How do Mockito matchers work?

Question

Mockito argument matchers (such as any, argThat, eq, same, and ArgumentCaptor.capture()) behave very differe

Answer 1

Mockito matchers are static methods and calls to those methods, which stand in for arguments during calls to when and verify.

Hamcrest matchers (archived version) (or Hamcrest-style matchers) are stateless, general-purpose object instances that implement Matcher<T> and expose a method matches(T) that returns true if the object matches the Matcher's criteria. They are intended to be free of side effects, and are generally used in assertions such as the one below.

/* Mockito */  verify(foo).setPowerLevel(gt(9000));
/* Hamcrest */ assertThat(foo.getPowerLevel(), is(greaterThan(9000)));

Mockito matchers exist, separate from Hamcrest-style matchers, so that descriptions of matching expressions fit directly into method invocations: Mockito matchers return T where Hamcrest matcher methods return Matcher objects (of type Matcher<T>).

Mockito matchers are invoked through static methods such as eq, any, gt, and startsWith on org.mockito.Matchers and org.mockito.AdditionalMatchers. There are also adapters, which have changed across Mockito versions:

• For Mockito 1.x, Matchers featured some calls (such as intThat or argThat) are Mockito matchers that directly accept Hamcrest matchers as parameters. ArgumentMatcher<T> extended org.hamcrest.Matcher<T>, which was used in the internal Hamcrest representation and was a Hamcrest matcher base class instead of any sort of Mockito matcher.
• For Mockito 2.0+, Mockito no longer has a direct dependency on Hamcrest. Matchers calls phrased as intThat or argThat wrap ArgumentMatcher<T> objects that no longer implement org.hamcrest.Matcher<T> but are used in similar ways. Hamcrest adapters such as argThat and intThat are still available, but have moved to MockitoHamcrest instead.

Regardless of whether the matchers are Hamcrest or simply Hamcrest-style, they can be adapted like so:

/* Mockito matcher intThat adapting Hamcrest-style matcher is(greaterThan(...)) */
verify(foo).setPowerLevel(intThat(is(greaterThan(9000))));

In the above statement: foo.setPowerLevel is a method that accepts an int. is(greaterThan(9000)) returns a Matcher<Integer>, which wouldn't work as a setPowerLevel argument. The Mockito matcher intThat wraps that Hamcrest-style Matcher and returns an int so it can appear as an argument; Mockito matchers like gt(9000) would wrap that entire expression into a single call, as in the first line of example code.

What matchers do/return

when(foo.quux(3, 5)).thenReturn(true);

When not using argument matchers, Mockito records your argument values and compares them with their equals methods.

when(foo.quux(eq(3), eq(5))).thenReturn(true);    // same as above
when(foo.quux(anyInt(), gt(5))).thenReturn(true); // this one's different

When you call a matcher like any or gt (greater than), Mockito stores a matcher object that causes Mockito to skip that equality check and apply your match of choice. In the case of argumentCaptor.capture() it stores a matcher that saves its argument instead for later inspection.

Matchers return dummy values such as zero, empty collections, or null. Mockito tries to return a safe, appropriate dummy value, like 0 for anyInt() or any(Integer.class) or an empty List<String> for anyListOf(String.class). Because of type erasure, though, Mockito lacks type information to return any value but null for any() or argThat(...), which can cause a NullPointerException if trying to "auto-unbox" a null primitive value.

Matchers like eq and gt take parameter values; ideally, these values should be computed before the stubbing/verification starts. Calling a mock in the middle of mocking another call can interfere with stubbing.

Matcher methods can't be used as return values; there is no way to phrase thenReturn(anyInt()) or thenReturn(any(Foo.class)) in Mockito, for instance. Mockito needs to know exactly which instance to return in stubbing calls, and will not choose an arbitrary return value for you.

Implementation details

Matchers are stored (as Hamcrest-style object matchers) in a stack contained in a class called ArgumentMatcherStorage. MockitoCore and Matchers each own a ThreadSafeMockingProgress instance, which statically contains a ThreadLocal holding MockingProgress instances. It's this MockingProgressImpl that holds a concrete ArgumentMatcherStorageImpl. Consequently, mock and matcher state is static but thread-scoped consistently between the Mockito and Matchers classes.

