May 20th, 2019
It's the year 2019. You're working on a new Django project for a client, a team-based Todo app. The app still has a lot of velocity and you're churning out features at a great rate. You've fully embraced the best practice of "fat-models" as encouraged by the Active Record pattern. But there is a (not so) small issue creeping in: complexity.
Maybe you have a
Todo model that needs to be marked as done. You add a method,
Todo.complete() and that works well enough. But then what if you need to send users an email when a
Todo is made complete? Now your
Todo model knows how to send emails, which seems counter-intuitive. Hey, maybe team leaders need to be notified, too...
This cycle continues and the painful conclusion is a monolithic, all powerful model if you're lucky, or worse: an untestable ball of methods that do way too much. The model has consumed your application, which is now nothing except merge conflicts on the same few files.
Photo by @yespanioly on Unsplash.
A great article from 2017 by Mitchel Cabuloy discussed Rails-like service objects in Django based on the core forms framework. This is where I personally got started with using the service pattern (and domain driven design in general).
You can check out this really great module, but I have found in some applications it may not be sufficient, especially if you are using Django Rest Framework, Graphene, or (ironically) Django forms.
The problem stems from Single Responsibility Principle (SRP).
service_objects as implemented in the model are both how to present and how to perform an action. You should be able to change your controller layer without changing or breaking the underlying service.
If your application is simple enough,
service_objects is a great way to break business logic out of your models. But if you need more flexibility or compatibility at a deeper level with other modules (like
rest_framework), you might want to roll your own.
Breaking It Out
There's three components to any Service we need to create:
- The actual data that goes into the Service that will be operated on,
- A handler that actually executes the action, containing all side effects, and,
- Some way to represent that handler to a user of the app, with a form, serializer, CLI, or what have you.
Let's talk implementation. Personally, I prefer representing these components in distinct classes: one command, one handler, (sometimes) many ways to create commands for the handler.
Command can be pretty easily represented with a
@dataclass in 3.7...
from dataclasses import dataclass @dataclass class CreateUserCommand: """Command class for creating a new user.""" email: str username: str password: str
...or with a
from typing import NamedTuple class CreateUserCommand(NamedTuple): """Command class for creating a new user.""" email: str username: str password: str
Service (or handler) can be represented as a class for easy extension:
class Service: """A service takes a command and performs an action with it.""" def __init__(self, cmd, manager=transaction.atomic, signal=None): self.cmd = cmd self.manager = manager # Django has the neat signals framework for message busses. self.signal = signal def execute(self): """Runs self.process() inside the manager object.""" # The default manager is django.db.transaction.atomic with self.manager(): return self.process() def process(self): """Run inside the manager object. Performs the action.""" raise NotImplemenetedError()
Our service named
CreateUser might look something like this:
class CreateUser(Service): def process(self): """Creates a user and sends self.signal.""" user = User.objects.create_user( username=self.cmd.username, email=self.cmd.email, password=self.cmd.password, ) # Using a message bus helps us avoid breaking SRP. self.signal.send_robust( sender=self.__class__, email=self.cmd.email, )
A note on using
signals: if you want your signal's errors to cause the transaction to fail, use
.send(). If you do not want the transaction to fail, use
.send_robust(). Protect critical changes from trivial failures, like email delivery, by using
Stop Writing Classes
You can use this pattern with a functional service handler. Here's a more functional implementation of
def create_user(cmd, manager=transaction.atomic, signal=user_created): """Create a user.""" with manager(): # As CreateUser above
This mostly comes down to personal preference. Either way, it's easier to test components in isolation when you can inject a different
signal, so I do suggest keeping them as keyword arguments with sensible defaults.
Strong unit test coverage is critical to maintaining a large codebase. A typical downfall in large Django apps is that it can be hard to write fast tests because of the reliance on models/the database.
However, with a Service pattern, it's simple to write fast unit tests that do not use the database.
# This example uses pytest. # Mocking our user model. @mock.patch('mybaseapp.models.User.create_user') def test_create_user_calls_creation_method(user_mock): manager_mock = mock.MagicMock() signal_mock = mock.MagicMock() cmd = CreateUserCommand( email="firstname.lastname@example.org", username="houseplantmom777", password="password secured" ) CreateUser( cmd, manager=manager_mock, signal=signal_mock ).execute() user_mock.assert_called_once_with( email="email@example.com", username="houseplantmom777", password="password secured", ) signal_mock.send_robust.assert_called_once_with( sender=CreateUser, email="firstname.lastname@example.org", )
Adding Another C to MVC
Commands have no way to gather data for themselves.
Services have no way to make or validate their commands. Instead, your
Controller will handle creating
Commands to send to your
In vanilla Django, a great representation of a
Controller for a service is a
from django import forms class ServiceForm(forms.Form): @property def command(self): """Getter for commands.""" if self.Meta.command is None: raise AttributeError("Set Meta.command to be a command.") return self.Meta.command def to_command(self, **kwargs): """Creates a command from the form data. Arguments: **kwargs: Injected into self.Meta.command during creation. """ self.cleaned_data.update(kwargs) return self.command(**self.cleaned_data) class Meta: command = None
ServiceForm is extended to accept specific fields, and its
.to_command() method is called to create the new command from the cleaned data and any extra arguments (
request.user is common).
Using it is easy:
class CreateUserForm(ServiceForm): email = forms.EmailField() username = forms.CharField() password = forms.CharField(widget=forms.PasswordInput) # Add your clean methods, etc... class Meta: command = CreateUserCommand
forms.Form as the controller is entirely arbitrary. Actions don't care about where they are from, that's the view and controller's jobs. If you are using Django Rest Framework, a
Serializer class works well. Similarly, if you're using the fabulous Graphene Django, a
Mutation class is the logical way to control
Actions don't care about where they are from. That's the View and Controller's job.
Typically, your choice of view usually dictates what controllers will create commands for services. In vanilla Django, I personally prefer sublassing the generic
FormView to create a
ServiceView. With Django Rest Framework, try creating a version of
APIView that calls services.
Creating a generic view for building commands from controllers to call out to services is a fantastic way to save time. This is a great way to add velocity back into your projects, without making a mess of the code.
For just calling
User.objects.create_user, this is certainly over-engineered. But don't discount the pattern because of a contrived example: Services are a great way to create code bases that are open to expansion and crystal clear. In reality, your business logic will be much more complicated than just creating a
User or two, and that's where services shine.
And best of all, your
models.py file isn't 1k lines anymore.
I hope you feel inspired to try out the service pattern in Django or (Python in general)! If you have tried it out before, I'd love to hear your thoughts and how it did or didn't work out for you. Thanks for reading!