo)
Here, bar is a static slot ("class variable") in foo. It is initialized by an expression which constructs a foo.
Why that works can be understood from the function-based API for the instantiation of a new type, where the static class initialization is performed by a function which is passed in. The defstruct macro compiles a call to make-struct-type in which the (new foo) expression ends up in the body of the anonymous function that is passed for the static-initfun argument. This function is called after the type is registered under the foo symbol already.
We could easily patch the C implementation of make_struct_type so that this breaks. The last few lines of that function are:
sethash(struct_type_hash, name, stype);
if (super) {
mpush(stype, mkloc(su->dvtypes, super));
memcpy(st->stslot, su->stslot, sizeof (val) * su->nstslots);
}
call_stinitfun_chain(st, stype);
return stype;
}
The call_stinifun_chain does the initialization which ends up evaluating (new foo) and storing it in the bar static slot, and the sethash call is what registers the type under its name.
If we simply reverse the order in which these functions are called, the language and type system will still be the same, and almost everything will work as before. Yet, the (:static bar (new foo)) slot specifier will fail.
I put the calls in that order because I wanted the language-controlled aspects of the type to be as complete as possible before exposing it to the user-definable initializations.
I can't think of any reason for foo not to be known at the time when that struct type is being initialized, let alone a good reason. It is legitimate for static construction to create an instance. For example, we could use it to create a "singleton".
This looks like a bug in Python.

Why is Python 3.x's super() magic?

In Python 3.x, super() can be called without arguments:
class A(object):
def x(self):
print("Hey now")
class B(A):
def x(self):
super().x()
>>> B().x()
Hey now
In order to make this work, some compile-time magic is performed, one consequence of which is that the following code (which rebinds super to super_) fails:
super_ = super
class A(object):
def x(self):
print("No flipping")
class B(A):
def x(self):
super_().x()
>>> B().x()
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "<stdin>", line 3, in x
RuntimeError: super(): __class__ cell not found
Why is super() unable to resolve the superclass at runtime without assistance from the compiler? Are there practical situations in which this behaviour, or the underlying reason for it, could bite an unwary programmer?
... and, as a side question: are there any other examples in Python of functions, methods etc. which can be broken by rebinding them to a different name?
The new magic super() behaviour was added to avoid violating the D.R.Y. (Don't Repeat Yourself) principle, see PEP 3135. Having to explicitly name the class by referencing it as a global is also prone to the same rebinding issues you discovered with super() itself:
class Foo(Bar):
def baz(self):
return super(Foo, self).baz() + 42
Spam = Foo
Foo = something_else()
Spam().baz() # liable to blow up
The same applies to using class decorators where the decorator returns a new object, which rebinds the class name:
#class_decorator_returning_new_class
class Foo(Bar):
def baz(self):
# Now `Foo` is a *different class*
return super(Foo, self).baz() + 42
The magic super() __class__ cell sidesteps these issues nicely by giving you access to the original class object.
The PEP was kicked off by Guido, who initially envisioned super becoming a keyword, and the idea of using a cell to look up the current class was also his. Certainly, the idea to make it a keyword was part of the first draft of the PEP.
However, it was in fact Guido himself who then stepped away from the keyword idea as 'too magical', proposing the current implementation instead. He anticipated that using a different name for super() could be a problem:
My patch uses an intermediate solution: it assumes you need __class__
whenever you use a variable named 'super'. Thus, if you (globally)
rename super to supper and use supper but not super, it won't work
without arguments (but it will still work if you pass it either
__class__ or the actual class object); if you have an unrelated
variable named super, things will work but the method will use the
slightly slower call path used for cell variables.
So, in the end, it was Guido himself that proclaimed that using a super keyword did not feel right, and that providing a magic __class__ cell was an acceptable compromise.
I agree that the magic, implicit behaviour of the implementation is somewhat surprising, but super() is one of the most mis-applied functions in the language. Just take a look at all the misapplied super(type(self), self) or super(self.__class__, self) invocations found on the Internet; if any of that code was ever called from a derived class you'd end up with an infinite recursion exception. At the very least the simplified super() call, without arguments, avoids that problem.
As for the renamed super_; just reference __class__ in your method as well and it'll work again. The cell is created if you reference either the super or __class__ names in your method:
>>> super_ = super
>>> class A(object):
... def x(self):
... print("No flipping")
...
>>> class B(A):
... def x(self):
... __class__ # just referencing it is enough
... super_().x()
...
>>> B().x()
No flipping

