When type is called with three arguments it behaves as the (meta)class it is, and creates a new instance, ie. it produces a new class/type.

Dummy = type('OtherDummy', (), dict(x=1))
Dummy.__class__              # <type 'type'> 
Dummy().__class__.__class__  # <type 'type'> 

It is possible to subclass type to create an custom metaclass.

class mytype(type):
    def __init__(cls, name, bases, dict):
        # call the base initializer
        type.__init__(cls, name, bases, dict)

        # perform custom initialization...
        cls.__custom_attribute__ = 2

Now, we have a new custom mytype metaclass which can be used to create classes in the same manner as type.

MyDummy = mytype('MyDummy', (), dict(x=2))
MyDummy.__class__              # <class '__main__.mytype'>
MyDummy().__class__.__class__  # <class '__main__.mytype'>
MyDummy.__custom_attribute__   # 2

When we create a new class using the class keyword the metaclass is by default chosen based on upon the baseclasses.

>>> class Foo(object):
...     pass

>>> type(Foo)
type

In the above example the only baseclass is object so our metaclass will be the type of object, which is type. It is possible override the default, however it depends on whether we use Python 2 or Python 3:

A special class-level attribute __metaclass__ can be used to specify the metaclass.

class MyDummy(object):
    __metaclass__ = mytype
type(MyDummy)  # <class '__main__.mytype'>

A special metaclass keyword argument specify the metaclass.

class MyDummy(metaclass=mytype):
    pass
type(MyDummy)  # <class '__main__.mytype'>

Any keyword arguments (except metaclass) in the class declaration will be passed to the metaclass. Thus class MyDummy(metaclass=mytype, x=2) will pass x=2 as a keyword argument to the mytype constructor.

Read this in-depth description of python meta-classes (opens new window) for more details.

A singleton is a pattern that restricts the instantiation of a class to one instance/object. For more info on python singleton design patterns, see here (opens new window).

class SingletonType(type):
    def __call__(cls, *args, **kwargs):
        try:
            return cls.__instance
        except AttributeError:
            cls.__instance = super(SingletonType, cls).__call__(*args, **kwargs)
            return cls.__instance

class MySingleton(object):
    __metaclass__ = SingletonType

class MySingleton(metaclass=SingletonType):
    pass

MySingleton() is MySingleton()  # True, only one instantiation occurs

Metaclass syntax

class MyClass(object):
    __metaclass__ = SomeMetaclass

class MyClass(metaclass=SomeMetaclass):
    pass

Python 2 and 3 compatibility with six

import six

class MyClass(six.with_metaclass(SomeMetaclass)):
    pass

What is a metaclass?

In Python, everything is an object: integers, strings, lists, even functions and classes themselves are objects. And every object is an instance of a class.

To check the class of an object x, one can call type(x), so:

>>> type(5)
<type 'int'>
>>> type(str)
<type 'type'>
>>> type([1, 2, 3])
<type 'list'>

>>> class C(object):
...     pass
...
>>> type(C)
<type 'type'>    

Most classes in python are instances of type. type itself is also a class. Such classes whose instances are also classes are called metaclasses.

The Simplest Metaclass

OK, so there is already one metaclass in Python: type. Can we create another one?

class SimplestMetaclass(type):
    pass

class MyClass(object):
    __metaclass__ = SimplestMetaclass

That does not add any functionality, but it is a new metaclass, see that MyClass is now an instance of SimplestMetaclass:

>>> type(MyClass)
<class '__main__.SimplestMetaclass'>

A Metaclass which does Something

A metaclass which does something usually overrides type's __new__, to modify some properties of the class to be created, before calling the original __new__ which creates the class:

class AnotherMetaclass(type):
    def __new__(cls, name, parents, dct):
        # cls is this class
        # name is the name of the class to be created
        # parents is the list of the class's parent classes
        # dct is the list of class's attributes (methods, static variables)

        # here all of the attributes can be modified before creating the class, e.g.

        dct['x'] = 8  # now the class will have a static variable x = 8

        # return value is the new class. super will take care of that
        return super(AnotherMetaclass, cls).__new__(cls, name, parents, dct)

Functionality in metaclasses can be changed so that whenever a class is built, a string is printed to standard output, or an exception is thrown. This metaclass will print the name of the class being built.

class VerboseMetaclass(type):

    def __new__(cls, class_name, class_parents, class_dict):
        print("Creating class ", class_name)
        new_class = super().__new__(cls, class_name, class_parents, class_dict)
        return new_class

You can use the metaclass like so:

class Spam(metaclass=VerboseMetaclass):
    def eggs(self):
        print("[insert example string here]")
s = Spam()
s.eggs()

The standard output will be:

Creating class Spam
[insert example string here]

You may have heard that everything in Python is an object. It is true, and all objects have a class:

>>> type(1)
int

The literal 1 is an instance of int. Lets declare a class:

>>> class Foo(object):
...    pass
...

Now lets instantiate it:

>>> bar = Foo()

What is the class of bar?

>>> type(bar)
Foo

Nice, bar is an instance of Foo. But what is the class of Foo itself?

>>> type(Foo)
type

Ok, Foo itself is an instance of type. How about type itself?

>>> type(type)
type

So what is a metaclass? For now lets pretend it is just a fancy name for the class of a class. Takeaways:

• Everything is an object in Python, so everything has a class
• The class of a class is called a metaclass
• The default metaclass is type, and by far it is the most common metaclass

But why should you know about metaclasses? Well, Python itself is quite "hackable", and the concept of metaclass is important if you are doing advanced stuff like meta-programming or if you want to control how your classes are initialized.

Remarks

When designing your architecture, consider that many things which can be accomplished with metaclasses can also be accomplished using more simple semantics:

• Traditional inheritance is often more than enough.
• Class decorators can mix-in functionality into a classes on a ad-hoc approach.
• Python 3.6 introduces __init_subclass__() which allows a class to partake in the creation of its subclass.