PEP: 3119 Title: Introducing Abstract Base Classes Version: $Revision$ Last-Modified: $Date$ Author: Guido van Rossum , Talin Status: Draft Type: Standards Track Content-Type: text/x-rst Created: 18-Apr-2007 Post-History: Not yet posted Abstract ======== This is a proposal to add Abstract Base Class (ABC) support to Python 3000. It proposes: * An "ABC support framework" which defines a built-in decorator that can be used to define abstract methods. A class containing an abstract method that isn't overridden cannot be instantiated. * Specific ABCs for containers and iterators, to be added to the collections module. Much of the thinking that went into the proposal is not about the specific mechanism of ABCs, as contrasted with Interfaces or Generic Functions (GFs), but about clarifying philosophical issues like "what makes a set", "what makes a mapping" and "what makes a sequence". Acknowledgements ---------------- Talin wrote the Rationale below [1]_ as well as most of the section on ABCs vs. Interfaces. For that alone he deserves co-authorship. The rest of the PEP uses "I" referring to the first author. Rationale ========= In the domain of object-oriented programming, the usage patterns for interacting with an object can be divided into two basic categories, which are 'invocation' and 'inspection'. Invocation means interacting with an object by invoking its methods. Usually this is combined with polymorphism, so that invoking a given method may run different code depending on the type of an object. Inspection means the ability for external code (outside of the object's methods) to examine the type or properties of that object, and make decisions on how to treat that object based on that information. Both usage patterns serve the same general end, which is to be able to support the processing of diverse and potentially novel objects in a uniform way, but at the same time allowing processing decisions to be customized for each different type of object. In classical OOP theory, invocation is the preferred usage pattern, and inspection is actively discouraged, being considered a relic of an earlier, procedural programming style. However, in practice this view is simply too dogmatic and inflexible, and leads to a kind of design rigidity that is very much at odds with the dynamic nature of a language like Python. In particular, there is often a need to process objects in a way that wasn't anticipated by the creator of the object class. It is not always the best solution to build in to every object methods that satisfy the needs of every possible user of that object. Moreover, there are many powerful dispatch philosophies that are in direct contrast to the classic OOP requirement of behavior being strictly encapsulated within an object, examples being rule or pattern-match driven logic. On the the other hand, one of the criticisms of inspection by classic OOP theorists is the lack of formalisms and the ad hoc nature of what is being inspected. In a language such as Python, in which almost any aspect of an object can be reflected and directly accessed by external code, there are many different ways to test whether an object conforms to a particular protocol or not. For example, if asking 'is this object a mutable sequence container?', one can look for a base class of 'list', or one can look for a method named '__getitem__'. But note that although these tests may seem obvious, neither of them are correct, as one generates false negatives, and the other false positives. The generally agreed-upon remedy is to standardize the tests, and group them into a formal arrangement. This is most easily done by associating with each class a set of standard testable properties, either via the inheritance mechanism or some other means. Each test carries with it a set of promises: it contains a promise about the general behavior of the class, and a promise as to what other class methods will be available. This PEP proposes a particular strategy for organizing these tests known as Abstract Base Classes, or ABC. ABCs are simply Python classes that are added into an object's inheritance tree to signal certain features of that object to an external inspector. Tests are done using isinstance(), and the presence of a particular ABC means that the test has passed. In addition, the ABCs define a minimal set of methods that establish the characteristic behavior of the type. Code that discriminates objects based on their ABC type can trust that those methods will always be present. Each of these methods are accompanied by an generalized abstract semantic definition that is described in the documentation for