The decorator module

Author:Michele Simionato
Version:4.3.1 (2018-08-04)
Supports:Python 2.6, 2.7, 3.0, 3.1, 3.2, 3.3, 3.4, 3.5, 3.6, 3.7
Download page:
Installation:pip install decorator
License:BSD license


The decorator module is over ten years old, but still alive and kicking. It is used by several frameworks (IPython, scipy, authkit, pylons, pycuda, sugar, …) and has been stable for a long time. It is your best option if you want to preserve the signature of decorated functions in a consistent way across Python releases. Version 4 is fully compatible with the past, except for one thing: support for Python 2.4 and 2.5 has been dropped. That decision made it possible to use a single code base both for Python 2.X and Python 3.X. This is a huge bonus, since I could remove over 2,000 lines of duplicated documentation/doctests. Having to maintain separate docs for Python 2 and Python 3 effectively stopped any development on the module for several years. Moreover, it is now trivial to distribute the module as an universal wheel since 2to3 is no more required. Since Python 2.5 has been released ages ago (in 2006), I felt that it was reasonable to drop the support for it. If you need to support ancient versions of Python, stick with the decorator module version 3.4.2. The current version supports all Python releases from 2.6 up to 3.6.

What’s New in version 4

  • New documentation There is now a single manual for all Python versions, so I took the opportunity to overhaul the documentation and to move it to Even if you are a long-time user, you may want to revisit the docs, since several examples have been improved.
  • Packaging improvements The code is now also available in wheel format. Integration with setuptools has improved and you can run the tests with the command python test too.
  • Code changes A new utility function decorate(func, caller) has been added. It does the same job that was performed by the older decorator(caller, func). The old functionality is now deprecated and no longer documented, but still available for now.
  • Multiple dispatch The decorator module now includes an implementation of generic functions (sometimes called “multiple dispatch functions”). The API is designed to mimic functools.singledispatch (added in Python 3.4), but the implementation is much simpler. Moreover, all decorators involved preserve the signature of the decorated functions. For now, this exists mostly to demonstrate the power of the module. In the future it could be enhanced/optimized. In any case, it is very short and compact (less then 100 lines), so you can extract it for your own use. Take it as food for thought.
  • Python 3.5 coroutines From version 4.1 it is possible to decorate coroutines, i.e. functions defined with the async def syntax, and to maintain the inspect.iscoroutinefunction check working for the decorated function.
  • Decorator factories From version 4.2 there is facility to define factories of decorators in a simple way, a feature requested by the users since a long time.

Usefulness of decorators

Python decorators are an interesting example of why syntactic sugar matters. In principle, their introduction in Python 2.4 changed nothing, since they did not provide any new functionality which was not already present in the language. In practice, their introduction has significantly changed the way we structure our programs in Python. I believe the change is for the best, and that decorators are a great idea since:

  • decorators help reducing boilerplate code;
  • decorators help separation of concerns;
  • decorators enhance readability and maintenability;
  • decorators are explicit.

Still, as of now, writing custom decorators correctly requires some experience and it is not as easy as it could be. For instance, typical implementations of decorators involve nested functions, and we all know that flat is better than nested.

The aim of the decorator module it to simplify the usage of decorators for the average programmer, and to popularize decorators by showing various non-trivial examples. Of course, as all techniques, decorators can be abused (I have seen that) and you should not try to solve every problem with a decorator, just because you can.

You may find the source code for all the examples discussed here in the file, which contains the documentation you are reading in the form of doctests.


Technically speaking, any Python object which can be called with one argument can be used as a decorator. However, this definition is somewhat too large to be really useful. It is more convenient to split the generic class of decorators in two subclasses:

signature-preserving decorators:
Callable objects which accept a function as input and return a function as output, with the same signature.
signature-changing decorators:
Decorators which change the signature of their input function, or decorators that return non-callable objects.

Signature-changing decorators have their use: for instance, the builtin classes staticmethod and classmethod are in this group. They take functions and return descriptor objects which are neither functions, nor callables.

Still, signature-preserving decorators are more common, and easier to reason about. In particular, they can be composed together, whereas other decorators generally cannot.

Writing signature-preserving decorators from scratch is not that obvious, especially if one wants to define proper decorators that can accept functions with any signature. A simple example will clarify the issue.

Statement of the problem

A very common use case for decorators is the memoization of functions. A memoize decorator works by caching the result of the function call in a dictionary, so that the next time the function is called with the same input parameters the result is retrieved from the cache and not recomputed.

There are many implementations of memoize in, but they do not preserve the signature. In recent versions of Python you can find a sophisticated lru_cache decorator in the standard library’s functools. Here I am just interested in giving an example.

