Testing Actor Systems

Testing Actor Systems

As with any piece of software, automated tests are a very important part of the development cycle. The actor model presents a different view on how units of code are delimited and how they interact, which has an influence on how to perform tests.

Akka comes with a dedicated module akka-testkit for supporting tests at different levels, which fall into two clearly distinct categories:

  • Testing isolated pieces of code without involving the actor model, meaning without multiple threads; this implies completely deterministic behavior concerning the ordering of events and no concurrency concerns and will be called Unit Testing in the following.
  • Testing (multiple) encapsulated actors including multi-threaded scheduling; this implies non-deterministic order of events but shielding from concurrency concerns by the actor model and will be called Integration Testing in the following.

There are of course variations on the granularity of tests in both categories, where unit testing reaches down to white-box tests and integration testing can encompass functional tests of complete actor networks. The important distinction lies in whether concurrency concerns are part of the test or not. The tools offered are described in detail in the following sections.

注釈

Be sure to add the module akka-testkit to your dependencies.

Synchronous Unit Testing with TestActorRef

Testing the business logic inside Actor classes can be divided into two parts: first, each atomic operation must work in isolation, then sequences of incoming events must be processed correctly, even in the presence of some possible variability in the ordering of events. The former is the primary use case for single-threaded unit testing, while the latter can only be verified in integration tests.

Normally, the ActorRef shields the underlying Actor instance from the outside, the only communications channel is the actor's mailbox. This restriction is an impediment to unit testing, which led to the inception of the TestActorRef. This special type of reference is designed specifically for test purposes and allows access to the actor in two ways: either by obtaining a reference to the underlying actor instance, or by invoking or querying the actor's behaviour (receive). Each one warrants its own section below.

注釈

It is highly recommended to stick to traditional behavioural testing (using messaging to ask the Actor to reply with the state you want to run assertions against), instead of using TestActorRef whenever possible.

警告

Due to the synchronous nature of TestActorRef it will not work with some support traits that Akka provides as they require asynchronous behaviours to function properly. Examples of traits that do not mix well with test actor refs are PersistentActor and AtLeastOnceDelivery provided by Akka Persistence.

Obtaining a Reference to an Actor

Having access to the actual Actor object allows application of all traditional unit testing techniques on the contained methods. Obtaining a reference is done like this:

static class MyActor extends UntypedActor {
  public void onReceive(Object o) throws Exception {
    if (o.equals("say42")) {
      getSender().tell(42, getSelf());
    } else if (o instanceof Exception) {
      throw (Exception) o;
    }
  }
  public boolean testMe() { return true; }
}

@Test
public void demonstrateTestActorRef() {
  final Props props = Props.create(MyActor.class);
  final TestActorRef<MyActor> ref = TestActorRef.create(system, props, "testA");
  final MyActor actor = ref.underlyingActor();
  assertTrue(actor.testMe());
}

Since TestActorRef is generic in the actor type it returns the underlying actor with its proper static type. From this point on you may bring any unit testing tool to bear on your actor as usual.

Testing the Actor's Behavior

When the dispatcher invokes the processing behavior of an actor on a message, it actually calls apply on the current behavior registered for the actor. This starts out with the return value of the declared receive method, but it may also be changed using become and unbecome in response to external messages. All of this contributes to the overall actor behavior and it does not lend itself to easy testing on the Actor itself. Therefore the TestActorRef offers a different mode of operation to complement the Actor testing: it supports all operations also valid on normal ActorRef. Messages sent to the actor are processed synchronously on the current thread and answers may be sent back as usual. This trick is made possible by the CallingThreadDispatcher described below (see CallingThreadDispatcher); this dispatcher is set implicitly for any actor instantiated into a TestActorRef.

final Props props = Props.create(MyActor.class);
final TestActorRef<MyActor> ref = TestActorRef.create(system, props, "testB");
final Future<Object> future = akka.pattern.Patterns.ask(ref, "say42", 3000);
assertTrue(future.isCompleted());
assertEquals(42, Await.result(future, Duration.Zero()));

As the TestActorRef is a subclass of LocalActorRef with a few special extras, also aspects like supervision and restarting work properly, but beware that execution is only strictly synchronous as long as all actors involved use the CallingThreadDispatcher. As soon as you add elements which include more sophisticated scheduling you leave the realm of unit testing as you then need to think about asynchronicity again (in most cases the problem will be to wait until the desired effect had a chance to happen).