Python's equivalent of .Net's sealed class

Does python have anything similar to a sealed class? I believe it's also known as final class, in java.
In other words, in python, can we mark a class so it can never be inherited or expanded upon? Did python ever considered having such a feature? Why?
Disclaimers
Actually trying to understand why sealed classes even exist. Answer here (and in many, many, many, many, many, really many other places) did not satisfy me at all, so I'm trying to look from a different angle. Please, avoid theoretical answers to this question, and focus on the title! Or, if you'd insist, at least please give one very good and practical example of a sealed class in csharp, pointing what would break big time if it was unsealed.
I'm no expert in either language, but I do know a bit of both. Just yesterday while coding on csharp I got to know about the existence of sealed classes. And now I'm wondering if python has anything equivalent to that. I believe there is a very good reason for its existence, but I'm really not getting it.
You can use a metaclass to prevent subclassing:
class Final(type):
def __new__(cls, name, bases, classdict):
for b in bases:
if isinstance(b, Final):
raise TypeError("type '{0}' is not an acceptable base type".format(b.__name__))
return type.__new__(cls, name, bases, dict(classdict))
class Foo:
__metaclass__ = Final
class Bar(Foo):
pass
gives:
>>> class Bar(Foo):
... pass
...
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "<stdin>", line 5, in __new__
TypeError: type 'Foo' is not an acceptable base type
The __metaclass__ = Final line makes the Foo class 'sealed'.
Note that you'd use a sealed class in .NET as a performance measure; since there won't be any subclassing methods can be addressed directly. Python method lookups work very differently, and there is no advantage or disadvantage, when it comes to method lookups, to using a metaclass like the above example.
Before we talk Python, let's talk "sealed":
I, too, have heard that the advantage of .Net sealed / Java final / C++ entirely-nonvirtual classes is performance. I heard it from a .Net dev at Microsoft, so maybe it's true. If you're building a heavy-use, highly-performance-sensitive app or framework, you may want to seal a handful of classes at or near the real, profiled bottleneck. Particularly classes that you are using within your own code.
For most applications of software, sealing a class that other teams consume as part of a framework/library/API is kinda...weird.
Mostly because there's a simple work-around for any sealed class, anyway.
I teach "Essential Test-Driven Development" courses, and in those three languages, I suggest consumers of such a sealed class wrap it in a delegating proxy that has the exact same method signatures, but they're override-able (virtual), so devs can create test-doubles for these slow, nondeterministic, or side-effect-inducing external dependencies.
[Warning: below snark intended as humor. Please read with your sense of humor subroutines activated. I do realize that there are cases where sealed/final are necessary.]
The proxy (which is not test code) effectively unseals (re-virtualizes) the class, resulting in v-table look-ups and possibly less efficient code (unless the compiler optimizer is competent enough to in-line the delegation). The advantages are that you can test your own code efficiently, saving living, breathing humans weeks of debugging time (in contrast to saving your app a few million microseconds) per month... [Disclaimer: that's just a WAG. Yeah, I know, your app is special. ;-]
So, my recommendations: (1) trust your compiler's optimizer, (2) stop creating unnecessary sealed/final/non-virtual classes that you built in order to either (a) eke out every microsecond of performance at a place that is likely not your bottleneck anyway (the keyboard, the Internet...), or (b) create some sort of misguided compile-time constraint on the "junior developers" on your team (yeah...I've seen that, too).
Oh, and (3) write the test first. ;-)
Okay, yes, there's always link-time mocking, too (e.g. TypeMock). You got me. Go ahead, seal your class. Whatevs.
