The Architecture of Open Source Applications (Volume 1)

Kim Moir

Implementing software modularity is a notoriously difficult task. Interoperability with a large code base written by a diverse community is also difficult to manage. At Eclipse, we have managed to succeed on both counts. In June 2010, the Eclipse Foundation made available its Helios coordinated release, with over 39 projects and 490 committers from over 40 companies working together to build upon the functionality of the base platform. What was the original architectural vision for Eclipse? How did it evolve? How does the architecture of an application serve to encourage community engagement and growth? Let's go back to the beginning.

On November 7, 2001, an open source project called Eclipse 1.0 was released. At the time, Eclipse was described as "an integrated development environment (IDE) for anything and nothing in particular." This description was purposely generic because the architectural vision was not just another set of tools, but a framework; a framework that was modular and scalable. Eclipse provided a component-based platform that could serve as the foundation for building tools for developers. This extensible architecture encouraged the community to build upon a core platform and extend it beyond the limits of the original vision. Eclipse started as a platform and the Eclipse SDK was the proof-of-concept product. The Eclipse SDK allowed the developers to self-host and use the Eclipse SDK itself to build newer versions of Eclipse.

The stereotypical image of an open source developer is that of an altruistic person toiling late into night fixing bugs and implementing fantastic new features to address their own personal interests. In contrast, if you look back at the early history of the Eclipse project, some of the initial code that was donated was based on VisualAge for Java, developed by IBM. The first committers who worked on this open source project were employees of an IBM subsidiary called Object Technology International (OTI). These committers were paid to work full time on the open source project, to answer questions on newsgroups, address bugs, and implement new features. A consortium of interested software vendors was formed to expand this open tooling effort. The initial members of the Eclipse consortium were Borland, IBM, Merant, QNX Software Systems, Rational Software, RedHat, SuSE, and TogetherSoft.

By investing in this effort, these companies would have the expertise to ship commercial products based on Eclipse. This is similar to investments that corporations make in contributing to the Linux kernel because it is in their self-interest to have employees improving the open source software that underlies their commercial offerings. In early 2004, the Eclipse Foundation was formed to manage and expand the growing Eclipse community. This not-for-profit foundation was funded by corporate membership dues and is governed by a board of directors. Today, the diversity of the Eclipse community has expanded to include over 170 member companies and almost 1000 committers.

Originally, people knew "Eclipse" as the SDK only but today it is much more. In July 2010, there were 250 diverse projects under development at There's tooling to support developing with C/C++, PHP, web services, model driven development, build tooling and many more. Each of these projects is included in a top-level project (TLP) which is managed by a project management committee (PMC) consisting of senior members of the project nominated for the responsibility of setting technical direction and release goals. In the interests of brevity, the scope of this chapter will be limited to the evolution of the architecture of the Eclipse SDK within Eclipse1 and Runtime Equinox2 projects. Since Eclipse has long history, I'll be focusing on early Eclipse, as well as the 3.0, 3.4 and 4.0 releases.

6.1. Early Eclipse

At the beginning of the 21st century, there were many tools for software developers, but few of them worked together. Eclipse sought to provide an open source platform for the creation of interoperable tools for application developers. This would allow developers to focus on writing new tools, instead of writing to code deal with infrastructure issues like interacting with the filesystem, providing software updates, and connecting to source code repositories. Eclipse is perhaps most famous for the Java Development Tools (JDT). The intent was that these exemplary Java development tools would serve as an example for people interested in providing tooling for other languages.

Before we delve into the architecture of Eclipse, let's look at what the Eclipse SDK looks like to a developer. Upon starting Eclipse and selecting the workbench, you'll be presented with the Java perspective. A perspective organizes the views and editors that are specific to the tooling that is currently in use.

[Java Perspective]

Figure 6.1: Java Perspective

Early versions of the Eclipse SDK architecture had three major elements, which corresponded to three major sub-projects: the Platform, the JDT (Java Development Tools) and the PDE (Plug-in Development Environment).

6.1.1. Platform

The Eclipse platform is written using Java and a Java VM is required to run it. It is built from small units of functionality called plugins. Plugins are the basis of the Eclipse component model. A plugin is essentially a JAR file with a manifest which describes itself, its dependencies, and how it can be utilized, or extended. This manifest information was initially stored in a plug-in.xml file which resides in the root of the plugin directory. The Java development tools provided plugins for developing in Java. The Plug-in Development Environment (PDE) provides tooling for developing plugins to extend Eclipse. Eclipse plugins are written in Java but could also contain non-code contributions such as HTML files for online documentation. Each plugin has its own class loader. Plugins can express dependencies on other plugins by the use of requires statements in the plugin.xml. Looking at the plugin.xml for the org.eclipse.ui plugin you can see its name and version specified, as well as the dependencies it needs to import from other plugins.

