w0rpzone

Blog

Immutable Interfaces and Covariant Return Types in Java

When writing data structures, you want your API for using your data structures to be as concise and easy to understand as possible, while providing as much safety as you can possibly get, without there being a huge amount of mental overhead. Today, let's look at a simple example which shows how the Immutable Interface pattern and also covariant return types can help you write code which is more understandable, while also providing safety benefits when working with concurrent code, and perhaps even some additional performance overhead benefits.

The problem with having only mutable objects.

Let's look at a simple example of a system you are trying to write. In your program domain, you have some Sprackets, which hold some a number for a count or something, and you have a Widget which holds a Spracket. So, you perhaps write a Spracket object, and it might look something like this.

public final class BadSpracket {
    private int count = 0;

    public int getCount() {
        return count;
    }

    public void setCount(int count) {
        this.count = count;
    }
}

So what we have looks pretty good thus far. You've got an object, it has some data in it, you can retrieve the data, change it. Everything seems okay. So, let's write our Widget which holds these Sprackets.

public final class BadWidget {
    BadSpracket spracket;

    public BadSpracket getSpracket() {
        BadSpracket copy = new BadSpracket();

        copy.setCount(spracket.getCount());

        return copy;
    }

    public void setSpracket(BadSpracket spracket) {
        this.spracket = spracket;
    }
}

Suddenly, eveything starts to fall apart a little, because you want to be able to control how your Spracket can be modified. So when you return a Spracket you make a defensive copy of the object. (Which could have been defined in say, a copy constructor, but that's another story.) So now every time you want to look at the details of the Spracket in the Widget, you need to create a copy of it. This also raises a design question, what do you copy? Just the contents of the object itself? Do you copy transitively, in other words, do you make a 'deep copy'? What kind of effect on memory will doing that have?

To make matters worse, there is a bug in this code. We copy the Spracket out of the object, but we don't copy one into the object. So now you could pass a reference to an existing Spracket object which you can modify elsewhere, which modifies the internals of the Widget without you knowing about it.

Immutable Interfaces help

I believe that this is actually typical Java code. If you've written any Java at all, you've surely come across this pattern of writing objects. You might not even see any problem with writing things this way. However, I assure you, there is a better way, and hopefully after you see this better way, you'll gain a desire to write better code. The better way is with the Immutable Interface pattern. Doubtless, someone else will have defined more 'officially' what the meaning of this phrase is, but let's make up our own definition, right on the spot.

The Immutable Interface pattern is a design pattern whereby objects are specified with interfaces which are immutable, and the mutable variations of the objects implement the immutable interfaces.

So there we go, that seems like a simple enough definition of what it is, and it should be pretty much right. So, what does all of that mean? Essentially, instead of having the BadSpracket above, you first define the immutable part of that class instead as an interface, like this.

public interface Spracket {
    int getCount();
}

That should seem simple enough. The immutable Spracket cannot be modified, so you can't set the count. So, let's create a mutable implementation of this interface.

public final class MutableSpracket implements Spracket {
    private int count = 0;

    @Override
    public int getCount() {
        return count;
    }

    public void setCount(int count) {
        this.count = count;
    }
}

So now, for our Spracket, we've got an implementation MutableSpracket which looks quite like our original BadSpracket example, only it implements some interface and has an @Override annotation. So what is the point of all of this? The point is, you can assign a reference to a Spracket with a reference to a MutableSpracket. This means you can write a function which creates the object with its mutable implementation, then calls functions with the immutable interface. However, the most important thing is how it changes the way we can write our Widgets. First, let's create another immutable interface.

public interface Widget {
    Spracket getSpracket();
}

So again, we have a pretty small and simple interface. The details of Spracket objects we just don't care about, all we care about in the immutable interface is that we can get one. So what might a mutable Widget look like? Let's see.

public final class MutableWidget implements Widget {
    private Spracket spracket;

    @Override
    public Spracket getSpracket() {
        return spracket;
    }

    public void setSpracket(Spracket spracket) {
        this.spracket = spracket;
    }
}

So great, we can create some MutableSpracket somewhere, pass it into the MutableWidget as an immutable Spracket, and maybe pass the MutableWidget along somewhere as an immutable Widget. So we now have a transitive system whereby we can control access to the details of our objects, or rather, how we modify them, as part of the object's interface. We can return references without having to create copies of objects, which can be a performance benefit. The best part is, we can pass an immutable reference to the object to another function and we can expect that function not to modify the object. This makes our code more predictable and easier to understand, and may lead to benefits in concurrent programming.