Back to Python: The fact that there's a hack rather than a keyword is probably a reflection of the pure-virtual nature of Python. It's just not "natural."
By the way, I came to this question because I had the exact same question. Working on the Python port of my ever-so-challenging and realistic legacy-code lab, and I wanted to know if Python had such an abominable keyword as sealed or final (I use them in the Java, C#, and C++ courses as a challenge to unit testing). Apparently it doesn't. Now I have to find something equally challenging about untested Python code. Hmmm...
Python does have classes that can't be extended, such as bool or NoneType:
>>> class ExtendedBool(bool):
... pass
...
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
TypeError: type 'bool' is not an acceptable base type
However, such classes cannot be created from Python code. (In the CPython C API, they are created by not setting the Py_TPFLAGS_BASETYPE flag.)
Python 3.6 will introduce the __init_subclass__ special method; raising an error from it will prevent creating subclasses. For older versions, a metaclass can be used.
Still, the most “Pythonic” way to limit usage of a class is to document how it should not be used.
Similar in purpose to a sealed class and useful to reduce memory usage (Usage of __slots__?) is the __slots__ attribute which prevents monkey patching a class. Because when the metaclass __new__ is called, it is too late to put a __slots__ into the class, we have to put it into the namespace at the first possible timepoint, i.e. during __prepare__. Additionally, this throws the TypeError a little bit earlier. Using mcs for the isinstance comparison removes the necessity to hardcode the metaclass name in itself. The disadvantage is that all unslotted attributes are read-only. Therefore, if we want to set specific attributes during initialization or later, they have to slotted specifically. This is feasible e.g. by using a dynamic metaclass taking slots as an argument.
def Final(slots=[]):
if "__dict__" in slots:
raise ValueError("Having __dict__ in __slots__ breaks the purpose")
class _Final(type):
#classmethod
def __prepare__(mcs, name, bases, **kwargs):
for b in bases:
if isinstance(b, mcs):
msg = "type '{0}' is not an acceptable base type"
raise TypeError(msg.format(b.__name__))
namespace = {"__slots__":slots}
return namespace
return _Final
class Foo(metaclass=Final(slots=["_z"])):
y = 1
def __init__(self, z=1):
self.z = 1
#property
def z(self):
return self._z
#z.setter
def z(self, val:int):
if not isinstance(val, int):
raise TypeError("Value must be an integer")
else:
self._z = val
def foo(self):
print("I am sealed against monkey patching")
where the attempt of overwriting foo.foo will throw AttributeError: 'Foo' object attribute 'foo' is read-only and attempting to add foo.x will throw AttributeError: 'Foo' object has no attribute 'x'. The limiting power of __slots__ would be broken when inheriting, but because Foo has the metaclass Final, you can't inherit from it. It would also be broken when dict is in slots, so we throw a ValueError in case. To conclude, defining setters and getters for slotted properties allows to limit how the user can overwrite them.
foo = Foo()
# attributes are accessible
foo.foo()
print(foo.y)
# changing slotted attributes is possible
foo.z = 2
# %%
# overwriting unslotted attributes won't work
foo.foo = lambda:print("Guerilla patching attempt")
# overwriting a accordingly defined property won't work
foo.z = foo.foo
# expanding won't work
foo.x = 1
# %% inheriting won't work
class Bar(Foo):
pass
In that regard, Foo could not be inherited or expanded upon. The disadvantage is that all attributes have to be explicitly slotted, or are limited to a read-only class variable.
Python 3.8 has that feature in the form of the typing.final decorator:
class Base:
#final
def done(self) -> None:
...
class Sub(Base):
def done(self) -> None: # Error reported by type checker
...
#final
class Leaf:
...
class Other(Leaf): # Error reported by type checker
See https://docs.python.org/3/library/typing.html#typing.final

