A note on Python (class) attributes

Without an in-depth reading of the Python Data Model, there are a few things about attributes in Python that might be surprising:

  • It is possible to add new attributes to any instances of “normal” user-defined classes after instantiation. This is because every class uses a __dict__ under the hood for keeping track of attributes.

  • Assigning to an instance’s class attribute name does not set the value of the class attribute. Instead, it creates a new instance attribute with the same name, which shadows the class attribute

  • The only way (apart from things like C extension modules) to define a class that does not allow to create new attributes after instantiation is to use the __slots__ mechanism. This results in __dict__ not being used.

Consider the following definition:

class MyClass(object):
    a = 1 # class attribute
    def __init__(self):
        self.b = 1 # instance attribute

inst1 = MyClass()
inst2 = MyClass()

All instances share the same class attribute, so any changes in the attribute value are reflected in all instances:

>>> MyClass.a = 2
>>> inst1.a == inst2.a == 2
True

As a class attribute, a (in contrast to the instance attribute b) is not in inst1.__dict__:

>>> inst1.__dict__
{'b': 1}

An important feature of instance attributes is that new attributes can be created not just in the __init__ method, but even after the class is instantiated. Needless to say, this just shows how important proper unit testing is for Python development – a simple typo could end up creating a new attribute instead of changing an existing attribute value.

>>> inst1.new_attr = 0
>>> inst1.__dict__
{'b': 1, 'new_attr': 0}

This, combined with the way Python resolves attribute names, means that the class attribute value can not be set as

>>> inst1.a = 3

Instead of setting the class attribute value, this creates a new instance attribute a:

>>> inst1.a
3
>>> inst2.a
2
>>> inst1.__dict__
{'b': 1, 'new_attr': 0, 'a': 3}

It is even possible do delete the instance attribute again, “un-shadowing” the class attribute:

>>> inst1.a
3
>>> del inst1.a
>>> inst1.a
2

There is a gist for a test that explores this more systematically, and includes a comparison to classes defined with __slots__ (you should not make a habit of needlessly preventing dynamic attribute creation with __slots__!).

As a final note, it is recommended to access class attributes and instance attributes in the same way whenever possible (i.e. inst1.a instead of inst1.__class__.a or MyClass.a when working with the instance, and self.a instead of self.__class__.a inside methods). This gives the flexibility to treat class attributes as defaults for instance attributes