Covariant return types give you even more power.

We can take this one step further with a feature available to both C++ and Java, among other langauges, which it turns out yields benefits for working with this design pattern. The whole reason for using this design pattern in Java is that Java doesn't have C++'s notion of const, so we cannot create methods which aren't permitted to modify the objects, and we can't create references to objects which only let us use those const methods. However, as we have just seen, we can emulate this behaviour through the use of clever design and interfaces. What this doesn't allow us to do is change from returning mutable references to objects to immutable references when we switch to having a mutable reference to the container object to an immutable reference to the container object. (In this case, the inner object would be Spracket and the container object would be Widget.) However, I suggest that there is a feature which ties in with inheritance which will allow us to accomplish even this. It is the ability to specify covariant return types.

Let's say you return a reference to some object from a function, and you have a subclass of that object's class. Very obviously, it is legal to return a reference to a subclass object from that function as well. So let's take things one step further and say that you have a subclass, or even an implementation of an interface, which only ever returns references to objects which are subclasses of the class. What if you change the return type of the method to the subclass in the method override? This works in both Java and C++. The reason being is that the return types are covariant. Every Bar which inherits Foo is also a Foo, so it should make perfect sense that every method override which returns a Foo could also be changed to return a Bar.

So, after some informal discussion, we should have a rough understanding of covariant return types. Let's try and apply it to our previous example.

public final class MutableWidget implements Widget {
    private MutableSpracket spracket;

    @Override
    public MutableSpracket getSpracket() {
        return spracket;
    }

    public void setSpracket(MutableSpracket spracket) {
        this.spracket = spracket;
    }
}

Now we have have exactly the same object, with one difference. When we pass a reference to a MutableWidget to some other function, and we want to mutate it, we can mutate the inner Spracket directly. Of course, we would only pass a mutable reference when we want to change the object, because why pass a mutable reference when we aren't changing anything? That's what an immutable reference is for. Note that we cannot have covariant argument types. In a base class interface, if setSpracket accepts a Spracket, that's the final say on what the method can accept. All subclasses have to accept Sprackets, too. However here, the setter isn't specified in any superclass or any interfaces. Its specified only in this class itself, so no overriding happens. Thus, the MutableSpracket reference can be used in both the getter and the setter.

Which this new implementation, we can now do something like this.

public abstract class WidgetFactory {
    private static void incrementSpracketCount(MutableWidget widget) {
        MutableSpracket spracket = widget.getSpracket();

        if (spracket != null) {
            spracket.setCount(spracket.getCount() + 1);
        }
    }

    public static Widget createWidget() {
        MutableWidget widget = new MutableWidget();

        widget.setSpracket(new MutableSpracket());

        incrementSpracketCount(widget);

        return widget;
    }
}

We can create some mutable object in a function somewhere, pass it to another side-effecting function which changes the object we give it, and then return it as an immutable reference. When we get our Widget from our createWidget function, and we call getSpracket, we'll get just a Spracket, which is an immutable reference. Thus, we can have this small world of mutable references where we can create a mutable object one piece at a time, and then give a immutable reference to the object to the outside world when we're done. Hopefully, this should make our code easier to reason about. The types we use in our programs will tell us something about where modifications to objects are taking place, interfaces accepting the immutable references will give us clues for when we really need to make copies, etc. By changing our class structure just a little bit, we will have hopefully created better code.