Consider the following simple implementation (note that it is generally impossible to correctly memoize something that depends on non-hashable arguments):

def memoize_uw(func):
    func.cache = {}

    def memoize(*args, **kw):
        if kw:  # frozenset is used to ensure hashability
            key = args, frozenset(kw.items())
            key = args
        if key not in func.cache:
            func.cache[key] = func(*args, **kw)
        return func.cache[key]
    return functools.update_wrapper(memoize, func)

Here I used the functools.update_wrapper utility, which was added in Python 2.5 to simplify the writing of decorators. (Previously, you needed to manually copy the function attributes __name__, __doc__, __module__, and __dict__ to the decorated function by hand).

Here is an example of usage:

def f1(x):
    "Simulate some long computation"
    return x

This works insofar as the decorator accepts functions with generic signatures. Unfortunately, it is not a signature-preserving decorator, since memoize_uw generally returns a function with a different signature from the original.

Consider for instance the following case:

def f1(x):
    "Simulate some long computation"
    return x

Here, the original function takes a single argument named x, but the decorated function takes any number of arguments and keyword arguments:

>>> from decorator import getfullargspec
>>> print(getfullargspec(f1))
FullArgSpec(args=[], varargs='args', varkw='kw', defaults=None, kwonlyargs=[], kwonlydefaults=None, annotations={})

This means that introspection tools (like pydoc) will give false information about the signature of f1 – unless you are using Python 3.5. This is pretty bad: pydoc will tell you that the function accepts the generic signature *args, **kw, but calling the function with more than one argument raises an error:

>>> f1(0, 1)
Traceback (most recent call last):
TypeError: f1() takes exactly 1 positional argument (2 given)

Notice that inspect.getfullargspec will give the wrong signature, even in the latest Python, i.e. version 3.6 at the time of writing.

The solution

The solution is to provide a generic factory of generators, which hides the complexity of making signature-preserving decorators from the application programmer. The decorate function in the decorator module is such a factory:

>>> from decorator import decorate

decorate takes two arguments:

  1. a caller function describing the functionality of the decorator, and
  2. a function to be decorated.

The caller function must have signature (f, *args, **kw), and it must call the original function f with arguments args and kw, implementing the wanted capability (in this case, memoization):

def _memoize(func, *args, **kw):
    if kw:  # frozenset is used to ensure hashability
        key = args, frozenset(kw.items())
        key = args
    cache = func.cache  # attribute added by memoize
    if key not in cache:
        cache[key] = func(*args, **kw)
    return cache[key]

Now, you can define your decorator as follows:

def memoize(f):
    A simple memoize implementation. It works by adding a .cache dictionary
    to the decorated function. The cache will grow indefinitely, so it is
    your responsibility to clear it, if needed.
    f.cache = {}
    return decorate(f, _memoize)

The difference from the nested function approach of memoize_uw is that the decorator module forces you to lift the inner function to the outer level. Moreover, you are forced to explicitly pass the function you want to decorate; there are no closures.

Here is a test of usage:

>>> @memoize
... def heavy_computation():
...     time.sleep(2)
...     return "done"

>>> print(heavy_computation()) # the first time it will take 2 seconds

>>> print(heavy_computation()) # the second time it will be instantaneous

The signature of heavy_computation is the one you would expect:

>>> print(getfullargspec(heavy_computation))
FullArgSpec(args=[], varargs=None, varkw=None, defaults=None, kwonlyargs=[], kwonlydefaults=None, annotations={})

A trace decorator

Here is an example of how to define a simple trace decorator, which prints a message whenever the traced function is called:

def _trace(f, *args, **kw):
    kwstr = ', '.join('%r: %r' % (k, kw[k]) for k in sorted(kw))
    print("calling %s with args %s, {%s}" % (f.__name__, args, kwstr))
    return f(*args, **kw)
def trace(f):
    return decorate(f, _trace)

Here is an example of usage:

>>> @trace
... def f1(x):
...     pass

It is immediate to verify that f1 works…

>>> f1(0)
calling f1 with args (0,), {}

…and it that it has the correct signature:

>>> print(getfullargspec(f1))
FullArgSpec(args=['x'], varargs=None, varkw=None, defaults=None, kwonlyargs=[], kwonlydefaults=None, annotations={})

The decorator works with functions of any signature:

>>> @trace
... def f(x, y=1, z=2, *args, **kw):
...     pass

>>> f(0, 3)
calling f with args (0, 3, 2), {}

>>> print(getfullargspec(f))
FullArgSpec(args=['x', 'y', 'z'], varargs='args', varkw='kw', defaults=(1, 2), kwonlyargs=[], kwonlydefaults=None, annotations={})

Function annotations

Python 3 introduced the concept of function annotations: the ability to annotate the signature of a function with additional information, stored in a dictionary named __annotations__. The decorator module (starting from release 3.3) will understand and preserve these annotations.

