[Python-3000] PEP 3124 - Overloading, Generic Functions, Interfaces, etc.
Phillip J. Eby
pje at telecommunity.com
Tue May 1 00:54:33 CEST 2007
This is just the first draft (also checked into SVN), and doesn't include
the details of how the extension API works (so that third-party interfaces
and generic functions can interoperate using the same decorators,
Comments and questions appreciated, as it'll help drive better explanations
of both the design and rationales. I'm usually not that good at guessing
what other people will want to know (or are likely to misunderstand) until
I get actual questions.
Title: Overloading, Generic Functions, Interfaces, and Adaptation
Version: $Revision: 55029 $
Last-Modified: $Date: 2007-04-30 18:48:06 -0400 (Mon, 30 Apr 2007) $
Author: Phillip J. Eby <pje at telecommunity.com>
Discussions-To: Python 3000 List <python-3000 at python.org>
Type: Standards Track
Requires: 3107, 3115, 3119
Replaces: 245, 246
This PEP proposes a new standard library module, ``overloading``, to
provide generic programming features including dynamic overloading
(aka generic functions), interfaces, adaptation, method combining (ala
CLOS and AspectJ), and simple forms of aspect-oriented programming.
The proposed API is also open to extension; that is, it will be
possible for library developers to implement their own specialized
interface types, generic function dispatchers, method combination
algorithms, etc., and those extensions will be treated as first-class
citizens by the proposed API.
The API will be implemented in pure Python with no C, but may have
some dependency on CPython-specific features such as ``sys._getframe``
and the ``func_code`` attribute of functions. It is expected that
e.g. Jython and IronPython will have other ways of implementing
similar functionality (perhaps using Java or C#).
Rationale and Goals
Python has always provided a variety of built-in and standard-library
generic functions, such as ``len()``, ``iter()``, ``pprint.pprint()``,
and most of the functions in the ``operator`` module. However, it
1. does not have a simple or straightforward way for developers to
create new generic functions,
2. does not have a standard way for methods to be added to existing
generic functions (i.e., some are added using registration
functions, others require defining ``__special__`` methods,
possibly by monkeypatching), and
3. does not allow dispatching on multiple argument types (except in
a limited form for arithmetic operators, where "right-hand"
(``__r*__``) methods can be used to do two-argument dispatch.
In addition, it is currently a common anti-pattern for Python code
to inspect the types of received arguments, in order to decide what
to do with the objects. For example, code may wish to accept either
an object of some type, or a sequence of objects of that type.
Currently, the "obvious way" to do this is by type inspection, but
this is brittle and closed to extension. A developer using an
already-written library may be unable to change how their objects are
treated by such code, especially if the objects they are using were
created by a third party.
Therefore, this PEP proposes a standard library module to address
these, and related issues, using decorators and argument annotations
(PEP 3107). The primary features to be provided are:
* a dynamic overloading facility, similar to the static overloading
found in languages such as Java and C++, but including optional
method combination features as found in CLOS and AspectJ.
* a simple "interfaces and adaptation" library inspired by Haskell's
typeclasses (but more dynamic, and without any static type-checking),
with an extension API to allow registering user-defined interface
types such as those found in PyProtocols and Zope.
* a simple "aspect" implementation to make it easy to create stateful
adapters and to do other stateful AOP.
These features are to be provided in such a way that extended
implementations can be created and used. For example, it should be
possible for libraries to define new dispatching criteria for
generic functions, and new kinds of interfaces, and use them in
place of the predefined features. For example, it should be possible
to use a ``zope.interface`` interface object to specify the desired
type of a function argument, as long as the ``zope.interface`` package
registered itself correctly (or a third party did the registration).
In this way, the proposed API simply offers a uniform way of accessing
the functionality within its scope, rather than prescribing a single
implementation to be used for all libraries, frameworks, and
The overloading API will be implemented as a single module, named
``overloading``, providing the following features:
The ``@overload`` decorator allows you to define alternate
implementations of a function, specialized by argument type(s). A
function with the same name must already exist in the local namespace.
