Using Reference Maps for Caches and Listeners

A while ago I wrote a blog post about the WeakHashMap. It then turned out that the WeakHashMap was not the optimal choice for that particular use case and I proposed a different solution. To make this a bit more seizable, I decided to implement a post a full code example. Let me again describe the use case.

Let's say you have a class wrapping some sort of event. Let's give the event class a name, write a Java interface and call it Auditable. Each Auditable subclass must implement two methods: validate and process. There is a invoker class called AuditableInvoker which receives a collection of Auditable's and invokes validate and process on each one of them. So far so good.



As an example, let's implement two Auditable subclasses which are pretty stupid. SleepingAuditable will just hold the current Thread for a few milliseconds. IteratingAuditable will run a small loop in it's validate and process methods.



In addition to that, there is a requirement that you need to know the execution time of the validate and process method in each Auditable subclass. Fortunately you can add listeners to AuditableInvoker. So all you have to do is to write a listener that measures the execution times. The listener need to start a stop watch before validate or process is invoked and stop this very stop watch after process and validate are finished. Once they are finished, the execution time can be computed and kept in a helper class that we call the StatsCollector. To keep things simple, our UnboundedStatsCollector will only increment a counter, completely ignoring the execution times.




The tricky part here is that you need to use the same stop watch before and after the invocations of an Auditable. A good use case for a map using weak referenced keys and object identity for comparison. Once an Auditable subclass has finished it's lifecycle and is no longer referenced somewhere else in the code, the garbage collection can collect the Auditable as well as the associated stop watch. This will prevent the Map from growing indefinitely. So here is a implementation using a ReferenceIdentityMap from the commons-collections project.


To verify that we really see the expected behavior, I have written a unit test that is stressing the ExecutionTimingAuditableLifecycleListener using multiple Threads. In this unit test I am re-using a class called MultithreadedStressTester which I stole from Nat Pryze's book "Growing Object Oriented Software guided by Tests".


The ExecutionTimingAuditableLifecycleListenerTest uses the MultithreadedStressTester to send a bunch of Threads over to the ExecutionTimingAuditableLifecycleListener, verifying that each invocation is properly timed using the ReferenceIdentityMap under the hood.

Finally, if you want to use google-guava instead of commons-collections, you can also use a LoadingCache with weak keys instead of the ReferenceIdentityMap. Here is a version of the ExecutionTimingAuditableLifecycleListener using google-guava.

The Static Final Inline Trap

Last week I was chasing a very interesting problem which we could only resolve with the help of a colleague. One of our testers found a bug in the current game that we are developing. If the user doesn't have Flashplayer installed, the No-Flash image is not shown properly. All code related to this, is pulled in from shared libraries which our project depends on. In the screen shot you will see something that looks like magic. There is a method which populates and returns a Map. Amongst others, one key called noflash_image_url is defining the location of the No-Flash image. In the debugger pane to the lower right, you can see the static final constant NOFLASH_IMAGE_URL evaluating to the correct value. This value is the correct location of the No-Flash image.

The strange thing however, is that the Map contains another (wrong) value, even though the Map value is also set using the same static final constant NOFLASH_IMAGE_URL. You can see that in the lower middle pane of the screen shot. So for some strange reason, the constant is evaluating to both the correct and the wrong value. I guess we could call it semi-constant in this particular case.

Anytime you see something weird like this, your first thought should be class loading problem. Based on my experience, the weirdest problems are often rooted in class loading. However this problem is of a different kind. For a better understanding, you have to know that the Class which initializes the Map is a dependency that ours project pulls in from a Maven dependency to library A. This library A then depends on another (transient) Maven library B which contains the NOFLASH_IMAGE_URL constant. Our project also defines a direct Maven dependency to library B. This is needed because library B changes quite often and we always want the latest version of B in our project. The latest version of B does contain the correct value the No-Flash image in the NOFLASH_IMAGE_URL constant. So one might think, that the Maven dependency resolution mechanism pulled in an old version of library B, but this wasn't the case.

A colleague then hinted me in the right direction. The Java compiler often does something called inlining for access to a static final variable. This is an optimization as the value of a final variable cannot change after it has been assigned. To verify this we ran the Java class file disassembler (javap) over the class file in library A.

   78: pop
   79: aload_2
   80: ldc #108; //String noflash_image_url
   82: ldc #109; //String http://static.playfish.com/shared/noflash.jpg
   84: invokeinterface #100,  3; //InterfaceMethod java/util/Map.put:(Ljava/lang/Object;Ljava/lang/Object;)Ljava/lang/Object;


As you can see, the compiler is really inlining the value in places where we, by running a debugger over the source code, expect an evaluation of the static final variable at runtime. This is definitely something you have to be aware of. We can fix this, by re-compiling library A against a newer version of library B, which will inline the correct No-Flash image. Another option to prevent inlining would be to declare the constant like this:

  public static final String NOFLASH_IMAGE_URL = new String ("...");


But then this might be really hard to understand for other developers and they might revert this change to a String literal if not properly documented. Update: I just realized that this problem is featured as Puzzle 93: Class Warfare in the Java Puzzlers book from Joshua Bloch. The compiler will inline all constant variables, such as Primitives and Strings which are initialized with a constant expression. Surprisingly, null is not inlined, neither are Java 5 enums.

Lost Virginity: WeakHashMap first timer

It was almost three years ago at OOPSLA in Nashville, where I heard about the WeakHashMap for the first time. The class is quite useful if you need a Map implementation, where the keys are compared using their memory references and not using equals. Another important property of the WeakHashMap is that the Map entries are removed "automagically" if no other object (than the WeakHashMap itself) holds a reference to the key object. The garbage collection will in that case remove Map entry and collect the key.

In the past 3 years I never used the WeakHashMap in any project. That changed yesterday. In the game that we are currently developing, we are using a mechanism where the game client is sending game events to the server. The server will then evaluate the game events and alter the users in memory before their state is made persistent in the database. Here is an example:

public interface GameEvent {



    /**

     * Subclasses may implement this to run validation logic, before the GameEvent is processed.

     *

     * @param user the {@link User} to apply this {@link GameEvent} on

     * @return {@link AuditResult} never <code>null</code>

     */

    AuditResult validate(User user);



    /**

     * Subclasses may implement this to run implementation specific logic, potentially altering the

     * given {@link User}. 

     *

     * @param user the {@link User} to apply this {@link GameEvent} on

     * @return {@link AuditResult} never <code>null</code>

     */

    AuditResult process(User user);

}





public class ConsumeFood implements GameEvent {



    private final int amount;



    public ConsumeFood(final int amount) {

        this.amount = amount;

    }



    @Override

    public AuditResult validate(final User user) {

        if (user.getFood() < this.amount) {

            return AuditResult("User doesn't have this amount of food.");

        }

        return AuditResult.SUCCESS;

    }



    @Override

    public AuditResult process(final User user) {

        user.addEnergy(this.amount);

        user.subtractFood(this.amount);

        return AuditResult.SUCCESS;

    }

}


Once the game client sends us the ConsumeFood game event, we subtract food from the player and add energy instead. We also have a wrapper class around a collection of game events and the execution logic looks like this:

public class GameEvents {

    

    .. other methods ...



    protected AuditResult process(final User user, final GameEvent change) {

        final AuditResult validateResult = change.validate(user);

        if (validateResult == AuditResult.SUCCESS) {

            return change.process(user);

        } else {

            return validateResult;

        }

    }

}


First we validate that we can apply the game event, then we process the event and alter the player. Since the number of different game events is continuously growing and growing, I thought it might be useful to measure the execution time of the validate and the process method in each game event. The way I implemented this a while ago, was through delegation. I added a wrapper class which was wrapping the real game event and timing the validate and process method:

import org.springframework.util.StopWatch;



public final class TimingGameEvent implements GameEvent {

    private final GameEvent gameEvent;



    public TimingGameEvent(final GameEvent gameEvent) {

        this.gameEvent = gameEvent;

    }



    /**

     * Delegates the processing to the encapsulated {@link GameEvent}. Uses a {@link StopWatch} to time the

     * execution time.

     */

    @Override

    public AuditResult process(final User user) {

        final StopWatch stopWatch = new StopWatch("process-stop-watch");

        stopWatch.start();

        try {

            return this.gameEvent.process(user);

        } finally {

            stopWatch.stop();

            this.processTimeInMs = stopWatch.getLastTaskTimeMillis();

        }

    }



    /**

     * Delegates the validation to the encapsulated {@link GameEvent}. Uses a {@link StopWatch} to time the

     * execution time.

