Tapestry Inversion of Control Container

The inner construction of the Tapestry framework is based on inversion of control, a design approach that allows a working system to be fabricated from many small, easily testable pieces.

An additional benefit of using IoC (Inversion of Control) is that, by breaking a complex system into small pieces, it becomes easier to modify and extend the system, by overriding or replacing selected parts of the system.

The use of IoC in Tapestry represents an evolution from Tapestry 3 to Tapestry 4 to Tapestry 5. Tapestry 3 did not use IoC, though it included some weaker mechanisms, such as extensions, that served a similar purpose. To make large scale changes to the behavior of Tapestry 3 required subclassing key classes and overriding methods.

Tapestry 4 introduced the use of the Apache HiveMind IoC container. In fact, the HiveMind project was created specifically for use as the IoC container for Tapestry 4. Tapestry 4 has met its goals for extensibility and configurability, largely because of HiveMind's flexibility.

Tapestry 5 extends on this, replacing HiveMind with a new container specifically build for Tapestry 5, designed for greater ease of use, expressiveness and performance. HiveMind itself has been subsequently shelved; T5 IoC can be considered a streamlined and improved HiveMind. And T5 IoC can be used separately from the rest of Tapestry!

Why Not Spring?

Spring is the most successful IoC container project. The Spring project combines a very good IoC container, integrated AspectJ support, and a large number of libraries built on top of the container. Spring is an excellent application container, but lacks a number of features necessary for a framework container:

• Spring beans can be wired together by name (or id), but it is not possible to introduce additional naming abstractions. Tapestry 4's "infrastructure:" abstraction was the key to allowing easy spot overrides of internal Tapestry services without having to duplicate the large web of interrelated services (nearly 200 in Tapestry 4.0).
• Although Spring allows beans to be intercepted, it does so in the form of a new bean, leaving the un-intercepted bean visible (and subject to misuse). Tapestry IoC "wraps" the service inside interceptors, preventing unintercepted access to the core service implementation.
• Spring's XML configuration files are quite verbose. This has improved with Spring 2.0, but still far more verbose that T5 IoC module classes.
• Spring has a simple map/list/value configuration scheme, but it is not distributed; it is part of a single bean definition.

Tapestry 5 IoC allows a service configuration to be assembled from multiple modules. This is very important for seamless extensibility of the framework, with zero configuration (just drop the module into the classpath and everything hooks together).

Why Not HiveMind?

The difficulty of managing the release schedules of two complex frameworks has proven to be an issue. HiveMind's 2.0 release will incorporate ideas similar to those present in Tapestry 5 IoC, but will also maintain legacy support for the existing XML-driven approach.

The use of HiveMind is also related to one of the common criticisms of Tapestry 4: startup time. The time it takes to parse and organize all that XML shows up as several seconds of startup time. It is hoped that creating a streamlined IoC container that is not driven by XML will alleviate those issues.

With the advent of new technologies (in particular, JDK 1.5 Annotations, and runtime class generation via Javassist) some of the precepts of HiveMind have been undermined. That is to say, in HiveMind (and Spring), all that XML is an awkward way to describe a few basic Java operations: instantiating classes and invoking methods on those classes (to inject dependencies into the instantiated instances). The central concept in Tapestry IoC is to eliminate XML and build an equivalent system around simple objects and methods.

Tapestry IoC also represents many simplifications of HiveMind, representing lessons learned while creating both HiveMind and Tapestry 4. HiveMind itself has wound down (it is not longer in active development), with the user base moving to Tapestry 5.

Why not Guice?

Google Guice is a newcomer to the IoC landscape. Guice and T5 IoC are very close and, in fact, T5 IoC expressly borrows many great and innovative ideas from Guice. Guice abandons not only XML but even any concept of a service id ... for injection, services are matched by type and perhaps filtered based on annotations.

Guice is still missing some core ideas needed in T5 IoC. There's no concept of configurations or anything similar. And there are limitations on injection based on scope (a request scoped value can't be injected into a global scope service; in T5 IoC, scope is internal to the proxy and never an issue).

Goals

As with Tapestry 5 in general, the goal of Tapestry IoC is greater simplicity, greater power, and an avoidance of XML.

Existing IoC containers such as HiveMind and Spring contain large amounts of XML configuration that exists to describe how and when to instantiate a particular JavaBean, and how to provide that bean with its dependencies (either by constructor injection, or by property injection). Other XML is used to hook objects into some form of lifecycle ... typically callback methods invoked when the object is instantiated and configured, or when it is being discarded.

The core concept of Tapestry IoC is that the Java language itself is the easiest and most succinct way to describe object creation and method invocation. Any approximation in XML is ultimately more verbose and unwieldy. As the [ examples|#service.htmlinjection] show, a small amount of Java code and a handful of naming conventions and annotations is far simpler and easier than a big chunk of XML.

In addition, moving from XML to Java code encourages testing; you can unit test the service builder methods of your module class, but you can't realistically unit test an XML descriptor.

