article banner (priority)

Effective Kotlin Item 36: Prefer composition over inheritance

This is a chapter from the book Effective Kotlin. You can find it on LeanPub or Amazon.

Inheritance is a powerful feature, but it is designed to create a hierarchy of objects with an "is a" relationship. When such a relationship is not clear, inheritance might be problematic and dangerous. When all we need is a simple code extraction or reuse, inheritance should be used with caution; instead, we should prefer a lighter alternative: class composition.

Simple behavior reuse

Let’s start with a simple problem: we have two classes with partially similar behavior. They should both display a progress bar before some action and hide it afterwards.

class ProfileLoader { fun load() { // show progress bar // load profile // hide progress bar } } class ImageLoader { fun load() { // show progress bar // load image // hide progress bar } }

In my experience, many developers would extract this common behavior by extracting a common superclass:

abstract class LoaderWithProgressBar { fun load() { // show progress bar action() // hide progress bar } abstract fun action() } class ProfileLoader: LoaderWithProgressBar() { override fun action() { // load profile } } class ImageLoader: LoaderWithProgressBar() { override fun action() { // load image } }

This approach works for such a simple case, but it has important downsides we should be aware of:

  • We can only extend one class. Extracting functionalities using inheritance often leads to either excessively complex hierarchies of types or to huge BaseXXX classes that accumulate many functionalities.

  • When we extend, we take everything from a class, which leads to classes that have functionalities and methods they don’t need (a violation of the Interface Segregation Principle).

  • Using superclass functionality is much less explicit. In general, it is a bad sign when a developer reads a method and needs to jump into superclasses many times to understand how this method works.

These are strong reasons that should make us think about an alternative, and a very good one is composition. By composition, we mean holding an object as a property (we compose it) and reusing its functionalities. This is an example of how we can use composition instead of inheritance to solve our problem:

class ProgressBar { fun show() { /* show progress bar */ } fun hide() { /* hide progress bar */ } } class ProfileLoader { val progressBar = ProgressBar() fun load() { progressBar.show() // load profile progressBar.hide() } } class ImageLoader { val progressBar = ProgressBar() fun load() { progressBar.show() // load image progressBar.hide() } }

Notice that composition is harder. We need to include the composed object and use it in every single class. This is the key reason why many prefer inheritance. However, this additional code is not useless: it informs the reader that a progress bar is used and how it is used. It also gives the developer more power over how this progress bar works.

Another thing to note is that composition is better when we want to extract multiple pieces of functionality. For instance, information that loading has finished:

class ImageLoader { private val progressBar = ProgressBar() private val finishedAlert = FinishedAlert() fun load() { progressBar.show() // load image progressBar.hide() finishedAlert.show() } }

We cannot extend more than a single class. Therefore, if we wanted to use inheritance instead, we would be forced to place both functionalities in a single superclass. This often leads to a complex hierarchy of the types that are used to add these functionalities. Such hierarchies are very hard to read and often also to modify. Just think about what happens if we need an alert in two subclasses but not in a third one? One way to deal with this problem is to use a parameterized constructor:

abstract class InternetLoader(val showAlert: Boolean) { fun load() { // show progress bar innerLoad() // hide progress bar if (showAlert) { // show alert } } abstract fun innerLoad() } class ProfileLoader : InternetLoader(showAlert = true) { override fun innerLoad() { // load profile } } class ImageLoader : InternetLoader(showAlert = false) { override fun innerLoad() { // load image } }

This is a bad solution. Having functionality blocked by a flag (showAlert in this case) is a bad sign. This problem is compounded when the subclass cannot block other unneeded functionality. This is a trait of inheritance: it takes everything from the superclass, not only what is needed.

