DEV Community

Cover image for Reconciling Dataclasses And Properties In Python
Florimond Manca
Florimond Manca

Posted on • Updated on • Originally published at blog.florimond.dev

Reconciling Dataclasses And Properties In Python

Originally published at blog.florimondmanca.com on Oct 6, 2018.

If you can't wait for the TL;DR, jump to Lessons Learned at the end of this article.

I've been playing around with Python 3.7's dataclasses — and so far they've been super awesome. For the most part, they're quite intuitive and easy to use and customize.

Yet, there's one thing I've been struggling with and couldn't find online help for, which is implementing Python properties on dataclasses. And chances are — I am not alone.

That's why I decided to solve the problem once and for all. I wanted to figure out how to reconcile dataclasses and properties.

So I sat down before my computer, fired up an interpreter and wrote down my thought process as I made my way towards a solution. After a night of trial and error and some wording enhancements, the result is this very blog post!

As a pleasant side effect, you'll get to see many features of dataclasses in action. I'll start with a quick overview of dataclasses, and we'll get to the code right after that.

Of course, you'll find all the code supporting this post on GitHub.

Dataclasses: a 10,000-ft overview

Dataclasses are, simply put, classes made to hold data. Their specification in PEP-557 was motivated by the fact that a lot of classes we write are merely used as editable data containers. When that happens, we spend time writing boilerplate code which most often results in an ugly __init__() method with tons of arguments and just as many lines for storing them as attributes — not to speak about handling default arguments

Dataclasses, featuring Raymond Hettinger in a GIF.

And there is! The answer is: dataclasses. 🎉

Python implements dataclasses in the well-named dataclasses module, whose superstar is the @dataclass decorator. This decorator is really just a code generator. It takes advantage of Python's type annotations (if you still don't use them, you really should) to automatically generate boilerplate code you'd have to mechanically type yourself otherwise.

As a point of comparison, here's how you would create a Vehicle class with a wheels attribute using a regular class declaration:

class Vehicle:
    def __init__(self, wheels: int):
        self.wheels = wheels
Enter fullscreen mode Exit fullscreen mode

Nothing fancy, really. Now, the @dataclass version:

# 0_initial.py
from dataclasses import dataclass

@dataclass
class Vehicle:
    wheels: int
Enter fullscreen mode Exit fullscreen mode

Believe it or not — these two code snippets are strictly equivalent! It's actually a win, because beyond __init__(), @dataclass generates a bunch of extra stuff for free, including a handsome __repr__():

>>> car = Vehicle(wheels=4)
>>> car
Vehicle(wheels=4)
Enter fullscreen mode Exit fullscreen mode

In short, dataclasses are a simple, elegant, Pythonic way of creating classes that hold data. 🐍

But…

The problem

I sometimes resort to the @property decorator to implement specific logic when getting/setting an attribute. That's really just the Pythonic way of implementing getters and setters.

Building upon the previous Vehicle class, I would make the wheels attribute private and put an @property on top of it:

class Vehicle:

    def __init__(self, wheels: int):
        self._wheels = wheels
        # note the underscore — it's now private! 👻

    @property
    def wheels(self) -> int:
        print('getting wheels')
        return self._wheels

    @wheels.setter
    def wheels(self, wheels: int):
        print('setting wheels to', wheels)
        self._wheels = wheels
Enter fullscreen mode Exit fullscreen mode

Here's what it looks like:

>>> car = Vehicle(wheels=4)
>>> car.wheels = 3
setting wheels to 3
>>> car.wheels
getting wheels
3
>>>
Enter fullscreen mode Exit fullscreen mode

Now the question is — how can I implement such a property on a dataclass?

Wait, so this is the problem?

You may think this to be a trivial question. Mind you, it is not trivial.

Dataclasses generate the __init__() method for you — and that's great. They even provide a __post_init__() hook method in case you want to do some more initialization (see Post-init processing).

However, this means you cannot do the same trick as with normal classes, i.e. storing a public-looking argument (e.g. wheels) into a private attribute (_wheels) that you'll build an @property out of.

That's where the problem comes from. And to be honest, it gave me a bit of a headache.