Most matcher calls only add to this stack, with an exception for matchers like and, or, and not. This perfectly corresponds to (and relies on) the evaluation order of Java, which evaluates arguments left-to-right before invoking a method:

when(foo.quux(anyInt(), and(gt(10), lt(20)))).thenReturn(true);
[6]      [5]  [1]       [4] [2]     [3]

This will:

1. Add anyInt() to the stack.
2. Add gt(10) to the stack.
3. Add lt(20) to the stack.
4. Remove gt(10) and lt(20) and add and(gt(10), lt(20)).
5. Call foo.quux(0, 0), which (unless otherwise stubbed) returns the default value false. Internally Mockito marks quux(int, int) as the most recent call.
6. Call when(false), which discards its argument and prepares to stub method quux(int, int) identified in 5. The only two valid states are with stack length 0 (equality) or 2 (matchers), and there are two matchers on the stack (steps 1 and 4), so Mockito stubs the method with an any() matcher for its first argument and and(gt(10), lt(20)) for its second argument and clears the stack.

This demonstrates a few rules:

• Mockito can't tell the difference between quux(anyInt(), 0) and quux(0, anyInt()). They both look like a call to quux(0, 0) with one int matcher on the stack. Consequently, if you use one matcher, you have to match all arguments.

• Call order isn't just important, it's what makes this all work. Extracting matchers to variables generally doesn't work, because it usually changes the call order. Extracting matchers to methods, however, works great.

  int between10And20 = and(gt(10), lt(20));
  /* BAD */ when(foo.quux(anyInt(), between10And20)).thenReturn(true);
  // Mockito sees the stack as the opposite: and(gt(10), lt(20)), anyInt().
  
  public static int anyIntBetween10And20() { return and(gt(10), lt(20)); }
  /* OK */  when(foo.quux(anyInt(), anyIntBetween10And20())).thenReturn(true);
  // The helper method calls the matcher methods in the right order.
  

• The stack changes often enough that Mockito can't police it very carefully. It can only check the stack when you interact with Mockito or a mock, and has to accept matchers without knowing whether they're used immediately or abandoned accidentally. In theory, the stack should always be empty outside of a call to when or verify, but Mockito can't check that automatically. You can check manually with Mockito.validateMockitoUsage().

• In a call to when, Mockito actually calls the method in question, which will throw an exception if you've stubbed the method to throw an exception (or require non-zero or non-null values). doReturn and doAnswer (etc) do not invoke the actual method and are often a useful alternative.

• If you had called a mock method in the middle of stubbing (e.g. to calculate an answer for an eq matcher), Mockito would check the stack length against that call instead, and likely fail.

• If you try to do something bad, like stubbing/verifying a final method, Mockito will call the real method and also leave extra matchers on the stack. The final method call may not throw an exception, but you may get an InvalidUseOfMatchersException from the stray matchers when you next interact with a mock.

Common problems

• InvalidUseOfMatchersException:

  • Check that every single argument has exactly one matcher call, if you use matchers at all, and that you haven't used a matcher outside of a when or verify call. Matchers should never be used as stubbed return values or fields/variables.

  • Check that you're not calling a mock as a part of providing a matcher argument.

  • Check that you're not trying to stub/verify a final method with a matcher. It's a great way to leave a matcher on the stack, and unless your final method throws an exception, this might be the only time you realize the method you're mocking is final.

• NullPointerException with primitive arguments: (Integer) any() returns null while any(Integer.class) returns 0; this can cause a NullPointerException if you're expecting an int instead of an Integer. In any case, prefer anyInt(), which will return zero and also skip the auto-boxing step.

• NullPointerException or other exceptions: Calls to when(foo.bar(any())).thenReturn(baz) will actually call foo.bar(null), which you might have stubbed to throw an exception when receiving a null argument. Switching to doReturn(baz).when(foo).bar(any()) skips the stubbed behavior.

General troubleshooting

• Use MockitoJUnitRunner, or explicitly call validateMockitoUsage in your tearDown or @After method (which the runner would do for you automatically). This will help determine whether you've misused matchers.

• For debugging purposes, add calls to validateMockitoUsage in your code directly. This will throw if you have anything on the stack, which is a good warning of a bad symptom.

Answer 2

Just a small addition to Jeff Bowman's excellent answer, as I found this question when searching for a solution to one of my own problems:

If a call to a method matches more than one mock's when trained calls, the order of the when calls is important, and should be from the most wider to the most specific. Starting from one of Jeff's examples:

when(foo.quux(anyInt(), anyInt())).thenReturn(true);
when(foo.quux(anyInt(), eq(5))).thenReturn(false);

is the order that ensures the (probably) desired result:

foo.quux(3 /*any int*/, 8 /*any other int than 5*/) //returns true
foo.quux(2 /*any int*/, 5) //returns false

If you inverse the when calls then the result would always be true.