The existing function is modified in-place by the decorator to add
the new implementation, and the modified function is returned by the
decorator. Thus, the following code::
from overloading import overload
from collections import Iterable
"""Flatten an object to its component iterables"""
def flatten(ob: Iterable):
for o in ob:
for ob in flatten(o):
def flatten(ob: basestring):
creates a single ``flatten()`` function whose implementation roughly
if isinstance(ob, basestring) or not isinstance(ob, Iterable):
for o in ob:
for ob in flatten(o):
**except** that the ``flatten()`` function defined by overloading
remains open to extension by adding more overloads, while the
hardcoded version cannot be extended.
For example, if someone wants to use ``flatten()`` with a string-like
type that doesn't subclass ``basestring``, they would be out of luck
with the second implementation. With the overloaded implementation,
however, they can either write this::
def flatten(ob: MyString):
or this (to avoid copying the implementation)::
from overloading import RuleSet
(Note also that, although PEP 3119 proposes that it should be possible
for abstract base classes like ``Iterable`` to allow classes like
``MyString`` to claim subclass-hood, such a claim is *global*,
throughout the application. In contrast, adding a specific overload
or copying a rule is specific to an individual function, and therefore
less likely to have undesired side effects.)
``@overload`` vs. ``@when``
The ``@overload`` decorator is a common-case shorthand for the more
general ``@when`` decorator. It allows you to leave out the name of
the function you are overloading, at the expense of requiring the
target function to be in the local namespace. It also doesn't support
adding additional criteria besides the ones specified via argument
annotations. The following function definitions have identical
effects, except for name binding side-effects (which will be described
def flatten(ob: basestring):
def flatten(ob: basestring):
def flatten_basestring(ob: basestring):
The first definition above will bind ``flatten`` to whatever it was
previously bound to. The second will do the same, if it was already
bound to the ``when`` decorator's first argument. If ``flatten`` is
unbound or bound to something else, it will be rebound to the function
definition as given. The last two definitions above will always bind
``flatten_basestring`` to the function definition as given.
Using this approach allows you to both give a method a descriptive
name (often useful in tracebacks!) and to reuse the method later.
Except as otherwise specified, all ``overloading`` decorators have the
same signature and binding rules as ``@when``. They accept a function
and an optional "predicate" object.
The default predicate implementation is a tuple of types with
positional matching to the overloaded function's arguments. However,
an arbitrary number of other kinds of of predicates can be created and
registered using the `Extension API`_, and will then be usable with
``@when`` and other decorators created by this module (like
``@before``, ``@after``, and ``@around``).
Method Combination and Overriding
When an overloaded function is invoked, the implementation with the
signature that *most specifically matches* the calling arguments is
the one used. If no implementation matches, a ``NoApplicableMethods``
error is raised. If more than one implementation matches, but none of
the signatures are more specific than the others, an ``AmbiguousMethods``
error is raised.
For example, the following pair of implementations are ambiguous, if
the ``foo()`` function is ever called with two integer arguments,
because both signatures would apply, but neither signature is more
*specific* than the other (i.e., neither implies the other)::
def foo(bar:int, baz:object):
def foo(bar:object, baz:int):
In contrast, the following pair of implementations can never be
ambiguous, because one signature always implies the other; the
``int/int`` signature is more specific than the ``object/object``
def foo(bar:object, baz:object):
def foo(bar:int, baz:int):
A signature S1 implies another signature S2, if whenever S1 would
apply, S2 would also. A signature S1 is "more specific" than another
signature S2, if S1 implies S2, but S2 does not imply S1.
Although the examples above have all used concrete or abstract types
as argument annotations, there is no requirement that the annotations
be such. They can also be "interface" objects (discussed in the
`Interfaces and Adaptation`_ section), including user-defined
interface types. (They can also be other objects whose types are
appropriately registered via the `Extension API`_.)
Proceeding to the "Next" Method
If the first parameter of an overloaded function is named
``__proceed__``, it will be passed a callable representing the next
most-specific method. For example, this code::
def foo(bar:object, baz:object):
print "got objects!"
def foo(__proceed__, bar:int, baz:int):
print "got integers!"
return __proceed__(bar, baz)
Will print "got integers!" followed by "got objects!".