Because I think the problem solving was interesting, I'll take you through 5 consecutive attempts to correctly implement that property on a dataclass version of the Vehicle class.

Attempt 1: declaring a private field

First, let's keep things simple. We want to store wheels in a private field and use it in the @property, right? So why not simply declare a _wheels field on the dataclass?

# 1_private_field.py
from dataclasses import dataclass

@dataclass
class Vehicle:
    _wheels: int

    # wheels @property as before
Enter fullscreen mode Exit fullscreen mode

Unfortunately, this won't work — otherwise this blog post wouldn't be of much use! 😙

The reason why is because the constructor now expects a _wheels argument instead of wheels.

>>> car = Vehicle(wheels=4)
Traceback (most recent call last)
<ipython-input-3-9c9de8fb1422> in <module>()
----> 1 car = Vehicle(wheels=4)
TypeError: __init__() got an unexpected keyword argument 'wheels'
Enter fullscreen mode Exit fullscreen mode

To be fair, that's just @dataclass doing its job. Still, that's not what we want.

Attempt 2: make use of InitVar

If you read through the documentation, you'll learn that InitVar allows you to implement init-only variables. These variables can be passed to the constructor, but won't be stored in an attribute on the class. Instead, the variable is passed as an argument to __post_init__().

Why not use this to create an init-only wheels variable and store that in a _wheels field? We just need to give the latter a default (e.g. None) so that it is not required by the constructor:

# 2_initvar.py
from dataclasses import dataclass, InitVar

@dataclass
class Vehicle:

    wheels: InitVar[int]
    _wheels: int = None  # default given => not required in __init__()

    def __post_init__(self, wheels: int):
        self._wheels = wheels

    # wheels @property as before
Enter fullscreen mode Exit fullscreen mode

Granted, __init__() now expects a wheels argument instead of _wheels, which is what we want.

However, @dataclass now generates other boilerplate code and magic methods using _wheels, which is problematic.

>>> car = Vehicle(wheels=4)
setting wheels to 4
>>> car
Vehicle(_wheels=4)  # 😕
Enter fullscreen mode Exit fullscreen mode

Attempt 3: make use of field()

Digging deeper into the docs, I found that one could fine-tune the field generation behavior using the field() function. You can pass it a default value and it accepts a repr argument to control whether the field should be included in the generated __repr__(). Here's how it looks when used on _wheels:

# 3_field.py
from dataclasses import dataclass, field

@dataclass
class Vehicle:

    wheels: InitVar[int]
    _wheels: int = field(default=None, repr=False)

    # __post_init__() as before
    # wheels @property as before
Enter fullscreen mode Exit fullscreen mode

Sweet — we don't have _wheels included in __repr__() anymore. But we still don't have wheels either!

>>> car = Vehicle(wheels=4)
setting wheels to 4
>>> car
Vehicle()  # Where is `wheels=4`? 😕😕😕
Enter fullscreen mode Exit fullscreen mode

Attempt 4: make wheels a proper field

In the previous attempts, wheels was an InitVar — not a field. This time, let's declare it as a field in its own right. It will be possible to pass it in the constructor, and it should be included in __repr__() this time around.

The good thing is, the @property definition of wheels declared later in the class will not interfere with @dataclass's generation process — because it is not a type annotation, which is what @dataclass relies on to generate the fields.

That might start to be a bit complicated, so let me show you some code. I'll reproduce the @property for wheels in full this time:

# 4_wheels_field.py
from dataclasses import dataclass, field

@dataclass
class Vehicle:

    wheels: int  # Now a regular dataclass field

    # The rest just as before:

    _wheels: int = field(default=None, repr=False)

    def __post_init__(self):
        # Note: wheels is not passed as an argument
        # here anymore, because it is not an
        # `InitVar` anymore.
        self._wheels = self.wheels  # (1)

    @property
    def wheels(self) -> int:
        print('getting wheels')
        return self._wheels

    @wheels.setter
    def wheels(self, wheels: int):
        print('setting wheels to', wheels)
        self._wheels = wheels
Enter fullscreen mode Exit fullscreen mode

It looks like we're getting there, aren't we?

Unfortunately, not quite. 😞 There's a catch in this implementation.