     */

    @Override

    public AuditResult validate(final User user) {

        final StopWatch stopWatch = new StopWatch("validate-stop-watch");

        stopWatch.start();

        try {

            return this.gameEvent.validate(user);

        } finally {

            stopWatch.stop();

            this.validationTimeInMs = stopWatch.getLastTaskTimeMillis();

        }

    }

}


This worked well. However, this week we got another requirement from business. We needed to implement some sort of gameplay recorder. Each game event that the server receives must be recorded, so we can replay these events later. My first idea was to add another wrapper around the already existing TimingGameEvent wrapper class but this would have made it difficult to serialize the real game event to a File. Yes we decided to serialize to and deserialize from a String, which is stored in a plain textfile and each line represents one game event. I discarded the idea to add other wrappers around the game event and suggested a refactoring. Instead of using delegating wrappers, why not use a listener mechanism. Each listener would be notified before and after execution of the validate and the process method in each game event. Listeners could register themselves and it would be easier to extend in the future. On the negative side, of course measuring the execution times would not be as accurate anymore, as there could be other listeners which want to be notified before the game event is validated and processed. This however was not a big issue, since we were not interested in the exact time in milliseconds but rather in long running methods of a couple of seconds. I also added a mechanism to make sure the execution timing listener would get notified just before the game event method was executed and right after it was returning. More on that later.

Here is the listener interface I came up with:

public interface GameEventLifecycleListener {



    void onValidationStart(final User user, final GameEvent gameEvent);



    void onValidationFinish(final User user, final GameEvent gameEvent, 

             final AuditResult auditResult);



    void onProcessStart(final User user, final GameEvent gameEvent);



    void onProcessFinish(final User user, final GameEvent gameEvent, 

             final AuditResult auditResult);

}


Refactoring the TimingGameEvent class from above to a TimingGameEventLifecycleListener wasn't straight forward. Each invocation of the validate or the process method will now result in two listener notifications. So how do you know when to “press” stop on the StopWatch?

This is where the WeakHashMap comes in handy. Remember that each game event is going through the same chain? First onValidationStart is called, then onValidationFinish, onProcessStart and finally onProcessFinish. So the Listener could maintain a Map of all event implemented using a WeakHashMap. The first notification callback will add the game event to this Map. Subsequent notifications can assume that the game event will be present in the WeakHashMap. After the game event has passed through the chain and no object is referencing the game event anymore, it will automatically be removed from the WeakHashMap. Here is a part of the TimingGameEventLifecycleListener which will show you the concept.

import org.springframework.core.Ordered;



public class TimingGameEventLifecycleListener extends AbstractGameEventLifecycleListener {



    /**

     * By default the WeakHashMap is not thread-safe, so it needs to be wrapped in a synchronizedMap. This however

     * is quite slow, hence the ExecutionTimingGameEventLifecycleListener should not be running in production

     * all the time.

     */

    private final Map<GameEvent, TimedExecution> timedExecutions = Collections.synchronizedMap(

        new WeakHashMap<GameEvent, TimedExecution>()

    );



    @Override

    public void onValidationStart(final User user, final GameEvent gameEvent) {

        final TimedExecution timeValidation = 

          new TimedExecution(gameEvent.getClass());

        this.timedExecutions.put(gameEvent, timeValidation);

        // other stuff

    }



    @Override

    public void onValidationFinish(final User user, final GameEvent gameEvent, 

          final AuditResult auditResult) {

        final TimedExecution timeValidation = this.timedExecutions.get(gameEvent);

        if (timeValidation != null) {

     timeValidation.stopTimer();

        }

    }



    ... other notification methods ...



    @Override

    public int getOrder() {

        return Ordered.HIGHEST_PRECEDENCE;

    }

}


So the WeakHashMap can be nice in the role of a cache between different Listener methods. Another thing you may notice in the code above, is that the Listener is deriving from AbstractGameEventLifecycleListener instead of implementing GameEventLifecycleListener. I added a abstract base class for two reasons. First it is better to provide empty default implementations of all notification methods. Concrete Listeners like the TimingGameEventLifecycleListener can then only overwrite the methods they are interested in (okay in this case we are interested in all four notification methods, but other Listeners might not be). The second reason is that we want to force the Listeners to be in a specific order. Every Listener can for himself decide "how important" he is by implementing the getOrder() method defined in the org.springframework.core.Ordered interface which the AbstractGameEventLifecycleListener is implementing. Normally this interface is used by Spring to apply an order to Aspects. Though you might choose to keep your domain clean of Spring framework classes. Here is the AbstractGameEventLifecycleListener:

public class GameEvents {



    private final GameEvent[] events;

    private final NavigableSet<AbstractGameEventLifecycleListener> listeners;



    public void addListeners(final

          Collection<AbstractGameEventLifecycleListener> listeners) {

        this.listeners.addAll(listeners);

    }



    public GameEvents(final GameEvent[] events) {

        final int length = events == null ? 0 : events.length;

        this.listeners = new TreeSet<AbstractGameEventLifecycleListener>();

        this.events = new GameEvent[length];

        if (length > 0) {

            System.arraycopy(events, 0, this.events, 0, length);

        }

    }



    // other methods



    protected AuditResult process(final User user, final GameEvent gameEvent) {

        final AuditResult validateResult = gameEvent.validate(user);

        if (validateResult == AuditResult.SUCCESS) {

            return gameEvent.process(user);

        } else {

            return validateResult;

        }

    } 





    /**

     * Runs the {@link GameEvent#validate(User)} function of the given 

     * {@code gameEvent}, notifying all {@link AbstractGameEventLifecycleListener}s

     * before and after. The listener having the highest 

     * precidence is notified last before and first after the validation method.

     * @param user the {@link User} to validate the game event for

     * @param gameEvent the gameEvent to validate

     * @return the result of the validation

     */

    AuditResult runValidate(final User user, final GameEvent gameEvent) {

        for (Iterator<AbstractGameEventLifecycleListener> 

            iterator = this.listeners.descendingIterator(); 

  iterator.hasNext(); ) {

            final AbstractGameEventLifecycleListener listener = iterator.next();

            listener.onValidationStart(user, gameEvent);

        }



        final AuditResult validateResult = gameEvent.validate(user);



        for (final AbstractGameEventLifecycleListener listener : this.listeners) {

            listener.onValidationFinish(user, gameEvent, validateResult);

        }



        return validateResult;

    }



    /**

     * Runs the {@link GameEvent#process(User)} function of the given 

     * {@code gameEvent}, notifying all {@link AbstractGameEventLifecycleListener}s

     * before and after. The listener having the highest 

     * precedence is notified last before and first after the validation method.

     * @param user the {@link User} to process the gameEvent for

     * @param gameEvent the audit gameEvent to process

     * @return the result of processing the gameEvent

     */

    AuditResult runProcess(final User user, final GameEvent gameEvent) {

        for (Iterator<AbstractGameEventLifecycleListener> iterator =

             this.listeners.descendingIterator(); 

  iterator.hasNext(); ) {

            final AbstractGameEventLifecycleListener listener = iterator.next();

            listener.onProcessStart(user, gameEvent);

        }



        final AuditResult validateResult = gameEvent.process(user);



        for (final AbstractGameEventLifecycleListener listener : this.listeners) {

            listener.onProcessFinish(user, gameEvent, validateResult);

        }



        return validateResult;

    }

}


I said earlier, it is desirable to notify the TimingGameEventLifecycleListener last before validation starts and first after it finishes (to get more accurate timings). The GameEvents class, which is notifying the listeners, will honor the order using a NavigableSet that can be iterated in forward and backward order. Take a look at the updated version of the GameEvents class to see how it is implemented:

public abstract class AbstractGameEventLifecycleListener 

 implements GameEventLifecycleListener, Ordered, Comparable<GameEventLifecycleListener> {



    @Override

    public void onValidationStart(final User user, final GameEvent gameEvent) { }



    @Override

    public void onValidationFinish(final User user, final GameEvent gameEvent, 

       final AuditResult auditResult) { }



    @Override

    public void onProcessStart(final User user, final GameEvent gameEvent) { }





    @Override

    public void onProcessFinish(final User user, final GameEvent gameEvent, 

       final AuditResult auditResult) { }



    /**

     * Compares the order of the two {@link GameEventLifecycleListener}s 

     * using {@link Ordered}.

     * @param other another {@link GameEventLifecycleListener}

     * @return int

     */

    @Override

    public int compareTo(final GameEventLifecycleListener other) {

        return Integer.valueOf(this.getOrder()).compareTo(other.getOrder());

    }

}


One thing I wasn't able to come up with, was a good unit test to verify that the WeakHashMap is indeed not holding key references forever. This is extremely difficult to test as it involves testing for garbage collection and no, I am not suggesting running System.gc() from your test. I found something similar on this blog post. Apparently the Netbeans API offers something called assertGC(..) but it wasn't really fitting for my use case. So if you have a good suggestion how to test the behavior of a WeakHashMap, I am happy to hear it.

