Carlos Chacin

Software Engineering Experiences

💉 Dependency Injection in Java ☕️

Posted at November 14, 2019 / Carlos Chacin

java dependency-injection di best-practices

14 Nov 2019 Carlos Chacin

Collaborators

Lee la versión en Español aquí

Java is an object-oriented language with some functional aspects included in its core. Like any other object-oriented language, classes and objects are the foundations of any functionality that we can write and use. The relationships between the classes/objects make it possible to extend and reuse functionality. However, the way that we choose to build those relationships determine how modular, decoupled, and reusable our codebase is, not only in terms of our production code but also in our test suites.

In this article, we are going to describe the concept of Dependency Injection in Java and how it helps us have a more modular and decoupled codebase, which makes our lives easier, even for testing, without the need of any sophisticated container or framework.

What is Dependency?

When a class ClassA uses any method of another class ClassB, we can say that ClassB is a dependency of ClassA.

class ClassA {

  ClassB classB = new ClassB();

  int tenPercent() {
    return classB.calculate() * 0.1d;
  }
}

In this example, ClassA is calculating 10% of the value, and calculating that value, it’s reusing the functionality exposed by ClassB.

And it can be used like this:

class Main {
  public static void main(String... args) {
    ClassA classA = new ClassA();

    System.out.println("Ten Percent: " + classA.tenPercent());
  }
}

Now, there is a big problem with this approach:

ClassA is tightly coupled with ClassB

If we needed to change/replace ClassB with ClassC because ClassC has an optimized version of the calculate() method, we need to recompile ClassA because we don’t have a way to change that dependency, it’s hardcoded inside of ClassA.

Collision

The Dependency Injection Principle

The Dependency Injection Principle is nothing but being able to pass (inject) the dependencies when required instead of initializing the dependencies inside of the recipient class.

Decouple the construction of your classes from the construction of your classes’ dependencies

Forms of Dependency Injection in Java

Setter Injection (Not recommended)

class ClassA {

  ClassB classB;

  /* Setter Injection */
  void setClassB(ClassB injected) {
    classB = injected;
  }

  int tenPercent() {
    return classB.calculate() * 0.1d;
  }
}

With this approach, we remove the new keyword from our ClassA. Thus, we move the responsibility for the creation of ClassB away from ClassA.

ClassA still has a hard dependency on ClassB but now it can be injected from the outside:

class Main {
  public static void main(String... args) {
    ClassA classA = new ClassA();
    ClassB classB = new ClassB();

    classA.setClassB(classB);

    System.out.println("Ten Percent: " + classA.tenPercent());
  }
}

The above example is better than the initial approach because now we can inject in ClassA an instance of ClassB or even better, a subclass of ClassB:

class ImprovedClassB extends ClassB {
  // content omitted
}

class Main {
  public static void main(String... args) {
    ClassA classA = new ClassA();
    ImprovedClassB improvedClassB = new ImprovedClassB();

    classA.setClassB(improvedClassB);

    System.out.println("Ten Percent: " + classA.tenPercent());
  }
}

But there is a significant problem with the Setter Injection approach:

We are hiding the ClassB dependency in ClassA because by reading the constructor signature, we cannot identify its dependencies right away. The code below causes a NullPointerException on runtime:

class Main {
  public static void main(String... args) {
    ClassA classA = new ClassA();

    System.out.println("Ten Percent: " + classA.tenPercent()); // NullPointerException here
  }
}

npe

In a statically typed language like Java, it’s always a good thing to let the compiler help us. See Constructor Injection

Constructor Injection (Highly recommended)

class ClassA {

  ClassB classB;

  /* Constructor Injection */
  ClassA(ClassB injected) {
    classB = injected;
  }

  int tenPercent() {
    return classB.calculate() * 0.1d;
  }
}

ClassA still has a hard dependency on ClassB but now it can be injected from the outside using the constructor:

class Main {
  public static void main(String... args) {
    /* Notice that we are creating ClassB fisrt */
    ClassB classB = new ImprovedClassB();

    /* Constructor Injection */
    ClassA classA = new ClassA(classB);

    System.out.println("Ten Percent: " + classA.tenPercent());
  }
}

ADVANTAGES:

The functionality remains intact compared with the Setter Injection approach
We removed the new initialization from the ClassA.
We still can inject a specialized subclass of ClassB to ClassA.
Now the compiler is going to ask us for the dependencies that we need in compile time.