<?xml version="1.0" encoding="UTF-8"?>

      <library name="ui.jar">
         <export name="*"/>
         <packages prefixes="org.eclipse.ui"/>
      <import plugin="org.apache.xerces"/>
      <import plugin="org.eclipse.core.resources"/>
      <import plugin="org.eclipse.update.core"/>
      :       :        :
      <import plugin="org.eclipse.text" export="true"/>
      <import plugin="org.eclipse.ui.workbench.texteditor" export="true"/>
      <import plugin="org.eclipse.ui.editors" export="true"/>

In order to encourage people to build upon the Eclipse platform, there needs to be a mechanism to make a contribution to the platform, and for the platform to accept this contribution. This is achieved through the use of extensions and extension points, another element of the Eclipse component model. The export identifies the interfaces that you expect others to use when writing their extensions, which limits the classes that are available outside your plugin to the ones that are exported. It also provides additional limitations on the resources that are available outside the plugin, as opposed to making all public methods or classes available to consumers. Exported plugins are considered public API. All others are considered private implementation details. To write a plugin that would contribute a menu item to the Eclipse toolbar, you can use the actionSets extension point in the org.eclipse.ui plugin.

<extension-point id="actionSets" name="%ExtPoint.actionSets"
<extension-point id="commands" name="%ExtPoint.commands"
<extension-point id="contexts" name="%ExtPoint.contexts"
<extension-point id="decorators" name="%ExtPoint.decorators"
<extension-point id="dropActions" name="%ExtPoint.dropActions"
                 schema="schema/dropActions.exsd"/> =

Your plugin's extension to contribute a menu item to the org.eclipse.ui.actionSet extension point would look like:

<?xml version="1.0" encoding="UTF-8"?>
      <library name="helloworld.jar"/>
      <import plugin="org.eclipse.ui"/>
            label="Example Action Set"
               label="Example &Menu"
               label="&Example Action"
               tooltip="Hello, Eclipse world"

When Eclipse is started, the runtime platform scans the manifests of the plugins in your install, and builds a plugin registry that is stored in memory. Extension points and the corresponding extensions are mapped by name. The resulting plugin registry can be referenced from the API provided by the Eclipse platform. The registry is cached to disk so that this information can be reloaded the next time Eclipse is restarted. All plugins are discovered upon startup to populate the registry but they are not activated (classes loaded) until the code is actually used. This approach is called lazy activation. The performance impact of adding additional bundles into your install is reduced by not actually loading the classes associated with the plugins until they are needed. For instance, the plugin that contributes to the org.eclipse.ui.actionSet extension point wouldn't be activated until the user selected the new menu item in the toolbar.

[Example Menu]

Figure 6.2: Example Menu

The code that generates this menu item looks like this:

package com.example.helloworld.actions;

import org.eclipse.jface.action.IAction;
import org.eclipse.jface.viewers.ISelection;
import org.eclipse.ui.IWorkbenchWindow;
import org.eclipse.ui.IWorkbenchWindowActionDelegate;
import org.eclipse.jface.dialogs.MessageDialog;

public class ExampleAction implements IWorkbenchWindowActionDelegate {
    private IWorkbenchWindow window;

    public ExampleAction() {

    public void run(IAction action) {
            "Hello, Eclipse architecture world");

    public void selectionChanged(IAction action, ISelection selection) {

    public void dispose() {

    public void init(IWorkbenchWindow window) {
        this.window = window;

Once the user selects the new item in the toolbar, the extension registry is queried by the plugin implementing the extension point. The plugin supplying the extension instantiates the contribution, and loads the plugin. Once the plugin is activated, the ExampleAction constructor in our example is run, and then initializes a Workbench action delegate. Since the selection in the workbench has changed and the delegate has been created, the action can change. The message dialog opens with the message "Hello, Eclipse architecture world".

This extensible architecture was one of the keys to the successful growth of the Eclipse ecosystem. Companies or individuals could develop new plugins, and either release them as open source or sell them commercially.

One of the most important concepts about Eclipse is that everything is a plugin. Whether the plugin is included in the Eclipse platform, or you write it yourself, plugins are all first class components of the assembled application. Figure 6.3 shows clusters of related functionality contributed by plugins in early versions of Eclipse.

[Early Eclipse Architecture]

Figure 6.3: Early Eclipse Architecture

The workbench is the most familiar UI element to users of the Eclipse platform, as it provides the structures that organize how Eclipse appears to the user on the desktop. The workbench consists of perspectives, views, and editors. Editors are associated with file types so the correct editor is launched when a file is opened. An example of a view is the "problems" view that indicates errors or warnings in your Java code. Together, editors and views form a perspective which presents the tooling to the user in an organized fashion.

