Tuesday, April 24, 2007

Duh! Annotation values as constants, not literals

One funny thing I discovered recently is that annotation values don't have to be literal values. You can reference constants. Thus I went from:

@Scope("perthread")
public class .... 

To:

@Scope(IOCConstants.PERTHREAD_SCOPE)
public class ...

I think this is a Good Thing, even though it's a bit more verbose (a static import helps there) because I know at compile time that there isn't a typo in my constant's variable name ... whereas @Scope("prthread") would not be caught until runtime.

Pleasing the Crowds, Improving IoC, Extending the Community

I've been quite busy of late: a nearby Tapestry 4 project to pay the bills, a bunch of chances to talk to crowds about Tapestry 5 (in Philadelphia, Minneapolis and at home in Portland, Oregon).

I've significantly changed my Tapestry 4 presentation; it now highlights the BeanForm component, then the Table component and lastly Trails. I'm showing the high end of what's possible in Tapestry, rather than showing people the gory little details up front. That usually gets their attention.

I then show what's going on in Tapestry 5 and that really get's peoples jaws dropping. People crave this. They want to use it today. That is the desired effect. Case-insensitive, pretty URLs. Live class reloading. No XML. Incredible performance. Best of breed exception reporting. And we've barely gotten started yet.

I've been busy: Guice has set a high bar for IoC containers, but it still doesn't have certain features that Tapestry is dependent upon. That hasn't slowed me from taking the best ideas from it. Tapestry 5 now has a similar approach to autobuilding, and in most cases, a Tapestry service no longer needs a service builder method. Instead, a module can define a static bind() method and use the ServiceBinder to tell Tapestry about service interfaces and service implementations: Here's an example from the Tapestry module itself:

    public static void bind(ServiceBinder binder)
    {
        binder.bind(ClasspathAssetAliasManager.class, ClasspathAssetAliasManagerImpl.class);
        binder.bind(PersistentLocale.class, PersistentLocaleImpl.class);
        binder.bind(ApplicationStateManager.class, ApplicationStateManagerImpl.class);
        binder.bind(
                ApplicationStatePersistenceStrategySource.class,
                ApplicationStatePersistenceStrategySourceImpl.class);
        binder.bind(BindingSource.class, BindingSourceImpl.class);
        binder.bind(TranslatorSource.class, TranslatorSourceImpl.class);
        binder.bind(PersistentFieldManager.class, PersistentFieldManagerImpl.class);
        binder.bind(FieldValidatorSource.class, FieldValidatorSourceImpl.class);
        binder.bind(ApplicationGlobals.class, ApplicationGlobalsImpl.class);
        binder.bind(AssetSource.class, AssetSourceImpl.class);
        binder.bind(Cookies.class, CookiesImpl.class);
        binder.bind(Environment.class, EnvironmentImpl.class);
        binder.bind(FieldValidatorDefaultSource.class, FieldValidatorDefaultSourceImpl.class);
        binder.bind(RequestGlobals.class, RequestGlobalsImpl.class);
        binder.bind(ResourceDigestGenerator.class, ResourceDigestGeneratorImpl.class);
        binder.bind(ValidationConstraintGenerator.class, ValidationConstraintGeneratorImpl.class);
        binder.bind(EnvironmentalShadowBuilder.class, EnvironmentalShadowBuilderImpl.class);
        binder.bind(ComponentSource.class, ComponentSourceImpl.class);
        binder.bind(BeanModelSource.class, BeanModelSourceImpl.class);
    }

The ServiceBinder interface is fluent, we could follow on with .withScope(), .withId() or .eagerLoad() if we wanted. In most cases, defaults from there are sensible or come from annotations on the implementation class (and therefore, rarely need to be overridden).

The primary mechanism for injection is via constructor parameters. In general, annotations are not necessary on those parameters any more, and Tapestry will find the correct object or service automatically (primarily by matching on parameter type).

Service builder methods are still useful for when a service involves more than just instantiating a class, such as registering it for some kind of notification from another service:

    public ComponentClassResolver buildComponentClassResolver(ServiceResources resources)
    {
        ComponentClassResolverImpl service = resources.autobuild(ComponentClassResolverImpl.class);

        // Allow the resolver to clean its cache when the source is invalidated

        _componentInstantiatorSource.addInvalidationListener(service);

        return service;
    }

The autobuild() method will construct an instance, performing necessary injections. We can then perform any additional realization before returning it.

The end result has been to simplify and minimize the amount of code in the module builder classes.

Tapestry IoC now supports services that do not have a service interface; the actual type is used as the service interface and the service is not proxied: it is created on first reference and can't have interceptors. Normal services are proxied on first reference, and only realized (converted into a full service with a core service implementation and interceptors) on first use (the first method call).