Happy

Field Injection (Kids don’t try this at home)

There is a 3rd way to inject dependencies in Java, and it is called Field Injection. The only way for field injection to work is:

Mutating the field because it’s a non-private and non-final field
Mutating a final/private field using reflection

This approach has the same problems exposed by the Setter Injection approach and additionally adds complexity due to the mutation/reflection required. Unfortunately, this is a pretty common pattern when people use a Dependency Injection Framework.

NOTE:

When a class ClassA uses any method of another class ClassB we can say that ClassB is a dependency of ClassA.

If ClassA has a dependency on ClassB, ClassA constructor should require ClassB.

Feedback

Realistic Example

Every single Hello World example for any idea, concept, or pattern is super simple to understand, and it just works fine. But when we need to implement it in a real project, things get more complicated, and often, as engineers, we tend to try to solve the problem by introducing new layers to the problem instead of understanding what the real problem is.

Now that we know the advantages of the Dependency Injection Principle using the Constructor Injection approach, let’s create a more realistic example to see some inconveniences and how can we solve it without introducing a new layer to the mix.

The Todo’s Application

Todos

Let’s design a Todo’s Application to perform CRUD operations (Create, Read, Update, Delete) to manage our todo list, and an original architecture can be like this:

TodoApp is the main class that is going to initialize our application; this can be an android app, web page, or a desktop application using any framework.
TodoView is the class that would display a view to interact with, this class is going to delegate the data-related aspects to the TodoHttpClient. It’s only responsibility is to paint/draw/render the information and get the input to perform actions against the data using the TodoHttpClient dependency.
TodoHttpClient is the class that contains a set of HTTP methods to persists Todo objects using a REST API.
Todo is a value object that represents a todo item in our data store.

Let’s write the Java classes for our design using the Constructor Injection approach that we just learned:

class Todo {
  /* Value Object class */
  // content omitted
}

class TodoApp {
  private final TodoView todoView;

  TodoApp(final TodoView todoView) {
    this.todoView = todoView;
  }
  // content omitted
}

class TodoView {
  private final TodoHttpClient todoHttpClient;

  TodoView(final TodoHttpClient todoHttpClient) {
    this.todoHttpClient = todoHttpClient;
  }
  // content omitted
}

class Main {
  public static void main(String... args) {
    new TodoApp(new TodoView(new TodoHttpClient("https://api.todos.io/")));
  }
}

Now let’s focus our attention on the relationship between the TodoView and TodoHttpClient classes and add more details to them:

Magic

class TodoHttpClient extends MyMagicalHttpAbstraction {

  TodoView(final String baseUrl) {
    super(baseUrl);
  }

  @GET
  List<Todo> getAll() {
    return super.get(Todo.class);
  }

  @GET
  Todo get(long id) {
    return super.get(Todo.class, id);
  }

  @POST
  long save(Todo todo) {
    return super.post(todo);
  }

  @PUT
  Todo update(Todo todo) {
    return super.put(todo, todo.getId());
  }

  @DELETE
  void delete(long id) {
    super.delete(Todo.class, id);
  }
}

class TodoView extends MyFrameworkView {
  private final TodoHttpClient httpClient;

  // View initialized by the view library/framework
  // or injected as a dependency as well
  private ListView listView;
  private DetailView detailView;

  TodoView(final TodoHttpClient httpClient) {
    this.httpClient = httpClient;
  }

  void showTodos() {
    listView.add(httpClient.getAll());
  }

  void showTodo(Todo selected) {
    detailView.print(httpClient.get(selected.getId()));
  }

  void save(Todo todo) {
    httpClient.save(todo);
    listView.add(todo)
  }

  void update(Todo todo) {
    httpClient.update(todo);
    detailView.refresh(todo);
  }

  void delete(long id) {
    httpClient.delete(id);
    listView.refresh();
  }
}

Testing our design

Scientist

Let’s create a unit test for the TodoView class where we test the class in isolation without instantiating any of its dependencies. In this case, the dependency is TodoHttpClient:

@ExtendWith(MockitoExtension.class)
class TodoViewTest {

  @Test
  void shouldBeEmptyWhenEmptyList(@Mock TodoHttpClient httpClient) {
    // Given
    Mockito.when(httpClient.getAll()).thenReturn(List.of());

    // When
    TodoView todoView = new TodoView(httpClient);
    todoView.showTodos();

    // Then
    Assertions.assertThat(todoView.getListView()).isEmpty();
  }
}

Now that we have our test case passing, let’s analyze how our design impacts the testing approach:

We introduced the Mockito framework to be able to create a fake instance of TodoHttpClient, and that adds much complexity.
We have to prepare our instance of TodoHttpClient to fake the return of an empty list when calling the getAll() method, now our unit test also contains implementation details about the TodoHttpClient.
Additionally, since TodoHttpClient is a concrete class, we cannot change the implementation to call a DB instead without having to change the TodoView class as well, and we would need to rewrite the unit tests even when they should isolate this implementation detail.

