DEV Community

loading...
Cover image for The Correct Way to Overload Functions in Python

The Correct Way to Overload Functions in Python

Martin Heinz
My name is Martin Heinz and I'm a software developer/DevOps engineer. I'm from Slovakia, living in Bratislava.
Originally published at martinheinz.dev ・6 min read

Function overloading is a common programming pattern which seems to be reserved to statically-typed, compiled languages. Yet there's an easy way to implement it in Python with help of Multiple Dispatch or as it's called in Python multimethods.

Overloading

First things first - you might be asking, how can we implement method overloading in Python when we all know that it's not possible? Well, even though Python is dynamically-typed language and therefore cannot have proper method overloading as that requires the language to be able to discriminate between types at compile-time, we can still implement it in a bit different way that is suitable for dynamically-typed languages.

This approach is called Multiple Dispatch or multimethods, where the interpreter differentiates between multiple implementations of a function/method at runtime based on dynamically determined types. To be more precise, the language uses types of arguments passed to a function during its invocation to dynamically choose which one of the multiple function implementations to use (or dispatch).

Now you might be thinking: "Do we really need this though? If it can't be implemented normally, maybe we shouldn't use it in Python..." Yea, valid point, but there are good reasons to want to implement some form of function/method overloading in Python. It's powerful tool that can make code more concise, readable and minimise its complexity. Without multimethods though, the "obvious way" to do this is using type inspection with isinstance(). This is very ugly, brittle solution that is closed to extension and I would call it an anti-pattern.

Besides that, there already is method overloading in Python for operators and methods like len() or new() using so-called dunder or magic methods (see docs here) and we all use that quite often, so why not use proper overloading for all the function, right?

So, now we know that we can kind-of implement overloading in Python, so how exactly do we do that?

Single Dispatch

Above we spoke about Multiple Dispatch, but Python doesn't support this out-of-the box, or in other words Multiple Dispatch is not a feature of Python standard library. What is available to us however, is called Single Dispatch, so let's begin with this simpler case first.

The only actual difference between multi and single dispatch is number of arguments which we can overload. So, for this implementation in standard library it's just one.

The function (and decorator) that provides this feature is called singledispatch and can be found in functools module.

This whole concept is best explained with some examples. There are many "academic" examples of overloading functions (geometric shapes, addition, subtraction...) that we've probably all seen already. Rather than going over that, let's see some practical examples. So, here's first example for singledispatch to format dates, times and datetimes:

from functools import singledispatch
from datetime import date, datetime, time

@singledispatch
def format(arg):
    return arg

@format.register
def _(arg: date):
    return f"{arg.day}-{arg.month}-{arg.year}"

@format.register
def _(arg: datetime):
    return f"{arg.day}-{arg.month}-{arg.year} {arg.hour}:{arg.minute}:{arg.second}"

@format.register(time)
def _(arg):
    return f"{arg.hour}:{arg.minute}:{arg.second}"

print(format("today"))
# today
print(format(date(2021, 5, 26)))
# 26-5-2021
print(format(datetime(2021, 5, 26, 17, 25, 10)))
# 26-5-2021 17:25:10
print(format(time(19, 22, 15)))
# 19:22:15
Enter fullscreen mode Exit fullscreen mode

We begin by defining the base format function that is going to be overloaded. This function is decorated with @singledispatch and provides base implementation, which is used if no better options is available. Next, we define individual functions for each type that we want to overload - in this case date, datetime and time - each of these have name _ (underscore) because they will be called (dispatched) through the format method anyway, so no need to give them useful names. Each of them is also decorated with @format.register which attaches them to the previously mentioned format function. Then, to make it possible to differentiate between types, we have two options - we can use type annotations - as demonstrated in first two cases or explicitly add the type to decorator as with the last one from the example.

In some cases it might make sense to use same implementation for multiple types - for example for number types such as int and float - for these situations decorator stacking is allowed, meaning that you can list (stack) multiple @format.register(type) lines to associate a function with all the valid types.

