In this chapter, we’ll start to make good on our promise to apply the dependency inversion principle as a way of decoupling our core logic from infrastructural concerns.
We’ll introduce the Repository, a simplifying abstraction over data storage, allowing us to decouple our model layer from the data layer. We’ll see a concrete example of how this simplifying abstraction makes our system more testable by hiding the complexities of the database.
Batch, Repository and SqlAlchemy shows a little preview of what we’re going to
build: a Repository class that sits between SqlAlchemy (our ORM) and our
Domain Model’s Batch classes.
[plantuml, chapter_02_class_diagram]
@startuml
package allocation {
class BatchRepository {
add (batch)
get (reference)
list ()
}
class Batch {
allocate ()
}
}
package SqlAlchemy {
class Session {
query ()
add ()
}
}
BatchRepository *-- Session : abstracts >
BatchRepository -> Batch : fetches >
@enduml
In the previous chapter we built a simple domain model that can allocate orders to batches of stock. It’s easy for us to write tests against this code because there aren’t any dependencies or infrastructure to set up. If we needed to run a database or an API and create test data, our tests would be harder to write and maintain.
Sadly, at some point we’ll need to put our perfect little model in the hands of users and we’ll need to contend with the real world of spreadsheets and web browsers and race conditions. For the next few chapters we’re going to look at how we can connect our idealised domain model to external state.
We expect to be working in an agile manner, so our priority is to get to an MVP
as quickly as possible. In our case, that’s going to be a web API. In a real project, you might dive straight in with some end-to-end tests and start plugging in a web framework, test-driving things outside-in.
But we know that, no matter what, we’re going to need some form of persistent storage, and this is a textbook, so we can allow ourselves a tiny bit more bottom-up development, and start to think about storage and databases.
When we build our first API endpoint, we know we’re going to have some code that looks more or less like What our first API endpoint will look like [1].
@flask.route.gubbins
def allocate_endpoint():
# extract order line from request
line = OrderLine(request.params, ...)
# load all batches from the DB
batches = ...
# call our domain service
allocate(line, batches)
# then save the allocation back to the database somehow
return 201We’ll need a way to retrieve batch info from the DB and instantiate our domain model objects from it, and we’ll also need a way of saving them back to the database.
As mentioned in [introduction], the "layered architecture" is a common approach to structuring a system that has a UI, some logic, and a database (see Layered Architecture).
Django’s Model-View-Template structure is closely related, as is Model-View-Controller (MVC). In any case, the aim is to keep the layers separate (which is a good thing), and to have each layer depend only on the one below…
But we want our domain model to have no dependencies whatsoever[2]. We don’t want infrastructure concerns bleeding over into our domain model and slowing down our unit tests or our ability to make changes.
Instead, as discussed in the introduction, we’ll think of our model as being on the "inside," and dependencies flowing inwards to it; what people sometimes call "onion architecture" (see Onion Architecture.)
[ditaa, onion_architecture]
+------------------------+
| Presentation Layer |
+------------------------+
|
V
+---------------------------------------------------+
| Domain Model |
+---------------------------------------------------+
^
|
+-------------------+
| Database Layer |
+-------------------+
If you’ve been reading around about architectural patterns, you may be asking yourself questions like this:
"Is this Ports and Adapters? Or is it Hexagonal Architecture? Is it the same as the Onion architecture? What about the Clean architecture? What’s a Port and what’s an Adapter? Why do you people have so many words for the same thing?
Although some people like to nitpick over the differences, all these are pretty much names for the same thing, and they all boil down to the dependency inversion principle—high-level modules (the domain) should not depend on low-level ones (the infrastructure).[3]
We’ll get into some of the nitty-gritty around "depending on abstractions," and whether there is a Pythonic equivalent of interfaces, later in the book.
Let’s remind ourselves of our domain model (see Our Model):
An allocation is the concept of linking an OrderLine to a Batch. We’re
storing the allocations as a collection on our Batch object:
Let’s see how we might translate this to a relational database.
In 2019 it’s unlikely that your team are hand-rolling their own SQL queries.
Instead, you’re almost certainly using some kind of framework to generate SQL for you based on your model objects.
These frameworks are called Object-Relational Mappers because they exist to bridge the conceptual gap between the world of objects and domain modelling, and the world of databases and relational algebra.