Here is an example:

>>> @trace
... def f(x: 'the first argument', y: 'default argument'=1, z=2,
...       *args: 'varargs', **kw: 'kwargs'):
...     pass

In order to introspect functions with annotations, one needs the utility inspect.getfullargspec (introduced in Python 3, then deprecated in Python 3.5, then undeprecated in Python 3.6):

>>> from inspect import getfullargspec
>>> argspec = getfullargspec(f)
>>> argspec.args
['x', 'y', 'z']
>>> argspec.varargs
>>> argspec.varkw
>>> argspec.defaults
(1, 2)
>>> argspec.kwonlyargs
>>> argspec.kwonlydefaults

You can check that the __annotations__ dictionary is preserved:

>>> f.__annotations__ is f.__wrapped__.__annotations__

Here f.__wrapped__ is the original undecorated function. This attribute exists for consistency with the behavior of functools.update_wrapper.

Another attribute copied from the original function is __qualname__, the qualified name. This attribute was introduced in Python 3.3.


It can become tedious to write a caller function (like the above _trace example) and then a trivial wrapper (def trace(f): return decorate(f, _trace)) every time. Not to worry! The decorator module provides an easy shortcut to convert the caller function into a signature-preserving decorator.

It is the decorator function:

>>> from decorator import decorator
>>> print(decorator.__doc__)
decorator(caller) converts a caller function into a decorator

The decorator function can be used as a signature-changing decorator, just like classmethod and staticmethod. But classmethod and staticmethod return generic objects which are not callable. Instead, decorator returns signature-preserving decorators (i.e. functions with a single argument).

For instance, you can write:

>>> @decorator
... def trace(f, *args, **kw):
...     kwstr = ', '.join('%r: %r' % (k, kw[k]) for k in sorted(kw))
...     print("calling %s with args %s, {%s}" % (f.__name__, args, kwstr))
...     return f(*args, **kw)

And trace is now a decorator!

>>> trace
<function trace at 0x...>

Here is an example of usage:

>>> @trace
... def func(): pass

>>> func()
calling func with args (), {}

The decorator function can also be used to define factories of decorators, i.e. functions returning decorators. In general you can just write something like this:

def decfactory(param1, param2, ...):
   def caller(f, *args, **kw):
       return somefunc(f, param1, param2, .., *args, **kw)
   return decorator(caller)

This is fully general but requires an additional level of nesting. For this reason since version 4.2 there is a facility to build decorator factories by using a single caller with default arguments i.e. writing something like this:

def caller(f, param1=default1, param2=default2, ..., *args, **kw):
    return somefunc(f, param1, param2, *args, **kw)
decfactory = decorator(caller)

Notice that this simplified approach only works with default arguments, i.e. param1, param2 etc must have known defaults. Thanks to this restriction, there exists an unique default decorator, i.e. the member of the family which uses the default values for all parameters. Such decorator can be written as decfactory() with no parameters specified; moreover, as a shortcut, it is also possible to elide the parenthesis, a feature much requested by the users. For years I have been opposite to this feature request, since having expliciti parenthesis to me is more clear and less magic; however once this feature entered in decorators of the Python standard library (I am referring to the dataclass decorator I finally gave up.

The example below will show how it works in practice.

Decorator factories

Sometimes one has to deal with blocking resources, such as stdin. Sometimes it is better to receive a “busy” message than just blocking everything. This can be accomplished with a suitable family of decorators (decorator factory), parameterize by a string, the busy message:

def blocking(f, msg='blocking', *args, **kw):
    if not hasattr(f, "thread"):  # no thread running
        def set_result():
            f.result = f(*args, **kw)
        f.thread = threading.Thread(None, set_result)
        return msg
    elif f.thread.isAlive():
        return msg
    else:  # the thread is ended, return the stored result
        del f.thread
        return f.result

Functions decorated with blocking will return a busy message if the resource is unavailable, and the intended result if the resource is available. For instance:

>>> @blocking("Please wait ...")
... def read_data():
...     time.sleep(3) # simulate a blocking resource
...     return "some data"

>>> print(read_data())  # data is not available yet
Please wait ...

>>> time.sleep(1)
>>> print(read_data())  # data is not available yet
Please wait ...

>>> time.sleep(1)
>>> print(read_data())  # data is not available yet
Please wait ...

>>> time.sleep(1.1)  # after 3.1 seconds, data is available
>>> print(read_data())
some data

Decorator factories are most useful to framework builders. Here is an example that gives an idea of how you could manage permissions in a framework:

class Action(object):
    def view(self):
        "Any user can view objects"

    def insert(self):
        "Only power users can insert objects"

    def delete(self):
        "Only the admin can delete objects"

where restricted is a decorator factory defined as follows

def restricted(func, user_class=User, *args, **kw):
    "Restrict access to a given class of users"
    self = args[0]
    if isinstance(self.user, user_class):
        return func(*args, **kw)
        raise PermissionError(
            '%s does not have the permission to run %s!'
            % (self.user, func.__name__))


The decorator facility can also produce a decorator starting from a class with the signature of a caller. In such a case the produced generator is able to convert functions into factories to create instances of that class.

As an example, here is a decorator which can convert a blocking function into an asynchronous function. When the function is called, it is executed in a separate thread.