One more special aspect which is overridden for single-threaded tests is the receiveTimeout, as including that would entail asynchronous queuing of ReceiveTimeout messages, violating the synchronous contract.

注釈

To summarize: TestActorRef overwrites two fields: it sets the dispatcher to CallingThreadDispatcher.global and it sets the receiveTimeout to None.

The Way In-Between: Expecting Exceptions

If you want to test the actor behavior, including hotswapping, but without involving a dispatcher and without having the TestActorRef swallow any thrown exceptions, then there is another mode available for you: just use the receive method on TestActorRef, which will be forwarded to the underlying actor:

final Props props = Props.create(MyActor.class);
final TestActorRef<MyActor> ref = TestActorRef.create(system, props, "myActor");
try {
  ref.receive(new Exception("expected"));
  fail("expected an exception to be thrown");
} catch (Exception e) {
  assertEquals("expected", e.getMessage());
}

Use Cases

You may of course mix and match both modi operandi of TestActorRef as suits your test needs:

  • one common use case is setting up the actor into a specific internal state before sending the test message
  • another is to verify correct internal state transitions after having sent the test message

Feel free to experiment with the possibilities, and if you find useful patterns, don't hesitate to let the Akka forums know about them! Who knows, common operations might even be worked into nice DSLs.

Asynchronous Integration Testing with JavaTestKit

When you are reasonably sure that your actor's business logic is correct, the next step is verifying that it works correctly within its intended environment. The definition of the environment depends of course very much on the problem at hand and the level at which you intend to test, ranging for functional/integration tests to full system tests. The minimal setup consists of the test procedure, which provides the desired stimuli, the actor under test, and an actor receiving replies. Bigger systems replace the actor under test with a network of actors, apply stimuli at varying injection points and arrange results to be sent from different emission points, but the basic principle stays the same in that a single procedure drives the test.

The JavaTestKit class contains a collection of tools which makes this common task easy.

import org.junit.AfterClass;
import org.junit.Assert;
import org.junit.BeforeClass;
import org.junit.Test;

import akka.actor.ActorRef;
import akka.actor.ActorSystem;
import akka.actor.Props;
import akka.actor.UntypedActor;
import akka.testkit.JavaTestKit;
import scala.concurrent.duration.Duration;

public class TestKitSampleTest {
  
  public static class SomeActor extends UntypedActor {
    ActorRef target = null;
    
    public void onReceive(Object msg) {
    
      if (msg.equals("hello")) {
        getSender().tell("world", getSelf());
        if (target != null) target.forward(msg, getContext());
      
      } else if (msg instanceof ActorRef) {
        target = (ActorRef) msg;
        getSender().tell("done", getSelf());
      }
    }
  }
  
  static ActorSystem system;
  
  @BeforeClass
  public static void setup() {
    system = ActorSystem.create();
  }
  
  @AfterClass
  public static void teardown() {
    JavaTestKit.shutdownActorSystem(system);
    system = null;
  }

  @Test
  public void testIt() {
    /*
     * Wrap the whole test procedure within a testkit constructor 
     * if you want to receive actor replies or use Within(), etc.
     */
    new JavaTestKit(system) {{
      final Props props = Props.create(SomeActor.class);
      final ActorRef subject = system.actorOf(props);

      // can also use JavaTestKit “from the outside”
      final JavaTestKit probe = new JavaTestKit(system);
      // “inject” the probe by passing it to the test subject
      // like a real resource would be passed in production
      subject.tell(probe.getRef(), getRef());
      // await the correct response
      expectMsgEquals(duration("1 second"), "done");
      
      // the run() method needs to finish within 3 seconds
      new Within(duration("3 seconds")) {
        protected void run() {

          subject.tell("hello", getRef());

          // This is a demo: would normally use expectMsgEquals().
          // Wait time is bounded by 3-second deadline above.
          new AwaitCond() {
            protected boolean cond() {
              return probe.msgAvailable();
            }
          };

          // response must have been enqueued to us before probe
          expectMsgEquals(Duration.Zero(), "world");
          // check that the probe we injected earlier got the msg
          probe.expectMsgEquals(Duration.Zero(), "hello");
          Assert.assertEquals(getRef(), probe.getLastSender());

          // Will wait for the rest of the 3 seconds
          expectNoMsg();
        }
      };
    }};
  }
  
}

The JavaTestKit contains an actor named testActor which is the entry point for messages to be examined with the various expectMsg... assertions detailed below. The test actor’s reference is obtained using the getRef method as demonstrated above. The testActor may also be passed to other actors as usual, usually subscribing it as notification listener. There is a whole set of examination methods, e.g. receiving all consecutive messages matching certain criteria, receiving a whole sequence of fixed messages or classes, receiving nothing for some time, etc.