* UPDATE * UPDATE * After a few weeks running this the WeakHashMap and seeing some weird errors in the logs every now and then, I realized it's not the right Map implementation to use. The WeakHashMap is not what you want to use here, because the keys are not really compared using object identity. Initially I thought this was the case, when reading through the Javadoc of the WeakHashMap. What you really want is a hybrid Map, that combines the WeakHashMap with a IdentityHashMap. This hybrid Map will compare the keys based on objects identity and also use weak key references. The bad news is, there is no such map in the JDK (Java 6 at least). The good news is, there is a WeakIdentityHashMap in the Hibernate Search project and a ReferenceIdentityMap in the Commons Collections Project which can be used.

Testing JMX between two Web Applications using Maven

The problem: you have two web applications and each is developed inside a separate Maven module. You need to communicate from one web application to the other and you don't want to implement a service but use JMX instead. This is a scenario we are having here at the moment. The first web application (application A) contains the game server logic of our new game. The second web application (application B) contains a debug tool which we will not deploy into production. I have selected JMX for the communication, mainly because I didn't wanted to add another technology and we are already using JMX in the first application. Both web application are Spring powered.

First here is a nice Spring feature, which completely hides the JMX complexity for the client application behind a proxy.


<bean id="clientConnector" class="org.springframework.jmx.support.MBeanServerConnectionFactoryBean">
<property name="serviceUrl" value="[SERVICE_URL]"/>
</bean>

<bean id="gameplayRecordable" class="org.springframework.jmx.access.MBeanProxyFactoryBean">
<property name="objectName" value="[MBEAN]" />
<property name="proxyInterface" value="any.java.Interface" />
<property name="server" ref="clientConnector" />
</bean>


First you define a client connector which connects you to the RMI server port of the other web application. Then you define a MBeanProxyFactoryBean using this client connector. The objectName must be the name of your MBean inside the MBean container. If you are not sure about the name, use jconsole to connect to the process of the first web application and look it up. Another important property of the MBeanProxyFactoryBean is the proxyInterface. This is an interface that the proxy will implement. The proxy will try to map each method call on that interface in application B to a JMX call in application A. I can really recommend to share the same Interface in both applications as it makes stuff really simple.

This was simple so far. Now lets say you want to write an integration test to automatically test the whole shebang. The test should start up a JMX enabled Jetty from Maven. This Jetty instance should explode the war file of application A (hosting the MBean you want to invoke). Once Jetty is up, the test should execute, connect to application A via the MBeanProxyFactoryBean and validate the results. First lets enable remote JMX access in the configuration of the maven-jetty-plugin:


<profiles>
<profile>
<id>itest</id>
<build>
<plugins>
<plugin>
<groupId>org.mortbay.jetty</groupId>
<artifactId>maven-jetty-plugin</artifactId>
<version>${version.jetty.plugin}</version>
<configuration>
<stopKey>stop_key</stopKey>
<stopPort>9999</stopPort>
<contextPath>/</contextPath>
<webApp>
${settings.localRepository}/com/package/../../your.war
</webApp>
<jettyConfig>
${basedir}/src/test/etc/jetty-jmx.xml</jettyConfig>
</configuration>
<executions>
<execution>
<id>start-jetty</id>
<phase>pre-integration-test</phase>
<goals>
<goal>deploy-war</goal>
</goals>
<configuration>
<daemon>true</daemon>
</configuration>
</execution>
<execution>
<id>stop-jetty</id>
<phase>prepare-package</phase>
<goals>
<goal>stop</goal>
</goals>
</execution>
</executions>
</plugin>
</plugins>
</build>
</profile>
</profiles>


As you can see, this plugin configuration is done in a Maven profile as we are defining this for application B which also has its own Jetty configuration. The important piece is the jettyConfig element which points to a jetty-jmx.xml file. To get this file, download the Jetty container that has the same version as your maven-jetty-plugin. For instance is you use version 6.1.26 of the maven-jetty-plugin, make sure you download jetty-6.1.26 from the codehaus download page. If you are using the the new jetty-maven-plugin and Jetty 7 or 8, you need to download the Jetty container from Eclipse. The configuration is the same for the maven-jetty-plugin and the jetty-maven-plugin. Just make sure you download the jetty-jmx.xml file from the right Jetty container as they are different. You don't need to specify any additional system properties like -Dcom.sun.management.jmxremote, -Dcom.sun.management.jmxremote.ssl, -Dcom.sun.management.jmxremote.authenticate or -Dcom.sun.management.jmxremote.port.

Once the jetty-jmx.xml is downloaded, put in somewhere inside your Maven module where it does not get packaged into the module artifact. In the example above you can see that we have the jetty-jmx.xml file in src/test/etc but any other location will do. Open the file and enable remote JMX access via RMI. In the Jetty 6 based jetty-jmx.xml file these elements should be commented in:


<Call id="rmiRegistry" class="java.rmi.registry.LocateRegistry" name="createRegistry">
<Arg type="int">2099</Arg>
</Call>

<Call id="jmxConnectorServer" class="javax.management.remote.JMXConnectorServerFactory"
name="newJMXConnectorServer">
<Arg>
<New class="javax.management.remote.JMXServiceURL">
<Arg>
service:jmx:rmi://localhost:17264/jndi/rmi://localhost:2099/jmxrmi
</Arg>
</New>
</Arg>
<Arg/>
<Arg>
<Ref id="MBeanServer"/>
</Arg>
<Call name="start"/>
</Call>


Note that we changed the port to 17264. You might want to use the default port instead. In the Jetty 7 based jetty-jmx.xml file these elements should be commented in:


<Call name="createRegistry" class="java.rmi.registry.LocateRegistry">
<Arg type="java.lang.Integer">1099</Arg>
<Call name="sleep" class="java.lang.Thread">
<Arg type="java.lang.Integer">1000</Arg>
</Call>
</Call>

<New id="ConnectorServer" class="org.eclipse.jetty.jmx.ConnectorServer">
<Arg>
<New class="javax.management.remote.JMXServiceURL">
<Arg type="java.lang.String">rmi</Arg>
<Arg type="java.lang.String" />
<Arg type="java.lang.Integer">0</Arg>
<Arg type="java.lang.String">/jndi/rmi://localhost:1099/jettyjmx</Arg>
</New>
</Arg>
<Arg>org.eclipse.jetty:name=rmiconnectorserver</Arg>
<Call name="start" />
</New>


To test the setup, run mvn -Pitest jetty:run and start jconsole. In jconsole you do not connect to a local process. Select Remote Process and enter the service URL. This URL can be copied from the jetty-jmx.xml file if you are using Jetty 6 (i.e. service:jmx:rmi://localhost:17264/jndi/rmi://localhost:2099/jmxrmi). If you are using Jetty 7 and the jetty-maven-plugin, there will be a info statement on the command line when Maven starts the Jetty container from where you can copy the service URL. Finally to execute the integration test, we use the maven-failsafe-plugin like this:


<plugin>
<groupId>org.apache.maven.plugins</groupId>
<artifactId>maven-failsafe-plugin</artifactId>
<version>2.9</version>
<configuration>
<includes>
<include>**/com/package/integration/*.java</include>
</includes>
</configuration>
<executions>
<execution>
<id>integration-test</id>
<goals>
<goal>integration-test</goal>
</goals>
</execution>
<execution>
<id>verify</id> to specify any additional system properties like -Dcom.sun.management.jmxremote, -Dcom.sun.management.jmxremote.ssl, -Dcom.sun.management.jmxremote.authenticate or -Dcom.sun.management.jmxremote.port.

Once the jetty-jmx.xml is downloaded, put in somewhere inside your Maven module where it does not get packaged into the module artifact. In the example above you can see that we have the jetty-jmx.xml file in src/test/etc but any other location will do. Open the file and enable remote JMX access via RMI. In the Jetty 6 based jetty-jmx.xml file these elements should be commented in:


<Call id="rmiRegistry" class="java.rmi.registry.LocateRegistry" name="createRegistry">
<Arg type="int">2099</Arg>
</Call>

<Call id="jmxConnectorServer" class="javax.management.remote.JMXConnectorServerFactory"
name="newJMXConnectorServer">
<Arg>
<New class="javax.management.remote.JMXServiceURL">
<Arg>service:jmx:rmi://localhost:17264/jndi/rmi://localhost:2099/jmxrmi</Arg>
</New>
</Arg>
<Arg/>
<Arg>
<Ref id="MBeanServer"/>
</Arg>
<Call name="start"/>
</Call>


Note that we changed the port to 17264. You might want to use the default port instead. In the Jetty 7 based jetty-jmx.xml file these elements should be commented in:


<Call name="createRegistry" class="java.rmi.registry.LocateRegistry">
<Arg type="java.lang.Integer">1099</Arg>
<Call name="sleep" class="java.lang.Thread">
<Arg type="java.lang.Integer">1000</Arg>
</Call>
</Call>

<New id="ConnectorServer" class="org.eclipse.jetty.jmx.ConnectorServer">
<Arg>
<New class="javax.management.remote.JMXServiceURL">
<Arg type="java.lang.String">rmi</Arg>
<Arg type="java.lang.String" />
<Arg type="java.lang.Integer">0</Arg>
<Arg type="java.lang.String">/jndi/rmi://localhost:1099/jettyjmx</Arg>
</New>
</Arg>
<Arg>org.eclipse.jetty:name=rmiconnectorserver</Arg>
<Call name="start" />
</New>