The Eclipse workbench is built on the Standard Widget Toolkit (SWT) and JFace, and SWT deserves a bit of exploration. Widget toolkits are generally classified as either native or emulated. A native widget toolkit uses operating system calls to build user interface components such as lists and push buttons. Interaction with components is handled by the operating system. An emulated widget toolkit implements components outside of the operating system, handling mouse and keyboard, drawing, focus and other widget functionality itself, rather than deferring to the operating system. Both designs have different strengths and weaknesses.

Native widget toolkits are "pixel perfect." Their widgets look and feel like their counterparts in other applications on the desktop. Operating system vendors constantly change the look and feel of their widgets and add new features. Native widget toolkits get these updates for free. Unfortunately, native toolkits are difficult to implement because their underlying operating system widget implementations are vastly different, leading to inconsistencies and programs that are not portable.

Emulated widget toolkits either provide their own look and feel, or try to draw and behave like the operating system. Their great strength over native toolkits is flexibility (although modern native widget toolkits such as Windows Presentation Framework (WPF) are equally as flexible). Because the code to implement a widget is part of the toolkit rather than embedded in the operating system, a widget can be made to draw and behave in any manner. Programs that use emulated widget toolkits are highly portable. Early emulated widget toolkits had a bad reputation. They were often slow and did a poor job of emulating the operating system, making them look out of place on the desktop. In particular, Smalltalk-80 programs at the time were easy to recognize due to their use of emulated widgets. Users were aware that they were running a "Smalltalk program" and this hurt acceptance of applications written in Smalltalk.

Unlike other computer languages such as C and C++, the first versions of Java came with a native widget toolkit library called the Abstract Window Toolkit (AWT). AWT was considered to be limited, buggy and inconsistent and was widely decried. At Sun and elsewhere, in part because of experience with AWT, a native widget toolkit that was portable and performant was considered to be unworkable. The solution was Swing, a full-featured emulated widget toolkit.

Around 1999, OTI was using Java to implement a product called VisualAge Micro Edition. The first version of VisualAge Micro Edition used Swing and OTI's experience with Swing was not positive. Early versions of Swing were buggy, had timing and memory issues and the hardware at the time was not powerful enough to give acceptable performance. OTI had successfully built a native widget toolkit for Smalltalk-80 and other Smalltalk implementations to gain acceptance of Smalltalk. This experience was used to build the first version of SWT. VisualAge Micro Edition and SWT were a success and SWT was the natural choice when work began on Eclipse. The use of SWT over Swing in Eclipse split the Java community. Some saw conspiracies, but Eclipse was a success and the use of SWT differentiated it from other Java programs. Eclipse was performant, pixel perfect and the general sentiment was, "I can't believe it's a Java program."

Early Eclipse SDKs ran on Linux and Windows. In 2010, there is support for over a dozen platforms. A developer can write an application for one platform, and deploy it to multiple platforms. Developing a new widget toolkit for Java was a contentious issue within the Java community at the time, but the Eclipse committers felt that it was worth the effort to provide the best native experience on the desktop. This assertion applies today, and there are millions of lines of code that depend on SWT.

JFace is a layer on top of SWT that provides tools for common UI programming tasks, such as frameworks for preferences and wizards. Like SWT, it was designed to work with many windowing systems. However, it is pure Java code and doesn't contain any native platform code.

The platform also provided an integrated help system based upon small units of information called topics. A topic consists of a label and a reference to its location. The location can be an HTML documentation file, or an XML document describing additional links. Topics are grouped together in table of contents (TOCs). Consider the topics as the leaves, and TOCs as the branches of organization. To add help content to your application, you can contribute to the extension point, as the org.eclipse.platform.doc.isv plugin.xml does below.

<?xml version="1.0" encoding="UTF-8"?>
<?eclipse version="3.0"?>

<!-- ===================================================================== -->
<!-- Define primary TOC                                                    -->
<!-- ===================================================================== -->
      <index path="index"/>
<!-- ===================================================================== -->
<!-- Define TOCs                                                           -->
<!-- ===================================================================== -->

Apache Lucene is used to index and search the online help content. In early versions of Eclipse, online help was served as a Tomcat web application. Additionally, by providing help within Eclipse itself, you can also use the subset of help plugins to provide a standalone help server.3

Eclipse also provides team support to interact with a source code repository, create patches and other common tasks. The workspace provided collection of files and metadata that stored your work on the filesystem. There was also a debugger to trace problems in the Java code, as well as a framework for building language specific debuggers.

One of the goals of the Eclipse project was to encourage open source and commercial consumers of this technology to extend the platform to meet their needs, and one way to encourage this adoption is to provide a stable API. An API can be thought of as a technical contract specifying the behavior of your application. It also can be thought of as a social contract. On the Eclipse project, the mantra is, "API is forever". Thus careful consideration must be given when writing an API given that it is meant to be used indefinitely. A stable API is a contract between the client or API consumer and the provider. This contract ensures that the client can depend on the Eclipse platform to provide the API for the long term without the need for painful refactoring on the part of the client. A good API is also flexible enough to allow the implementation to evolve.