The ActorSystem passed in to the constructor of JavaTestKit is accessible via the getSystem method.

注釈

Remember to shut down the actor system after the test is finished (also in case of failure) so that all actors—including the test actor—are stopped.

Built-In Assertions

The above mentioned expectMsgEquals is not the only method for formulating assertions concerning received messages, the full set is this:

final String hello = expectMsgEquals("hello");
final Object   any = expectMsgAnyOf("hello", "world");
final Object[] all = expectMsgAllOf("hello", "world");
final int i        = expectMsgClass(Integer.class);
final Number j     = expectMsgAnyClassOf(Integer.class, Long.class);
expectNoMsg();
final Object[] two = receiveN(2);

In these examples, the maximum durations you will find mentioned below are left out, in which case they use the default value from configuration item akka.test.single-expect-default which itself defaults to 3 seconds (or they obey the innermost enclosing Within as detailed below). The full signatures are:

  • public <T> T expectMsgEquals(Duration max, T msg)

    The given message object must be received within the specified time; the object will be returned.

  • public Object expectMsgAnyOf(Duration max, Object... msg)

    An object must be received within the given time, and it must be equal (compared with equals()) to at least one of the passed reference objects; the received object will be returned.

  • public Object[] expectMsgAllOf(Duration max, Object... msg)

    A number of objects matching the size of the supplied object array must be received within the given time, and for each of the given objects there must exist at least one among the received ones which equals it (compared with equals()). The full sequence of received objects is returned in the order received.

  • public <T> T expectMsgClass(Duration max, Class<T> c)

    An object which is an instance of the given Class must be received within the allotted time frame; the object will be returned. Note that this does a conformance check, if you need the class to be equal you need to verify that afterwards.

  • public <T> T expectMsgAnyClassOf(Duration max, Class<? extends T>... c)

    An object must be received within the given time, and it must be an instance of at least one of the supplied Class objects; the received object will be returned. Note that this does a conformance check, if you need the class to be equal you need to verify that afterwards.

    注釈

    Because of a limitation in Java’s type system it may be necessary to add @SuppressWarnings("unchecked") when using this method.

  • public void expectNoMsg(Duration max)

    No message must be received within the given time. This also fails if a message has been received before calling this method which has not been removed from the queue using one of the other methods.

  • Object[] receiveN(int n, Duration max)

    n messages must be received within the given time; the received messages are returned.

For cases which require more refined conditions there are constructs which take code blocks:

  • ExpectMsg<T>

    new JavaTestKit(system) {{
      getRef().tell(42, ActorRef.noSender());
      final String out = new ExpectMsg<String>("match hint") {
          // do not put code outside this method, will run afterwards
          protected String match(Object in) {
            if (in instanceof Integer) {
              return "match";
            } else {
              throw noMatch();
            }
          }
        }.get(); // this extracts the received message
      assertEquals("match", out);
    }};
    

    The match(Object in) method will be evaluated once a message has been received within the allotted time (which may be given as constructor argument). If it throws noMatch() (where it is sufficient to call that method; the throw keyword is only needed in cases where the compiler would otherwise complain about wrong return types—Java is lacking Scala’s notion of a type which signifies “will not ever return normally”), then the expectation fails with an AssertionError, otherwise the matched and possibly transformed object is stored for retrieval using the get method.