In a step away from the code, we are running a vote to add Dan Gredler as a Tapestry committer. I expect that to run successfully, and I also expect to add a few more people to the roles soon.

I'm getting very excited. Things are coming together nicely (but never fast enough).

Saturday, April 07, 2007

Ouch! VMWare Fusion is not compatible with VMWare Player

Turns out that VMWare Fusion, the VMWare for Mac OS X, creates an image that isn't compatible with the VMWare player on Windows. This is why I'm happy I do test runs, but it does mean I have to spend another few hours recreating my workshop environment on Windows, using VMWare Workstation. So much for catching a movie tonight, or leaving the house tomorrow.

Hopefully this is something that'll be corrected in a later release.

Tapestry 5 at JavaOne and OSCON

My session at JavaOne has been scheduled:

BOF-6563: Tapestry 5: Java Programming Language Power, Scripting Ease
Esplanade 307-310
May 9 20-07, 8:55pm

In addition to JavaOne and NFJS, I've also been accepted to speak again at OSCON this year. I'm looking forward to spreading the Tapestry 5 message and meeting people using Tapestry, as always!

Wednesday, April 04, 2007

More Servlet Spec Wish-List Ideas

I'm thinking back to an earlier post about Servlet Spec 2.5 and what's missing.

Here's a few more ideas, based on implementing Tapestry 5 and other things besides. My previous ideas still stand.

web.xml introspection: It would be great to know what's going on inside web.xml. An API to query the web.xml meta-data and learn the names and mappings of the other servlets and filters would be great. Tapestry 5 acts as a filter and it would be handy if it could know what other servlets were in the WAR to help it decide whether to process a URL as inside Tapestry, or let another servlet handle it.

Event registration: Many of the listeners possible with the servlet API only apply to objects created by the servlet container's brain-dead IoC container. If my code needs to recerive notifications as a say, HttpSessionAttributeListener, I currently have to put together a Rube Goldberg machine in order to receive those notifications. There should be an API, perhaps attached to the ServletContext object, for registering for such notifications.

Quality of Service: Currently we have single server or clustered. Single server is not scalable beyond a certain point. Clustered causes a lot of problems managing data and introduces lots of overhead in terms of copying session attributes around the cluster.

I would like at least one additional level of service: non-replicated cluster. It assumes sticky sessions such that all requests for a given session are handled by a single box within the cluster. If that box fails, then future requests will be sent to a different box within the cluster, and the application will receive a notification that data has been lost.

Really, most non-financial applications can survive a rare loss of data midway through a session, and having this intermediate service level would allow the vast majority of applications to scale much, much higher than is possible today. I've previously implemented something like this by storing an object in the HttpSession that contained a Map in a transient variable.

Concurrency documentation: Many objects in the Servlet API are shared between threads (ServletContext, HttpSession) and it would be good form, and very useful, to know the proper semantics there. Futher, there needs to be better documentation for application and framework developers about the implications thereof.

In light of what I've learned about the Java Memory Model from Brian Goetz (and his book, Java Concurrency in Practice) any non-synchronized access to a shared object, say a mutable object stored into the HttpSession, is suspect: a potential source of scary, non-reproduceable, intermittent concurrency bugs.

Should all methods of such objects be synchronized? Can the servlet container acquire a lock when the HttpSession is accessed and hold it for the application for the remainder of the request (that would be handy!). Could there be an API for allowing the framework/application to manage access to HttpSession variables?

It's clear that the HttpSession API was designed to store small, immutable objects such as Strings and Integers. It is also clear that most applications don't use it this way (though Tapestry is a little better because it mostly stores individual page fields in the session, which are often small immutables, whereas most frameworks and applications store more complex state holding mutable objects).

... well, I've run out of steam, but I'll be racking my brains for more ideas like these.

Tuesday, April 03, 2007

Tapestry 5 IoC from a Guice Perspective

So Guice has a few good ideas, and with Tapestry 5 still in a fluid state, I've absorbed a few of them into Tapestry 5 IoC. At the core, is the key ObjectProvider interface:

public interface ObjectProvider
{
    <T> T provide(Class<T> objectType, AnnotationProvider annotationProvider, ServiceLocator locator);
}