Besides ability to overload basic functions, functools module contains also singledispatchmethod that can be applied to methods of a class. Example of that could be the following:

from functools import singledispatchmethod
from datetime import date, time

class Formatter:
    @singledispatchmethod
    def format(self, arg):
        raise NotImplementedError(f"Cannot format value of type {type(arg)}")

    @format.register
    def _(self, arg: date):
        return f"{arg.day}-{arg.month}-{arg.year}"

    @format.register
    def _(self, arg: time):
        return f"{arg.hour}:{arg.minute}:{arg.second}"

f = Formatter()
print(f.format(date(2021, 5, 26)))
# 26-5-2021
print(f.format(time(19, 22, 15)))
# 19:22:15
Enter fullscreen mode Exit fullscreen mode

Multiple Dispatch

Oftentimes Single Dispatch won't be sufficient and you might need the proper Multiple Dispatch functionality. This is available from multipledispatch module which can be found here and can be installed with pip install multipledispatch.

This module and it's decorator - @dispatch, behaves very similarly to the @singledispatch in the standard library. Only actual difference is that it can take multiple types as arguments:

from multipledispatch import dispatch

@dispatch(list, str)
def concatenate(a, b):
    a.append(b)
    return a

@dispatch(str, str)
def concatenate(a, b):
    return a + b

@dispatch(str, int)
def concatenate(a, b):
    return a + str(b)


print(concatenate(["a", "b"], "c"))
# ['a', 'b', 'c']
print(concatenate("Hello", "World"))
# HelloWorld
print(concatenate("a", 1))
# a1
Enter fullscreen mode Exit fullscreen mode

The above snippet shows how we can use @dispatch decorator to overload multiple arguments, for example to implement concatenation of various types. As you probably noticed, with multipledispatch library we didn't need to define and register base function, rather we created multiple functions with same name. If we wanted to provide base implementation, we could use @dispatch(object, object) which would catch any non-specific argument types.

The previous examples shows proof-of-concept, but if we wanted to really implement such concatenate function, we would need to make it much more generic. This can be solved with use of union types. In this specific example we could change the first function as follows:

@dispatch((list, tuple), (str, int))
def concatenate(a, b):
    return list(a) + [b]

print(concatenate(["a", "b"], "c"))
# ['a', 'b', 'c']
print(concatenate(("a", "b"), 1))
# ['a', 'b', 1]
Enter fullscreen mode Exit fullscreen mode

This would make it so that first argument of the function could be any of list or tuple, while second one would be str or int. This is already much better than the previous solution, but it can be further improved using abstract types. Instead of listing all the possible sequences, we can use Sequence abstract type (assuming that our implementation can handle it) which covers things like list, tuple or range:

from collections.abc import Sequence

@dispatch(Sequence, (str, int))
def concatenate(a, b):
    return list(a) + [b]
Enter fullscreen mode Exit fullscreen mode

If you want to take this approach, then it's good to take a look at collections.abc module and see which container data-type best suits your needs. Mostly to make sure that your function will be able to handle all the types that fall into the chosen container.

All this mixing and matching of argument types is convenient, but can also cause ambiguities when choosing suitable function for some specific set of parameters. Fortunately, multipledispatch provides AmbiguityWarning which is raised if ambiguous behaviour is possible:

test_multipledispatch:10: AmbiguityWarning:
Ambiguities exist in dispatched function some_func

The following signatures may result in ambiguous behavior:
    [str, object], [object, str]


Consider making the following additions:

@dispatch(str, str)
def some_func(...)
Enter fullscreen mode Exit fullscreen mode

Closing Thoughts

In this article we went over a simple, yet powerful concept which I rarely see being used in Python, which is a shame considering that it can greatly improve code readability and get rid of anti-patters like type inspection using isinstance(). Also, I hope you would agree that this approach to function overloading should be considered the "obvious way" and I hope that you will make use of it when needed.

If you want to dive deeper into this topic and get your hands dirty you can implement multimethods yourselves as shown in Guido's article - this can be a good exercise to understand how multiple dispatch actually works.

Finally, I should also probably mention that this article omits examples of the well-known operator overloading which I mentioned in the beginning as well as some approaches for overloading constructors for example using factories. So, in case that's what you're looking for, go check out these links/resources, which give good overview on there topics.

Discussion (1)

Collapse
hanpari profile image
Pavel Morava

I thought you were going to talk about this overloading. Nice article, though.

I agree it is a shame that single dispatch is not used more. I wrote an article related to singledispatch here.

I believe the greatest advantage of this approach is that one can define a new implementation anywhere. The caveat is that VS Code had trouble with IntelliSense. PyCharm worked fine though.