6.1.2. Java Development Tools (JDT)

The JDT provides Java editors, wizards, refactoring support, debugger, compiler and an incremental builder. The compiler is also used for content assist, navigation and other editing features. A Java SDK isn't shipped with Eclipse so it's up to the user to choose which SDK to install on their desktop. Why did the JDT team write a separate compiler to compile your Java code within Eclipse? They had an initial compiler code contribution from VisualAge Micro Edition. They planned to build tooling on top of the compiler, so writing the compiler itself was a logical decision. This approach also allowed the JDT committers to provide extension points for extending the compiler. This would be difficult if the compiler was a command line application provided by a third party.

Writing their own compiler provided a mechanism to provide support for an incremental builder within the IDE. An incremental builder provides better performance because it only recompiles files that have changed or their dependencies. How does the incremental builder work? When you create a Java project within Eclipse, you are creating resources in the workspace to store your files. A builder within Eclipse takes the inputs within your workspace (.java files), and creates an output (.class files). Through the build state, the builder knows about the types (classes or interfaces) in the workspace, and how they reference each other. The build state is provided to the builder by the compiler each time a source file is compiled. When an incremental build is invoked, the builder is supplied with a resource delta, which describes any new, modified or deleted files. Deleted source files have their corresponding class files deleted. New or modified types are added to a queue. The files in the queue are compiled in sequence and compared with the old class file to determine if there are structural changes. Structural changes are modifications to the class that can impact another type that references it. For example, changing a method signature, or adding or removing a method. If there are structural changes, all the types that reference it are also added to the queue. If the type has changed at all, the new class file is written to the build output folder. The build state is updated with reference information for the compiled type. This process is repeated for all the types in the queue until empty. If there are compilation errors, the Java editor will create problem markers. Over the years, the tooling that JDT provides has expanded tremendously in concert with new versions of the Java runtime itself.

6.1.3. Plug-in Development Environment (PDE)

The Plug-in Development Environment (PDE) provided the tooling to develop, build, deploy and test plugins and other artifacts that are used to extend the functionality of Eclipse. Since Eclipse plugins were a new type of artifact in the Java world there wasn't a build system that could transform the source into plugins. Thus the PDE team wrote a component called PDE Build which examined the dependencies of the plugins and generated Ant scripts to construct the build artifacts.

6.2. Eclipse 3.0: Runtime, RCP and Robots

6.2.1. Runtime

Eclipse 3.0 was probably one of the most important Eclipse releases due to the number of significant changes that occurred during this release cycle. In the pre-3.0 Eclipse architecture, the Eclipse component model consisted of plugins that could interact with each other in two ways. First, they could express their dependencies by the use of the requires statement in their plugin.xml. If plugin A requires plugin B, plugin A can see all the Java classes and resources from B, respecting Java class visibility conventions. Each plugin had a version, and they could also specify the versions of their dependencies. Secondly, the component model provided extensions and extension points. Historically, Eclipse committers wrote their own runtime for the Eclipse SDK to manage classloading, plugin dependencies and extensions and extension points.

The Equinox project was created as a new incubator project at Eclipse. The goal of the Equinox project was to replace the Eclipse component model with one that already existed, as well as provide support for dynamic plugins. The solutions under consideration included JMX, Jakarta Avalon and OSGi. JMX was not a fully developed component model so it was not deemed appropriate. Jakarta Avalon wasn't chosen because it seemed to be losing momentum as a project. In addition to the technical requirements, it was also important to consider the community that supported these technologies. Would they be willing to incorporate Eclipse-specific changes? Was it actively developed and gaining new adopters? The Equinox team felt that the community around their final choice of technology was just as important as the technical considerations.

After researching and evaluating the available alternatives, the committers selected OSGi. Why OSGi? It had a semantic versioning scheme for managing dependencies. It provided a framework for modularity that the JDK itself lacked. Packages that were available to other bundles must be explicitly exported, and all others were hidden. OSGi provided its own classloader so the Equinox team didn't have to continue to maintain their own. By standardizing on a component model that had wider adoption outside the Eclipse ecosystem, they felt they could appeal to a broader community and further drive the adoption of Eclipse.

The Equinox team felt comfortable that since OSGi already had an existing and vibrant community, they could work with that community to help include the functionality that Eclipse required in a component model. For instance, at the time, OSGi only supported listing requirements at a package level, not a plugin level as Eclipse required. In addition, OSGi did not yet include the concept of fragments, which were Eclipse's preferred mechanism for supplying platform or environment specific code to an existing plugin. For example, fragments provide code for working with Linux and Windows filesystems as well as fragments which contribute language translations. Once the decision was made to proceed with OSGi as the new runtime, the committers needed an open source framework implementation. They evaluated Oscar, the precursor to Apache Felix, and the Service Management Framework (SMF) developed by IBM. At the time, Oscar was a research project with limited deployment. SMF was ultimately chosen since it was already used in shipping products and thus was deemed enterprise-ready. The Equinox implementation serves as the reference implementation of the OSGi specification.