If there is no next most-specific method, ``__proceed__`` will be
bound to a ``NoApplicableMethods`` instance. When called, a new
``NoApplicableMethods`` instance will be raised, with the arguments
passed to the first instance.
Similarly, if the next most-specific methods have ambiguous precedence
with respect to each other, ``__proceed__`` will be bound to an
``AmbiguousMethods`` instance, and if called, it will raise a new
Thus, a method can either check if ``__proceed__`` is an error
instance, or simply invoke it. The ``NoApplicableMethods`` and
``AmbiguousMethods`` error classes have a common ``DispatchError``
base class, so ``isinstance(__proceed__, overloading.DispatchError)``
is sufficient to identify whether ``__proceed__`` can be safely
(Implementation note: using a magic argument name like ``__proceed__``
could potentially be replaced by a magic function that would be called
to obtain the next method. A magic function, however, would degrade
performance and might be more difficult to implement on non-CPython
platforms. Method chaining via magic argument names, however, can be
efficiently implemented on any Python platform that supports creating
bound methods from functions -- one simply recursively binds each
function to be chained, using the following function or error as the
``im_self`` of the bound method.)
"Before" and "After" Methods
In addition to the simple next-method chaining shown above, it is
sometimes useful to have other ways of combining methods. For
example, the "observer pattern" can sometimes be implemented by adding
extra methods to a function, that execute before or after the normal
To support these use cases, the ``overloading`` module will supply
``@before``, ``@after``, and ``@around`` decorators, that roughly
correspond to the same types of methods in the Common Lisp Object
System (CLOS), or the corresponding "advice" types in AspectJ.
Like ``@when``, all of these decorators must be passed the function to
be overloaded, and can optionally accept a predicate as well::
print "Beginning the actual transaction"
def check_single_access(db: SingletonDB):
raise TransactionError("Database already in use")
def start_logging(db: LoggableDB):
``@before`` and ``@after`` methods are invoked either before or after
the main function body, and are *never considered ambiguous*. That
is, it will not cause any errors to have multiple "before" or "after"
methods with identical or overlapping signatures. Ambiguities are
resolved using the order in which the methods were added to the
"Before" methods are invoked most-specific method first, with
ambiguous methods being executed in the order they were added. All
"before" methods are called before any of the function's "primary"
methods (i.e. normal ``@overload`` methods) are executed.
"After" methods are invoked in the *reverse* order, after all of the
function's "primary" methods are executed. That is, they are executed
least-specific methods first, with ambiguous methods being executed in
the reverse of the order in which they were added.
The return values of both "before" and "after" methods are ignored,
and any uncaught exceptions raised by *any* methods (primary or other)
immediately end the dispatching process. "Before" and "after" methods
cannot have ``__proceed__`` arguments, as they are not responsible
for calling any other methods. They are simply called as a
notification before or after the primary methods.
Thus, "before" and "after" methods can be used to check or establish
preconditions (e.g. by raising an error if the conditions aren't met)
or to ensure postconditions, without needing to duplicate any existing
The ``@around`` decorator declares a method as an "around" method.
"Around" methods are much like primary methods, except that the
least-specific "around" method has higher precedence than the
most-specific "before" or method.
Unlike "before" and "after" methods, however, "Around" methods *are*
responsible for calling their ``__proceed__`` argument, in order to
continue the invocation process. "Around" methods are usually used
to transform input arguments or return values, or to wrap specific
cases with special error handling or try/finally conditions, e.g.::
def lock_while_committing(__proceed__, db: SingletonDB):
They can also be used to replace the normal handling for a specific
case, by *not* invoking the ``__proceed__`` function.
The ``__proceed__`` given to an "around" method will either be the
next applicable "around" method, a ``DispatchError`` instance,
or a synthetic method object that will call all the "before" methods,
followed by the primary method chain, followed by all the "after"
methods, and return the result from the primary method chain.
Thus, just as with normal methods, ``__proceed__`` can be checked for
``DispatchError``-ness, or simply invoked. The "around" method should
return the value returned by ``__proceed__``, unless of course it
wishes to modify or replace it with a different return value for the
function as a whole.