The most important thing an ORM gives us is "persistence ignorance": the idea that our fancy domain model doesn’t need to know anything about how data are loaded or persisted. This helps to keep our domain clean of direct dependencies on particular databases technologies.[4]
But if you follow the typical SQLAlchemy tutorial, you’ll end up with something like this:
from sqlalchemy import Column, ForeignKey, Integer, String
from sqlalchemy.ext.declarative import declarative_base
from sqlalchemy.orm import relationship
Base = declarative_base()
class Order(Base):
id = Column(Integer, primary_key=True)
class OrderLine(Base):
id = Column(Integer, primary_key=True)
sku = Column(String(250))
qty = Integer(String(250))
order_id = Column(Integer, ForeignKey('order.id'))
order = relationship(Order)
class Allocation(Base):
...You don’t need to understand SQLAlchemy to see that our pristine model is now full of dependencies on the ORM, and is starting to look ugly as hell besides. Can we really say this model is ignorant of the database? How can it be separate from storage concerns when our model properties are directly coupled to database columns?
If you’re more used to Django, the SQLAlchemy snippet above translates to something like this:
class Order(models.Model):
pass
class OrderLine(models.Model):
sku = models.CharField(max_length=255)
qty = models.IntegerField()
order = models.ForeignKey(Order)
class Allocation(models.Model):
...The point is the same — our model classes inherit directly from ORM classes, so our model depends on the ORM. We want it to be the other way around.
Django doesn’t provide an equivalent for SQLAlchemy’s "classical mapper," but see [appendix_django] for some examples of how you apply dependency inversion and the Repository pattern to Django.
Well, thankfully, that’s not the only way to use SQLAlchemy. The alternative is to define your schema separately, and an explicit mapper for how to convert between the schema and our domain model:
from sqlalchemy.orm import mapper, relationship
import model #(1)
metadata = MetaData()
order_lines = Table( #(2)
'order_lines', metadata,
Column('id', Integer, primary_key=True, autoincrement=True),
Column('sku', String(255)),
Column('qty', Integer, nullable=False),
Column('orderid', String(255)),
)
...
def start_mappers():
lines_mapper = mapper(model.OrderLine, order_lines) #(3)-
The ORM imports (or "depends on" or "knows about") the domain model, and not the other way around.
-
We define our database tables and columns using SQLAlchemy’s abstractions.
-
And when we call the
mapperfunction, SQLAlchemy does its magic to bind our domain model classes to the various tables we’ve defined.
The end result will be that, if we call start_mappers(), we will be able to
easily load and save domain model instances from and to the database. But if
we never call that function, then our domain model classes stay blissfully
unaware of the database.
This gives us all the benefits of SQLAlchemy, including the ability to use
alembic for migrations, and the ability to transparently query using our
domain classes, as we’ll see.
When you’re first trying to build your ORM config, it can be useful to write some tests for it, as in Testing the ORM directly (throwaway tests) (test_orm.py):
def test_orderline_mapper_can_load_lines(session): #(1)
session.execute( #(1)
'INSERT INTO order_lines (orderid, sku, qty) VALUES '
'("order1", "RED-CHAIR", 12),'
'("order1", "RED-TABLE", 13),'
'("order2", "BLUE-LIPSTICK", 14)'
)
expected = [
model.OrderLine("order1", "RED-CHAIR", 12),
model.OrderLine("order1", "RED-TABLE", 13),
model.OrderLine("order2", "BLUE-LIPSTICK", 14),
]
assert session.query(model.OrderLine).all() == expected
def test_orderline_mapper_can_save_lines(session):
new_line = model.OrderLine("order1", "DECORATIVE-WIDGET", 12)
session.add(new_line)
session.commit()
rows = list(session.execute('SELECT orderid, sku, qty FROM "order_lines"'))
assert rows == [("order1", "DECORATIVE-WIDGET", 12)]-
If you’ve not used pytest, the
sessionargument to this test needs explaining. You don’t need to worry about the details of pytest or its fixtures for the purposes of this book, but the short explanation is that you can define common dependencies for your tests as "fixtures," and pytest will inject them to the tests that need them by looking at their function arguments. In this case, it’s a SQLAlchemy database session.