A compatibility layer was also provided so that existing plugins would still work in a 3.0 install. Asking developers to rewrite their plugins to accommodate changes in the underlying infrastructure of Eclipse 3.0 would have stalled the momentum on Eclipse as a tooling platform. The expectation from Eclipse consumers was that the platform should just continue to work.

With the switch to OSGi, Eclipse plugins became known as bundles. A plugin and a bundle are the same thing: They both provide a modular subset of functionality that describes itself with metadata in a manifest. Previously, dependencies, exported packages and the extensions and extension points were described in plugin.xml. With the move to OSGi bundles, the extensions and extension points continued to be described in plugin.xml since they are Eclipse concepts. The remaining information was described in the META-INF/MANIFEST.MF, OSGi's version of the bundle manifest. To support this change, PDE provided a new manifest editor within Eclipse. Each bundle has a name and version. The manifest for the org.eclipse.ui bundle looks like this:

Manifest-Version: 1.0
Bundle-ManifestVersion: 2
Bundle-SymbolicName: org.eclipse.ui; singleton:=true
Bundle-Version: 3.3.0.qualifier
Bundle-ClassPath: .
Bundle-Activator: org.eclipse.ui.internal.UIPlugin
Bundle-Vendor: %Plugin.providerName
Bundle-Localization: plugin
Export-Package: org.eclipse.ui.internal;x-internal:=true
Require-Bundle: org.eclipse.core.runtime;bundle-version="[3.2.0,4.0.0)",
Eclipse-LazyStart: true
Bundle-RequiredExecutionEnvironment: CDC-1.0/Foundation-1.0, J2SE-1.3

As of Eclipse 3.1, the manifest can also specify a bundle required execution environment (BREE). Execution environments specify the minimum Java environment required for the bundle to run. The Java compiler does not understand bundles and OSGi manifests. PDE provides tooling for developing OSGi bundles. Thus, PDE parses the bundle's manifest, and generates the classpath for that bundle. If you specified an execution environment of J2SE-1.4 in your manifest, and then wrote some code that included generics, you would be advised of compile errors in your code. This ensures that your code adheres to the contract you have specified in the manifest.

OSGi provides a modularity framework for Java. The OSGi framework manages collections of self-describing bundles and manages their classloading. Each bundle has its own classloader. The classpath available to a bundle is constructed by examining the dependencies of the manifest and generating a classpath available to the bundle. OSGi applications are collections of bundles. In order to fully embrace of modularity, you must be able to express your dependencies in a reliable format for consumers. Thus the manifest describes exported packages that are available to clients of this bundle which corresponds to the public API that was available for consumption. The bundle that is consuming that API must have a corresponding import of the package they are consuming. The manifest also allows you to express version ranges for your dependencies. Looking at the Require-Bundle heading in the above manifest, you will note that the org.eclipse.core.runtime bundle that org.eclipse.ui depends on must be at least 3.2.0 and less than 4.0.0.

[OSGi Bundle Lifecycle]

Figure 6.4: OSGi Bundle Lifecycle

OSGi is a dynamic framework which supports the installation, starting, stopping, or uninstallation of bundles. As mentioned before, lazy activation was a core advantage to Eclipse because plugin classes were not loaded until they were needed. The OSGi bundle lifecycle also enables this approach. When you start an OSGi application, the bundles are in the installed state. If its dependencies are met, the bundle changes to the resolved state. Once resolved, the classes within that bundle can be loaded and run. The starting state means that the bundle is being activated according to its activation policy. Once activated, the bundle is in the active state, it can acquire required resources and interact with other bundles. A bundle is in the stopping state when it is executing its activator stop method to clean up any resources that were opened when it was active. Finally, a bundle may be uninstalled, which means that it's not available for use.

As the API evolves, there needs to be a way to signal changes to your consumers. One approach is to use semantic versioning of your bundles and version ranges in your manifests to specify the version ranges for your dependencies. OSGi uses a four-part versioning naming scheme as shown in Figure 6.5.

[Versioning Naming Scheme]

Figure 6.5: Versioning Naming Scheme

With the OSGi version numbering scheme, each bundle has a unique identifier consisting of a name and a four part version number. An id and version together denote a unique set of bytes to the consumer. By Eclipse convention, if you're making changes to a bundle, each segment of the version signifies to the consumer the type of change being made. Thus, if you want to indicate that you intend to break API, you increment the first (major) segment. If you have just added API, you increment the second (minor) segment. If you fix a small bug that doesn't impact API, the third (service) segment is incremented. Finally, the fourth or qualifier segment is incremented to indicate a build id source control repository tag.