Let’s improve our design

VideoLearning

One thing that we can do to decouple our classes is to introduce an interface since the Java language is always a good thing to rely on abstractions instead of relying on actual implementations.

Let’s put an interface between TodoView and TodoHttpClient:

TodoProvider

interface TodoProvider {
  List<Todo> getAll();
  Todo get(long id);
  long save(Todo todo);
  Todo update(Todo todo);
  void delete(long id);
}

Let’s make the TodoHttpClient to implement that interface:

class TodoHttpClient extends MyMagicalHttpAbstraction implements TodoProvider {

  TodoView(final String baseUrl) {
    super(baseUrl);
  }

  @GET
  List<Todo> getAll() {
    return super.get(Todo.class);
  }

  @GET
  Todo get(long id) {
    return super.get(Todo.class, id);
  }

  @POST
  long save(Todo todo) {
    return super.post(todo);
  }

  @PUT
  Todo update(Todo todo) {
    return super.put(todo, todo.getId());
  }

  @DELETE
  void delete(long id) {
    super.delete(Todo.class, id);
  }
}

Now the TodoView class looks like this:

class TodoView extends MyFrameworkView {
  private final TodoProvider provider;

  // View initialized by the view library/framework
  // or injected as a dependency as well
  private ListView listView;
  private DetailView detailView;

  TodoView(final TodoProvider httpClient) {
    this.provider = provider;
  }

  void showTodos() {
    listView.add(provider.getAll());
  }

  void showTodo(Todo selected) {
    detailView.print(provider.get(selected.getId()));
  }

  void save(Todo todo) {
    provider.save(todo);
    listView.add(todo)
  }

  void update(Todo todo) {
    provider.update(todo);
    detailView.refresh(todo);
  }

  void delete(long id) {
    provider.delete(id);
    listView.refresh();
  }
}

What do we gain with these changes?

We are able to change the TodoHttpClient with something like TodoDBProvider in the TodoApp and the application behavior would remain the same:

new TodoApp(new TodoView(new TodoDbProvider("dbName", "dbUser", "dbPassword")));

Let’s see how that helps in unit tests

Library

@ExtendWith(MockitoExtension.class)
class TodoViewTest {

  @Test
  void shouldBeEmptyWhenEmptyList(@Mock TodoProvider provider) {
    // Given
    Mockito.when(provider.getAll()).thenReturn(List.of());

    // When
    TodoView todoView = new TodoView(httpClient);
    todoView.showTodos();

    // Then
    Assertions.assertThat(todoView.getListView()).isEmpty();
  }
}

The test is still green which is great, but wait… nothing changed actually :(

The only changes were related to naming:

TodoHttpClient -> TodoProvider no value for the test.
httpClient -> provider no value for the test here.
We are still relying on the mocking framework.
We are still coupling the test to the interface’s name: TodoProvider.
We are still coupling the test to the method name: getAll()

Can we remove the mocking framework?

If we have now an interface, why are we coupled to the mocking framework to create a fake object that we can manually create using an anonymous class? Let’s change that:

@ExtendWith(MockitoExtension.class)
class TodoViewTest {

  // Given
  TodoProvider provider = new TodoProvider() {
    @Override
    public List<TodoItem> getAll() {
      return List.of();
    }

    @Override
    public TodoItem get(long id) {
      return null;
    }

    @Override
    public long save(TodoItem todo) {
      return 0;
    }

    @Override
    public TodoItem update(TodoItem todo) {
      return null;
    }

    @Override
    public void delete(long id) {

    }
  };

  // When
  var todoView = new TodoView(provider);
  todoView.displayListView();

  // Then
  assertThat(todoView.getTodoItemList()).isEmpty();
}

Sweet, now our design is more flexible since we can inject a different TodoProvider implementation, and we can do the same in our tests without using a mocking framework. But, we are paying the price: Verbosity, the mocking framework removes the need for implementing every single method from the interfaces.

Only the beginning

In the next article, let’s remove the verbosity from the tests and write an even better design.

Stay tuned for more posts like this.