How to add property to a class dynamically?

Anthony Kong picture Anthony Kong · Aug 25, 2009 · Viewed 247.8k times · Source

The goal is to create a mock class which behaves like a db resultset.

So for example, if a database query returns, using a dict expression, {'ab':100, 'cd':200}, then I would like to see:

>>> dummy.ab
100

At first I thought maybe I could do it this way:

ks = ['ab', 'cd']
vs = [12, 34]
class C(dict):
    def __init__(self, ks, vs):
        for i, k in enumerate(ks):
            self[k] = vs[i]
            setattr(self, k, property(lambda x: vs[i], self.fn_readyonly))

    def fn_readonly(self, v)
        raise "It is ready only"

if __name__ == "__main__":
    c = C(ks, vs)
    print c.ab

but c.ab returns a property object instead.

Replacing the setattr line with k = property(lambda x: vs[i]) is of no use at all.

So what is the right way to create an instance property at runtime?

P.S. I am aware of an alternative presented in How is the __getattribute__ method used?

Answer

Eevee picture Eevee · Aug 31, 2009

I suppose I should expand this answer, now that I'm older and wiser and know what's going on. Better late than never.

You can add a property to a class dynamically. But that's the catch: you have to add it to the class.

>>> class Foo(object):
...     pass
... 
>>> foo = Foo()
>>> foo.a = 3
>>> Foo.b = property(lambda self: self.a + 1)
>>> foo.b
4

A property is actually a simple implementation of a thing called a descriptor. It's an object that provides custom handling for a given attribute, on a given class. Kinda like a way to factor a huge if tree out of __getattribute__.

When I ask for foo.b in the example above, Python sees that the b defined on the class implements the descriptor protocol—which just means it's an object with a __get__, __set__, or __delete__ method. The descriptor claims responsibility for handling that attribute, so Python calls Foo.b.__get__(foo, Foo), and the return value is passed back to you as the value of the attribute. In the case of property, each of these methods just calls the fget, fset, or fdel you passed to the property constructor.

Descriptors are really Python's way of exposing the plumbing of its entire OO implementation. In fact, there's another type of descriptor even more common than property.

>>> class Foo(object):
...     def bar(self):
...         pass
... 
>>> Foo().bar
<bound method Foo.bar of <__main__.Foo object at 0x7f2a439d5dd0>>
>>> Foo().bar.__get__
<method-wrapper '__get__' of instancemethod object at 0x7f2a43a8a5a0>

The humble method is just another kind of descriptor. Its __get__ tacks on the calling instance as the first argument; in effect, it does this:

def __get__(self, instance, owner):
    return functools.partial(self.function, instance)

Anyway, I suspect this is why descriptors only work on classes: they're a formalization of the stuff that powers classes in the first place. They're even the exception to the rule: you can obviously assign descriptors to a class, and classes are themselves instances of type! In fact, trying to read Foo.bar still calls property.__get__; it's just idiomatic for descriptors to return themselves when accessed as class attributes.

I think it's pretty cool that virtually all of Python's OO system can be expressed in Python. :)

Oh, and I wrote a wordy blog post about descriptors a while back if you're interested.