Proceed with caution, nothing is perfect

I do not believe that any single pattern in programming is perfect, and every system in software has some kind of flaw in it. Thus, it is important for us to note the flaws with the solutions discussed above, and understand how to avoid these flaws, or when not to apply these solutions.

Casting all your cares away, for casting will cause problems

The first and foremost problem with the Immutable Interface pattern is that you're only a cast away from creating problems for yourself. And worse still, it might not be you creating the casts, but your users. Say you get an immutable reference to your widget above with Widget. Consider what happens if you do this.

public void letsMakeAMess(Widget widget) {
    MutableWidget mutableWidget = (MutableWidget) widget;

    // Oh no! We changed it anyway!
    mutableWidget.getSpracket().setCount(347);
}

The Immutable Inteface pattern makes the above issue a possibility. Supposing MutableWidget is the only implementer of Widget, and widget and the Spracket above are not null, the above code example will compile and run perfectly, without errors or warnings, and your object will be changed. So, this is not very good. What are our options here?

Discourage casting, in fact, make casting into a code smell.

I believe that we should try to use casting as little as possible. I believe casting references is something we can design around, and when you do write a cast, that line of code which performs a cast should read, "I ran out of ideas." There are design patterns, such as the Vistor Pattern, which may reduce your need for casting. You might reduce the need for casting by using stronger type information. Think, "Do I really need an Object, or should I use some strong interface?" Regarding a need for designing code with stronger type information, the type theorist Philip Wadler had this to say about Object.

If you see the class Object mentioned in a Java program, it is usually a sign that the program would benefit from the use of generic types. You might say that the class is well named — when you see it you should "object" and use generic types instead.

If you like the cut of his jib from that comment, you might like this video, where he talks about generics, erasure, and type theory.

However, we can't always assume the uses of our interfaces will do the right thing, and so we might decide instead to take the second approach.

For reasons of safety, copy immutable objects at application boundaries.

So inside of your library code, you might take all of the advantages of your mutable and immutable objects, but then once you get to some boundary of your API, you create a copy of your objects to avoid letting users of the API break things internally. Of course, this is for when you actually apply the pattern.

Is it really immutable?

The last thing you want to do is to be able to return references to objects which can change internally, in places where they shouldn't be changing at all. Every part of the state should operate transitively. So if you have an object which holds immutable references, those objects held shouldn't change. If you have to create a copy of a mutable object, you should make it possible to create a deep copy of the object where needed. In addition to being a multi-threading concern, this problem can appear in even single-threaded code.

This is a pattern which is designed to tell you when you don't need to create copies, but it can be misleading. Think about how your objects are used, and don't return immutable references to mutable objects when it's going to be possible to write to those objects later, outside of the casting problem mentioned before.

The Immutable Interface pattern isn't always useful.

The immutable interface pattern doesn't always make sense. Let us consider two example objects, one which only makes sense if it's immutable, another which only makes sense if it's mutable.

For the immutable example, let us consider a Pair<T, U> class. These pair types are tuples of size 2, designed to represent the combination of two types at once. So you construct a pair with two values, and you pass the pair along. Because our language would allow it, we might create setters for the two elements of the pairs. However, I would beg the question, why would you ever want to do such a thing? Tuples are by definition immutable objects. Tuples collect many values in an immutable sequence, an immutable combination of values, if you will. There is no reason to modify the values a tuple holds, you just create another tuple with different values. So for this case, with our pair class, there's no reason to create an immutable interface. Just write the types in the constructor, and make the inner private variables final.

So now we have an example of something that only makes sense as being immutable, so what about only mutable objects? The most obvious example, I think, is an InputStream. The purpose of an InputStream is that it is an abstraction over some data. You open up a stream, you read data from it, once, forwards, and you close the stream when you're done with it. Reading data from an input stream mutates it, as after you've read a byte from the InputStream, you don't see it again. Unless of course you decide to somehow reset the InputStream, which is again another mutation. Supposing you had an immutable interface for an InputStream, what would you do with it? I struggle to imagine anything meaningful. The object only makes sense when it's mutable. You don't need to use the pattern here.

You don't always want the mutable covariant return types

The mutable covariant return type trick was shown as an example of something which could be useful, under certain circumstances. This isn't always the case. Consider an object with a mutable variation, perhaps a Monkey and you've got one greedy monkey. Your monkey holds some Peanuts, and as all monkeys know, not every Peanut is as big as the last, so the Peanuts have a getSize method for retrieving their size, among other attributes, such as the type of the peanuts. Suppose you have some means of getting the peanuts from your MutableMonkey, which gives you MutablePeanuts, which have setSize methods. You will then have the power to change the details of the peanuts the monkey holds, outside of the careful attention and control of your PeanutFactory. You'll have access to details you shouldn't really have access to. You really only need to get the immutable Peanuts from your Monkeys.

So in other words, if you use the covariant return types to give access to inner mutable objects, you may end up with having some unhappy monkeys. And nobody wants unhappy monkeys.

Generic objects cannot be totally covariant

Consider an object Tool<Widget> and another object Tool<MutableWidget>. Given a method Tool<Widget> getTool() in an interface, it is illegal to write a method Tool<MutableWidget> getTool() as a method override. The reasons for this are numerous, but we can simply accept that it cannot be done. Thus, covariant return types will not help you when you have to deal with generic objects.

There are some ways to get around this.

  1. Don't use covariant return types in some cases.
  2. Don't use generic objects in some cases.
  3. Write immutable wrappers for generic objects.
  4. Write the generic objects themselves with the immutable interface pattern.

Option 1 states that we should just return the immutable/mutable generic type, and not worry about it too much. This can be a valid solution in a few cases. Option 2 states that we should avoid generic objects so we can get this behaviour. It might be that Tool only really makes sense for Widgets, so it really only needs to hold Widgets. It might also be the case that Tool is useful for more than one kind of object, but that object either better represented with an interface, or Tool is itself is better represented with an interface, instead of a generic type parameter. Option 3 tells us that, supposing Tool is specified with an interface, we could write a wrapper class which makes the mutable method calls illegal at runtime. This isn't great, because we lose compile time checks, but this is exactly what happens with List<E> and Collections.unmodifiableList in the Java standard library.

You might wonder why the Java standard library doesn't use the immutable interface pattern for the list types. A vendor for another object oriented language did make this decision.

Option 4 tells us that we might instead have an immutable Tool interface, and a MutableTool concrete class. In this case, then every MutableTool<T> which implements a Tool<T> creates the covariance relation we want, and we can indeed perform the override we want. However, the generic type parameter itself will still not be covariant, so mileage may vary. It might be the case that you don't need to modify the T inside of the Tool, so in that case Option 4 becomes a perfect solution.

Conclusion

We have looked at common patterns of writing code with mutable objects in Java, and how these patterns can create code which is harder to understand and more error-prone. We have examined a design pattern which improves on this, and a language feature which can aide use of the design pattern in certain circumstances. We have closed with some words of advice on when these solutions might fail, and when these solutions might not be useful. Nevertheless, the Immutable Interface pattern and covariant return types are useful tools which can help you write better code.

This article makes no discussion of how write software that is heavily protected against problems like NullPointerExceptions. Perhaps means of avoiding problems such as those can be discussed in a future article.

Hopefully, after reading this article, you will have gained a few ideas about how you plan to write your next codebase which may prove useful. Let us forever be seeking ways to improve ourselves, and therefore better ways to create software.

Post Your Comment

Need help with comment formatting?

Okay, here's a cheat sheet for you. Click here to get rid of this.

Choose a help topic below. Raw HTML is not allowed and will be replaced with ugly text. Other Markdown syntax is valid, but probably not relevant.

*italic* **bold**
_italic_ __bold__
1. Item #1
2. Item #2

* Unordered list item
* Unordered list item
Three backticks are your friend here: ```square = lambda x: x * x```

Also in a block like so:

```
def square(x):
  return x * x
```