(This is similar to the approach used in the concurrent.futures package. But I don’t recommend that you implement futures this way; this is just an example.)

class Future(threading.Thread):
    A class converting blocking functions into asynchronous
    functions by using threads.
    def __init__(self, func, *args, **kw):
            counter = func.counter
        except AttributeError:  # instantiate the counter at the first call
            counter = func.counter = itertools.count(1)
        name = '%s-%s' % (func.__name__, next(counter))

        def func_wrapper():
            self._result = func(*args, **kw)
        super(Future, self).__init__(target=func_wrapper, name=name)

    def result(self):
        return self._result

The decorated function returns a Future object. It has a .result() method which blocks until the underlying thread finishes and returns the final result.

Here is the minimalistic usage:

>>> @decorator(Future)
... def long_running(x):
...     time.sleep(.5)
...     return x

>>> fut1 = long_running(1)
>>> fut2 = long_running(2)
>>> fut1.result() + fut2.result()


Python’s standard library has the contextmanager decorator, which converts a generator function into a GeneratorContextManager factory. For instance, if you write this…

>>> from contextlib import contextmanager
>>> @contextmanager
... def before_after(before, after):
...     print(before)
...     yield
...     print(after)

…then before_after is a factory function that returns GeneratorContextManager objects, which provide the use of the with statement:

>>> with before_after('BEFORE', 'AFTER'):
...     print('hello')

Basically, it is as if the content of the with block was executed in the place of the yield expression in the generator function.

In Python 3.2, GeneratorContextManager objects were enhanced with a __call__ method, so that they can be used as decorators, like so:

>>> @ba
... def hello():
...     print('hello')
>>> hello()

The ba decorator basically inserts a with ba: block inside the function.

However, there are two issues:

  1. GeneratorContextManager objects are only callable in Python 3.2, so the previous example breaks in older versions of Python. (You can solve this by installing contextlib2, which backports the Python 3 functionality to Python 2.)
  2. GeneratorContextManager objects do not preserve the signature of the decorated functions. The decorated hello function above will have the generic signature hello(*args, **kwargs), but fails if called with more than zero arguments.

For these reasons, the decorator module, starting from release 3.4, offers a decorator.contextmanager decorator that solves both problems, and works in all supported Python versions. Its usage is identical, and factories decorated with decorator.contextmanager will return instances of ContextManager, a subclass of the standard library’s contextlib.GeneratorContextManager class. The subclass includes an improved __call__ method, which acts as a signature-preserving decorator.

The FunctionMaker class

You may wonder how the functionality of the decorator module is implemented. The basic building block is a FunctionMaker class. It generates on-the-fly functions with a given name and signature from a function template passed as a string.

If you’re just writing ordinary decorators, then you probably won’t need to use FunctionMaker directly. But in some circumstances, it can be handy. You will see an example shortly–in the implementation of a cool decorator utility (decorator_apply).

FunctionMaker provides the .create classmethod, which accepts the name, signature, and body of the function you want to generate, as well as the execution environment where the function is generated by exec.

Here’s an example:

>>> def f(*args, **kw): # a function with a generic signature
...     print(args, kw)

>>> f1 = FunctionMaker.create('f1(a, b)', 'f(a, b)', dict(f=f))
>>> f1(1,2)
(1, 2) {}

It is important to notice that the function body is interpolated before being executed; be careful with the % sign!

FunctionMaker.create also accepts keyword arguments. The keyword arguments are attached to the generated function. This is useful if you want to set some function attributes (e.g., the docstring __doc__).

For debugging/introspection purposes, it may be useful to see the source code of the generated function. To do this, just pass addsource=True, and the generated function will get a __source__ attribute:

>>> f1 = FunctionMaker.create(
...     'f1(a, b)', 'f(a, b)', dict(f=f), addsource=True)
>>> print(f1.__source__)
def f1(a, b):
    f(a, b)

The first argument to FunctionMaker.create can be a string (as above), or a function. This is the most common usage, since you typically decorate pre-existing functions.

If you’re writing a framework, however, you may want to use FunctionMaker.create directly, rather than decorator, because it gives you direct access to the body of the generated function.

For instance, suppose you want to instrument the __init__ methods of a set of classes, by preserving their signature. (This use case is not made up. This is done by SQAlchemy, and other frameworks, too.) Here is what happens:

  • If first argument of FunctionMaker.create is a function, an instance of FunctionMaker is created with the attributes args, varargs, keywords, and defaults. (These mirror the return values of the standard library’s inspect.getfullargspec.)
  • For each item in args (a list of strings of the names of all required arguments), an attribute arg0, arg1, …, argN is also generated.
  • Finally, there is a signature attribute, which is a string with the signature of the original function.