In addition to expressing the fixed dependencies between bundles, there is also a mechanism within OSGi called services which provides further decoupling between bundles. Services are objects with a set of properties that are registered with the OSGi service registry. Unlike extensions, which are registered in the extension registry when Eclipse scans bundles during startup, services are registered dynamically. A bundle that is consuming a service needs to import the package defining the service contract, and the framework determines the service implementation from the service registry.

Like a main method in a Java class file, there is a specific application defined to start Eclipse. Eclipse applications are defined using extensions. For instance, the application to start the Eclipse IDE itself is org.eclipse.ui.ide.workbench which is defined in the org.eclipse.ui.ide.application bundle.


There are many applications provided by Eclipse such as those to run standalone help servers, Ant tasks, and JUnit tests.

6.2.2. Rich Client Platform (RCP)

One of the most interesting things about working in an open source community is that people use the software in totally unexpected ways. The original intent of Eclipse was to provide a platform and tooling to create and extend IDEs. However, in the time leading up to the 3.0 release, bug reports revealed that the community was taking a subset of the platform bundles and using them to build Rich Client Platform (RCP) applications, which many people would recognize as Java applications. Since Eclipse was initially constructed with an IDE-centric focus, there had to be some refactoring of the bundles to allow this use case to be more easily adopted by the user community. RCP applications didn't require all the functionality in the IDE, so several bundles were split into smaller ones that could be consumed by the community for building RCP applications.

Examples of RCP applications in the wild include the use of RCP to monitor the Mars Rover robots developed by NASA at the Jet Propulsion Laboratory, Bioclipse for data visualization of bioinformatics and Dutch Railway for monitoring train performance. The common thread that ran through many of these applications was that these teams decided that they could take the utility provided by the RCP platform and concentrate on building their specialized tools on top of it. They could save development time and money by focusing on building their tools on a platform with a stable API that guaranteed that their technology choice would have long term support.

[Eclipse 3.0 Architecture]

Figure 6.6: Eclipse 3.0 Architecture

Looking at the 3.0 architecture in Figure 6.6, you will note that the Eclipse Runtime still exists to provide the application model and extension registry. Managing the dependencies between components, the plugin model is now managed by OSGi. In addition to continuing to be able to extend Eclipse for their own IDEs, consumers can also build upon the RCP application framework for more generic applications.

6.3. Eclipse 3.4

The ability to easily update an application to a new version and add new content is taken for granted. In Firefox it happens seamlessly. For Eclipse it hasn't been so easy. Update Manager was the original mechanism that was used to add new content to the Eclipse install or update to a new version.

To understand what changes during an update or install operation, it's necessary to understand what Eclipse means by "features". A feature is a PDE artifact that defines a set of bundles that are packaged together in a format that can be built or installed. Features can also include other features. (See Figure 6.7.)

[Eclipse 3.3 SDK Feature Hierarchy]

Figure 6.7: Eclipse 3.3 SDK Feature Hierarchy

If you wished to update your Eclipse install to a new build that only incorporated one new bundle, the entire feature had to be updated since this was the coarse grained mechanism that was used by update manager. Updating a feature to fix a single bundle is inefficient.

There are PDE wizards to create features, and build them in your workspace. The feature.xml file defines the bundles included in the feature, and some simple properties of the bundles. A feature, like a bundle, has a name and a version. Features can include other features, and specify version ranges for the features they include. The bundles that are included in a feature are listed, along with specific properties. For instance, you can see that the org.eclipse.launcher.gtk.linux.x86_64 fragment specifies the operating system (os), windowing system (ws) and architecture (arch) where it should be used. Thus upgrading to a new release, this fragment would only be installed on this platform. These platform filters are included in the OSGi manifest of this bundle.

<?xml version="1.0" encoding="UTF-8"?>



   <license url="%licenseURL">



An Eclipse application consists of more than just features and bundles. There are platform specific executables to start Eclipse itself, license files, and platform specific libraries, as shown in this list of files included in the Eclipse application.

These files couldn't be updated via update manager, because again, it only dealt with features. Since many of these files were updated every major release, this meant that users had to download a new zip each time there was a new release instead of updating their existing install. This wasn't acceptable to the Eclipse community. PDE provided support for product files, which specified all the files needed to build an Eclipse RCP application. However, update manager didn't have a mechanism to provision these files into your install which was very frustrating for users and product developers alike. In March 2008, p2 was released into the SDK as the new provisioning solution. In the interest of backward compatibility, Update Manager was still available for use, but p2 was enabled by default.