/// [SG] I set up the conftest to have a session, and could only get the tests to work if I dropped the (frozen=True) on the OrderLine dataclass, otherwise I would get dataclasses.FrozenInstanceError: cannot assign to field '_sa_instance_state' I feel I am having to work quite hard to follow along ;-(. Is not spelling everything out a deliberate tactic to make the reader learn ///
You probably wouldn’t keep these tests around—as we’ll see shortly, once you’ve taken the step of inverting the dependency of ORM and Domain Model, it’s only a small additional step to implement an additional abstraction called the repository pattern, which will be easier to write tests against, and will provide a simple, common interface for faking out later in tests.
But we’ve already achieved our objective of inverting the traditional dependency: the domain model stays "pure" and free from infrastructure concerns. We could throw away SQLAlchemy and use a different ORM, or a totally different persistence system, and the domain model doesn’t need to change at all.
Depending on what you’re doing in your domain model, and especially if you stray far from the OO paradigm, you may find it increasingly hard to get the ORM to produce the exact behavior you need, and you may need to modify your domain model[5]. As so often with architectural decisions, there is a trade-off you’ll need to consider. As the Zen of Python says, "Practicality beats purity!"
At this point though, our API endpoint might look something like Using SQLAlchemy directly in our API endpoint, and we could get it to work just fine.
@flask.route.gubbins
def allocate_endpoint():
session = start_session()
# extract order line from request
line = OrderLine(
request.params['order_id'],
request.params['sku'],
request.params['qty'],
)
# load all batches from the DB
batches = session.query(Batch).all()
# call our domain service
allocate(line, batches)
# save the allocation back to the database
session.commit()
return 201The Repository pattern is an abstraction over persistent storage. It hides the boring details of data access by pretending that all of our data is in memory.
If we had infinite memory in our laptops, we’d have no need for clumsy databases. Instead, we could just use our objects whenever we liked. What would that look like?
import all_my_data
def create_a_batch():
batch = Batch(...)
all_my_data.batches.add(batch)
def modify_a_batch(batch_id, new_quantity):
batch = all_my_data.batches.get(batch_id)
batch.change_initial_quantity(new_quantity)Even though our objects are in memory, we need to put them somewhere so we can find them again. Our in memory data would let us add new objects, just like a list or a set, and since the objects are in memory we never need to call a "Save" method, we just fetch the object we care about, and modify it in memory.
The simplest repository has just two methods: add to put a new item in the
repository, and get to return a previously added item.[6].
We stick rigidly to using these methods for data access in our domain and our
service layer. This self-imposed simplicity stops us from coupling our domain
model to the database.
Here’s what an abstract base class for our repository would look like:
class AbstractRepository(abc.ABC):
@abc.abstractmethod #(1)
def add(self, batch: model.Batch):
raise NotImplementedError #(2)
@abc.abstractmethod
def get(self, reference) -> model.Batch:
raise NotImplementedError|
Warning
|
We’re using abstract base classes in this book for didactic reasons:
we hope they help explain what the interface of the repository abstraction
is. In real life, we’ve often found ourselves deleting ABCs from our
production code, because Python makes it too easy to ignore them, and
they end up unmaintained and, at worst, misleading.
In practice we tend to rely on Python’s duck-typing to enable abstractions.
To a Pythonista, a repository is any object that has add(thing) and
get(id) methods.
|
-
Python tip:
@abc.abstractmethodis one of the only things that makes ABCs actually "work" in Python. Python will refuse to let you instantiate a class that does not implement all theabstractmethodsdefined in its parent class -
raise NotImplementedErroris nice but neither necessary nor sufficient. In fact, your abstract methods can have real behavior which subclasses can call out to, if you want.
|
Note
|
To really reap the benefits of ABCs (such as they may be) you’ll want to
be running some helpers like pylint and mypy.
|
You know they say economists know the price of everything and the value of nothing? Well, Programmers know the benefits of everything and the tradeoffs of nothing.
Whenever we introduce an architectural pattern in this book, we’ll always be trying to ask: "what do we get for this? And what does it cost us?."
Usually at the very least we’ll be introducing an extra layer of abstraction, and although we may hope it will be reducing complexity overall, it does add complexity locally, and it has a cost in terms raw numbers of moving parts and ongoing maintenance.