Tapestry IoC modules are easily packaged into JAR files, supporting zero-configuration usage: just drop the JAR onto the classpath.

Another goal is "developer friendliness". This is a true cross-cutting concern, and one not likely to be packaged into an aspect any time soon. The Tapestry IoC framework is designed to be easy to use and easy to understand. Further, when things go wrong, it actively attempts to help you via comprehensive checks and carefully composed error messages. Further, all user-visible objects implement a reasonable toString() method, to help you understand what's going when you inevitably try to figure things out in the debugger.

In terms of building services using Tapestry IoC ... the objective here is "lightness", a term borrowed from the board game Go. In Go, two players place stones on an initially empty board, creating walls to enclose territory or eliminate the encroaching stones played by the opponent. The winner at the end of the game controls the most territory, and it is the constant tension between taking territory and defending existing territory that drives the game. In Go, groups of playing stones are "light" (or have "good shape") when the minimum number of them control the maximum area on the board. Playing "heavy" just gives your opponent a free chance to take control of another section of the board.

In software development, we are also attempting to create complex systems from simple pieces, but our tension is derived from the need to add functionality balanced against the need to test and maintain existing code. Too often in the world of software development, the need to add functionality trumps all, and testing and maintenance is deferred ... until too late.

IoC containers in general, and Tapestry IoC very specifically, exist to address this issue, to provide the foundations for balancing the need to quickly add functionality against the need to test new functionality and maintain existing functionality. IoC containers provide the means to break large, complex, monolithic blocks into light, small, testable pieces.

When building a registry of services, lightness refers to the proper division of responsibility, the seperation of concerns, and the limiting of dependencies between different parts of the system. This style is often called Law of Demeter. Using an IoC container makes it easier to embrace this approach, since one critical concern, which objects are responsible for instantiating which others, is entirely managed by the container. With this lifecycle concern removed, it becomes very easy to reduce complex chunks of code into small, testable, reusable services.

"Light" means:

• Small interfaces of two or three methods.
• Small methods, with two or three parameters (because dependencies are injected in behind the scenes, rather than passed into the method).
• Anonymous communication via events, rather than explicit method invocations. The service implementation can implement an event listener interface.

See The Pragmatic Programmer for more insights into building solid code.

Overview

The Tapestry IoC container takes over all the plumbing necessary for a highly scalable, extensible, thread-safe, testable application. Please see the main Tapestry IoC ovcerview for more details.

Terminology

The basic unit in Tapestry IoC is a service. A service consists of a service interface and a service implementation. The service interface is an ordinary Java interface. The service implementation is a Java object that implements the service interface. Often there will only be a single service per service interface, but in some situations, there may be many different services and service implementations all sharing the same service interface.

Services are identified by a unique id. Typically, a service id matches the unqualified name of the service interface, but this is simply a convention.

Services are aggregated into modules:

• A module is defined by a module class, a specific class containing a mix of static or instance methods, used to define services, decorate them (see below), or contribute to service configurations (again, more below).
• Methods of the module class define the services provided by the module, and the same methods are responsible for instantiating the service implementation.

The methods which define and construct services are called service builder methods.

The registry is the outside world's view of the modules and services. From the registry, it is possible to obtain a service, via its unique id or by its service interface. Access by unique id is caseless (meaning, a match will be found even the case of the search key doesn't match the case of the service id itself).

Services may be decorated by service decorator methods. These methods create interceptor objects that wrap around core service implementations, adding behavior such as logging, security access, or transaction management. Interceptors implement the same service interface as the service. Control is given over the order in which decorators are applied to a service.

A service may have a configuration. The configuration is either a map, a collection, or an ordered list. The service defines the type of object allowed to be contributed into the configuration. The configuration is constructed from contributions provided by one or more modules. Service contributor methods are invoked to contribute objects into configurations.

Services are instantiated as needed. In this case, "need" translates to "when a method of the service is invoked". A service is represented (to the outside world, or to other services) as a proxy that implements the service interface. The first time a method is invoked on the proxy, the full service (consisting of the core service implementation wrapped with any interceptors) is constructed. This occurs in a completely thread-safe manner. Just-in-time instantiation allows for more complex, more finely grained networks of services, and improves startup time.

Instantiating a service, injecting dependencies, and decorating the service are all parts of service realization, the point at which a service transitions from virtual (just a proxy) to real (fully instantiated and ready to operate).

Services define a scope that controls when the service is constructed, as well as its visibility. The default scope is singleton, meaning a single global instance created as needed. Other scopes allow service implementations to be bound to the current thread (i.e., the current request in a servlet application).

Dependencies are other services (or other objects) that are needed by a service implementation. These dependencies can be injected into a service builder method and provided, from there, to a service implementation via its constructor, or via methods on the service implementation. These may also be referred to as collaborators, especially in the context of writing unit tests.

The point of Injection is a field, method parameter, or constructor parameter that receives an injected value. The type of service (or other dependency) is determined by the type of the field or parameter. Often, annotations further identify what is to be injected, or in the case of field injection, that an injection is required.