Implementations of this interface plug into a chain-of-command within Tapestry and, via the AnnotationProvider, can query for the presence of annotations to help decide exactly what is to be injected. A lot of Guice's examples, things like the @Blue service, are easily replicated in Tapestry 5 IoC, there's just not a lot of need for it. I supposed it would look something like:

public class BlueObjectProvider {
  private final Map<Class,Object> _blueObjects;

  public BlueObjectProvider(...) { ... }

    <T> T provide(Class<T> objectType, AnnotationProvider annotationProvider, ServiceLocator locator) {
      if (annotationProvider.getAnnotation(Blue.class) == null) return null;

      return objectType.cast(_blueObjects.get(objectType));
  }
}

Implicit here is a service configuration of implementations, keyed on object type, used when the @Blue annotation is present. Here's what the module would look like:

public class BlueModule {

  public ObjectProvider buildBlueProvider(Map<Class, Object> configuration)
  {
    return new BlueObjectProvider(configuration);
  }

  public void contributeMasterObjectProvider(
    @InjectService("BlueProvider") ObjectProvider blueProvider,
    OrderedConfiguration<ObjectProvider> configuration) {
    
    configuration.add("Blue", blueProvider);
  }

  public void contributeBlueProvider(MappedConfiguration<Class, Object> configuration,
    @InjectService("MyBlueService") Service myBlueService) {
    configuration.add(Service.class, myBlueService);
  }
}

public class OtherModule {

  public OtherService build(@Inject @Blue Service service) {
    return new OtherServiceImpl(service);
  }
}

On the one hand, this is more verbose, since its somewhat more procedurally based than Guice's approach, which is pattern based. On the other hand, it demonstrates a couple of key features of Tapestry 5 IoC:

  • Service construction occurs inside service builder methods; there's no layer of abstraction around service instantiation, configuration and initialization ... you just do it in Java code. There simply isn't a better language for descibing instantiating Java objects and invoking Java methods than Java itself.
  • Dependency injection is a concern of the module not the service implementation. The special annotations go in the module class, and the service implementation class is blissfully unaware of where those dependencies come from or are identified and obtained.
  • Dependencies are passed to the service builder methods, which do whatever is appropriate with those dependencies; in many cases, the dependencies are not passed to the instantiated class, but are used for other concerns, such as registering the new service for event notifications from some other service.
  • The contributeBlueProvider() method can appear in many different modules, and the results are cumulative. This allows one module to "lay the groundwork" and for other modules to "fill in the details". And, again, if this style of injection ("flavored" with the annotations) really became popular, it would be easy to integrate it into the central Tapestry 5 IoC module so it could be leveraged everywhere else.

In fact, in the OtherModule class, you can see how the injection lays out once the groundwork is prepared: Just an @Inject qualified with @Blue and you're done. Again, I can't emphasize enough how well moving construction concerns into the service builder methods works; it gives you complete control, it's unit testable in a couple of different ways, it keeps the service implementations themselves clean (of dependencies on the IoC container -- even in the form of annotations). It also keeps explicit what order operations occur in. HiveMind chased this with ever more complex XML markup to describe how to instantiate, configure, and initialize services.

That last feature, the interaction of multiple modules discovered at runtime, is the key distinguishing feature, and the one that's hardest to grasp. For me, it is the point where an IoC container transitions from an application container to a framework container. In an application container, you explicitly know what modules are present and how they will interact. It's your cake, starting from raw flour, sugar and eggs. Nothings in the mix that you didn't add yourself.

With HiveMind and Tapestry 5 IoC, the goal is to build off of an existing framework, and its interwoven set of services. The framework, or layers of framework, provide the cake, and you are trying to add just the icing. Configurations are the way to inject that icing without disturbing the whole cake, or even knowing the cake's recipe.

So Guice is a good kick in the pants, and Tapestry 5 IoC has evolved; the old @Inject annotation, which used to include a string attribute to name the thing to inject, has been simplified. The @Inject annotation is still there, but the string part is gone. Tapestry primarily works from the object type to match to the lone service implementing that interface. When that is insufficient, there's the MasterObjectProvider and Alias service configurations, which provide plenty of room to disambiguate (possibly by adding additional annotations to the field being injected).

The big advantage is type safety; refactor the name of a service interface and you'll see little or no disruption of the service network, because dependencies are almost always expressed in terms of just the (refactored) service interface, rather than any additional layer of "logical name".

I think there's a growing issue with Google-worship. It was interesting and disturbing to see so many people announce "Spring is dead! Long live Guice!" on various blogs (and I'll even admit to a little envy ... how come Tapestry 5 IoC doesn't generate this kind of interest?). Guice itself is pretty useless by itself, just as Spring's IoC container is, by itself, useless. The value of Spring is in what's built on top of the container: Wrappers for JDBC, Hibernate, JPA, JMS and every other acronym you can name. Transaction support and interceptors. And, more recently, the really solid AspectJ layer. If Guice wants to be a Spring-killer, or even a real player in the IoC space, it needs to embrace and extend Spring: make it easier to use those Spring integrations than it would be using Spring natively.