Orignally published: May 8, 2005
Table of Contents
- Why New-Style Classes?
- Static Methods
- Class Methods
- Attribute Slots
- The Constructor __new__
- Cooperative Super Call
New-style classes are part of an effort to unify built-in types and user-defined classes in the Python programming language. New-style classes have been around since Python 2.2 (not that new anymore), so it’s definitely time to take advantage of the new possibilities.
A new-style class is one that is derived, either directly or indirectly, from a built-in type. (Something that wasn’t possible at all before Python 2.2.) Built-in types include types such as:
- and others
The base class for all new-style classes is called
object. All of the following are new-style classes:
class NewStyleUserDefinedClass(object): pass class DerivedFromBuiltInType(list): pass class IndirectlyDerivedFromType(DerivedFromBuiltInType): pass
Here’s what new-style classes have to offer:
- Properties: Attributes that are defined by get/set methods
- Static methods and class methods
- The new
__getattribute__hook, which, unlike
__getattr__, is called for every attribute access, not just when the attribute can’t be found in the instance
- Descriptors: A protocol to define the behavior of attribute access through objects
- Overriding the constructor
- Metaclasses (not discussed)
I’ll try to be very brief, yet to give you enough information so that you can start using these language features. Examples are presented in place of long descriptions. Once your interest is awakened, you can refer to the References section for more detailed material on these topics.
A property is an attribute that is defined by get/set methods. The concept is simple and well-known from other languages. A property is defined like this:
class ClassWithProperty(object): ... TheProperty = property(fget=<the get method>, fset=<the set method>, fdel=<the del method>, doc=<the docstring>)
The signature of the
property descriptor is
property(fget=None, fset=None, fdel=None, doc=None). If any of the methods is not specified, an exception of type
AttributeError is raised when the respective operation is attempted. For example, to define a read-only property, you would specify
fget but not
fset. (Write-only properties are also possible, although a regular method achieves the same thing in a less awkward way.)
Here’s a more complete example:
class ClassWithProperty(object): def __SetTheProperty(self, value): print "Setting the property" self.__m_the_property = value def __GetTheProperty(self): print "Getting the property" return self.__m_the_property def __DelTheProperty(self): print "Deleting the property" del self.__m_the_property TheProperty = property(fget=__GetTheProperty, fset=__SetTheProperty, fdel=__DelTheProperty, doc="The property description.") def __GetReadOnlyProperty(self): return "This is a calculated value." ReadOnlyProperty = property(fget=__GetReadOnlyProperty)
The property is used like this:
>>> c = ClassWithProperty() >>> c.TheProperty = 10 Setting the property >>> print c.TheProperty Getting the property 10 >>> del c.TheProperty Deleting the property >>> # The property itself is still there after deleting >>> c.TheProperty = 5 Setting the property >>> c.TheProperty.__doc__ 'The property description.' >>> print c.ReadOnlyProperty This is a calculated value. >>> c.ReadOnlyProperty = 100 Traceback (most recent call last): File "<interactive input>", line 1, in ? AttributeError: can't set attribute
Note: Don’t forget to derive your class from
object, otherwise properties won’t work.
There’s not much to explain about static methods; they behave just like in C++. A static method is defined using the
class MyClass(object): def SomeMethod(x): print x SomeMethod = staticmethod(SomeMethod) >>> MyClass.SomeMethod(15) 15 >>> obj = MyClass() >>> obj.SomeMethod(15) 15
You should really consider creating a static method whenever a method does not make substantial use of the instance (
A class method is similar to a static method in that it has no
self argument. Instead, it receives a class as its first argument. By convention, this argument is called
cls. A class method is defined using the
class MyClass(object): def SomeMethod(cls, x): print cls, x SomeMethod = classmethod(SomeMethod) class DerivedClass(MyClass): pass >>> MyClass.SomeMethod(15) <class '__main__.MyClass'> 15 >>> obj = MyClass() >>> obj.SomeMethod(15) <class '__main__.MyClass'> 15 >>> DerivedClass.SomeMethod(150) <class '__main__.DerivedClass'> 150
In the last call, you can see that only the class involved in making the method call defines the value of the
cls argument. This is despite the fact that the method has been defined in a different class.
We have already seen three different descriptors:
But what exactly is a descriptor?
In general, a descriptor is an object attribute with “binding behavior”, one whose attribute access has been overridden by methods in the descriptor protocol. Those methods are
__delete__. If any of those methods are defined for an object, it is said to be a descriptor. 
When executing the assignment
x.m = y, then
m may be an object that defines a
__set__ method. If that’s the case, that method is called to perform the assignment.
The following example shows how to define each of the three methods from the descriptor protocol:
class MyDescriptor(object): def __get__(self, obj, type=None): print "get", self, obj, type return "The value" def __set__(self, obj, value): print "set", self, obj, val return None def __delete__(self, obj): print "delete", self, obj return None class SomeClass(object): m = MyDescriptor()
Note: Both classes must be derived from
Now we can start using the descriptor:
>>> x = SomeClass() >>> print x.m get <__main__.MyDescriptor object at 0x12345678> <__main__.SomeClass object at 0x23456789> <class '__main__.SomeClass'> The value >>> x.m = 1000 set <__main__.MyDescriptor object at 0x12345678> <__main__.SomeClass object at 0x23456789> 1000 >>> del x.m delete <__main__.MyDescriptor object at 0x12345678> <__main__.SomeClass object at 0x23456789>
Here’s how the descriptor methods get called:
- When writing an attribute, the
__setattr__method invokes the descriptor’s
- When reading an attribute, the
__getattribute__method invokes the descriptor’s
- When deleting an attribute, the
__delattr__method invokes the descriptor’s
A few caveats:
- Both the descriptor class and the class using it must be new-style classes.
- When overriding
__delattr__, make sure to invoke the inherited method. (That is, extend these methods; don’t override them.) Otherwise, the descriptor mechanism will stop working.
As a sidenote: In his How-To Guide , Raymond Hettinger has this to say about the difference between data and non-data descriptors:
If an object defines both
__set__, it is considered a data descriptor. Descriptors that only define
__get__ are called non-data descriptors (they are typically used for methods but other uses are possible).
Data and non-data descriptors differ in how overrides are calculated with respect to entries in an instance’s dictionary. If an instance’s dictionary has an entry with the same name as a data descriptor, the data descriptor takes precedence. If an instance’s dictionary has an entry with the same name as a non-data descriptor, the dictionary entry takes precedence.
To make a read-only data descriptor, define both
__set__ with the
__set__ raising an
AttributeError when called. Defining the
__set__ method with an exception raising placeholder is enough to make it a data descriptor.
However, I did not manage to add a dictionary entry in such a way that a non-data descriptor’s
__get__ stopped being called. Maybe I tried the wrong things, maybe I’m misunderstanding something. In the meantime, I’ll just follow Raymond’s advice about defining both
__set__ for read-only data descriptors.
An interesting new feature that I discovered in Guido van Rossum’s paper on new-style classes  is that of “slots”. Here’s how it works:
class X(object): __slots__ = ["m", "n"] >>> x = X() >>> x.m = 10 >>> x.n = 10 >>> x.k = 3 Traceback (most recent call last): File "<interactive input>", line 1, in ? AttributeError: 'X' object has no attribute 'k'
__slots__ reserves space for the listed variables directly in the instance. Classes that define slots don’t have an instance dictionary (
__dict__). If you try to assign to an attribute that’s not in
__slots__, you receive an error. This may be quite useful for
struct-like classes, because it prevents problems with misspelled attribute names.
The main purpose seems to be with classes that derive from built-in types. For example, a derived
dict can have a few slots for all additional attributes that it needs. No second
__dict__ has to be created for these attributes, which saves space.
Just be warned that a slot in a derived class hides a slot of the same name in the base class.
If you are like me, then you probably always thought of the
__init__ method as the Python equivalent of what is called a constructor in C++. This isn’t the whole story.
When an instance of a class is created, Python first calls the
__new__ method of the class.
__new__ is a static method that is called with the class as its first argument.
__new__ returns a new instance of the class.
__init__ method is called afterwards to initialize the instance. In some situations (think “unplickling”!), no initialization is performed. Also, immutable types like
str are completely constructed by the
__new__ method; their
__init__ method does nothing. This way, it is impossible to circumvent immutability by explicitly calling the
__init__ method after construction.
There’s a new way in which Python handles method resolution in connection with multiple inheritance. Personally, I try to avoid multiple inheritance whenever possible, so I won’t go into detail here. However, what I did learn is that when someone else derives multiply from my own classes, it would be nice if my classes performed what’s called “cooperative super calls”. In short, instead of the old
<base-class>.<inherited-method>(self), one should call
class BaseClass: def Method(self): pass class DerivedClass(BaseClass): def Method(self): super(DerivedClass, self).Method()
Seems to be fairly easy, and if it helps…
We’ve seen how new-style classes can be used to
- Define properties
- Define static and class methods
- Define descriptors
- Assign attribute slots
- Override the constructor
I haven’t mentioned metaclasses at all. I haven’t used metaclasses myself, so I better refer you to the experts. There’s a number of links to metaclass-related articles on the New-style Classes page at python.org .