Indeed, you may think line (1) puts the value of wheels that was given to the constructor (and stored into self.wheels) into _wheels. For example, calling Vehicle(wheels=4) would result in having _wheels == 4. Sadly, that is not the case!

Here's why: when executing __post_init__(), self.wheels is the value returned by the wheels property's getter — not the value initially stored during __init__()! And that getter returns self._wheels, which is None by default.

I know, it's getting all tangled up, but please bear with me:

>>> car = Vehicle(wheels=4)
setting wheels to 4
getting wheels  # hint: this is (1) being executed
>>> print(car.wheels)
getting wheels
None  # nope, nothing in there…
Enter fullscreen mode Exit fullscreen mode

If you think about it, what we're doing in (1) is just replacing _wheels with its own value. Quite useless, if you ask me. We would actually get the same result if we didn't even implement __post_init__().

Duh! So what can we do? 😩

Fortunately, there's hope!

Attempt 5: exclude _wheels from the constructor

Let me warn you — this fifth and final attempt will work, and the reason why, which I'll explain in a minute, is outrageous.

At this point, you'd be right to feel sad — I felt sad myself. But fear not! There is one thing from the documentation that we haven't tried yet.

So far, the _wheels attribute has been declared using field(default=None, repr=False). Using default=None here means that we are able to omit passing a value for _wheels in the constructor — it will be given the value of None during __init__(). However, it is still possible to give it a value in the constructor, and everything will work as expected:

>>> car = Vehicle(wheels=4, _wheels=3)
setting wheels to 4
getting wheels
>>> car.wheels
getting wheels
3
Enter fullscreen mode Exit fullscreen mode

Well, how about we find a way to remove the _wheels argument from the constructor? Will it solve our problem? (Spoiler alert: it will.)

Guess what: field() accepts an init argument for that exact purpose. The docs on field() read:

init: If true (the default), this field is included as a parameter to the generated __init__ method.

Sounds trivial, right? Well, let's try using it on _wheels (I removed the __post_init__() hook because we previously showed that it was actually useless):

# 5_init_false.py
from dataclasses import dataclass, field

@dataclass
class Vehicle:

    wheels: int
    _wheels: int = field(init=False, repr=False)

    @property
    def wheels(self) -> int:
        print('getting wheels')
        return self._wheels

    @wheels.setter
    def wheels(self, wheels: int):
        print('setting wheels to', wheels)
        self._wheels = wheels
Enter fullscreen mode Exit fullscreen mode

Well, guess what? This has just solved all of our problems.

Because we used init=False, the constructor generated by @dataclass will not initialize _wheels at all.

However, it will initialize wheels with the value passed to the constructor. If we could extract the generated code, one of the instructions in __init__() would look like this:

self.wheels = wheels
Enter fullscreen mode Exit fullscreen mode

Now, you tell me — what does this execute exactly?

Yep, that's right. It will execute the setter! 🙀

Look! We've actually been seeing the print statement in the setter since attempt 4!

>>> car = Vehicle(wheels=4)
setting wheels to 4  # the `wheels` setter being called
Enter fullscreen mode Exit fullscreen mode

Let me remind you of the code for that setter:

@wheels.setter
def wheels(self, wheels: int):
    print('setting wheels to', wheels)
    self._wheels = wheels
Enter fullscreen mode Exit fullscreen mode

It sets the value of _wheels!

As a result, the value for wheels passed in the constructor is put into _wheels — and nowhere else, because after the class has been generated, wheels only refers to the @property, not to a field on the dataclass.

If you think about it, this is exactly what we were doing when implementing the property on good ol' regular classes. Remember?

class Vehicle:

    def __init__(self, wheels: int):
        self._wheels = wheels
        # This is equivalent to calling:
        # `self.wheels = wheels`
        # which *is* what the __init__() method
        # now generated by @dataclass actually does.
Enter fullscreen mode Exit fullscreen mode

Caveat: this approach only holds because the property's setter is implemented. If we only implemented a getter (i.e. to make a read-only field), the __init__() method wouldn't be able to assign the attribute and would crash. This is intended behavior, though, because dataclasses were designed to be editable data containers. If you really need read-only fields, you shouldn't be resorting to dataclasses in the first place. Perhaps NamedTuples would be a viable alternative — they are the read-only equivalent of dataclasses.