The decorators described above (``@overload``, ``@when``, ``@before``,
``@after``, and ``@around``) collectively implement what in CLOS is
called the "standard method combination" -- the most common patterns
used in combining methods.
Sometimes, however, an application or library may have use for a more
sophisticated type of method combination. For example, if you
would like to have "discount" methods that return a percentage off,
to be subtracted from the value returned by the primary method(s),
you might write something like this::
from overloading import always_overrides, merge_by_default
from overloading import Around, Before, After, Method, MethodList
"""Apply return values as discounts"""
def __call__(self, *args, **kw):
retval = self.tail(*args, **kw)
for sig, body in self.sorted():
retval -= retval * body(*args, **kw)
# merge discounts by priority
# discounts have precedence over before/after/primary methods
# but not over "around" methods
# Make a decorator called "discount" that works just like the
# standard decorators...
discount = Discount.make_decorator('discount')
# and now let's use it...
def ten_percent_off_shoes(product: Shoe)
Similar techniques can be used to implement a wide variety of
CLOS-style method qualifiers and combination rules. The process of
creating custom method combination objects and their corresponding
decorators is described in more detail under the `Extension API`_
Note, by the way, that the ``@discount`` decorator shown will work
correctly with any new predicates defined by other code. For example,
if ``zope.interface`` were to register its interface types to work
correctly as argument annotations, you would be able to specify
discounts on the basis of its interface types, not just classes or
``overloading``-defined interface types.
Similarly, if a library like RuleDispatch or PEAK-Rules were to
register an appropriate predicate implementation and dispatch engine,
one would then be able to use those predicates for discounts as well,
from somewhere import Pred # some predicate implementation
" product.material.name=='Blue Suede'")
The process of defining custom predicate types and dispatching engines
is also described in more detail under the `Extension API`_ section.
Overloading Inside Classes
All of the decorators above have a special additional behavior when
they are directly invoked within a class body: the first parameter
(other than ``__proceed__``, if present) of the decorated function
will be treated as though it had an annotation equal to the class
in which it was defined.
That is, this code::
produces the same effect as this (apart from the existence of a
def get_conjuncts_of_and(ob: And):
This behavior is both a convenience enhancement when defining lots of
methods, and a requirement for safely distinguishing multi-argument
overloads in subclasses. Consider, for example, the following code::
def foo(self, ob):
print "got an object"
def foo(__proceed__, self, ob:Iterable):
print "it's iterable!"
return __proceed__(self, ob)
foo = A.foo # foo must be defined in local namespace
def foo(__proceed__, self, ob:Iterable):
print "B got an iterable!"
return __proceed__(self, ob)
Due to the implicit class rule, calling ``B().foo()`` will print
"B got an iterable!" followed by "it's iterable!", and finally,
"got an object", while ``A().foo()`` would print only the messages
defined in ``A``.
Conversely, without the implicit class rule, the two "Iterable"
methods would have the exact same applicability conditions, so calling
either ``A().foo()`` or ``B().foo()`` would result in an
It is currently an open issue to determine the best way to implement
this rule in Python 3.0. Under Python 2.x, a class' metaclass was
not chosen until the end of the class body, which means that
decorators could insert a custom metaclass to do processing of this
sort. (This is how RuleDispatch, for example, implements the implicit
PEP 3115, however, requires that a class' metaclass be determined
*before* the class body has executed, making it impossible to use this
technique for class decoration any more.
At this writing, discussion on this issue is ongoing.
Interfaces and Adaptation
The ``overloading`` module provides a simple implementation of
interfaces and adaptation. The following example defines an
``IStack`` interface, and declares that ``list`` objects support it::
from overloading import abstract, Interface
def push(self, ob)
"""Push 'ob' onto the stack"""
"""Pop a value and return it"""
when(IStack.push, (list, object))(list.append)
mylist = 
mystack = IStack(mylist)
The ``Interface`` class is a kind of "universal adapter". It accepts
a single argument: an object to adapt. It then binds all its methods
to the target object, in place of itself. Thus, calling
``mystack.push(42``) is the same as calling
The ``@abstract`` decorator marks a function as being abstract: i.e.,
having no implementation. If an ``@abstract`` function is called,
it raises ``NoApplicableMethods``. To become executable, overloaded
methods must be added using the techniques previously described. (That
is, methods can be added using ``@when``, ``@before``, ``@after``,
``@around``, or any custom method combination decorators.)