the ABC. These standard semantic definitions are not enforced, but are strongly recommended. Like all other things in Python, these promises are in the nature of a gentlemen's agreement, which in this case means that while the language does enforce some of the promises made in the ABC, it is up to the implementer of the concrete class to insure that the remaining ones are kept. Specification ============= The specification follows the categories listed in the abstract: * An "ABC support framework" which defines a built-in decorator that make it easy to define ABCs, and mechanisms to support it. * Specific ABCs for containers and iterators, to be added to the collections module. ABC Support Framework --------------------- We define a new built-in decorator, ``@abstractmethod``, to be used to declare abstract methods. A class containing at least one method declared with this decorator that hasn't been overridden yet cannot be instantiated. Such a methods may be called from the overriding method in the subclass (using ``super`` or direct invocation). For example:: class A: @abstractmethod def foo(self): pass A() # raises TypeError class B(A): pass B() # raises TypeError class C(A): def foo(self): print(42) C() # works **Implementation:** The ``@abstractmethod`` decorator sets the function attribute ``__isabstractmethod__`` to the value ``True``. The ``type.__new__`` method computes the type attribute ``__abstractmethods__`` as the set of all method names that have an ``__isabstractmethod__`` attribute whose value is true. It does this by combining the ``__abstractmethods__` attributes of the base classes, adding the names of all methods in the new class dict that have a true ``__isabstractmethod__`` attribute, and removing the names of all methods in the new class dict that don't have a true ``__isabstractmethod__`` attribute. If the resulting ``__abstractmethods__`` set is non-empty, the class is considered abstract, and attempts to instantiate it will raise ``TypeError``. (CPython can uses an internal flag ``Py_TPFLAGS_ABSTRACT`` to speed up this check [6]_.) **Discussion:** Unlike C++ or Java, abstract methods as defined here may have an implementation. This implementation can be called via the ``super`` mechanism from the class that overrides it. This could be useful as an end-point for a super-call in framework using a cooperative multiple-inheritance [7]_, [8]_. ABCs for Containers and Iterators --------------------------------- The ``collections`` module will define ABCs necessary and sufficient to work with sets, mappings, sequences, and some helper types such as iterators and dictionary views. The ABCs provide implementations of their abstract methods that are technically valid but fairly useless; e.g. ``__hash__`` returns 0, and ``__iter__`` returns an empty iterator. In general, the abstract methods represent the behavior of an empty container of the indicated type. Some ABCs also provide concrete (i.e. non-abstract) methods; for example, the ``Iterator`` class has an ``__iter__`` method returning itself, fulfilling an important invariant of iterators (which in Python 2 has to be implemented anew by each iterator class). No ABCs override ``__init__``, ``__new__``, ``__str__`` or ``__repr__``. Defining a standard constructor signature would unnecessarily constrain custom container types, for example Patricia trees or gdbm files. Defining a specific string representation for a collection is similarly left up to individual implementations. Ordering ABCs ''''''''''''' These ABCs are closer to ``object`` in the ABC hierarchy. ``PartiallyOrdered`` This ABC defines the 4 inequality operations ``<``, ``<=``, ``>=``, ``>``. (Note that ``==`` and ``!=`` are defined by ``object``.) Classes deriving from this ABC should satisfy weak invariants such as ``a < b < c`` implies ``a < c`` but don't require that for any two instances ``x`` and ``y`` exactly one of ``x < y``, ``x == y`` or ``x >= y`` apply. ``TotallyOrdered`` This ABC derives from ``PartiallyOrdered``. It adds no new operations but implies a promise of stronger invariants. **Open issues:** Should ``float`` derive from ``TotallyOrdered`` even though for ``NaN`` this isn't strictly correct? **Open issues:** Where should these live? The ``collections`` module doesn't seem right. One Trick Ponies '''''''''''''''' These abstract classes represent single methods like ``__iter__`` or ``__len__``. ``Hashable`` The base class for classes defining ``__hash__``. The ``__hash__`` method should return an ``Integer`` (see "Numbers" below). The abstract ``__hash__`` method always returns 0, which is a valid (albeit