Repository pattern is probably one of the easiest choices in the book though,
if you’ve already heading down the DDD and dependency inversion route. As far
as our code is concerned, we’re really just swapping the SQLAlchemy abstraction
(session.query(Batch)) for a different one (batches_repo.get) which we
designed.
We will have to write a few lines of code in our repository class each time we add a new domain object that we want to retrieve, but in return we get a very simple abstraction over our storage layer, which we control. It would make it very easy to make fundamental changes to the way we store things (see [appendix_csvs]), and as we’ll see, it is very easy to fake out for unit tests.
In addition, repository pattern is so common in the DDD world that, if you do collaborate with programmers that have come to Python from the Java and C# worlds, they’re likely to recognise it. Repository pattern shows an illustration.
[ditaa, repository_pattern_diagram]
+-----------------------------+
| Presentation Layer |
+-----------------------------+
|^
|| +------------------+
||----------| Domain Model |
|| | objects |
|| +------------------+
V|
+------------------------------+
| Repository |
+------------------------------+
|
V
+------------------------------+
| Database Layer |
+------------------------------+
As always, we start with a test. This would probably be classified as an integration test, since we’re checking that our code (the repository) is correctly integrated with the database; hence, the tests tend to mix raw SQL with calls and assertions on our own code.
|
Tip
|
Unlike the ORM tests from earlier, these tests are good candidates for staying part of your codebase longer term, particularly if any parts of your domain model mean the object-relational map is nontrivial. |
def test_repository_can_save_a_batch(session):
batch = model.Batch("batch1", "RUSTY-SOAPDISH", 100, eta=None)
repo = repository.SqlAlchemyRepository(session)
repo.add(batch) #(1)
session.commit() #(2)
rows = list(session.execute(
'SELECT reference, sku, _purchased_quantity, eta FROM "batches"' #(3)
))
assert rows == [("batch1", "RUSTY-SOAPDISH", 100, None)]-
repo.add()is the method under test here -
We keep the
.commit()outside of the repository, and make it the responsibility of the caller. There are pros and cons for this, some of our reasons will become clearer when we get to [chapter_05_uow]. -
And we use the raw SQL to verify that the right data has been saved.
The next test involves retrieving batches and allocations so it’s more complex:
def insert_order_line(session):
session.execute( #(1)
'INSERT INTO order_lines (orderid, sku, qty) VALUES ("order1", "GENERIC-SOFA", 12)'
)
[[orderline_id]] = session.execute(
'SELECT id FROM order_lines WHERE orderid=:orderid AND sku=:sku',
dict(orderid="order1", sku="GENERIC-SOFA")
)
return orderline_id
def insert_batch(session, batch_id): #(2)
...
def test_repository_can_retrieve_a_batch_with_allocations(session):
orderline_id = insert_order_line(session)
batch1_id = insert_batch(session, "batch1")
insert_batch(session, "batch2")
insert_allocation(session, orderline_id, batch1_id) #(3)
repo = repository.SqlAlchemyRepository(session)
retrieved = repo.get("batch1")
expected = model.Batch("batch1", "GENERIC-SOFA", 100, eta=None) #(3)
assert retrieved == expected # Batch.__eq__ only compares reference
assert retrieved.sku == expected.sku
assert retrieved._purchased_quantity == expected._purchased_quantity
assert retrieved._allocations == {model.OrderLine("order1", "GENERIC-SOFA", 12)}-
This tests the read side, so the raw SQL is preparing data to be read by the
repo.get() -
We’ll spare you the details of
insert_batchandinsert_allocation, the point is to create a couple of different batches, and for the batch we’re interested in to have one existing order line allocated to it. -
And that’s what we verify here.
Whether or not you painstakingly write tests for every model is a judgement
call. Once you have one class tested for create/modify/save, you might be
happy to go on and do the others with a minimal roundtrip test, or even nothing
at all, if they all follow a similar pattern. In our case, the ORM config
that sets up the ._allocations set is a little complex, so it merited a
specific test.