To test the setup, run mvn -Pitest jetty:run and start jconsole. In jconsole you do not connect to a local process. Select Remote Process and enter the service URL. This URL can be copied from the jetty-jmx.xml file if you are using Jetty 6 (i.e. service:jmx:rmi://localhost:17264/jndi/rmi://localhost:2099/jmxrmi). If you are using Jetty 7 and the jetty-maven-plugin, there will be a info statement on the command line when Maven starts the Jetty container from where you can copy the service URL. Finally to execute the integration test, we use the maven-failsafe-plugin like this:


<plugin>
<groupId>org.apache.maven.plugins</groupId>
<artifactId>maven-failsafe-plugin</artifactId>
<version>2.9</version>
<configuration>
<includes>
<include>**/com/package/integration/*.java</include>
</includes>
</configuration>
<executions>
<execution>
<id>integration-test</id>
<goals>
<goal>integration-test</goal>
</goals>
</execution>
<execution>
<id>verify</id>
<goals>
<goal>verify</goal>
</goals>
</execution>
</executions>
</plugin>

<goals>
<goal>verify</goal>
</goals>
</execution>
</executions>
</plugin>

Sharing configuration files from a Maven Parent Project

Okay this post is probably not much news for people who know Maven in and out. I am planning to use this as a reference for myself, in the case that I have to solve a similar problem again in the future. The current project I am working with, is set up as a Maven multi module project. There is a parent pom which is set to pom-packaging. There are several child modules, set to either jar- or war-packaging. Within the pom.xml file of the parent project, we use the pluginManagement section to define plugins that should be available to the child modules. The pluginManagement mechanism is an excellent way to avoid the DRY problem and not to duplicate Maven configuration within the inheriting projects.

In most cases configuring plugins within the pluginManagement section is straight forward. It can however get a bit problematic if the plugin depends on (or is reading from external) configuration files. Lets have a look at one example from this parent project of ours.


<project xmlns="http://maven.apache.org/POM/4.0.0"
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
xsi:schemaLocation="http://maven.apache.org/POM/4.0.0
http://maven.apache.org/maven-v4_0_0.xsd">
<modelVersion>4.0.0</modelVersion>

<groupId>com.package</groupId>
<artifactId>project-parent</artifactId>
<packaging>pom</packaging>
<version>0.1-SNAPSHOT</version>

<modules>
<module>child-a</module>
<module>child-b</module>
<module>child-c</module>
</modules>

<properties>
<version.mysql.connector>5.1.12</version.mysql.connector>
</properties>

<build>

<pluginManagement>
<plugins>
<plugin>
<groupId>org.codehaus.mojo</groupId>
<artifactId>sql-maven-plugin</artifactId>
<version>1.4</version>
<dependencies>
<dependency>
<groupId>mysql</groupId>
<artifactId>mysql-connector-java</artifactId>
<version>${version.mysql.connector}</version>
</dependency>
</dependencies>
<configuration>
<driver>com.mysql.jdbc.Driver</driver>
<url>jdbc:mysql://localhost/</url>
<username>xyz</username>
<password>xyz</password>
</configuration>

<executions>
<execution>
<id>drop-and-recreate-db</id>
<phase>process-test-resources</phase>
<goals>
<goal>execute</goal>
</goals>
<configuration>
<autocommit>true</autocommit>
<srcFiles>
<srcFile>
${project.build.directory}/sql/schema/user.sql
</srcFile>
<srcFile>
${project.build.directory}/sql/schema/core.sql
</srcFile>
<srcFile>
${project.build.directory}/sql/schema/game.sql
</srcFile>
</srcFiles>
<onError>abort</onError>
</configuration>
</execution>
</executions>
</plugin>
</plugins>
</pluginManagement>

</build>
</project>


Here we use the sql-maven-plugin to setup the database before tests are being run. The sql-maven-plugin will execute a bunch of *.sql files which are stored in a subfolder in the parent project. As you deploy the parent project to your repository, these files won't be published along with the pom.xml as the packaging is set to pom-packaging. Therefore, if you run the inherited sql-maven-plugin the *.sql files will not be available and the plugin will fail. This will certainly be a problem if your continuous integration server has a build plan for each Maven child module and not a single build plan for the entire project.

To overcome this problem, there are 2 things you have to do. First, the parent project needs to publish the *.sql files (or other static files which are needed) to your repository, so that the inheriting modules have access to these files. For this to work, we use the maven-assembly-plugin like this:


<project xmlns="http://maven.apache.org/POM/4.0.0"
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
xsi:schemaLocation="http://maven.apache.org/POM/4.0.0
http://maven.apache.org/maven-v4_0_0.xsd">
<modelVersion>4.0.0</modelVersion>

... as before ...

<build>
<plugins>
<plugin>
<artifactId>maven-assembly-plugin</artifactId>
<inherited>false</inherited>
<configuration>
<descriptors>
<descriptor>
${project.basedir}/assembly/zip.xml
</descriptor>
</descriptors>
</configuration>
<executions>
<execution>
<id>make-assembly</id>
<phase>package</phase>
<goals>
<goal>single</goal>
</goals>
</execution>
</executions>
</plugin>
</plugins>

<pluginManagement>
... as before ...
</pluginManagement>

</build>
</project>


Note that the maven-assembly-plugin in this case is not configured within the pluginManagement section of the parent pom, as we don't want to make this functionality available to child modules. In the configuration you can see that the plugin is set up to be executed during the package phase and that the plugin configuration is defined in the zip.xml file. This zip.xml file looks like this:


<assembly xmlns="http://maven.apache.org/plugins/maven-assembly-plugin/assembly/1.1.0"
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
xsi:schemaLocation="http://maven.apache.org/plugins/maven-assembly-plugin/assembly/1.1.0 http://maven.apache.org/xsd/assembly-1.1.0.xsd">
<id>sql-files</id>
<formats>
<format>zip</format>
</formats>
<includeBaseDirectory>false</includeBaseDirectory>
<fileSets>
<fileSet>
<directory>${project.basedir}/sql/schema</directory>
<outputDirectory/>
<includes>
<include>**/*</include>
</includes>
</fileSet>
</fileSets>
</assembly>


This configuration will create a zip-file of all *.sql files found in ${project.basedir}/sql/schema and publish this zip-file along with the pom.xml when mvn deploy is executed. The id of this configuration is "sql-files". This id will be used as a suffix and become part of the filename of the zip-file.

Now that we publish the zip-file to the repository, we need a way for the child modules to grab and extract the zip-file before the maven-sql-plugin is executed. This is where the maven-dependency-plugin comes in handy. Again, the maven-dependency-plugin is configured in the pluginManagement section of the parent pom.xml, as this time we want to inherit the functionality to child modules. Here is what the configuration of the maven-dependency-plugin looks like:


<project xmlns="http://maven.apache.org/POM/4.0.0"
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
xsi:schemaLocation="http://maven.apache.org/POM/4.0.0
http://maven.apache.org/maven-v4_0_0.xsd">
<modelVersion>4.0.0</modelVersion>

... as before ...

<build>
<plugins>
... as before ...
</plugins>

<pluginManagement>
<plugin>
<groupId>org.apache.maven.plugins</groupId>
<artifactId>maven-dependency-plugin</artifactId>
<version>2.3</version>
<executions>
<execution>
<id>unpack-sql-files</id>
<phase>process-test-resources</phase>
<goals>
<goal>unpack</goal>
</goals>
<configuration>
<artifactItems>
<artifactItem>
<groupId>com.package</groupId>
<artifactId>project-parent</artifactId>
<version>
${parent.version}
</version>
<type>zip</type>
<classifier>sql-files</classifier>
<overWrite>true</overWrite>
<outputDirectory>
${project.build.directory}/sql/schema
</outputDirectory>
<includes>**/*.sql</includes>
</artifactItem>
</artifactItems>
<includes>**/*</includes>
<overWriteReleases>true</overWriteReleases>
<overWriteSnapshots>true</overWriteSnapshots>
</configuration>
</execution>
</executions>
</plugin>

... as before ...

</pluginManagement>

</build>
</project>


The plugin (if a child module decides to use this) will be executed during the process-test-resources phase of the build. We are getting the zip file by specifying the groupId, artifactId, version and the type. Also the classifier value must match the id which we used earlier in the zip.xml file for configuring the maven-assembly-plugin. The zip-file is extracted to ${project.build.directory}/sql/schema and we are only extracting files having the *.sql extension (well there shouldn't be any other files but ok). This concludes what needs to be done to extract the zip-file and child modules are now ready to use the extracted files. Here is a snippet from a pom.xml file in a Maven child module. This is everything needed to run the sql-maven-plugin defined in the parent pom and extract the required configuration files upfront.