  • ReceiveWhile<T>

    new JavaTestKit(system) {{
      getRef().tell(42, ActorRef.noSender());
      getRef().tell(43, ActorRef.noSender());
      getRef().tell("hello", ActorRef.noSender());
      final String[] out =
        new ReceiveWhile<String>(String.class, duration("1 second")) {
          // do not put code outside this method, will run afterwards
          protected String match(Object in) {
            if (in instanceof Integer) {
              return in.toString();
            } else {
              throw noMatch();
            }
          }
        }.get(); // this extracts the received messages
      assertArrayEquals(new String[] {"42", "43"}, out);
      expectMsgEquals("hello");
    }};
    

    This construct works like ExpectMsg, but it continually collects messages as long as they match the criteria, and it does not fail when a non-matching one is encountered. Collecting messages also ends when the time is up, when too much time passes between messages or when enough messages have been received.

    new ReceiveWhile<String>(     // type of array to be created must match ...
          String.class,           // ... this class which is needed to that end
          duration("100 millis"), // maximum collect time
          duration("50 millis"),  // maximum time between messages
          12                      // maximum number of messages to collect
          ) {
      // match elided ...
    };
    

    The need to specify the String result type twice results from the need to create a correctly typed array and Java’s inability to infer the class’s type argument.

  • AwaitCond

    new JavaTestKit(system) {{
      getRef().tell(42, ActorRef.noSender());
      new AwaitCond(
            duration("1 second"),  // maximum wait time
            duration("100 millis") // interval at which to check the condition
            ) {
        // do not put code outside this method, will run afterwards
        protected boolean cond() {
          // typically used to wait for something to start up
          return msgAvailable();
        }
      };
    }};
    

    This general construct is not connected with the test kit’s message reception, the embedded condition can compute the boolean result from anything in scope.

    • AwaitAssert
    new JavaTestKit(system) {{
      getRef().tell(42, ActorRef.noSender());
      new AwaitAssert(
            duration("1 second"),  // maximum wait time
            duration("100 millis") // interval at which to check the condition
            ) {
        // do not put code outside this method, will run afterwards
        protected void check() {
          assertEquals(msgAvailable(), true);
        }
      };
    }};
    

    This general construct is not connected with the test kit’s message reception, the embedded assert can check anything in scope.

There are also cases where not all messages sent to the test kit are actually relevant to the test, but removing them would mean altering the actors under test. For this purpose it is possible to ignore certain messages:

  • IgnoreMsg

    new JavaTestKit(system) {{
      // ignore all Strings
      new IgnoreMsg() {
        protected boolean ignore(Object msg) {
          return msg instanceof String;
        }
      };
      getRef().tell("hello", ActorRef.noSender());
      getRef().tell(42, ActorRef.noSender());
      expectMsgEquals(42);
      // remove message filter
      ignoreNoMsg();
      getRef().tell("hello", ActorRef.noSender());
      expectMsgEquals("hello");
    }};
    

Expecting Log Messages

Since an integration test does not allow to the internal processing of the participating actors, verifying expected exceptions cannot be done directly. Instead, use the logging system for this purpose: replacing the normal event handler with the TestEventListener and using an EventFilter allows assertions on log messages, including those which are generated by exceptions:

new JavaTestKit(system) {{
  assertEquals("TestKitDocTest", system.name());
  final ActorRef victim = system.actorOf(Props.empty(), "victim");

  final int result = new EventFilter<Integer>(ActorKilledException.class) {
    protected Integer run() {
      victim.tell(Kill.getInstance(), ActorRef.noSender());
      return 42;
    }
  }.from("akka://TestKitDocTest/user/victim").occurrences(1).exec();

  assertEquals(42, result);
}};

If a number of occurrences is specific—as demonstrated above—then exec() will block until that number of matching messages have been received or the timeout configured in akka.test.filter-leeway is used up (time starts counting after the run() method returns). In case of a timeout the test fails.

注釈

Be sure to exchange the default logger with the TestEventListener in your application.conf to enable this function:

akka.loggers = [akka.testkit.TestEventListener]

Timing Assertions

Another important part of functional testing concerns timing: certain events must not happen immediately (like a timer), others need to happen before a deadline. Therefore, all examination methods accept an upper time limit within the positive or negative result must be obtained. Lower time limits need to be checked external to the examination, which is facilitated by a new construct for managing time constraints:

new JavaTestKit(system) {{
  getRef().tell(42, ActorRef.noSender());
  new Within(Duration.Zero(), Duration.create(1, "second")) {
    // do not put code outside this method, will run afterwards
    public void run() {
      assertEquals((Integer) 42, expectMsgClass(Integer.class));
    }
  };
}};

The block in Within.run must complete after a Duration which is between min and max, where the former defaults to zero. The deadline calculated by adding the max parameter to the block's start time is implicitly available within the block to all examination methods, if you do not specify it, it is inherited from the innermost enclosing within block.