6.3.1. p2 Concepts

Equinox p2 is all about installation units (IU). An IU is a description of the name and id of the artifact you are installing. This metadata also describes the capabilities of the artifact (what is provided) and its requirements (its dependencies). Metadata can also express applicability filters if an artifact is only applicable to a certain environment. For instance, the org.eclipse.swt.gtk.linux.x86 fragment is only applicable if you're installing on a Linux gtk x86 machine. Fundamentally, metadata is an expression of the information in the bundle's manifest. Artifacts are simply the binary bits being installed. A separation of concerns is achieved by separating the metadata and the artifacts that they describe. A p2 repository consists of both metadata and artifact repositories.

[P2 Concepts]

Figure 6.8: P2 Concepts

A profile is a list of IUs in your install. For instance, your Eclipse SDK has a profile that describes your current install. From within Eclipse, you can request an update to a newer version of the build which will create a new profile with a different set of IUs. A profile also provides a list of properties associated with the installation, such as the operating system, windowing system, and architecture parameters. Profiles also store the installation directory and the location. Profiles are held by a profile registry, which can store multiple profiles. The director is responsible for invoking provisioning operations. It works with the planner and the engine. The planner examines the existing profile, and determines the operations that must occur to transform the install into its new state. The engine is responsible for carrying out the actual provisioning operations and installing the new artifacts on disk. Touchpoints are part of the engine that work with the runtime implementation of the system being installed. For instance, for the Eclipse SDK, there is an Eclipse touchpoint which knows how to install bundles. For a Linux system where Eclipse is installed from RPM binaries, the engine would deal with an RPM touchpoint. Also, p2 can perform installs in-process or outside in a separate process, such as a build.

There were many benefits to the new p2 provisioning system. Eclipse install artifacts could be updated from release to release. Since previous profiles were stored on disk, there was also a way to revert to a previous Eclipse install. Additionally, given a profile and a repository, you could recreate the Eclipse install of a user that was reporting a bug to try to reproduce the problem on your own desktop. Provisioning with p2 provided a way to update and install more than just the Eclipse SDK, it was a platform that applied to RCP and OSGi use cases as well. The Equinox team also worked with the members of another Eclipse project, the Eclipse Communication Framework (ECF) to provide reliable transport for consuming artifacts and metadata in p2 repositories.

There were many spirited discussions within the Eclipse community when p2 was released into the SDK. Since update manager was a less than optimal solution for provisioning your Eclipse install, Eclipse consumers had the habit of unzipping bundles into their install and restarting Eclipse. This approach resolves your bundles on a best effort basis. It also meant that any conflicts in your install were being resolved at runtime, not install time. Constraints should be resolved at install time, not run time. However, users were often oblivious to these issues and assumed since the bundles existed on disk, they were working. Previously, the update sites that Eclipse provided were a simple directory consisting of JARred bundles and features. A simple site.xml file provided the names of the features that were available to be consumed in the site. With the advent of p2, the metadata that was provided in the p2 repositories was much more complex. To create metadata, the build process needed to be tweaked to either generate metadata at build time or run a generator task over the existing bundles. Initially, there was a lack of documentation available describing how to make these changes. As well, as is always the case, exposing new technology to a wider audience exposed unexpected bugs that had to be addressed. However, by writing more documentation and working long hours to address these bugs, the Equinox team was able to address these concerns and now p2 is the underlying provision engine behind many commercial offerings. As well, the Eclipse Foundation ships its coordinated release every year using a p2 aggregate repository of all the contributing projects.

6.4. Eclipse 4.0

Architecture must continually be examined to evaluate if it is still appropriate. Is it able to incorporate new technology? Does it encourage growth of the community? Is it easy to attract new contributors? In late 2007, the Eclipse project committers decided that the answers to these questions were no and they embarked on designing a new vision for Eclipse. At the same time, they realized that there were thousands of Eclipse applications that depended on the existing API. An incubator technology project was created in late 2008 with three specific goals: simplify the Eclipse programming model, attract new committers and enable the platform to take advantage of new web-based technologies while providing an open architecture.

[Eclipse 4.0 SDK Early Adopter Release]

Figure 6.9: Eclipse 4.0 SDK Early Adopter Release

Eclipse 4.0 was first released in July 2010 for early adopters to provide feedback. It consisted of a combination of SDK bundles that were part of the 3.6 release, and new bundles that graduated from the technology project. Like 3.0, there was a compatibility layer so that existing bundles could work with the new release. As always, there was the caveat that consumers needed to be using the public API in order to be assured of that compatibility. There was no such guarantee if your bundle used internal code. The 4.0 release provided the Eclipse 4 Application Platform which provided the following features.