Taking the whole package

When we use inheritance, we take everything from the superclass: methods, expectations (contract), and behavior. Therefore, it is a great tool for representing a hierarchy of objects, but it’s not so great when we just want to reuse some common parts. For such cases, composition is better because we can choose the behavior we need. As an example, let’s say that in our system we have decided to represent a Dog that can bark and sniff:

abstract class Dog { open fun bark() { /*...*/ } open fun sniff() { /*...*/ } }

What if then we need to create a robot dog that can bark but can't sniff?

class Labrador: Dog() class RobotDog : Dog() { override fun sniff() { throw Error("Operation not supported") // Do you really want that? } }

Notice that such a solution violates the interface-segregation principle as RobotDog has a method it doesn’t need. It also violates the Liskov Substitution Principle as it breaks superclass behavior. On the other hand, what if your RobotDog also needs to be a Robot class because Robot can calculate (i.e., it has the calculate method)? Multiple inheritance is not supported in Kotlin.

abstract class Robot { open fun calculate() { /*...*/ } } class RobotDog : Dog(), Robot() // Error

These are serious design problems and limitations that do not occur when you use composition instead. When we use composition, we choose what we want to reuse. To represent type hierarchy, it is safer to use interfaces, and we can implement multiple interfaces. What has not yet been shown is that inheritance can lead to unexpected behavior.

Inheritance breaks encapsulation

To some degree, when we extend a class, we depend not only on how it works from the outside but also on how it is implemented inside. This is why we say that inheritance breaks encapsulation. Let’s look at an example inspired by the book Effective Java by Joshua Bloch. Let’s say that we need a set that will know how many elements have been added to it during its lifetime. Such a set can be created using inheritance from HashSet:

class CounterSet<T>: HashSet<T>() { var elementsAdded: Int = 0 private set override fun add(element: T): Boolean { elementsAdded++ return super.add(element) } override fun addAll(elements: Collection<T>): Boolean { elementsAdded += elements.size return super.addAll(elements) } }

This implementation might look good, but it doesn’t work correctly:

val counterList = CounterSet<String>() counterList.addAll(listOf("A", "B", "C")) print(counterList.elementsAdded) // 6

Why is that? The reason is that HashSet uses the add method under the hood of addAll. The counter is then incremented twice for each element added using addAll. This problem can be naively solved by removing the custom addAll function:

class CounterSet<T>: HashSet<T>() { var elementsAdded: Int = 0 private set override fun add(element: T): Boolean { elementsAdded++ return super.add(element) } }

However, this solution is dangerous. What if the creators of Java decided to optimize HashSet.addAll and implement it in a way that doesn't depend on the add method? If they did that, this implementation would break with a Java update. Together with this implementation, any other libraries which depend on our current implementation would break as well. The Java creators know this, so they are cautious of making changes to these types of implementations. The same problem affects any library creator or even developers of large projects. So, how can we solve this problem? We should use composition instead of inheritance:

class CounterSet<T> { private val innerSet = HashSet<T>() var elementsAdded: Int = 0 private set fun add(element: T) { elementsAdded++ innerSet.add(element) } fun addAll(elements: Collection<T>) { elementsAdded += elements.size innerSet.addAll(elements) } } val counterList = CounterSet<String>() counterList.addAll(listOf("A", "B", "C")) print(counterList.elementsAdded) // 3

One problem is that in this case we lose polymorphic behavior because CounterSet is not a Set anymore. To keep this behavior, we can use the delegation pattern. The delegation pattern means our class implements an interface, composes an object that implements the same interface, and forwards methods defined in the interface to this composed object. Such methods are called forwarding methods. Take a look at the following example:

class CounterSet<T> : MutableSet<T> { private val innerSet = HashSet<T>() var elementsAdded: Int = 0 private set override fun add(element: T): Boolean { elementsAdded++ return innerSet.add(element) } override fun addAll(elements: Collection<T>): Boolean { elementsAdded += elements.size return innerSet.addAll(elements) } override val size: Int get() = innerSet.size override fun contains(element: T): Boolean = innerSet.contains(element) override fun containsAll(elements: Collection<T>): Boolean = innerSet.containsAll(elements) override fun isEmpty(): Boolean = innerSet.isEmpty() override fun iterator() = innerSet.iterator() override fun clear() = innerSet.clear() override fun remove(element: T): Boolean = innerSet.remove(element) override fun removeAll(elements: Collection<T>): Boolean = innerSet.removeAll(elements) override fun retainAll(elements: Collection<T>): Boolean = innerSet.retainAll(elements) }