It should be noted that if the last message-receiving assertion of the block is expectNoMsg or receiveWhile, the final check of the within is skipped in order to avoid false positives due to wake-up latencies. This means that while individual contained assertions still use the maximum time bound, the overall block may take arbitrarily longer in this case.

注釈

All times are measured using System.nanoTime, meaning that they describe wall time, not CPU time or system time.

Accounting for Slow Test Systems

The tight timeouts you use during testing on your lightning-fast notebook will invariably lead to spurious test failures on the heavily loaded Jenkins server (or similar). To account for this situation, all maximum durations are internally scaled by a factor taken from the 環境設定, akka.test.timefactor, which defaults to 1.

You can scale other durations with the same factor by using dilated method in JavaTestKit.

new JavaTestKit(system) {{
  final Duration original = duration("1 second");
  final Duration stretched = dilated(original);
  assertTrue("dilated", stretched.gteq(original));
}};

Using Multiple Probe Actors

When the actors under test are supposed to send various messages to different destinations, it may be difficult distinguishing the message streams arriving at the testActor when using the JavaTestKit as shown until now. Another approach is to use it for creation of simple probe actors to be inserted in the message flows. The functionality is best explained using a small example:

new JavaTestKit(system) {{
  // simple actor which just forwards messages
  class Forwarder extends UntypedActor {
    final ActorRef target;
    @SuppressWarnings("unused")
    public Forwarder(ActorRef target) {
      this.target = target;
    }
    public void onReceive(Object msg) {
      target.forward(msg, getContext());
    }
  }
  
  // create a test probe
  final JavaTestKit probe = new JavaTestKit(system);

  // create a forwarder, injecting the probe’s testActor
  final Props props = Props.create(Forwarder.class, this, probe.getRef());
  final ActorRef forwarder = system.actorOf(props, "forwarder");

  // verify correct forwarding
  forwarder.tell(42, getRef());
  probe.expectMsgEquals(42);
  assertEquals(getRef(), probe.getLastSender());
}};

This simple test verifies an equally simple Forwarder actor by injecting a probe as the forwarder’s target. Another example would be two actors A and B which collaborate by A sending messages to B. In order to verify this message flow, a TestProbe could be inserted as target of A, using the forwarding capabilities or auto-pilot described below to include a real B in the test setup.

If you have many test probes, you can name them to get meaningful actor names in test logs and assertions:

new JavaTestKit(system) {{
  final TestProbe worker = new TestProbe(system, "worker");
  final TestProbe aggregator = new TestProbe(system, "aggregator");

  assertTrue(worker.ref().path().name().startsWith("worker"));
  assertTrue(aggregator.ref().path().name().startsWith("aggregator"));
}};

Probes may also be equipped with custom assertions to make your test code even more concise and clear:

new JavaTestKit(system) {{
  class MyProbe extends JavaTestKit {
    public MyProbe() {
      super(system);
    }
    public void assertHello() {
      expectMsgEquals("hello");
    }
  }

  final MyProbe probe = new MyProbe();
  probe.getRef().tell("hello", ActorRef.noSender());
  probe.assertHello();
}};

You have complete flexibility here in mixing and matching the JavaTestKit facilities with your own checks and choosing an intuitive name for it. In real life your code will probably be a bit more complicated than the example given above; just use the power!

警告

Any message send from a TestProbe to another actor which runs on the CallingThreadDispatcher runs the risk of dead-lock, if that other actor might also send to this probe. The implementation of TestProbe.watch and TestProbe.unwatch will also send a message to the watchee, which means that it is dangerous to try watching e.g. TestActorRef from a TestProbe.