Anyway, long story short…

Success!

We have successfully implemented properties on dataclasses. 🎉

Memes will never let you down.

To be honest, it was surprisingly not easy. We've been through five different attempts, navigating through the documentation and painstakingly coding our way towards dataclass properties.

So after all this hassle, can we at least derive a quick recipe for implementing properties on dataclasses?

The answer is: yes, we can. ✌️

Lessons learned

Have you noticed a pattern between using an @property on a regular class vs. on a dataclass?

Look, here's the regular class version:

class Vehicle:

    def __init__(self, wheels: int):
        self._wheels = wheels

    @property
    def wheels(self) -> int:
        return self._wheels

    @wheels.setter
    def wheels(self, wheels: int):
        self._wheels = wheels
Enter fullscreen mode Exit fullscreen mode

And the dataclass version, using a diff syntax to highlight the differences:

+ from dataclasses import dataclass, field

+ @dataclass
  class Vehicle:

+     wheels: int
+     _wheels: field(init=False, repr=False)

-     def __init__(self, wheels: int):
-         self._wheels = wheels

      @property
      def wheels(self) -> int:
          return self._wheels

      @wheels.setter
      def wheels(self, wheels: int):
          self._wheels = wheels
Enter fullscreen mode Exit fullscreen mode

Written in plain words, for you litterature freaks:

How to implement a property on a dataclass:

  1. Remove __init__()
  2. Declare the property as a field
  3. Add an associated private variable using field(init=False, repr=False).

Way to go!

If you managed to read up to here, congratulations! This blog post dealt with a highly specific and technical topic, yet I've had a lot of fun writing it and figuring out how to implement dataclass properties.

For those wondering — the series on Apache Kafka still goes on! I figured a small break never hurts, and I felt like it gave me the opportunity to write something spontaneous.

I hope you enjoyed this post and, as mentioned in introduction, you can find all the code on GitHub. See you next time! 💻

Stay in touch!

If you enjoyed this post, you can find me on Twitter for updates, announcements and news. 🐤

Top comments (6)

Collapse
 
rhymes profile image
rhymes

Thanks Florimond, this is very interesting!

I'm not sure if I'll need data classes in future projects but I'll definitely come back to this post if I do :)

I'm also pratically sold on the idea of type annotations in Python

ps. Raymond Hettinger is one of the coolest Python developers in the business :-D

Collapse
 
florimondmanca profile image
Florimond Manca

Sweet! I didn’t think data classes would be useful to me before I began using them... ;) Also glad to hear you’re adopting type annotations! I enjoy the extra safety and readability they provide, especially when working with other developers.

Collapse
 
rhymes profile image
rhymes

Yeah, they also help tooling/IDEs a lot.

Collapse
 
paddyroddy profile image
Patrick Roddy

do you how to combine this with abstract classes? If I do the following:

from dataclasses import dataclass, field
from abc import abstractmethod

@dataclass # type: ignore
class Vehicle:

    wheels: int
    _wheels: int = field(init=False, repr=False)

    @property
    def wheels(self) -> int:
        print('getting wheels')
        return self._wheels

    @wheels.setter
    def wheels(self, wheels: int):
        print('setting wheels to', wheels)
        self._wheels = wheels

    @abstractmethod
    def sound_horn(self):
        raise NotImplementedError

then I get the warning test.py:10: error: Name 'wheels' already defined on line 7

Collapse
 
sandordargo profile image
Sandor Dargo

So do you think it's worth using dataclasses with properties? To me, it seems less readable. On the other hand, you have repr...

Collapse
 
florimondmanca profile image
Florimond Manca

Hi! I do agree with you — it is probably less readable, and not very obvious. The reason being that there are many little details whose interplay allow this "trick" to even work.

So my answer would be: it depends on your priorities. Is it okay for you to sacrifice some readability over the advantages of dataclasses?

For example, it might not be okay because other people working on the project may not be familiar with dataclasses; they'd probably end up clueless about how it can even work.

Also, remember the Zen of Python: "there should be one — and preferably only one — way to do it." Consider alternatives; if they look more obvious, use them. Otherwise, go for @property. :-)