inefficient) implementation. **Invariant:** If classes ``C1`` and ``C2`` both derive from ``Hashable``, the condition ``o1 == o2`` must imply ``hash(o1) == hash(o2)`` for all instances ``o1`` of ``C1`` and all instances ``o2`` of ``C2``. IOW, two objects shouldn't compare equal but have different hash values. Another constraint is that hashable objects, once created, should never change their value (as compared by ``==``) or their hash value. If a class cannot guarantee this, it should not derive from ``Hashable``; if it cannot guarantee this for certain instances only, ``__hash__`` for those instances should raise a ``TypeError`` exception. Note: being an instance of this class does not imply that an object is immutable; e.g. a tuple containing a list as a member is not immutable; its ``__hash__`` method raises ``TypeError``. ``Iterable`` The base class for classes defining ``__iter__``. The ``__iter__`` method should always return an instance of ``Iterator`` (see below). The abstract ``__iter__`` method returns an empty iterator. ``Iterator`` The base class for classes defining ``__next__``. This derives from ``Iterable``. The abstract ``__next__`` method raises ``StopIteration``. The concrete ``__iter__`` method returns ``self``. (Note: this assumes PEP 3114 is implemented.) ``Sized`` The base class for classes defining ``__len__``. The ``__len__`` method should return an ``Integer`` (see "Numbers" below) >= 0. The abstract ``__len__`` method returns 0. **Invariant:** If a class ``C`` derives from ``Sized`` as well as from ``Iterable``, the invariant ``sum(1 for x in o) == len(o)`` should hold for any instance ``o`` of ``C``. **Open issues:** Is ``Sized`` the best name? Proposed alternatives already tentatively rejected: ``Finite`` (nobody understood it), ``Lengthy``, ``Sizeable`` (both too cute), ``Countable`` (the set of natural numbers is a countable set in math), ``Enumerable`` (sounds like a sysnonym for ``Iterable``), ``Dimension``, ``Extent`` (sound like numbers to me), ``Bounded`` (probably just as confusing as ``Fininte``). ``Container`` The base class for classes defining ``__contains__``. The ``__contains__`` method should return a ``bool``. The abstract ``__contains__`` method returns ``False``. **Invariant:** If a class ``C`` derives from ``Container`` as well as from ``Iterable``, then ``(x in o for x in o)`` should be a generator yielding only True values for any instance ``o`` of ``C``. Note: strictly speaking, there are three variants of this method's semantics. The first one is for sets and mappings, which is fast: O(1) or O(log N). The second one is for membership checking on sequences, which is slow: O(N). The third one is for subsequence checking on (character or byte) strings, which is also slow: O(N). Would it make sense to distinguish these? The signature of the third variant is different, since it takes a sequence (typically of the same type as the method's target) intead of an element. For now, I'm using the same type for all three. This means that is is possible for ``x in o`` to be True even though ``x`` is never yielded by ``iter(o)``. A suggested name for the third form is ``Searchable``. Sets '''' These abstract classes represent various stages of "set-ness". The most fundamental set operation is the membership test, written as ``x in s`` and implemented by ``s.__contains__(x)``. This is already taken care of by the `Container`` class defined above. Therefore, we define a set as finite, iterable container for which certain invariants from mathematical set theory hold. The built-in type ``set`` derives from ``MutableSet``. The built-in type ``frozenset`` derives from ``HashableSet``. You might wonder why we require a set to be finite -- surely certain infinite sets can be represented just fine in Python. For example, the set of even integers could be defined like this:: class EvenIntegers(Container): def __contains__(self, x): return x % 2 == 0 However, such sets have rather limited practical value, and deciding whether one such set is a subset of another would be difficult in general without using a symbolic algebra package. So I consider this out of the scope of a pragmatic proposal like this. ``Set`` This is a finite, iterable, partially ordered container, i.e. a subclass of ``Sized``, ``Iterable``, ``Container`` and ``PartiallyOrdered``. Not every subset of those three classes is a set though! Sets have the additional invariant that each element occurs only once (as can be determined by iteration), and in addition sets define concrete operators that implement the inequality operations as subclass/superclass tests. Sets with different implementations