Watching Other Actors from Probes

A JavaTestKit can register itself for DeathWatch of any other actor:

new JavaTestKit(system) {{
  final JavaTestKit probe = new JavaTestKit(system);
  probe.watch(target);
  target.tell(PoisonPill.getInstance(), ActorRef.noSender());
  final Terminated msg = probe.expectMsgClass(Terminated.class);
  assertEquals(msg.getActor(), target);
}};

Replying to Messages Received by Probes

The probe stores the sender of the last dequeued message (i.e. after its expectMsg* reception), which may be retrieved using the getLastSender method. This information can also implicitly be used for having the probe reply to the last received message:

new JavaTestKit(system) {{
  final JavaTestKit probe = new JavaTestKit(system);
  probe.getRef().tell("hello", getRef());
  probe.expectMsgEquals("hello");
  probe.reply("world");
  expectMsgEquals("world");
  assertEquals(probe.getRef(), getLastSender());
}};

Forwarding Messages Received by Probes

The probe can also forward a received message (i.e. after its expectMsg* reception), retaining the original sender:

new JavaTestKit(system) {{
  final JavaTestKit probe = new JavaTestKit(system);
  probe.getRef().tell("hello", getRef());
  probe.expectMsgEquals("hello");
  probe.forward(getRef());
  expectMsgEquals("hello");
  assertEquals(getRef(), getLastSender());
}};

Auto-Pilot

Receiving messages in a queue for later inspection is nice, but in order to keep a test running and verify traces later you can also install an AutoPilot in the participating test probes (actually in any TestKit) which is invoked before enqueueing to the inspection queue. This code can be used to forward messages, e.g. in a chain A --> Probe --> B, as long as a certain protocol is obeyed.

new JavaTestKit(system) {{
  final JavaTestKit probe = new JavaTestKit(system);
  // install auto-pilot
  probe.setAutoPilot(new TestActor.AutoPilot() {
    public AutoPilot run(ActorRef sender, Object msg) {
      sender.tell(msg, ActorRef.noSender());
      return noAutoPilot();
    }
  });
  // first one is replied to directly ...
  probe.getRef().tell("hello", getRef());
  expectMsgEquals("hello");
  // ... but then the auto-pilot switched itself off
  probe.getRef().tell("world", getRef());
  expectNoMsg();
}};

The run method must return the auto-pilot for the next message, wrapped in an Option; setting it to None terminates the auto-pilot.

Caution about Timing Assertions

The behavior of within blocks when using test probes might be perceived as counter-intuitive: you need to remember that the nicely scoped deadline as described above is local to each probe. Hence, probes do not react to each other's deadlines or to the deadline set in an enclosing JavaTestKit instance:

new JavaTestKit(system) {{
  final JavaTestKit probe = new JavaTestKit(system);
  new Within(duration("1 second")) {
    public void run() {
      probe.expectMsgEquals("hello");
    }
  };
}};

Here, the expectMsgEquals call will use the default timeout.

Testing parent-child relationships

The parent of an actor is always the actor that created it. At times this leads to a coupling between the two that may not be straightforward to test. There are several approaches to improve testability of a child actor that needs to refer to its parent:

  1. when creating a child, pass an explicit reference to its parent
  2. create the child with a TestProbe as parent
  3. create a fabricated parent when testing

Conversely, a parent's binding to its child can be lessened as follows:

  1. when creating a parent, tell the parent how to create its child

For example, the structure of the code you want to test may follow this pattern:

static class Parent extends UntypedActor {
  final ActorRef child = context().actorOf(Props.create(Child.class), "child");
  boolean ponged = false;

  @Override public void onReceive(Object message) throws Exception {
    if ("pingit".equals(message)) {
      child.tell("ping", self());
    } else if ("pong".equals(message)) {
      ponged = true;
    } else {
      unhandled(message);
    }
  }
}

static class Child extends UntypedActor {
  @Override public void onReceive(Object message) throws Exception {
    if ("ping".equals(message)) {
      context().parent().tell("pong", self());
    } else {
      unhandled(message);
    }
  }
}

Introduce child to its parent

The first option is to avoid use of the context.parent function and create a child with a custom parent by passing an explicit reference to its parent instead.

class DependentChild extends UntypedActor {
  private final ActorRef parent;

  public DependentChild(ActorRef parent) {
    this.parent = parent;
  }

  @Override public void onReceive(Object message) throws Exception {
    if ("ping".equals(message)) {
      parent.tell("pong", self());
    } else {
      unhandled(message);
    }
  }
}

Create the child using JavaTestKit

The JavaTestKit class can in fact create actors that will run with the test probe as parent. This will cause any messages the the child actor sends to context().getParent() to end up in the test probe.

JavaTestKit parent = new JavaTestKit(system);
ActorRef child = parent.childActorOf(Props.create(Child.class));

parent.send(child, "ping");
parent.expectMsgEquals("pong");

Using a fabricated parent

If you prefer to avoid modifying the child constructor you can create a fabricated parent in your test. This, however, does not enable you to test the parent actor in isolation.

class FabricatedParentCreator implements Creator<Actor> {
  private final TestProbe proxy;

  public FabricatedParentCreator(TestProbe proxy) {
    this.proxy = proxy;
  }

  @Override public Actor create() throws Exception {
    return new UntypedActor() {
      final ActorRef child = context().actorOf(Props.create(Child.class), "child");

      @Override public void onReceive(Object x) throws Exception {
        if (sender().equals(child)) {
          proxy.ref().forward(x, context());
        } else {
          child.forward(x, context());
        }
      }
    };
  }
}
TestProbe proxy = new TestProbe(system);
ActorRef parent = system.actorOf(Props.create(new FabricatedParentCreator(proxy)));

proxy.send(parent, "ping");
proxy.expectMsg("pong");

Externalize child making from the parent

Alternatively, you can tell the parent how to create its child. There are two ways to do this: by giving it a Props object or by giving it a function which takes care of creating the child actor:

class DependentParent extends UntypedActor {
  final ActorRef child;
  boolean ponged = false;

  public DependentParent(Props childProps) {
    child = context().actorOf(childProps, "child");
  }

  @Override public void onReceive(Object message) throws Exception {
    if ("pingit".equals(message)) {
      child.tell("ping", self());
    } else if ("pong".equals(message)) {
      ponged = true;
    } else {
      unhandled(message);
    }
  }
}
class GenericDependentParent extends UntypedActor {
  final ActorRef child;
  boolean ponged = false;

  public GenericDependentParent(Function<ActorRefFactory, ActorRef> childMaker)
    throws Exception {
    child = childMaker.apply(context());
  }

  @Override public void onReceive(Object message) throws Exception {
    if ("pingit".equals(message)) {
      child.tell("ping", self());
    } else if ("pong".equals(message)) {
      ponged = true;
    } else {
      unhandled(message);
    }
  }
}

Creating the Actor is straightforward and the function may look like this in your test code:

Function<ActorRefFactory, ActorRef> maker = new Function<ActorRefFactory, ActorRef>() {
  @Override public ActorRef apply(ActorRefFactory param) throws Exception {
    return probe.ref();
  }
};
ActorRef parent = system.actorOf(Props.create(GenericDependentParent.class, maker));

And like this in your application code:

Function<ActorRefFactory, ActorRef> maker = new Function<ActorRefFactory, ActorRef>() {
  @Override public ActorRef apply(ActorRefFactory f) throws Exception {
    return f.actorOf(Props.create(Child.class));
  }
};
ActorRef parent = system.actorOf(Props.create(GenericDependentParent.class, maker));

Which of these methods is the best depends on what is most important to test. The most generic option is to create the parent actor by passing it a function that is responsible for the Actor creation, but using TestProbe or having a fabricated parent is often sufficient.

CallingThreadDispatcher

The CallingThreadDispatcher serves good purposes in unit testing, as described above, but originally it was conceived in order to allow contiguous stack traces to be generated in case of an error. As this special dispatcher runs everything which would normally be queued directly on the current thread, the full history of a message's processing chain is recorded on the call stack, so long as all intervening actors run on this dispatcher.

How to use it

Just set the dispatcher as you normally would:

system.actorOf(
  Props.create(MyActor.class)
    .withDispatcher(CallingThreadDispatcher.Id()));

How it works

When receiving an invocation, the CallingThreadDispatcher checks whether the receiving actor is already active on the current thread. The simplest example for this situation is an actor which sends a message to itself. In this case, processing cannot continue immediately as that would violate the actor model, so the invocation is queued and will be processed when the active invocation on that actor finishes its processing; thus, it will be processed on the calling thread, but simply after the actor finishes its previous work. In the other case, the invocation is simply processed immediately on the current thread. Futures scheduled via this dispatcher are also executed immediately.

This scheme makes the CallingThreadDispatcher work like a general purpose dispatcher for any actors which never block on external events.