<project xmlns="http://maven.apache.org/POM/4.0.0" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
xsi:schemaLocation="http://maven.apache.org/POM/4.0.0 http://maven.apache.org/maven-v4_0_0.xsd">

<modelVersion>4.0.0</modelVersion>
<parent>
<groupId>com.package</groupId>
<artifactId>project-parent</artifactId>
<version>0.1-SNAPSHOT</version>
</parent>

<artifactId>child-a</artifactId>
<packaging>war</packaging>

<build>
<plugins>
<plugin>
<groupId>org.apache.maven.plugins</groupId>
<artifactId>maven-dependency-plugin</artifactId>
</plugin>

<plugin>
<groupId>org.codehaus.mojo</groupId>
<artifactId>sql-maven-plugin</artifactId>
</plugin>
</plugins>
</build>
</project>


For the sake of completeness, once again the full parent pom.xml file.


<project xmlns="http://maven.apache.org/POM/4.0.0"
xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
xsi:schemaLocation="http://maven.apache.org/POM/4.0.0
http://maven.apache.org/maven-v4_0_0.xsd">
<modelVersion>4.0.0</modelVersion>

<groupId>com.package</groupId>
<artifactId>project-parent</artifactId>
<packaging>pom</packaging>
<version>0.1-SNAPSHOT</version>

<modules>
<module>child-a</module>
<module>child-b</module>
<module>child-c</module>
</modules>

<properties>
<version.mysql.connector>5.1.12</version.mysql.connector>
</properties>

<build>
<plugins>
<plugin>
<artifactId>maven-assembly-plugin</artifactId>
<inherited>false</inherited>
<configuration>
<descriptors>
<descriptor>
${project.basedir}/assembly/zip.xml
</descriptor>
</descriptors>
</configuration>
<executions>
<execution>
<id>make-assembly</id>
<phase>package</phase>
<goals>
<goal>single</goal>
</goals>
</execution>
</executions>
</plugin>
</plugins>

<pluginManagement>
<plugin>
<groupId>org.apache.maven.plugins</groupId>
<artifactId>maven-dependency-plugin</artifactId>
<version>2.3</version>
<executions>
<execution>
<id>unpack-sql-files</id>
<phase>process-test-resources</phase>
<goals>
<goal>unpack</goal>
</goals>
<configuration>
<artifactItems>
<artifactItem>
<groupId>com.package</groupId>
<artifactId>project-parent</artifactId>
<version>
${parent.version}
</version>
<type>zip</type>
<classifier>sql-files</classifier>
<overWrite>true</overWrite>
<outputDirectory>
${project.build.directory}/sql/schema
</outputDirectory>
<includes>**/*.sql</includes>
</artifactItem>
</artifactItems>
<includes>**/*</includes>
<overWriteReleases>true</overWriteReleases>
<overWriteSnapshots>true</overWriteSnapshots>
</configuration>
</execution>
</executions>
</plugin>

<plugin>
<groupId>org.codehaus.mojo</groupId>
<artifactId>sql-maven-plugin</artifactId>
<version>1.4</version>
<dependencies>
<dependency>
<groupId>mysql</groupId>
<artifactId>mysql-connector-java</artifactId>
<version>${version.mysql.connector}</version>
</dependency>
</dependencies>
<configuration>
<driver>com.mysql.jdbc.Driver</driver>
<url>jdbc:mysql://localhost/</url>
<username>xyz</username>
<password>xyz</password>
</configuration>

<executions>
<execution>
<id>drop-and-recreate-db</id>
<phase>process-test-resources</phase>
<goals>
<goal>execute</goal>
</goals>
<configuration>
<autocommit>true</autocommit>
<srcFiles>
<srcFile>
${project.build.directory}/sql/schema/user.sql
</srcFile>
<srcFile>
${project.build.directory}/sql/schema/core.sql
</srcFile>
<srcFile>
${project.build.directory}/sql/schema/game.sql
</srcFile>
</srcFiles>
<onError>abort</onError>
</configuration>
</execution>
</executions>
</plugin>
</pluginManagement>

</build>
</project>

Heating up the Code Generator

It has been a couple of years since I used a code generator in one of my projects. It must have been 2004 or 2005, when MDA was a big buzzword. Back then, we used a self-made code generator, which was written by a very smart developer who I used to work with at Aperto AG in Berlin. He even contributed the code generator to the open source community later on. Those were the days.

Since then, there wasn't much use for a code generator anymore. IDE's getting better and better at helping you with code generation and auto-completion etc. And of course, adding a generator to your build process is always a bit of work. So it is often faster to write the boilerplate code yourself, if that doesn't takes forever. However, last week it was time to bring code generation back from the grave. The game we are currently developing for EA here in Norway, has a mechanism where the game client sends game events to the server. In our domain we call these events also audit changes. A typical audit change can be that the player has found a treasure, that the player has consumed food or that the player has discovered a new scenery. On the client side, the audit change is implemented in ActionScript 3. On the server side the audit change is implemented in Java. There is a transport layer in between which serializes the AS3 object, sends it over the network and deserializes it back into a Java object. For us server developers, this meant that every new audit change also needed a transport definition about how to serialize and deserialize the audit change. This definition was always wrapped into a new audit change type class. The type definition class was written manually, which sort of was okay until we had more than 20 audit changes in the game. Thats when I started to look into generating the transport layer on the server side.

In Java 5 along with the new Annotation language feature, Sun added a command-line utility called Annotation Processing Tool (apt). This was later merged into the standard javac compiler with the release of Java 6. There is also the apt-jelly project which provides an interface to apt and can be used to generate code artifacts based on templates written with Freemarker or Jelly. Finally to glue everything together, there is the maven-apt-plugin which can be used to execute a AnnotationProcessorFactory during your build and therefore integrate apt into your project. The maven-apt-plugin looks sort of dead however. I think nowadays even the standard maven-compiler-plugin or the maven-annotation-plugin can be used to process your annotations and generate code artifacts. Since I got our generator working using the maven-apt-plugin, I did not bother looking at the other 2 plugins. If someone has a working example on how they are used with a AnnotationProcessorFactory, I would be really happy to see it. Let's look into some code now.

Here is an example of an audit change as it could exist in the game:



/**
* Audit change telling that the {@link User} has bought a {@link House}.
*/
@DatatypeDefinition(minSize = 7)
public class BoughtHouseForGold implements AuditChange {

private int itemId;

@DatatypeCollection(elementType = Integer.class)
private List<Integer> sceneries;

@DatatypeIgnore
private User friend;

... other stuff not relevant ...
}


The transport class that we want to generate need to serialize and deserialize every field of the audit change. As you can guess from above, we do not want to transport the friend field in the example. The meta data, that isn't accessible via reflection within the template which we will write later, needs to be given to the generator in another way - for instance via Annotations. Thats why I created a bunch of Annotations just to instruct the generator.


/**
* Marks the type annotated by this annotation as something that can be
* serialized and deserialized using a Datatype.
*/
@Retention(RetentionPolicy.SOURCE)
@Target({ElementType.TYPE})
public @interface DatatypeDefinition {
int minSize() default 0;
}


/**
* Any {@link Field} annotated this way, will be rendered as a Collection of the
* specified type when the Datatype is generated.
*/
@Retention(RetentionPolicy.SOURCE)
@Target({ElementType.FIELD})
public @interface DatatypeCollection {
Class<?> elementType();
}


/**
* Any {@link Field} annotated this way, will be ignored in when the Datatype is generated.
*/
@Retention(RetentionPolicy.SOURCE)
@Target({ElementType.FIELD})
public @interface DatatypeIgnore {
}


Now for the hardest part, the template. I can recommend, that you start writing the code for the first class (the one that should be generated later) manually before you work on the template. Add the class into the default location in Maven, i.e. src/main/java/com/whatever/package. The generated classes will end up in a different location later (under target/generated-sources/) so it will be easy to compare the expected outcome and the generated outcome while working on the template. Here is a template example in which I use Freemarker directives.