can be compared safely, efficiently and correctly. Because ``Set`` derives from ``Sized``, ``__eq__`` takes a shortcut and returns ``False`` immediately if two sets of unequal length are compared. Similarly, ``__le__`` returns ``False`` immediately if the first set has more members than the second set. Note that set inclusion implements only a partial ordering; e.g. ``{1, 2}`` and ``{1, 3}`` are not ordered (all three of ``<``, ``==`` and ``>`` return ``False`` for these arguments). Sets cannot be ordered relative to mappings or sequences, but they can be compared for equality (and then they always compare unequal). Note: the ``issubset`` and ``issuperset`` methods found on the set type in Python 2 are not supported, as these are mostly just aliases for ``__le__`` and ``__ge__``. **Open issues:** Spell out the invariants and method definitions. ``ComposableSet`` This is a subclass of ``Set`` that defines abstract operators to compute union, intersection, symmetric and asymmetric difference, respectively ``__or__``, ``__and__``, ``__xor__`` and ``__sub__``. These operators should return instances of ``ComposableSet``. The abstract implementations return no meaningful values but raise ``NotImplementedError``; this is because any generic implementation would have to create new instances but the ABCs don't (and shouldn't, IMO) provide an API for creating new instances. **Invariants:** The implementations of these operators should ensure that the results match the mathematical definition of set composition. **Open issues:** Should I spell out the invariants? Should we define an API for creating new instances (e.g. a class method or a fixed constructor signature)? Should we just pick a concrete return type (e.g. ``set``)? ``HashableSet`` This is a subclass of both ``ComposableSet`` and ``Hashable``. It implements a concrete ``__hash__`` method that subclasses should not override; or if they do, the subclass should compute the same hash value. This is so that sets with different implementations still hash to the same value, so they can be used interchangeably as dictionary keys. (A similar constraint exists on the hash values for different types of numbers and strings.) **Open issues:** Should I spell out the hash algorithm? Should there be another ABC that derives from Set and Hashable (but not from Composable)? ``MutableSet`` This is a subclass of ``ComposableSet`` implementing additional operations to add and remove elements. The supported methods have the semantics known from the ``set`` type in Python 2 (except for ``discard``, which is modeled after Java): ``.add(x)`` Abstract method returning a ``bool`` that adds the element ``x`` if it isn't already in the set. It should return ``True`` if ``x`` was added, ``False`` if it was already there. The abstract implementation raises ``NotImplementedError``. ``.discard(x)`` Abstract method returning a ``bool`` that removes the element ``x`` if present. It should return ``True`` if the element was present and ``False`` if it wasn't. The abstract implementation raises ``NotImplementedError``. ``.clear()`` Abstract method that empties the set. The abstract implementation raises ``NotImplementedError``. (Making this concrete would just add a slow, cumbersome default implementation.) ``.pop()`` Concrete method that removes an arbitrary item. If the set is empty, it raises ``KeyError``. The default implementation removes the first item returned by the set's iterator. ``.toggle(x)`` Concrete method returning a ``bool`` that adds x to the set if it wasn't there, but removes it if it was there. It should return ``True`` if ``x`` was added, ``False`` if it was removed. This also supports the in-place mutating operations ``|=``, ``&=``, ``^=``, ``-=``. These are concrete methods whose right operand can be an arbitrary ``Iterable``. It does not support the named methods that perform (almost) the same operations. Mappings '''''''' These abstract classes represent various stages of mapping-ness. The ``Mapping`` class represents the most common read-only mapping API. However, code *accepting* a mapping is encouraged to check for the ``BasicMapping`` ABC when iteration is not used. This allows for certain "black-box" implementations that can look up values by key but don't provide a convenient iteration API. A hypothetical example would be an interface to a hierarchical filesystem, where keys are pathnames relative to some root directory. Iterating over all pathnames would presumably take forever, as would counting the number of valid pathnames. The built-in type ``dict`` derives from ``MutableMapping``. ``BasicMapping`` A subclass of ``Container`` defining the following methods: ``.