Why Is The property Decorator Only Defined For Classes?

tl;dr: How come property decorators work with class-level function definitions, but not with module-level definitions?
I was applying property decorators to some module-level functions, thinking they would allow me to invoke the methods by mere attribute lookup.
This was particularly tempting because I was defining a set of configuration functions, like get_port, get_hostname, etc., all of which could have been replaced with their simpler, more terse property counterparts: port, hostname, etc.
Thus, config.get_port() would just be the much nicer config.port
I was surprised when I found the following traceback, proving that this was not a viable option:
TypeError: int() argument must be a string or a number, not 'property'
I knew I had seen some precedant for property-like functionality at module-level, as I had used it for scripting shell commands using the elegant but hacky pbs library.
The interesting hack below can be found in the pbs library source code. It enables the ability to do property-like attribute lookups at module-level, but it's horribly, horribly hackish.
# this is a thin wrapper around THIS module (we patch sys.modules[__name__]).
# this is in the case that the user does a "from pbs import whatever"
# in other words, they only want to import certain programs, not the whole
# system PATH worth of commands. in this case, we just proxy the
# import lookup to our Environment class
class SelfWrapper(ModuleType):
def __init__(self, self_module):
# this is super ugly to have to copy attributes like this,
# but it seems to be the only way to make reload() behave
# nicely. if i make these attributes dynamic lookups in
# __getattr__, reload sometimes chokes in weird ways...
for attr in ["__builtins__", "__doc__", "__name__", "__package__"]:
setattr(self, attr, getattr(self_module, attr))
self.self_module = self_module
self.env = Environment(globals())
def __getattr__(self, name):
return self.env[name]
Below is the code for inserting this class into the import namespace. It actually patches sys.modules directly!
# we're being run as a stand-alone script, fire up a REPL
if __name__ == "__main__":
globs = globals()
f_globals = {}
for k in ["__builtins__", "__doc__", "__name__", "__package__"]:
f_globals[k] = globs[k]
env = Environment(f_globals)
run_repl(env)
# we're being imported from somewhere
else:
self = sys.modules[__name__]
sys.modules[__name__] = SelfWrapper(self)
Now that I've seen what lengths pbs has to go through, I'm left wondering why this facility of Python isn't built into the language directly. The property decorator in particular seems like a natural place to add such functionality.
Is there any partiuclar reason or motivation for why this isn't built directly in?
This is related to a combination of two factors: first, that properties are implemented using the descriptor protocol, and second that modules are always instances of a particular class rather than being instantiable classes.
This part of the descriptor protocol is implemented in object.__getattribute__ (the relevant code is PyObject_GenericGetAttr starting at line 1319). The lookup rules go like this:
Search through the class mro for a type dictionary that has name
If the first matching item is a data descriptor, call its __get__ and return its result
If name is in the instance dictionary, return its associated value
If there was a matching item from the class dictionaries and it was a non-data descriptor, call its __get__ and return the result
If there was a matching item from the class dictionaries, return it
raise AttributeError
The key to this is at number 3 - if name is found in the instance dictionary (as it will be with modules), then its value will just be returned - it won't be tested for descriptorness, and its __get__ won't be called. This leads to this situation (using Python 3):
>>> class F:
... def __getattribute__(self, attr):
... print('hi')
... return object.__getattribute__(self, attr)
...
>>> f = F()
>>> f.blah = property(lambda: 5)
>>> f.blah
hi
<property object at 0xbfa1b0>
You can see that .__getattribute__ is being invoked, but isn't treating f.blah as a descriptor.
It is likely that the reason for the rules being structured this way is an explicit tradeoff between the usefulness of allowing descriptors on instances (and, therefore, in modules) and the extra code complexity that this would lead to.
Properties are a feature specific to classes (new-style classes specifically) so by extension the property decorator can only be applied to class methods.
A new-style class is one that derives from object, i.e. class Foo(object):
Further info: Can modules have properties the same way that objects can?

Categories