<#-- for each type annotated with DatatypeDefinition -->
<@forAllTypes var="type" annotationVar="datatypeDefinition" annotation="package.DatatypeDefinition">
<#-- tell apt-jelly that the outcome will be a java source artifact -->
<@javaSource name="package.types.${type.simpleName}Type">
package package.types;

<#-- all imports go here -->
import java.io.IOException;

/**
* This class contains the {@link Datatype} for {@link ${type.simpleName}}.
*/
public class ${type.simpleName}Type extends AbstractAuditableType<${type.simpleName}> { <#-- class name based on type that was annotated with DatatypeDefinition -->
public ${type.simpleName}Type() {
super(
<#-- replace camel case with underscores -->
TypeCodes.${type.simpleName?replace("(?<=[a-z0-9])[A-Z]|(?<=[a-zA-Z])[0-9]|(?<=[A-Z])[A-Z](?=[a-z])", "_$0", 'r')?upper_case}_TYPE_CODE,
new Datatype<${type.simpleName}>(${type.simpleName}.class, ${datatypeDefinition.minSize}) {

@Override
public ${type.simpleName} read(final DatatypeInput in) throws DataFormatException {
final ${type.simpleName} value = new ${type.simpleName}();
<@forAllFields var="field">
<#assign useField = true>
<#-- do not do anything if field is a constant -->
<#if field.static = true><#assign useField = false></#if>
<#-- do not do anything if annotated with @DatatypeIgnore -->
<@ifHasAnnotation declaration=field annotation="package.DatatypeIgnore"><#assign useField = false></@ifHasAnnotation>
<#if useField = true>
<#-- build name of the setter method -->
<#assign setter = "set${field?cap_first}">
<#assign useCollection = false>
<@ifHasAnnotation var="datatypeCollectionAnnotation" declaration=field annotation="package.DatatypeCollection"><#assign useCollection = true></@ifHasAnnotation>
<#if useCollection = true>
<#if datatypeCollectionAnnotation.elementType = "java.lang.Integer">
value.${setter}(in.readList(Datatype.uintvar31));
<#else>
System.out.println('Cannot read collections of type: ${datatypeCollectionAnnotation.elementType}. Extend auditable-type.fmt');
</#if>
<#else>
<#-- Handling for fields without extra annotations -->
<#if field.type = "int" || field.type = "java.lang.Integer">
value.${setter}(in.readUintvar31());
<#elseif field.type = "boolean" || field.type = "java.lang.Boolean">
value.${setter}(in.readBoolean());
<#elseif field.type = "java.lang.String">
value.${setter}(in.readString());
</#if>
</#if>
<#assign useCollection = false>
</#if>
<#assign useField = false>
</@forAllFields>
return value;
}

@Override
public void write(final DatatypeOutput out, final ${type.simpleName} value) throws IOException {
<@forAllFields var="field">
<#assign useField = true>
<#-- do not do anything if field is a constant -->
<#if field.static = true><#assign useField = false></#if>
<#-- do not do anything if annotated with @DatatypeIgnore -->
<@ifHasAnnotation declaration=field annotation="DatatypeIgnore"><#assign useField = false></@ifHasAnnotation>
<#if useField>
<#-- build name of the getter method -->
<#assign getter = "get${field?cap_first}">
<#if field.type = "boolean" || field.type = "java.lang.Boolean">
<#-- boolean getter starts with is -->
<#assign getter = "is${field?cap_first}">
</#if>
<#assign useCollection = false>
<@ifHasAnnotation var="datatypeCollectionAnnotation" declaration=field annotation="DatatypeCollection"><#assign useCollection = true></@ifHasAnnotation>
<#if useCollection = true>
<#if datatypeCollectionAnnotation.elementType = "java.lang.Integer">
out.writeCollection(Datatype.uintvar31, value.${getter}());
<#else>
System.out.println('Cannot write collections of type: ${datatypeCollectionAnnotation.elementType}. Extend auditable-type.fmt');
</#if>
<#else>
<#if field.type = "int" || field.type = "java.lang.Integer">
out.writeUintvar31(value.${getter}());
<#elseif field.type = "boolean" || field.type = "java.lang.Boolean">
out.writeBoolean(value.${getter}());
<#elseif field.type = "java.lang.String">
out.writeString(value.${getter}());
</#if>
</#if>
<#assign useCollection = false>
</#if>
<#assign useField = false>
</@forAllFields>
}
}
);
}
}
</@javaSource>
</@forAllTypes>


My apologies, this is incredibly hard to read here on the blog. It helps to click the "view source" button in the upper right corner of the code above and copy everything to a text editor. I also added comments in the template to explain what I am doing.

Finally here is the configuration for the maven-apt-plugin, so that it will generate your code artifacts before compiling your project (note that target/generated-sources will be merged with the real sources during compile time).


<plugin>
<groupId>org.codehaus.mojo</groupId>
<artifactId>apt-maven-plugin</artifactId>
<version>1.0-alpha-4</version>
<configuration>
<factory>net.sf.jelly.apt.freemarker.FreemarkerProcessorFactory</factory>
<options>
<option>template=${basedir}/src/main/resources/apt/auditable-type.fmt
</option>
</options>
<fork>true</fork>
</configuration>
<dependencies>
<dependency>
<groupId>net.sf.apt-jelly</groupId>
<artifactId>apt-jelly-core</artifactId>
<version>2.14</version>
</dependency>
<dependency>
<groupId>net.sf.apt-jelly</groupId>
<artifactId>apt-jelly-freemarker</artifactId>
<version>2.14</version>
</dependency>
</dependencies>
<executions>
<execution>
<goals>
<goal>process</goal>
</goals>
</execution>
</executions>
</plugin>


And voilà, here is our generated BoughtHouseForGoldType class fresh out of the oven:


package package.types;

import java.io.IOException;

/**
* This class contains the {@link Datatype} for {@link BoughtHouseForGold}.
*/
public class BoughtHouseForGoldType extends AbstractAuditableType<BoughtHouseForGold> {
public BoughtHouseForGoldType() {
super(
TypeCodes.BOUGHT_HOUSE_FOR_GOLD_TYPE_CODE,
new Datatype<BoughtHouseForGold>(BoughtHouseForGold.class, 7) {

@Override
public BoughtHouseForGold read(final DatatypeInput in) throws DataFormatException {
final BoughtHouseForGold value = new BoughtHouseForGold();
value.setItemId(in.readUintvar31());
value.setSceneries(in.readList(Datatype.uintvar31));
return value;
}

@Override
public void write(final DatatypeOutput out, final BoughtHouseForGold value) throws IOException {
out.writeUintvar31(value.getItemId());
out.writeCollection(Datatype.uintvar31, value.getSceneries());
}
}
);
}
}

The Beauties and Pitfalls of ThreadLocals

ThreadLocal must be one of the Java features, which not many developers have used. It doesn't happen so often, that you stand in front of a problem, for which ThreadLocal is the best solution. In the current game I am developing, I have however decided to use a ThreadLocal to store per-thread slash per-request context information. Actually this is one of the use cases that Brian Goetz wrote about in his great article about ThreadLocals.

Other applications for ThreadLocal include storing or accumulating per-thread context information for later retrieval.


This game uses the Spring Framework and the Servlet which is handling all requests, is a Spring Bean. When the request comes in, the Servlet will delegate handling of the request data to something that we call a RpcHandler. Once the appropriate RpcHandler is selected by the Servlet, the request has to go through 3 stages and each stage is mapped to a method in the RpcHandler.


// stage 1
public void readRequest(final DatatypeInput in) {

}

// stage 2
public void performOperation() {

}

// stage 3
public void writeResponse(final DatatypeOutput out) {

}


First the request data is read, then the RpcHandler is performing it's main operation and finally some response data might be written. In other games developers have decided to create a new instance of the RpcHandler subclass for each request. This makes it easy as the per-request context information can be stored in private fields of the RpcHandler. The negative part is that this approach is using more memory as there will be 1 instance of RpcHandler per request. This might sound acceptable if the RpcHandler was simple and lightweight but often they contain further object-trees. Because our game was Spring based, I was striving for a solution were we only had 1 instance per RpcHandler sublass. Each instance is injected as a Spring Bean into the Servlet. Since the RpcHandler interface could not be changed, so that per-request context information was given as a parameter to the 3 methods above, I needed to use a ThreadLocal.

First I create a wrapper class for the context information called Message. The RpcHandler would then contain a private field of type ThreadLocal encapsulating a Message on a per-request basis. Even though ServletContainers like Jetty or Tomcat use a thread pool were Threads are recycled, this would still be okay as one Thread is used to service an entire request (actually this might be different in Servlet 3.0 where requests can be suspended and I'am assuming Threads are given back to the pool?). Quoting Brian Goetz:

ThreadLocal is also useful in servlet-based applications or any multithreaded server application in which the unit of work is an entire request, because then a single thread will be used during the entire course of handling the request.


Since usually each Spring Bean does only exist as 1 instance per ApplicationContext (unless specified otherwise), I decided to have a non-static ThreadLocal field which is sort of different from what the JavaDoc class comment suggests:

ThreadLocal instances are typically private static fields in classes that wish to associate state with a thread.