NOTE: You should not pass signature strings with default arguments (e.g., something like 'f1(a, b=None)'). Just pass 'f1(a, b)', followed by a tuple of defaults:

>>> f1 = FunctionMaker.create(
...     'f1(a, b)', 'f(a, b)', dict(f=f), addsource=True, defaults=(None,))
>>> print(getfullargspec(f1))
FullArgSpec(args=['a', 'b'], varargs=None, varkw=None, defaults=(None,), kwonlyargs=[], kwonlydefaults=None, annotations={})

Getting the source code

Internally, FunctionMaker.create uses exec to generate the decorated function. Therefore inspect.getsource will not work for decorated functions. In IPython, this means that the usual ?? trick will give you the (right on the spot) message Dynamically generated function. No source code available.

In the past, I considered this acceptable, since inspect.getsource does not really work with “regular” decorators. In those cases, inspect.getsource gives you the wrapper source code, which is probably not what you want:

def identity_dec(func):
    def wrapper(*args, **kw):
        return func(*args, **kw)
    return wrapper
def wrapper(*args, **kw):
    return func(*args, **kw)
>>> import inspect
>>> print(inspect.getsource(example))
    def wrapper(*args, **kw):
        return func(*args, **kw)

(See bug report 1764286 for an explanation of what is happening). Unfortunately the bug still exists in all versions of Python < 3.5.

However, there is a workaround. The decorated function has the __wrapped__ attribute, pointing to the original function. The simplest way to get the source code is to call inspect.getsource on the undecorated function:

>>> print(inspect.getsource(factorial.__wrapped__))
def factorial(n, acc=1):
    "The good old factorial"
    if n == 0:
        return acc
    return factorial(n-1, n*acc)

Dealing with third-party decorators

Sometimes on the net you find some cool decorator that you would like to include in your code. However, more often than not, the cool decorator is not signature-preserving. What you need is an easy way to upgrade third party decorators to signature-preserving decorators… without having to rewrite them in terms of decorator.

You can use a FunctionMaker to implement that functionality as follows:

def decorator_apply(dec, func):
    Decorate a function by preserving the signature even if dec
    is not a signature-preserving decorator.
    return FunctionMaker.create(
        func, 'return decfunc(%(signature)s)',
        dict(decfunc=dec(func)), __wrapped__=func)

decorator_apply sets the generated function’s __wrapped__ attribute to the original function, so you can get the right source code. If you are using a Python later than 3.2, you should also set the __qualname__ attribute to preserve the qualified name of the original function.

Notice that I am not providing this functionality in the decorator module directly, since I think it is best to rewrite the decorator instead of adding another level of indirection. However, practicality beats purity, so you can add decorator_apply to your toolbox and use it if you need to.

To give a good example for decorator_apply, I will show a pretty slick decorator that converts a tail-recursive function into an iterative function. I have shamelessly stolen the core concept from Kay Schluehr’s recipe in the Python Cookbook,

class TailRecursive(object):
    tail_recursive decorator based on Kay Schluehr's recipe
    with improvements by me and George Sakkis.

    def __init__(self, func):
        self.func = func
        self.firstcall = True
        self.CONTINUE = object()  # sentinel

    def __call__(self, *args, **kwd):
        CONTINUE = self.CONTINUE
        if self.firstcall:
            func = self.func
            self.firstcall = False
                while True:
                    result = func(*args, **kwd)
                    if result is CONTINUE:  # update arguments
                        args, kwd = self.argskwd
                    else:  # last call
                        return result
                self.firstcall = True
        else:  # return the arguments of the tail call
            self.argskwd = args, kwd
            return CONTINUE

Here the decorator is implemented as a class returning callable objects.

def tail_recursive(func):
    return decorator_apply(TailRecursive, func)

Here is how you apply the upgraded decorator to the good old factorial:

def factorial(n, acc=1):
    "The good old factorial"
    if n == 0:
        return acc
    return factorial(n-1, n*acc)
>>> print(factorial(4))

This decorator is pretty impressive, and should give you some food for thought! ;)

Notice that there is no recursion limit now; you can easily compute factorial(1001) (or larger) without filling the stack frame.

Notice also that the decorator will not work on functions which are not tail recursive, such as the following:

def fact(n):  # this is not tail-recursive
    if n == 0:
        return 1
    return n * fact(n-1)

Reminder: A function is tail recursive if it does either of the following:

  • returns a value without making a recursive call; or,
  • returns directly the result of a recursive call.

Python 3.5 coroutines

I am personally not using Python 3.5 coroutines yet, because at work we are still maintaining compatibility with Python 2.7. However, some users requested support for coroutines and since version 4.1 the decorator module has it. You should consider the support experimental and kindly report issues if you find any.