6.4.1. Model Workbench

In 4.0, a model workbench is generated using the Eclipse Modeling Framework (EMFgc). There is a separation of concerns between the model and the rendering of the view, since the renderer talks to the model and then generates the SWT code. The default is to use the SWT renderers, but other solutions are possible. If you create an example 4.x application, an XMI file will be created for the default workbench model. The model can be modified and the workbench will be instantly updated to reflect the changes in the model. Figure 6.10 is an example of a model generated for an example 4.x application.

[Model Generated for Example 4.x Application]

Figure 6.10: Model Generated for Example 4.x Application

6.4.2. Cascading Style Sheets Styling

Eclipse was released in 2001, before the era of rich Internet applications that could be skinned via CSS to provide a different look and feel. Eclipse 4.0 provides the ability to use stylesheets to easily change the look and feel of the Eclipse application. The default CSS stylesheets can be found in the css folder of the org.eclipse.platform bundle.

6.4.3. Dependency Injection

Both the Eclipse extensions registry and OSGi services are examples of service programming models. By convention, a service programming model contains service producers and consumers. The broker is responsible for managing the relationship between producers and consumers.

[Relationship Between Producers and Consumers]

Figure 6.11: Relationship Between Producers and Consumers

Traditionally, in Eclipse 3.4.x applications, the consumer needed to know the location of the implementation, and to understand inheritance within the framework to consume services. The consumer code was therefore less reusable because people couldn't override which implementation the consumer receives. For example, if you wanted to update the message on the status line in Eclipse 3.x, the code would look like:


Eclipse 3.6 is built from components, but many of these components are too tightly coupled. To assemble applications of more loosely coupled components, Eclipse 4.0 uses dependency injection to provide services to clients. Dependency injection in Eclipse 4.x is through the use of a custom framework that uses the the concept of a context that serves as a generic mechanism to locate services for consumers. The context exists between the application and the framework. Contexts are hierarchical. If a context has a request that cannot be satisfied, it will delegate the request to the parent context. The Eclipse context, called IEclipseContext, stores the available services and provides OSGi services lookup. Basically, the context is similar to a Java map in that it provides a mapping of a name or class to an object. The context handles model elements and services. Every element of the model, will have a context. Services are published in 4.x by means of the OSGi service mechanism.

[Service Broker Context]

Figure 6.12: Service Broker Context

Producers add services and objects to the context which stores them. Services are injected into consumer objects by the context. The consumer declares what it wants, and the context determines how to satisfy this request. This approach has made consuming dynamic service easier. In Eclipse 3.x, a consumer had to attach listeners to be notified when services were available or unavailable. With Eclipse 4.x, once a context has been injected into a consumer object, any change is automatically delivered to that object again. In other words, dependency injection occurs again. The consumer indicates that it will use the context by the use of Java 5 annotations which adhere to the JSR 330 standard, such as @inject, as well as some custom Eclipse annotations. Constructor, method, and field injection are supported. The 4.x runtime scans the objects for these annotations. The action that is performed depends on the annotation that's found.

This separation of concerns between context and application allows for better reuse of components, and absolves the consumer from understanding the implementation. In 4.x, the code to update the status line would look like this:

IStatusLineManager statusLine;
⋮    ⋮    ⋮

6.4.4. Application Services

One of the main goals in Eclipse 4.0 was to simplify the API for consumers so that it was easy to implement common services. The list of simple services came to be known as "the twenty things" and are known as the Eclipse Application services. The goal is to offer standalone APIs that clients can use without having to have a deep understanding of all the APIs available. They are structured as individual services so that they can also be used in other languages other than Java, such as Javascript. For example, there is an API to access the application model, to read and modify preferences and report errors and warnings.

6.5. Conclusion

The component-based architecture of Eclipse has evolved to incorporate new technology while maintaining backward compatibility. This has been costly, but the reward is the growth of the Eclipse community because of the trust established that consumers can continue to ship products based on a stable API.

Eclipse has so many consumers with diverse use cases and our expansive API became difficult for new consumers to adopt and understand. In retrospect, we should have kept our API simpler. If 80% of consumers only use 20% of the API, there is a need for simplification which was one of the reasons that the Eclipse 4.x stream was created.

The wisdom of crowds does reveal interesting use cases, such as disaggregating the IDE into bundles that could be used to construct RCP applications. Conversely, crowds often generate a lot of noise with requests for edge case scenarios that take a significant amount of time to implement.

In the early days of the Eclipse project, committers had the luxury of dedicating significant amounts of time to documentation, examples and answering community questions. Over time, this responsibility has shifted to the Eclipse community as a whole. We could have been better at providing documentation and use cases to help out the community, but this has been difficult given the large number of items planned for every release. Contrary to the expectation that software release dates slip, at Eclipse we consistently deliver our releases on time which allows our consumers to trust that they will be able to do the same.

By adopting new technology, and reinventing how Eclipse looks and works, we continue the conversation with our consumers and keep them engaged in the community. If you're interested in becoming involved with Eclipse, please visit


  3. For example: