Glossary

dict classes

A regular class whose attributes are stored in the object.__dict__ attribute of every single instance. This is quite wasteful especially for objects with very few data attributes and the space consumption can become significant when creating large numbers of instances.

This is the type of class you get by default both with and without attrs (except with the next APIs attr.define, attr.mutable, and attr.frozen).

slotted classes

A class whose instances have no object.__dict__ attribute and define their attributes in a object.__slots__ attribute instead. In attrs, they are created by passing slots=True to @attr.s (and are on by default in attr.define/attr.mutable/attr.frozen).

Their main advantage is that they use less memory on CPython 1 and are slightly faster.

However they also come with several possibly surprising gotchas:

  • Slotted classes don’t allow for any other attribute to be set except for those defined in one of the class’ hierarchies __slots__:

    >>> import attr
>>> @attr.s(slots=True)
... class Coordinates(object):
...     x = attr.ib()
...     y = attr.ib()
...
>>> c = Coordinates(x=1, y=2)
>>> c.z = 3
Traceback (most recent call last):
    ...
AttributeError: 'Coordinates' object has no attribute 'z'

  • Slotted classes can inherit from other classes just like non-slotted classes, but some of the benefits of slotted classes are lost if you do that. If you must inherit from other classes, try to inherit only from other slotted classes.

  • However, it’s not possible to inherit from more than one class that has attributes in __slots__ (you will get an TypeError: multiple bases have instance lay-out conflict).

  • It’s not possible to monkeypatch methods on slotted classes. This can feel limiting in test code, however the need to monkeypatch your own classes is usually a design smell.

    If you really need to monkeypatch an instance in your tests, but don’t want to give up on the advantages of slotted classes in production code, you can always subclass a slotted class as a dict class with no further changes and all the limitations go away:

    >>> import attr, unittest.mock
>>> @attr.s(slots=True)
... class Slotted(object):
...     x = attr.ib()
...
...     def method(self):
...         return self.x
>>> s = Slotted(42)
>>> s.method()
42
>>> with unittest.mock.patch.object(s, "method", return_value=23):
...     pass
Traceback (most recent call last):
   ...
AttributeError: 'Slotted' object attribute 'method' is read-only
>>> @attr.s  # implies 'slots=False'
... class Dicted(Slotted):
...     pass
>>> d = Dicted(42)
>>> d.method()
42
>>> with unittest.mock.patch.object(d, "method", return_value=23):
...     assert 23 == d.method()

  • Slotted classes must implement __getstate__ and __setstate__ to be serializable with pickle protocol 0 and 1. Therefore, attrs creates these methods automatically for slots=True classes (Python 2 uses protocol 0 by default).

    Note

    If the @attr.s(slots=True) decorated class already implements the __getstate__ and __setstate__ methods, they will be overwritten by attrs autogenerated implementation by default.

    This can be avoided by setting @attr.s(getstate_setstate=False) or by setting @attr.s(auto_detect=True).

    Also, think twice before using pickle.

  • Slotted classes are weak-referenceable by default. This can be disabled in CPython by passing weakref_slot=False to @attr.s 2.

  • Since it’s currently impossible to make a class slotted after it’s been created, attrs has to replace your class with a new one. While it tries to do that as graciously as possible, certain metaclass features like object.__init_subclass__ do not work with slotted classes.

  • The class.__subclasses__ attribute needs a garbage collection run (which can be manually triggered using gc.collect), for the original class to be removed. See issue #407 for more details.

1

On PyPy, there is no memory advantage in using slotted classes.

2

On PyPy, slotted classes are naturally weak-referenceable so weakref_slot=False has no effect.