You end up with something like A typical repository (repository.py):
class SqlAlchemyRepository(AbstractRepository):
def __init__(self, session):
self.session = session
def add(self, batch):
self.session.add(batch)
def get(self, reference):
return self.session.query(model.Batch).filter_by(reference=reference).one()
def list(self):
return self.session.query(model.Batch).all()And now our flask endpoint might look something like Using our repository directly in our API endpoint:
@flask.route.gubbins
def allocate_endpoint():
batches = SqlAlchemyRepository.list()
lines = [
OrderLine(l['orderid'], l['sku'], l['qty'])
for l in request.params...
]
allocate(lines, batches)
session.commit()
return 201We bumped into a friend at a DDD conference the other day who said "I haven’t used an ORM in 10 years." Repository pattern and an ORM both act as abstractions in front of raw SQL, so using one behind the other isn’t really necessary. Why not have a go at implementing our repository without using the ORM?
We’ve left the repository tests, but figuring out what SQL to write is up to you. Perhaps it’ll be harder than you think, perhaps it’ll be easier, but the nice thing is—the rest of your application just doesn’t care.
Here’s one of the biggest benefits of Repository Pattern.
class FakeRepository(AbstractRepository):
def __init__(self, batches):
self._batches = set(batches)
def add(self, batch):
self._batches.add(batch)
def get(self, reference):
return next(b for b in self._batches if b.reference == reference)
def list(self):
return list(self._batches)Because it’s a simple wrapper around a set, all the methods are one-liners.
Using a fake repo in tests is really easy, and we have a simple abstraction that’s easy to use and reason about:
fake_repo = FakeRepository([batch1, batch2, batch3])You’ll see this fake in action in the next chapter.
|
Tip
|
Building fakes for your abstractions is an excellent way to get design feedback: if it’s hard to fake, then the abstraction is probably too complicated. |
- Apply Dependency Inversion to your ORM
-
Our domain model should be free of infrastructure concerns, so your ORM should import your model, and not the other way around.
- Repository pattern is a simple abstraction around permanent storage
-
The repository gives you the illusion of a collection of in-memory objects. It makes it very easy to create a
FakeRepositoryfor testing, and it makes it easy to swap fundamental details of your infrastructure without disrupting your core application. See [appendix_csvs] for an example.
Bearing the Rich Hickey quote in mind, in each chapter we’re going to try and summarise the costs and benefits of each architectural pattern we introduce. We want to be very clear that we’re not saying every single application needs to be built this way; only sometimes does the complexity of the app and domain make it worth investing the time and effort in adding these extra layers of indirection. With that in mind, Repository Pattern and Peristence Ignorance: The Trade-Offs shows some of the pros and cons of Repository Pattern and our persistence ignorant model.
| Pros | Cons |
|---|---|
|
|
In a single diagram[7] (Domain Model Tradeoffs as a Diagram), the central conceit of is: yes, for simple cases, a decoupled domain model is harder work than a simple ORM / ActiveRecord pattern. But the more complex the domain, the more an investment in freeing yourself from infrastructure concerns will pay off, in terms of the ease of making changes:
[ditaa, domain_model_tradeoffs_diagram]
Cost of changes
^ /
| ActiveRecord / ORM |
| | ----/
| / ----/
| | ----/
| / ----/
| | ----/ Domain Model w/ Repository Pattern
| / ----/
| | ----/
| ----/
| ----/ /
| ----/ /
| ----/ -/
|----/ --/
| ---/
| ----/
|------/
|
+--------------------------------------------------------------->
Complexity of Business Domain/Logic
Our example code isn’t really complex enough to give more than a hint of what
the right-hand side of the graph looks like, but the hints are there.
Imagine, for example, if we decide one day that we want to change allocations
to live on the OrderLine instead of on the Batch object: if we were using
Django, say, we’d have to define and think through the database migration
before we could run any tests. As it is, because our model is just plain
Python objects, we can change a set() to being a new attribute, without
needing to think about the database until later.
You’ll be wondering, how do we actually instantiate these repositories, fake or real? What will our flask app actually look like? We’ll find out in the next exciting instalment, the Service Layer pattern.
But first, a brief digression.
list or delete or update, but in the ideal world, we only modify our model objects one at a time, and delete is usually handled as a soft-delete, ie batch.cancel(). Finally, update is taken care of by the Unit of Work, as we’ll see in [chapter_05_uow].