__getitem__(key)`` Abstract method that returns the value corresponding to ``key``, or raises ``KeyError``. The implementation always raises ``KeyError``. ``.get(key, default=None)`` Concrete method returning ``self[key]`` if this does not raise ``KeyError``, and the ``default`` value if it does. ``.__contains__()`` Concrete method returning ``True`` if ``self[key]`` does not raise ``KeyError``, and ``False`` if it does. ``Mapping`` A subclass of ``BasicMapping``, ``iterable`` and ``Sized``. It defines concrete methods ``__eq__``, ``keys``, ``items``, ``values``. Iterating over a mapping should return all the valid keys (i.e. those keys for which ``.__getitem__()`` returns a value), once each, and nothing else. The lengh of a mapping should equal to the number of elements returned by iterating over the object until the end of the iterator is reached (this is implied by the invariant listed above for ``Sized``). Two mappings, even with different implementations, can be compared for equality, and are considered equal if and only iff their items compare equal when converted to sets. The ``keys``, ``items`` and ``values`` methods return views; ``keys`` and ``items`` return ``Set`` views, ``values`` returns a ``Container`` view. The following invariant should hold: m.items() == set(zip(m.keys(), m.values())). ``HashableMapping`` A subclass of ``Mapping`` and ``Hashable``. The values should be instances of ``Hashable``. The concrete ``__hash__`` method should be equal to ``hash(m.items())``. ``MutableMapping`` A subclass of ``Mapping`` that also implements some standard mutating methods. Abstract methods include ``__setitem__``, ``__delitem__``, ``clear``, ``update``. Concrete methods include ``pop``, ``popitem``. Note: ``setdefault`` is *not* included. **Open issues:** * We should say more about mapping view types. Sequences ''''''''' These abstract classes represent various stages of sequence-ness. The built-in ``list`` and ``bytes`` types derive from ``MutableSequence``. The built-in ``tuple`` and ``str`` types derive from ``HashableSequence``. ``Sequence`` A subclass of ``Iterable``, ``Sized``, ``Container``. It defines a new abstract method ``__getitem__`` that has a complicated signature: when called with an integer, it returns an element of the sequence or raises ``IndexError``; when called with a ``slice`` object, it returns another ``Sequence``. The concrete ``__iter__`` method iterates over the elements using ``__getitem__`` with integer arguments 0, 1, and so on, until ``IndexError`` is raised. The length should be equal to the number of values returned by the iterator. **Open issues:** Other candidate methods, which can all have default concrete implementations that only depend on ``__len__`` and ``__getitem__`` with an integer argument: __reversed__, index, count, __add__, __mul__, __eq__, __lt__, __le__. ``HashableSequence`` A subclass of ``Sequence`` and ``Hashable``. The concrete ``__hash__`` method should implements the hashing algorithms used by tuples in Python 2. ``MutableSequence`` A subclass of ``Sequence`` adding some standard mutating methods. Abstract mutating methods: ``__setitem__`` (for integer indices as well as slices), ``__delitem__`` (ditto), ``insert``, ``append``, ``reverse``. Concrete mutating methods: ``extend``, ``pop``, ``remove``. Concrete mutating operators: ``+=``, ``*=`` (these mutate the object in place). Note: this does not define ``sort()`` -- that is only required to exist on genuine ``list`` instances. Strings ------- Python 3000 has two built-in string types: byte strings (``bytes``), deriving from ``MutableSequence``, and (Unicode) character strings (``str``), deriving from ``HashableSequence``. **Open issues:** define the base interfaces for these so alternative implementations and subclasses know what they are in for. This may be the subject of a new PEP or PEPs (PEP 358 should be co-opted for the ``bytes`` type). ABCs for Numbers ---------------- **Open issues:** Define: ``Number``, ``Complex``, ``Real``, ``Rational``, ``Integer``. Maybe also ``Cardinal`` (``Integer`` >= 0)? We probably also need ``Index``, which converts to ``Integer`` using ``__index__``. This should be moved out to a separate PEP. Guidelines for Writing ABCs --------------------------- Some suggestions for writing ABCs: * Use the ``@abstractmethod`` decorator. * Define abstract methods that could be useful as an end point when called via a super chain. * Define concrete methods that are very simple permutations of abstract methods (e.g. ``Mapping.get``). * Keep abstract classes small, one per use case instead of one per concept. ABCs vs. Alternatives ===================== In this section I will attempt to compare and contrast ABCs to other approaches that have been proposed. ABCs vs. Duck Typing -------------------- Does the introduction of ABCs mean the end of Duck Typing? I don't think so. Python will not require that a class derives from ``BasicMapping`` or ``Sequence`` when it defines a ``__getitem__`` method, nor will the ``x[y]`` syntax require that ``x`` is an instance of either ABC. You will still be able to assign any "file-like" object to ``sys.stdout``, as long as it has a ``write`` method. Of course, there will be some carrots to encourage users to derive from the appropriate base classes; these vary from default implementations for certain functionality to an improved ability to distinguish between mappings and sequences. But there are no sticks. If ``hasattr(x, __len__)`` works for you, great! ABCs are intended to solve problems that don't have a good solution at all in Python 2, such as distinguishing between mappings and sequences. ABCs vs. Generic Functions -------------------------- ABCs are compatible with Generic Functions (GFs). For example, my own Generic Functions implementation [4]_ uses the classes (types) of the arguments as the dispatch key, allowing derived classes to override base classes. Since (from Python's perspective) ABCs are quite ordinary classes, using an ABC in the default implementation for a GF can be quite appropriate. For example, if I have an overloaded ``prettyprint`` function, it would make total sense to define pretty-printing of sets like this:: @prettyprint.register(Set) def pp_set(s): return "{" + ... + "}" # Details left as an exercise and implementations for specific subclasses of Set could be added easily. I believe ABCs also won't present any problems for RuleDispatch, Phillip Eby's GF implementation in PEAK [5]_. Of course, GF proponents might claim that GFs (and concrete, or implementation, classes) are all you need. But even they will not deny the usefulness of inheritance; and one can easily consider the ABCs proposed in this PEP as optional implementation base classes; there is no requirement that all user-defined mappings derive from ``BasicMapping``. ABCs vs. Interfaces ------------------- ABCs are not intrinsically incompatible with Interfaces, but there is considerable overlap. For now, I'll leave it to proponents of Interfaces to explain why Interfaces are better. I expect that much of the work that went into e.g. defining the various shades of "mapping-ness" and the nomenclature could easily be adapted for a proposal to use Interfaces instead of ABCs. "Interfaces" in this context refers to a set of proposals for additional metadata elements attached to a class which are not part of the regular class hierarchy, but do allow for certain types of inheritance testing. Such metadata would be designed, at least in some proposals, so as to be easily mutable by an application, allowing application writers to override the normal classification of an object. The drawback to this idea of attaching mutable metadata to a class is that classes are shared state, and mutating them may lead to conflicts of intent. Additionally, the need to override the classification of an object can be done more cleanly using generic functions: In the simplest case, one can define a "category membership" generic function that simply returns False in the base implementation, and then provide overrides that return True for any classes of interest. References ========== .. [1] An Introduction to ABC's, by Talin (http://mail.python.org/pipermail/python-3000/2007-April/006614.html) .. [2] Incomplete implementation prototype, by GvR (http://svn.python.org/view/sandbox/trunk/abc/) .. [3] Possible Python 3K Class Tree?, wiki page created by Bill Janssen (http://wiki.python.org/moin/AbstractBaseClasses) .. [4] Generic Functions implementation, by GvR (http://svn.python.org/view/sandbox/trunk/overload/) .. [5] Charming Python: Scaling a new PEAK, by David Mertz (http://www-128.ibm.com/developerworks/library/l-cppeak2/) .. [6] Implementation of @abstractmethod (http://python.org/sf/1706989) .. [7] Unifying types and classes in Python 2.2, by GvR (http://www.python.org/download/releases/2.2.3/descrintro/) .. [8] "Putting Metaclasses to Work: A New Dimension in Object-Oriented Programming", by Ira R. Forman and Scott H. Danforth (http://www.amazon.com/gp/product/0201433052) Copyright ========= This document has been placed in the public domain. .. Local Variables: mode: indented-text indent-tabs-mode: nil sentence-end-double-space: t fill-column: 70 coding: utf-8 End: