Never say final: mocking Kotlin classes in unit tests

Never say final: mocking Kotlin classes in unit tests

Danny Preussler

Test doubles are one of the most essential part of unit testing. A test double replaces a class that we need in our test but that is not the focus of our test. We remove that dependency by using a “fake” object. This way our test becomes more robust due to a reduced dependency on other classes. Such a double could be a stub, a dummy or a mock. All of these need the ability to replace an object with something artificial that looks and behaves like the object we need. In Java, the common approach for test doubles is using interfaces: one implementation of the class is the “real” one, and the other one is used for testing.

Writing a lot of unit tests means that all your classes need interfaces. This creates a lot of boilerplate. Therefore it became a common rule for many developers that if you have only one implementation for your interface the interface is useless and can be dropped.

This rule can easily be applied for testing also. As long as we only test the “public contract” of a class we can still replace this by a double. This can be done via extension and overriding or via a mocking framework like Mockito. Mocking classes not only interfaces became pretty common and are a powerful feature. In other languages, like ObjectiveC, where developers do not focus on interfaces as much as in Java, this is the default approach.

The Kotlin problem

Everything seemed fine, until Kotlin came around, a great modern language. Given how well Kotlin and Java work together, the Android developer community was hooked almost from the beginning. But Kotlin brought a new concept: everything is final by default. The rule: “closed for modifications, open for extension” is baked into the language. Every class is final by default, every method is final by default! Inheritance was widely overused in the last decades and Kotlin tries to make this a bit better.

From a testing perspective this is a problem: if you can not extend a class you can not replace it with any kind of double, not even a generated mock:

org.mockito.exceptions.base.MockitoException:
Cannot mock/spy class com.mypackage.MyKotlinClass
Mockito cannot mock/spy following:
 — final classes
 — anonymous classes
 — primitive types

And if the class would not be final you still have all the final methods. But as final methods can not be overridden, real code will suddenly run in your mock. This is something you want to avoid. The problem itself is not new, especially APIs or different third party code were sealed with finals. The developers from Jayway solved this solution a long time ago by creating PowerMock. PowerMock allows to mock finals and even statics. The question for us Kotlin developers is: is there a better way, easier way? Using PowerMock is like using a sledge hammer to crack a nut.

What else could we do? Going back to interfaces? This would lead us to the situation described above. We could open all the classes we need. As we (hopefully) test all the classes, could this lead to opening all our classes!? And would this mean modifying our code just for testing? This sounds wrong! The main feature of the language would be broken.

An idea to solve it

In an ideal world we would just have something like:

controller = mockFinal()

which would be an extension of normal mock() method. Everything this call would need to do, is removing the final modifier from the class and all its public methods.

With Java reflection is very easy to remove the final modifier of a field. For example:

Field modifiersField = Field.class.getDeclaredField(“modifiers”); modifiersField.setAccessible(true);
modifiersField.setInt(field, field.getModifiers() & ~Modifier.FINAL);

But for classes and methods things tend to be a bit trickier. You can not achieve this with pure Java reflection API.

But thanks to frameworks like javassist there is a way. Javassist is a large toolkit around byte code. It allows the manipulation and even creation of Java classes at runtime. And it also includes a way to change the class and method modifiers:

CtClass clazz = ...
int notFinalModifier = Modifier.clear(clazz.getModifiers(), Modifier.FINAL);
clazz.setModifiers(notFinalModifier);
return clazz.toClass();  // returns new non final class

With this, a new non-final class can be easily created. But this leads to the next problem: javassist creates a new class but we can not load the new class into the current test. Because the class is already loaded by the class loader. Two version of it are not allowed:

javassist.CannotCompileException: by java.lang.LinkageError: loader (instance of  sun/misc/Launcher$AppClassLoader): attempted  duplicate class definition for name: "com/myPackage/MyKotlinClass"

So we just need a different class loader right? The problem: an assignment like

controller = mockFinal()

would not work if two different class loaders would be used. So even if we could get passed that Linker error for duplicate class from above we would get a ClassCastException! This is because for the JVM classes are identified by name, package and _class loade_r. So even without modification, instances of two identical classes cannot be assigned to each other if the class loader differs.

This means we need to aim a bit larger than the initial idea: there must be only one class loader for our full test class. Good news! With JUnit there is way to achieve this. We can change the TestRunner that is used for our class.

public class MyTestRunner extends BlockJUnit4ClassRunner {

    public MyTestRunner(Class<?> clazz) throws InitializationError {
        super(getFromMyClassloader(clazz));
    }

    private static Class<?> getFromMyClassloader(Class<?> clazz) ...

then we would write our own ClassLoader and override:

@Override
public Class<?> loadClass(String name) throws ClassNotFoundException { ...

Here you can remove the final modifier with javassist as shown above and return the new “better” class.

Have a look on how all of these look when put together: https://github.com/dpreussler/kotlin-testrunner

If you want to use it, all you need to do is to add **@RunWith(KotlinTestRunner.class) **for Java tests or **@RunWith(KotlinTestRunner::class) **for Kotlin tests

@RunWith(KotlinTestRunner::class)
class MyKotlinTestclass {
   @Test
   fun test() {
   ...
   }
}

With this you will never have to care about if your Kotlin class is final or open. Simply mock them, it will work :-)

Wrap up

For now it seems there is no way around without changing the Testrunner. PowerMock uses the same approach but requires more configuration. So for what we wanted to achieve this is a reasonable solution. Give it a try and let me know how it works for you.

And of course the Runner is not limited to Kotlin. It can also be used for any Java test where a final method or class needs to be mocked.