Here is an example of a typical RpcHandler subclass in the game:


public class SaveChangesHandler extends AbstractRpcHandler {

@Override
public void readRequest(final DatatypeInput in) {
// extract Message from DatatypeInput
setMessage(...);
}

@Override
public void performOperation() throws DAOException, RpcFailureException {
final Message message = getMessage();
// use Message to update the User
userService.save(message.getUser(), ...);
}

@Override
public void writeResponse(final DatatypeOutput out) throws IOException {
final Message message = getMessage();
// use Message to write something back to repsonse
out.write(message.getUser(), ...)
}

}


The AbstractRpcHandler class looks like this:


public abstract class AbstractRpcHandler extends RpcHandler<GameDaoSupport> {

private ThreadLocal<Message> threadLocalMessage;

public Message getMessage() {
return threadLocalMessage.get();
}

public void setMessage(final Message message) {
this.threadLocalMessage = new ThreadLocal<Message>();
this.threadLocalMessage.set(message);

// add a cleanup callback which will be
// executed from the ServletRequestCleanupFilter
ServletRequestCleanupService.addCleanupCallback(
new ServletRequestThreadLocalCleanupCallback(
this.threadLocalMessage
)
);
}
}


This worked well for some time but sometimes I could see the following errors in the logs:


2011-07-26 13:50:50 ERROR c.p.s.RpcServlet:handleRequest threw exception:
java.lang.NullPointerException: null
at com.foo.save.SaveChangesHandler.writeResponse(SaveChangesHandler.java:21)


Apparently the Message in writeResponse(..) was null, which was a bit strange because it was not null when performOperation(..) was invoked by the RpcServlet in stage 2. It took a bit but I found at least one problem with the code in AbstractRpcHandler:


public void setMessage(final Message message) {
this.threadLocalMessage = new ThreadLocal<Message>();
this.threadLocalMessage.set(message);

// add a cleanup callback which will be
// executed from the ServletRequestCleanupFilter
ServletRequestCleanupService.addCleanupCallback(
new ServletRequestThreadLocalCleanupCallback(
this.threadLocalMessage
)
);
}


The intention was to update the thread-local Message for the current Thread. Instead the code was reassigning the shared field to a new ThreadLocal instance and the previously assigned ThreadLocal instance was probably garbage collected. Here is a visualization of the problem flow:


Thread1 → setMessage(m1)
assigning new ThreadLocal1
ThreadLocal1.set(m1)
Thread1 → getMessage
ThreadLocal1.get → m1
Thread2 → setMessage(m2)
assigning new ThreadLocal2
ThreadLocal2.set(m2)
Thread1 → getMessage
ThreadLocal2.get → null *bang*


Took a while to spot the problem. I can really recommend to mark you static or non-static ThreadLocals as final, so this problem could not have occurred. I changed the AbstractRpcHandler to the following code and the errors disappeared.


public abstract class AbstractRpcHandler extends RpcHandler<GameDaoSupport> {

private final ThreadLocal<Message> threadLocalMessage = new ThreadLocal<Message>();

public Message getMessage() {
return threadLocalMessage.get();
}

public void setMessage(final Message message) {
this.threadLocalMessage.set(message);

// add a cleanup callback which will be
// executed from the ServletRequestCleanupFilter
ServletRequestCleanupService.addCleanupCallback(
new ServletRequestThreadLocalCleanupCallback(
this.threadLocalMessage
)
);
}
}


On a side note, I found it was a good practice to cleanup the ThreadLocals after the request was finished. Not really because this represented a memory problem as the number of Threads was limited by the Jetty thread pool. Rather I felt safer, if the recycled Thread serving a new Request would not see the Message of the previously handled HTTP Request. To do the cleanup using a Servlet Filter, I stole some code from the Apache Jetspeed 2 project (ServletRequestCleanupFilter, ServletRequestCleanupService, ServletRequestThreadLocalCleanupCallback and ServletRequestCleanupCallback).

Interviewed at Google Pt.2

Some people asked me why I wasn't publishing the second part of my Google interview experience. Well I didn't care so much about Google and job interviews in general after I moved to Norway. Anyway, I finished the whole blog post back in September at the airport in Zürich. So you might as well read the rest.

So lunch was great. Google has his own canteen with dedicated cooking staff etc. There is a salad bar, free drinks, desert and 3 dishes to choose from every day. Not bad. The canteen was packed when we went. It must have been 200 people or so.

Anyway, after lunch the next interview was again with two people, of which one was talking and the other one listening. This time I had two coding exercises in the interview. The first one, was to come up with an implementation of a class, which you could send a function or a block of executable code to, that should then be executed after 34 seconds. The second exercise was to suggest a unit test for a function producing random floats values between 0 and 1. For the first task I picked Java as concrete language and suggested using a ThreadPoolExecutor, Callable and Thread.sleep. The interviewer asked me, how we could bring down the test execution time, as we did not want to wait 34 seconds every time the test runs. I suggested having the time span configurable, so that it would fire after the configured amount of seconds. A better answer would have been to refactor the code, so that it was given an instance of a class representing a clock. I knew the clock test pattern but I was not able to pull it out in this situation - too bad. As for second task, testing random float values, I suggested creating a test set and group/count the values using some array indices. The test validation should only allow a certain degree of deviation. The interviewer then asked me how to calculate the deviation based on the input size but it was out of my league and I could not answer. For the correct answer you need some good knowledge in statistics I suppose. This was probably the worst interview I gave this day. It was also a bit harder for me because the interviewer had a very strong French accent. I had to ask him to repeat a couple of times, which might have been annoying for him. This is particular maddening because a co-worker I worked with for 3 years was French, so I should have been used to French-English.

The next interviewer was from Poland. He was a very cool guy and very friendly. We talked a bit about Apache Cassandra, which I had mentioned in my CV, since he knew one of the Facebook developers working with Cassandra. The coding exercise he had for me was about a graph problem. He gave me a class data-structure holding a list of File objects. The Files were supposed to have references to each other, pretty much like edges in a graph. I was given the task to print out the Files in a specific order. Unfortunately I understood him wrong and suggested a standard DFS traversal. But he was interested in a very specific order, which required a modified DFS approach. It was a nice interview anyways and he told me I was actually the first guy who had implemented a DFS using a stack and that he liked it. Well, I thought using a Stack for DFS was the default way, strange.

Time went by and after 45 minutes I found myself again in a very test-orientated session. The first 20 minutes or so, I was asked a lot of test related questions like "How would you do performance testing and let the development team know about the results?". My gut feeling was that I didn't give very great answers to these questions. Basically in my previous companies we never had an organization, where there was a unit of Software test engineers supporting the main developers. It is often the developer itself who does the performance testing and has the results at hand.. At least at this point, I realized I should have prepared myself more for test related questions since this was a Software Engineer in Test position after all. In my preparations I spent most time focusing on solving technical questions. But okay you never know this in advance. The final coding exercise in this interview was to come up with a suggestion on how to verify a file backup (directory structure) so that it matches a given source directory. I suggested to create an abstraction for each physical file or directory which would store the filename and a checksum. Then you could do a traversal of each tree, building up two linked lists. Finally you start removing the first item in each list and compare the items. If they don't match, it is either the directory structure that was wrong or something with the files. As the last step, I had to implement this in a language of my choice. I used again Java.

After a long day, the final interview started. A very nice girl, who worked with the Google search as a Software engineer, started to ask questions about my current job position. Surprise, this was actually the only time someone showed interest in what I was currently doing. I spoke about our system, the problems which we had and have. She asked me questions like "What happens if the database fails in this case?" and so on. I think she was testing my overall understanding of how different parts of a software system affect each other. The second part of the interview was performed as a code review. She had printed a function in Java and I my task was to review it. It was easy enough to find most of the potential issues like using static members (testability), not validating input, thread safety, going over the index bound of an array etc. I mixed up the explanation for bitwise XOR with AND, embarrassing but hey you are nervous. After doing the code review I had to rewrite a recursive function call into a iterative one, which I somehow managed. When the time was up, we said goodbye and she guided me out of the Zürich office building.

So what do I think about the interview? I think it was a great experience. I met nice people and they showed me some of my weaknesses. The preparations I had done upfront helped me only to some extend. There were less technical questions, no behavioral questions and definitely no questions about me or Google's history. A better preparation would have been to read some books on Software testing and less books about algorithms and data-structures.

So much for the second part. Here is something I am adding to the text I wrote 9 months ago. Never add technologies to your CV which you only know superficial. This can get you into embarrassing situations. Less is more. I have seen CV's for candidates who have worked with everything that exists out there - not good. For a Google interview, I think it is a plus if you know more than one language. Back then I was a Java only guy more or less. It's great if you can Python, C or C++. Some of my whiteboard examples where too Java specific, i.e. when using something like ThreadPoolExecutor take a step back and talk more about concepts than JDK classes. Don't expect every interviewer to know the language you are good at. Finally, expect to be disappointed if you fail. This can hit you hard, especially if you are used to getting a job offer every time you been to an interview.

Fluent Flexunit tests

As some of you might know, I work for a company which is doing Facebook games. Currently almost all of our Facebook games use Adobe Flash a the client and Java as the backend technology. There are different ways to have a Flex client communicate with a Java server application. For instance BlazeDS, which I really like. Unfortunately we are not using BlazeDS for the games I have been involved with. Instead we are using a framework, also based on a binary format similar to AMF, that someone at my company has written some time ago. There are some things about this framework which I don't like but overall it is performing really good. Anyway, last weekend I started to read a great book called Flex on Java, which was published by Manning almost half a year ago.