Here I will give a single example of usage. Suppose you want to log the moment a coroutine starts and the moment it stops for debugging purposes. You could write code like the following:

import time
import logging
from asyncio import get_event_loop, sleep, wait
from decorator import decorator

async def log_start_stop(coro, *args, **kwargs):'Starting %s%s', coro.__name__, args)
    t0 = time.time()
    await coro(*args, **kwargs)
    dt = time.time() - t0'Ending %s%s after %d seconds', coro.__name__, args, dt)

async def make_task(n):
    for i in range(n):
        await sleep(1)

if __name__ == '__main__':
    tasks = [make_task(3), make_task(2), make_task(1)]

and you will get an output like this:

INFO:root:Starting make_task(1,)
INFO:root:Starting make_task(3,)
INFO:root:Starting make_task(2,)
INFO:root:Ending make_task(1,) after 1 seconds
INFO:root:Ending make_task(2,) after 2 seconds
INFO:root:Ending make_task(3,) after 3 seconds

This may be handy if you have trouble understanding what it going on with a particularly complex chain of coroutines. With a single line you can decorate the troubling coroutine function, understand what happens, fix the issue and then remove the decorator (or keep it if continuous monitoring of the coroutines makes sense). Notice that inspect.iscoroutinefunction(make_task) will return the right answer (i.e. True).

It is also possible to define decorators converting coroutine functions into regular functions, such as the following:

def coro_to_func(coro, *args, **kw):
    "Convert a coroutine into a function"
    return get_event_loop().run_until_complete(coro(*args, **kw))

Notice the diffence: the caller in log_start_stop was a coroutine function and the associate decorator was converting coroutines->coroutines; the caller in coro_to_func is a regular function and converts coroutines -> functions.

Multiple dispatch

There has been talk of implementing multiple dispatch functions (i.e. “generic functions”) in Python for over ten years. Last year, something concrete was done for the first time. As of Python 3.4, we have the decorator functools.singledispatch to implement generic functions!

As its name implies, it is limited to single dispatch; in other words, it is able to dispatch on the first argument of the function only.

The decorator module provides the decorator factory dispatch_on, which can be used to implement generic functions dispatching on any argument. Moreover, it can manage dispatching on more than one argument. (And, of course, it is signature-preserving.)

Here is a concrete example (from a real-life use case) where it is desiderable to dispatch on the second argument.

Suppose you have an XMLWriter class, which is instantiated with some configuration parameters, and has the .write method which serializes objects to XML:

class XMLWriter(object):
    def __init__(self, **config):
        self.cfg = config

    def write(self, obj):
        raise NotImplementedError(type(obj))

Here, you want to dispatch on the second argument; the first is already taken by self. The dispatch_on decorator factory allows you to specify the dispatch argument simply by passing its name as a string. (Note that if you misspell the name you will get an error.)

The decorated function write is turned into a generic function ( write is a function at the idea it is decorated; it will be turned into a method later, at class instantiation time), and it is called if there are no more specialized implementations.

Usually, default functions should raise a NotImplementedError, thus forcing people to register some implementation. You can perform the registration with a decorator:

def writefloat(self, obj):
    return '<float>%s</float>' % obj

Now XMLWriter can serialize floats:

>>> writer = XMLWriter()
>>> writer.write(2.3)

I could give a down-to-earth example of situations in which it is desiderable to dispatch on more than one argument–for instance, I once implemented a database-access library where the first dispatching argument was the the database driver, and the second was the database record–but here I will follow tradition, and show the time-honored Rock-Paper-Scissors example:

class Rock(object):
    ordinal = 0
class Paper(object):
    ordinal = 1
class Scissors(object):
    ordinal = 2

I have added an ordinal to the Rock-Paper-Scissors classes to simplify the implementation. The idea is to define a generic function (win(a, b)) of two arguments corresponding to the moves of the first and second players. The moves are instances of the classes Rock, Paper, and Scissors:

  • Paper wins over Rock
  • Scissors wins over Paper
  • Rock wins over Scissors

The function will return +1 for a win, -1 for a loss, and 0 for parity. There are 9 combinations, but combinations with the same ordinal (i.e. the same class) return 0. Moreover, by exchanging the order of the arguments, the sign of the result changes. Therefore, it is sufficient to directly specify only 3 implementations:

@dispatch_on('a', 'b')
def win(a, b):
    if a.ordinal == b.ordinal:
        return 0
    elif a.ordinal > b.ordinal:
        return -win(b, a)
    raise NotImplementedError((type(a), type(b)))
@win.register(Rock, Paper)
def winRockPaper(a, b):
    return -1
@win.register(Paper, Scissors)
def winPaperScissors(a, b):
    return -1
@win.register(Rock, Scissors)
def winRockScissors(a, b):
    return 1

Here is the result:

>>> win(Paper(), Rock())
>>> win(Scissors(), Paper())
>>> win(Rock(), Scissors())
>>> win(Paper(), Paper())
>>> win(Rock(), Rock())
>>> win(Scissors(), Scissors())
>>> win(Rock(), Paper())
>>> win(Paper(), Scissors())
>>> win(Scissors(), Rock())

The point of generic functions is that they play well with subclassing. For instance, suppose we define a StrongRock, which does not lose against Paper:

class StrongRock(Rock):
@win.register(StrongRock, Paper)
def winStrongRockPaper(a, b):
    return 0

Then you do not need to define other implementations; they are inherited from the parent:

>>> win(StrongRock(), Scissors())

You can introspect the precedence used by the dispath algorithm by calling .dispatch_info(*types):

>>> win.dispatch_info(StrongRock, Scissors)
[('StrongRock', 'Scissors'), ('Rock', 'Scissors')]

Since there is no direct implementation for (StrongRock, Scissors), the dispatcher will look at the implementation for (Rock, Scissors) which is available. Internally, the algorithm is doing a cross product of the class precedence lists (or Method Resolution Orders, MRO for short) of StrongRock and Scissors, respectively.