The problem now is that we need to implement a lot of forwarding methods (nine, in this case). Thankfully, Kotlin introduced interface delegation support that is designed to help in this kind of scenario. When we delegate an interface to an object, Kotlin will generate all the required forwarding methods during compilation. Here is Kotlin’s interface delegation in action:

class CounterSet<T>( private val innerSet: MutableSet<T> = mutableSetOf() ) : MutableSet<T> by innerSet { var elementsAdded: Int = 0 private set override fun add(element: T): Boolean { elementsAdded++ return innerSet.add(element) } override fun addAll(elements: Collection<T>): Boolean { elementsAdded += elements.size return innerSet.addAll(elements) } }

This is a case in which delegation is a good choice: we need polymorphic behavior and inheritance would be dangerous. However, delegation is not common. In most cases, polymorphic behavior is not needed or we use it in a different way, so composition without delegation is more suitable.

The fact that inheritance breaks encapsulation is a security concern, but in many cases the behavior is specified in a contract or we don’t depend on it in subclasses (this is generally true when methods are designed for inheritance). There are other reasons to choose the composition pattern, one of which is that it is easier to reuse and gives us more flexibility.

Restricting overriding

To prevent developers from extending classes that are not designed for inheritance, we can just keep them final. However, if for some reason we need to allow inheritance, all methods are still final by default. To let developers override them, they must be set to open:

open class Parent { fun a() {} open fun b() {} } class Child: Parent() { override fun a() {} // Error override fun b() {} }

Use this mechanism wisely and open only those methods that are designed for inheritance. Also, remember that when you override a method, you can make it final for all subclasses:

open class ProfileLoader: InternetLoader() { final override fun load() { // load profile } }

In this way, you can limit the number of methods that can be overridden in subclasses.

Summary

There are a few important differences between composition and inheritance:

  • Composition is more secure - We depend not on how a class is implemented but only on its externally observable behavior.

  • Composition is more flexible - We can extend only a single class but we can compose many. When we inherit, we take everything; but when we compose, we can choose what we need. When we change the behavior of a superclass, we change the behavior of all subclasses. It is hard to change the behavior of only some subclasses. When a class we have composed changes, it will only change our behavior if it has changed its contract with the outside world.

  • Composition is more explicit - This is both an advantage and a disadvantage. When we use a method from a superclass, we can do so implicitly, like methods from the same class. This requires less work, but it can be confusing and is more dangerous as it is easy to confuse where a method comes from (is it from the same class, a superclass, the top level, or is it an extension?). When we call a method on a composed object, we know where it comes from.

  • Composition is more demanding - We need to use a composed object explicitly. When we add some functionalities to a superclass, we often do not need to modify the subclasses. When we use composition, we more often need to adjust usages.

  • Inheritance gives us strong polymorphic behavior - This is also a double-edged sword. On one hand, it is convenient that a dog can be treated like an animal. On the other hand, it is very constraining: it must be an animal. Every subclass of an animal should be consistent with animal behavior. The superclass defines the contract, and the subclasses should respect it.

It is a general OOP rule to prefer composition over inheritance, but Kotlin encourages composition even more by making all classes and methods final by default and by making interface delegation a first-class citizen. This makes this rule even more important in Kotlin projects.

So, when is composition more reasonable? The rule of thumb is that we should use inheritance when there is a definite "is a" relationship. In other words, every class that uses inheritance needs to be its superclass. All unit tests written for superclasses should also pass for their subclasses (Liskov substitution principle). Object-oriented frameworks for displaying views are good examples: Application in JavaFX, Activity in Android, UIViewController in iOS, and React.Component in React. The same is true when we define our own special kind of view element that always has the same set of functionalities and characteristics. Just remember to design these classes with inheritance in mind, and specify how inheritance should be used. Also, keep as final methods that are not designed for inheritance.