I am not the best Action Script developer, in fact I had never written a single line of Action Script when I started at my current company. The language is quite similar to JavaScript however, so Java developers are able to learn Action Script rather fast. From the first chapters in the book I could learn some stuff about Flex that I didn't knew before. Also it woke my interest in asynchronous testing using Flexunit. Similar to the Webservice example from the book, our unit-tests are also using asynchronous service calls. Well these tests are not really unit but integration tests. The first thing that is done is to start a embedded jetty using Maven and the maven-jetty-plugin:


<build>
<plugins>
<plugin>
<groupId>org.mortbay.jetty</groupId>
<artifactId>maven-jetty-plugin</artifactId>
<version>6.1.26</version>
<configuration>
<stopKey>foo</stopKey>
<stopPort>9999</stopPort>
<contextPath>/</contextPath>
<webApp>
${settings.localRepository}/package/goes/here/1.0-SNAPSHOT/game-1.0-SNAPSHOT.war
</webApp>
</configuration>
<executions>
<execution>
<id>start-jetty</id>
<phase>process-test-classes</phase>
<goals>
<goal>deploy-war</goal>
</goals>
<configuration>
<daemon>true</daemon>
</configuration>
</execution>
<execution>
<id>stopconfigurationReport-jetty</id>
<phase>prepare-package</phase>
<goals>
<goal>stop</goal>
</goals>
</execution>
</executions>
</plugin>
</plugins>
</build>


It is a bit dirty that we are starting a WAR file straight out of the Maven repository using an absolute file path. Not sure there is a way to tell the maven-jetty-plugin to start a WAR file of a Maven dependency instead. This does the job for now.

When embedded Jetty is running, our Flexunit integration tests will fire a sequence of service calls against the server. The sequence starts with resetting the test user, then some calls altering the player data and finally a call to load the test user back from the server. For the test user loaded at the end of the test, we run some Flexunit assert statements to validate that we are in the expected state. Since each of the service calls in the test sequence must be fired asynchronously, we use Async.handleEvent to wait until a certain Event has occurred otherwise we fail the test. Async.handleEvent has a Function argument called eventHandler, which will be the Function that is invoked if the Event has been fired successfully. Given that your service call sequence has a few steps, the test quickly mutates into a chain of Function calls of private Functions within the same test class. If you are coming from a JUnit or TestNG background, these sort of tests are not very attractive. For another developer, just from reading the code, it might be really hard to follow the test flow. I found myself adding comments all over the place, documenting where the test is continuing and what we were testing.

Here is a typical test which I made up to give you an example:


public class AlterPlayerTest {

private static const TIMEOUT : int = 2000;

private static const INITIALIZED : String = "initialized";
private static const SAVED : String = "saved";
private static const LOADED : String = "loaded";
private static const FAILED : String = "failed";

private var eventDispatcher : EventDispatcher = new EventDispatcher();
private var rpcClient : RpcClient;
private var user : User = null;

public function AlterPlayerTest() { }

[Test(async)]
public function testAlteringUser() : void {
this.rpcClient = new RpcClient(ConnectionParameters.getValues());

// make sure the user is created on the server side
this.rpcClient.init(dispatchInitializeEvent, dispatchFailEvent);
Async.handleEvent(this, this.eventDispatcher, INITIALIZED, alterUser, TIMEOUT); // flow continues in alterUser function
}

private function alterUser(evt : Event, passThru : Object) : void {
var user : User = new User();
user.id = new NetworkId(NetworkId.FACEBOOK, "189891015686187");

// make sure all objects are gone
var removeAll : RemoveAllObjects = new RemoveAllObjects();

var changeBatch : ChangeBatch = new ChangeBatch(new ArrayCollection([ removeAll ]).source);

this.rpcClient.saveUserUsingChangeBatch(dispatchSavedEvent, dispatchFailEvent, user, changeBatch);
Async.handleEvent(this, this.eventDispatcher, SAVED, reloadChangedUser, TIMEOUT); // flow continues in the reloadChangedUser function
}

private function reloadChangedUser(evt : Event, passThru : Object) : void {
// load back the user and verify that the RemoveAllObjects has been applied
this.rpcClient.load(dispatchUserLoadedEvent, dispatchFailEvent);
Async.handleEvent(this, this, LOADED, assertChanges, TIMEOUT);
}

private function assertChanges(evt : Event, passThru : Object) : void {
var user : User = this.user;
Assert.assertNotNull("user should not be null", user);

var objects : WorldObjects = this.user.worldObjects;
Assert.assertTrue("user should not have any objects left but had " + objects.length, objects.length == 0);
}

private function dispatchInitializeEvent(... args) : void {
dispatchEvent(new Event(INITIALIZED));
}

private function dispatchSavedEvent(results : Array) : void {
dispatchEvent(new Event(SAVED));
}

private function dispatchUserLoadedEvent(user : User) : void {
this.user = user;
dispatchEvent(new Event(LOADED));
}

private function dispatchFailEvent() : void {
dispatchEvent(new Event("Failed"));
}

}


In the AlterPlayerTest the service call sequence has contains only 3 service calls. The initial Flexunit test method is called testAlteringUser from which rpcClient.init is invoked. The test will break if no INITIALIZED Event is fired within 2 seconds. If the INITIALIZED Event is dispatched properly, the test flow continues with the alterUser function in which the service call rpcClient.saveUserUsingChangeBatch is made. Here the test expects the SAVED Event or it will fail. Being successful, the test will finally continue in reloadChangedUser, where the last service call rpcClient.load is made. Here the test user is retrieved from the Java server backend and again, a LOADED Event is expected. At the end the test is asserting that the test user has been altered.

While this test might be fine, we often have test scenarios where different game players interact with each other and the service call sequence is 10+ calls long, which makes the test really hard to read.

Fortunately today I found a nice framework called Fluint Sequences which has been contributed and become part of Flexunit. As the name suggests, use this if you want to write a fluent sequence of service calls within your test case. So here is a rewrite of the test above using Fluint Sequences:


public class AlterPlayerSequentiallyTest {

private static const TIMEOUT : int = 2000;

private static const INITIALIZED : String = "initialized";
private static const SAVED : String = "saved";
private static const LOADED : String = "loaded";
private static const FAILED : String = "failed";

private var eventDispatcher : EventDispatcher = new EventDispatcher();

public function AlterPlayerSequentiallyTest() { }

[Test(async)]
public function testAlteringUser() : void {
var rpcClient = new RpcClient(ConnectionParameters.getValues());

var user : User = new User();
user.id = new NetworkId(NetworkId.FACEBOOK, "189891015686187");

var passThroughData : Object = new Object();
passThroughData.usr = null;

var sequence:SequenceRunner = new SequenceRunner(this);

// make sure the user is created on the server side
sequence.addStep(sequenceCaller(rpcClient.init,
[function() : void {
eventDispatcher.dispatchEvent(new Event(INITIALIZED));
}, dispatchFailEvent]
)
);
sequence.addStep(new SequenceWaiter(this.eventDispatcher, INITIALIZED, TIMEOUT));

// make sure all objects are gone
sequence.addStep(sequenceCaller(rpcClient.saveUserUsingChangeBatch,
[function() : void {
eventDispatcher.dispatchEvent(new Event(SAVED));
}, dispatchFailEvent, user, new ChangeBatch(new ArrayCollection([ new RemoveAllObjects() ]).source)]
)
);
sequence.addStep(new SequenceWaiter(this.eventDispatcher, SAVED, TIMEOUT));

// reload user and store in passThroughData
sequence.addStep(sequenceCaller(rpcClient.load,
[function(user : User) : void {
passThroughData.usr = user;
eventDispatcher.dispatchEvent(new Event(LOADED));
}, dispatchFailEvent]
)
);
sequence.addStep(new SequenceWaiter(this.eventDispatcher, LOADED, TIMEOUT));

// assert changes
sequence.addAssertHandler(assertChanges, passThroughData);

// start the test
sequence.run();
}

private function sequenceCaller(fnc : Function, args : Array) {
return new SequenceCaller(
this.eventDispatcher,
fnc,
args
)
}

private function assertChanges(evt : Event, passThroughData : Object) : void {
var user : User = passThroughData.usr;
Assert.assertNotNull("user should not be null", user);

var objects : WorldObjects = this.user.worldObjects;
Assert.assertTrue("user should not have any objects left but had " + objects.length, objects.length == 0);
}

private function dispatchFailEvent() : void {
dispatchEvent(new Event("Failed"));
}
}


This test is not at all shorter than the first one but I think it is much easier to read and understand. Also some private member fields could be changed to local variables. The bread and butter of this test is the SequenceRunner, which acts as a container for steps which should be executed in order. No step will be executed until you invoke SequenceRunner.run() however. The first step is a SequenceCaller wrapping the rpcClient.init call followed by a SequenceWaiter step which waits the specified time for the INITIALIZED Event to occur. I do the same thing for altering and reloading the user. Finally I define and execute a Function to contain my assertions.