In the presence of multiple threads it may happen that two invocations of an actor running on this dispatcher happen on two different threads at the same time. In this case, both will be processed directly on their respective threads, where both compete for the actor's lock and the loser has to wait. Thus, the actor model is left intact, but the price is loss of concurrency due to limited scheduling. In a sense this is equivalent to traditional mutex style concurrency.

The other remaining difficulty is correct handling of suspend and resume: when an actor is suspended, subsequent invocations will be queued in thread-local queues (the same ones used for queuing in the normal case). The call to resume, however, is done by one specific thread, and all other threads in the system will probably not be executing this specific actor, which leads to the problem that the thread-local queues cannot be emptied by their native threads. Hence, the thread calling resume will collect all currently queued invocations from all threads into its own queue and process them.

Limitations

警告

In case the CallingThreadDispatcher is used for top-level actors, but without going through TestActorRef, then there is a time window during which the actor is awaiting construction by the user guardian actor. Sending messages to the actor during this time period will result in them being enqueued and then executed on the guardian’s thread instead of the caller’s thread. To avoid this, use TestActorRef.

If an actor's behavior blocks on a something which would normally be affected by the calling actor after having sent the message, this will obviously dead-lock when using this dispatcher. This is a common scenario in actor tests based on CountDownLatch for synchronization:

val latch = new CountDownLatch(1)
actor ! startWorkAfter(latch)   // actor will call latch.await() before proceeding
doSomeSetupStuff()
latch.countDown()

The example would hang indefinitely within the message processing initiated on the second line and never reach the fourth line, which would unblock it on a normal dispatcher.

Thus, keep in mind that the CallingThreadDispatcher is not a general-purpose replacement for the normal dispatchers. On the other hand it may be quite useful to run your actor network on it for testing, because if it runs without dead-locking chances are very high that it will not dead-lock in production.

警告

The above sentence is unfortunately not a strong guarantee, because your code might directly or indirectly change its behavior when running on a different dispatcher. If you are looking for a tool to help you debug dead-locks, the CallingThreadDispatcher may help with certain error scenarios, but keep in mind that it has may give false negatives as well as false positives.

Thread Interruptions

If the CallingThreadDispatcher sees that the current thread has its isInterrupted() flag set when message processing returns, it will throw an InterruptedException after finishing all its processing (i.e. all messages which need processing as described above are processed before this happens). As tell cannot throw exceptions due to its contract, this exception will then be caught and logged, and the thread’s interrupted status will be set again.

If during message processing an InterruptedException is thrown then it will be caught inside the CallingThreadDispatcher’s message handling loop, the thread’s interrupted flag will be set and processing continues normally.

注釈

The summary of these two paragraphs is that if the current thread is interrupted while doing work under the CallingThreadDispatcher, then that will result in the isInterrupted flag to be true when the message send returns and no InterruptedException will be thrown.

Benefits

To summarize, these are the features with the CallingThreadDispatcher has to offer:

  • Deterministic execution of single-threaded tests while retaining nearly full actor semantics
  • Full message processing history leading up to the point of failure in exception stack traces
  • Exclusion of certain classes of dead-lock scenarios

Tracing Actor Invocations

The testing facilities described up to this point were aiming at formulating assertions about a system’s behavior. If a test fails, it is usually your job to find the cause, fix it and verify the test again. This process is supported by debuggers as well as logging, where the Akka toolkit offers the following options:

  • Logging of exceptions thrown within Actor instances

    This is always on; in contrast to the other logging mechanisms, this logs at ERROR level.

  • Logging of special messages

    Actors handle certain special messages automatically, e.g. Kill, PoisonPill, etc. Tracing of these message invocations is enabled by the setting akka.actor.debug.autoreceive, which enables this on all actors.

  • Logging of the actor lifecycle

    Actor creation, start, restart, monitor start, monitor stop and stop may be traced by enabling the setting akka.actor.debug.lifecycle; this, too, is enabled uniformly on all actors.

All these messages are logged at DEBUG level. To summarize, you can enable full logging of actor activities using this configuration fragment:

akka {
  loglevel = "DEBUG"
  actor {
    debug {
      autoreceive = on
      lifecycle = on
    }
  }
}

Configuration

There are several configuration properties for the TestKit module, please refer to the reference configuration.

Contents