In the example above, the ``list.append`` method is added as a method
for ``IStack.push()`` when its arguments are a list and an arbitrary
object. Thus, ``IStack.push(mylist, 42)`` is translated to
``list.append(mylist, 42)``, thereby implementing the desired
(Note: the ``@abstract`` decorator is not limited to use in interface
definitions; it can be used anywhere that you wish to create an
"empty" generic function that initially has no methods. In
particular, it need not be used inside a class.)
Subclassing and Re-assembly
Interfaces can be subclassed::
"""Return the number of items on the stack"""
# define __len__ support for ISizedStack
Or assembled by combining functions from existing interfaces::
__len__ = ISizedStack.__len__
# list now implements Sizable as well as ISizedStack, without
# making any new declarations!
A class can be considered to "adapt to" an interface at a given
point in time, if no method defined in the interface is guaranteed to
raise a ``NoApplicableMethods`` error if invoked on an instance of
that class at that point in time.
In normal usage, however, it is "easier to ask forgiveness than
permission". That is, it is easier to simply use an interface on
an object by adapting it to the interface (e.g. ``IStack(mylist)``)
or invoking interface methods directly (e.g. ``IStack.push(mylist,
42)``), than to try to figure out whether the object is adaptable to
(or directly implements) the interface.
Implementing an Interface in a Class
It is possible to declare that a class directly implements an
interface, using the ``declare_implementation()`` function::
from overloading import declare_implementation
self.data = 
def push(self, ob):
The ``declare_implementation()`` call above is roughly equivalent to
the following steps::
when(IStack.push, (Stack,object))(lambda self, ob: self.push(ob))
when(IStack.pop, (Stack,))(lambda self, ob: self.pop())
That is, calling ``IStack.push()`` or ``IStack.pop()`` on an instance
of any subclass of ``Stack``, will simply delegate to the actual
``push()`` or ``pop()`` methods thereof.
For the sake of efficiency, calling ``IStack(s)`` where ``s`` is an
instance of ``Stack``, **may** return ``s`` rather than an ``IStack``
adapter. (Note that calling ``IStack(x)`` where ``x`` is already an
``IStack`` adapter will always return ``x`` unchanged; this is an
additional optimization allowed in cases where the adaptee is known
to *directly* implement the interface, without adaptation.)
For convenience, it may be useful to declare implementations in the
class header, e.g.::
class Stack(metaclass=Implementer, implements=IStack):
Instead of calling ``declare_implementation()`` after the end of the
Interfaces as Type Specifiers
``Interface`` subclasses can be used as argument annotations to
indicate what type of objects are acceptable to an overload, e.g.::
def traverse(g: IGraph, s: IStack):
g = IGraph(g)
s = IStack(s)
Note, however, that the actual arguments are *not* changed or adapted
in any way by the mere use of an interface as a type specifier. You
must explicitly cast the objects to the appropriate interface, as
Note, however, that other patterns of interface use are possible.
For example, other interface implementations might not support
adaptation, or might require that function arguments already be
adapted to the specified interface. So the exact semantics of using
an interface as a type specifier are dependent on the interface
objects you actually use.
For the interface objects defined by this PEP, however, the semantics
are as described above. An interface I1 is considered "more specific"
than another interface I2, if the set of descriptors in I1's
inheritance hierarchy are a proper superset of the descriptors in I2's
So, for example, ``ISizedStack`` is more specific than both
``ISizable`` and ``ISizedStack``, irrespective of the inheritance
relationships between these interfaces. It is purely a question of
what operations are included within those interfaces -- and the
*names* of the operations are unimportant.
Interfaces (at least the ones provided by ``overloading``) are always
considered less-specific than concrete classes. Other interface
implementations can decide on their own specificity rules, both
between interfaces and other interfaces, and between interfaces and
Non-Method Attributes in Interfaces
The ``Interface`` implementation actually treats all attributes and
methods (i.e. descriptors) in the same way: their ``__get__`` (and
``__set__`` and ``__delete__``, if present) methods are called with
the wrapped (adapted) object as "self". For functions, this has the
effect of creating a bound method linking the generic function to the
For non-function attributes, it may be easiest to specify them using
the ``property`` built-in, and the corresponding ``fget``, ``fset``,
and ``fdel`` attributes::
"""Read-only length attribute"""
# ILength(aList).length == list.__len__(aList)
Alternatively, methods such as ``_get_foo()`` and ``_set_foo()``
may be defined as part of the interface, and the property defined
in terms of those methods, but this a bit more difficult for users
to implement correctly when creating a class that directly implements
the interface, as they would then need to match all the individual
method names, not just the name of the property or attribute.
The adaptation system provided assumes that adapters are "stateless",
which is to say that adapters have no attributes or storage apart from
those of the adapted object. This follows the "typeclass/instance"
model of Haskell, and the concept of "pure" (i.e., transitively
However, there are occasionally cases where, to provide a complete
implementation of some interface, some sort of additional state is
One possibility of course, would be to attach monkeypatched "private"
attributes to the adaptee. But this is subject to name collisions,
and complicates the process of initialization. It also doesn't work
on objects that don't have a ``__dict__`` attribute.
So the ``Aspect`` class is provided to make it easy to attach extra
information to objects that either:
1. have a ``__dict__`` attribute (so aspect instances can be stored
in it, keyed by aspect class),
2. support weak referencing (so aspect instances can be managed using
a global but thread-safe weak-reference dictionary), or
3. implement or can be adapt to the ``overloading.IAspectOwner``
interface (technically, #1 or #2 imply this)
Subclassing ``Aspect`` creates an adapter class whose state is tied
to the life of the adapted object.
For example, suppose you would like to count all the times a certain
method is called on instances of ``Target`` (a classic AOP example).
You might do something like::
from overloading import Aspect
count = 0
def count_after_call(self, *args, **kw):
Count(self).count += 1
The above code will keep track of the number of times that
``Target.some_method()`` is successfully called (i.e., it will not
count errors). Other code can then access the count using
``Aspect`` instances can of course have ``__init__`` methods, to
initialize any data structures. They can use either ``__slots__``
or dictionary-based attributes for storage.
While this facility is rather primitive compared to a full-featured
AOP tool like AspectJ, persons who wish to build pointcut libraries
or other AspectJ-like features can certainly use ``Aspect`` objects
and method-combination decorators as a base for more expressive AOP
XXX spec out full aspect API, including keys, N-to-1 aspects, manual
attach/detach/delete of aspect instances, and the ``IAspectOwner``
TODO: explain how all of these work
parse_rule(ruleset, body, predicate, actiontype, localdict, globaldict)
Most of the functionality described in this PEP is already implemented
in the in-development version of the PEAK-Rules framework. In
particular, the basic overloading and method combination framework
(minus the ``@overload`` decorator) already exists there. The
implementation of all of these features in ``peak.rules.core`` is 656
lines of Python at this writing.
``peak.rules.core`` currently relies on the DecoratorTools and
BytecodeAssembler modules, but both of these dependencies can be
replaced, as DecoratorTools is used mainly for Python 2.3
compatibility and to implement structure types (which can be done
with named tuples in later versions of Python). The use of
BytecodeAssembler can be replaced using an "exec" or "compile"
workaround, given a reasonable effort. (It would be easier to do this
if the ``func_closure`` attribute of function objects was writable.)
The ``Interface`` class has been previously prototyped, but is not
included in PEAK-Rules at the present time.
The "implicit class rule" has previously been implemented in the
RuleDispatch library. However, it relies on the ``__metaclass__``
hook that is currently eliminated in PEP 3115.
I don't currently know how to make ``@overload`` play nicely with
``classmethod`` and ``staticmethod`` in class bodies. It's not really
clear if it needs to, however.
This document has been placed in the public domain.
More information about the Python-3000