Generic functions and virtual ancestors

In Python, generic functions are complicated by the existence of “virtual ancestors”: superclasses which are not in the class hierarchy.

Consider this class:

class WithLength(object):
    def __len__(self):
        return 0

This class defines a __len__ method, and is therefore considered to be a subclass of the abstract base class collections.Sized:

>>> issubclass(WithLength, collections.Sized)

However, collections.Sized is not in the MRO of WithLength; it is not a true ancestor. Any implementation of generic functions (even with single dispatch) must go through some contorsion to take into account the virtual ancestors.

In particular, if we define a generic function…

def get_length(obj):
    raise NotImplementedError(type(obj))

…implemented on all classes with a length…

def get_length_sized(obj):
    return len(obj)

…then get_length must be defined on WithLength instances…

>>> get_length(WithLength())

…even if collections.Sized is not a true ancestor of WithLength.

Of course, this is a contrived example–you could just use the builtin len–but you should get the idea.

Since in Python it is possible to consider any instance of ABCMeta as a virtual ancestor of any other class (it is enough to register it as ancestor.register(cls)), any implementation of generic functions must be aware of the registration mechanism.

For example, suppose you are using a third-party set-like class, like the following:

class SomeSet(collections.Sized):
    # methods that make SomeSet set-like
    # not shown ...
    def __len__(self):
        return 0

Here, the author of SomeSet made a mistake by inheriting from collections.Sized (instead of collections.Set).

This is not a problem. You can register a posteriori collections.Set as a virtual ancestor of SomeSet:

>>> _ = collections.Set.register(SomeSet)
>>> issubclass(SomeSet, collections.Set)

Now, let’s define an implementation of get_length specific to set:

def get_length_set(obj):
    return 1

The current implementation (and functools.singledispatch too) is able to discern that a Set is a Sized object, by looking at the class registry, so it uses the more specific implementation for Set:

>>> get_length(SomeSet())  # NB: the implementation for Sized would give 0

Sometimes it is not clear how to dispatch. For instance, consider a class C registered both as collections.Iterable and collections.Sized, and defines a generic function g with implementations for both collections.Iterable and collections.Sized:

def singledispatch_example1():
    singledispatch = dispatch_on('obj')

    def g(obj):
        raise NotImplementedError(type(g))

    def g_sized(object):
        return "sized"

    def g_iterable(object):
        return "iterable"

    g(C())  # RuntimeError: Ambiguous dispatch: Iterable or Sized?

It is impossible to decide which implementation to use, since the ancestors are independent. The following function will raise a RuntimeError when called. This is consistent with the “refuse the temptation to guess” philosophy. functools.singledispatch would raise a similar error.

It would be easy to rely on the order of registration to decide the precedence order. This is reasonable, but also fragile:

  • if, during some refactoring, you change the registration order by mistake, a different implementation could be taken;
  • if implementations of the generic functions are distributed across modules, and you change the import order, a different implementation could be taken.

So the decorator module prefers to raise an error in the face of ambiguity. This is the same approach taken by the standard library.

However, it should be noted that the dispatch algorithm used by the decorator module is different from the one used by the standard library, so in certain cases you will get different answers. The difference is that functools.singledispatch tries to insert the virtual ancestors before the base classes, whereas decorator.dispatch_on tries to insert them after the base classes.

Here’s an example that shows the difference:

def singledispatch_example2():
    # adapted from functools.singledispatch test case
    singledispatch = dispatch_on('arg')

    class S(object):

    class V(c.Sized, S):
        def __len__(self):
            return 0

    def g(arg):
        return "base"

    def g_s(arg):
        return "s"

    def g_container(arg):
        return "container"

    v = V()
    assert g(v) == "s"
    c.Container.register(V)  # add c.Container to the virtual mro of V
    assert g(v) == "s"  # since the virtual mro is V, Sized, S, Container
    return g, V

If you play with this example and replace the singledispatch definition with functools.singledispatch, the assertion will break: g will return "container" instead of "s", because functools.singledispatch will insert the Container class right before S.

Notice that here I am not making any bold claim such as “the standard library algorithm is wrong and my algorithm is right” or viceversa. It just point out that there are some subtle differences. The only way to understand what is really happening here is to scratch your head by looking at the implementations. I will just notice that .dispatch_info is quite essential to see the class precedence list used by algorithm:

>>> g, V = singledispatch_example2()
>>> g.dispatch_info(V)
[('V',), ('Sized',), ('S',), ('Container',)]

The current implementation does not implement any kind of cooperation between implementations. In other words, nothing is akin either to call-next-method in Lisp, or to super in Python.

Finally, let me notice that the decorator module implementation does not use any cache, whereas the singledispatch implementation does.

Caveats and limitations

One thing you should be aware of, is the performance penalty of decorators. The worse case is shown by the following example:

$ cat
python3 -m timeit -s "
from decorator import decorator

def do_nothing(func, *args, **kw):
    return func(*args, **kw)

def f():
" "f()"

python3 -m timeit -s "
def f():
" "f()"

On my laptop, using the do_nothing decorator instead of the plain function is five times slower:

$ bash
1000000 loops, best of 3: 1.39 usec per loop
1000000 loops, best of 3: 0.278 usec per loop

Of course, a real life function probably does something more useful than the function f here, so the real life performance penalty could be negligible. As always, the only way to know if there is a penalty in your specific use case is to measure it.

More importantly, you should be aware that decorators will make your tracebacks longer and more difficult to understand.

Consider this example:

>>> @trace
... def f():
...     1/0

Calling f() gives you a ZeroDivisionError. But since the function is decorated, the traceback is longer:

>>> f()
Traceback (most recent call last):
     File "<string>", line 2, in f
     File "<doctest __main__[22]>", line 4, in trace
       return f(*args, **kw)
     File "<doctest __main__[51]>", line 3, in f
ZeroDivisionError: ...

You see here the inner call to the decorator trace, which calls f(*args, **kw), and a reference to File "<string>", line 2, in f.

This latter reference is due to the fact that, internally, the decorator module uses exec to generate the decorated function. Notice that exec is not responsible for the performance penalty, since is the called only once (at function decoration time); it is not called each time the decorated function is called.

Presently, there is no clean way to avoid exec. A clean solution would require changing the CPython implementation, by adding a hook to functions (to allow changing their signature directly).

Even in Python 3.5, it is impossible to change the function signature directly. Thus, the decorator module is still useful! As a matter of fact, this is the main reason why I still maintain the module and release new versions.

It should be noted that in Python 3.5, a lot of improvements have been made: you can decorate a function with func_tools.update_wrapper, and pydoc will see the correct signature. Unfortunately, the function will still have an incorrect signature internally, as you can see by using inspect.getfullargspec; so, all documentation tools using inspect.getfullargspec - which has been rightly deprecated - will see the wrong signature.

In the present implementation, decorators generated by decorator can only be used on user-defined Python functions or methods. They cannot be used on generic callable objects or built-in functions, due to limitations of the standard library’s inspect module, especially for Python 2. In Python 3.5, many such limitations have been removed, but I still think that it is cleaner and safer to decorate only functions. If you want to decorate things like classmethods/staticmethods and general callables - which I will never support in the decorator module - I suggest you to look at the wrapt project by Graeme Dumpleton.

There is a strange quirk when decorating functions with keyword arguments, if one of the arguments has the same name used in the caller function for the first argument. The quirk was reported by David Goldstein.

Here is an example where it is manifest:

>>> @memoize
... def getkeys(**kw):
...     return kw.keys()
>>> getkeys(func='a')
Traceback (most recent call last):
TypeError: _memoize() got multiple values for ... 'func'

The error message looks really strange… until you realize that the caller function _memoize uses func as first argument, so there is a confusion between the positional argument and the keywork arguments.

The solution is to change the name of the first argument in _memoize, or to change the implementation like so:

def _memoize(*all_args, **kw):
    func = all_args[0]
    args = all_args[1:]
    if kw:  # frozenset is used to ensure hashability
        key = args, frozenset(kw.items())
        key = args
    cache = func.cache  # attribute added by memoize
    if key not in cache:
        cache[key] = func(*args, **kw)
    return cache[key]

This avoids the need to name the first argument, so the problem simply disappears. This is a technique that you should keep in mind when writing decorators for functions with keyword arguments. Also, notice that lately I have come to believe that decorating functions with keyword arguments is not such a good idea, and you may want not to do that.

On a similar note, there is a restriction on argument names. For instance, if you name an argument _call_ or _func_, you will get a NameError:

>>> @trace
... def f(_func_): print(f)
Traceback (most recent call last):
NameError: _func_ is overridden in
def f(_func_):
    return _call_(_func_, _func_)

Finally, the implementation is such that the decorated function makes a (shallow) copy of the original function dictionary:

>>> def f(): pass # the original function
>>> f.attr1 = "something" # setting an attribute
>>> f.attr2 = "something else" # setting another attribute

>>> traced_f = trace(f) # the decorated function

>>> traced_f.attr1
>>> traced_f.attr2 = "something different" # setting attr
>>> f.attr2 # the original attribute did not change
'something else'

LICENSE (2-clause BSD)

Copyright (c) 2005-2017, Michele Simionato All rights reserved.

Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:

Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. Redistributions in bytecode form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.


If you use this software and you are happy with it, consider sending me a note, just to gratify my ego. On the other hand, if you use this software and you are unhappy with it, send me a patch!