To Couple or Not to Couple; That is the Question

I’ve reached another point of contemplation in my efforts to build the greatest media player ever. This is an architectural matter about application development.

The puzzle in question is the use of loosely coupled application architecture versus event-driven application architecture.

Loosely Coupled Applications

In a loosely coupled application, interfaces define contracts for various components of the application. Dependent code is then given (preferably injected with) a reference to an instance of that interface. The implementation may or may not be a singleton, but it does not matter to the dependent code. It has a handle on an instance of the interface it requires to do its job and can execute accordingly.

Event-Driven Applications

At the other end of the spectrum, an event-based application would have one component broadcasting a message while another component listens for that message. The two components know nothing about each other and neither have any direct interaction. In this case, it is up to a messaging bus to deliver messages from one component to another. The responsibility of the sender ends when it sends the message, but the sender may also be a receiver, listening for a response message of some kind.

Comparison

The biggest difference in the two schools of thought here is that the event-driven architecture is asynchronous. When a sender broadcasts a message, there’s no guarantee that the response will come anytime soon. This works well for environments like JavaScript where services often cross the client/server boundary in order to fulfill a request, such as in AJAX.

A second advantage of event-driven architecture is the ease with which additional behaviors can be added to an ecosystem. Adding behaviors in response to a message is as simple as adding a new listener for that message; neither the existing receiver(s) nor the sender need to be modified.

Although the asynchronous nature of event-driven applications sounds promising, in actuality, it can create a bit of burdon on the developer. You no longer have any coupling whatsoever (not even loose) between senders and the receivers. This can make it difficult to read and understand code; there are no contextual clues as to who might be listening for a message. This can make it difficult to refactor code if you aren’t able to track down all of the listeners of a particular type of message.

An Example

Since this is in reference to Project: Apollo, I’ll use it as a reference.

The underlying media player in Project: Apollo supports some obvious controls: play, pause, next, and so on. To capture this, I created an interface:

public interface TransportControl {
    void next();
    void pause();
    void play();
    void previous();
    void seek(double percent);
    void stop();
}

There are several components throughout Project: Apollo that need to be able to control playback. The GUI, for example, has buttons that correspond to these functions. The playlist also invokes these controls to advance the player to the next track.

In a loosely-coupled application, an instance of TransportControl would be injected into the GUI as a collaborator. Then, the GUI could respond to a user click by directly invoking the next() method on the TransportControl. This keeps the code clear and easy to read, but it actually creates another problem. The playlist still needs to be informed when the GUI has invoked a “next” command. Some mechanism must exist to instruct the underlying player what to play next.

In the event-driven architecture, the GUI would simply broadcast a message requesting that the next item be played. In this case, the TransportControl would receive the “next” message and know that it should stop playing the current item. The PlaylistManager would also receive this message and instruct the TransportControl to load a new item. Further, a debug logger can log out the entire process, providing detailed messages about what is happening.

Conclusion and Comments

At this point, I am unsure of the best solution. Clearly there is more than one way to solve this problem and some folks I know would recommend I simply make something work and move on; it can always be refactored later. Good advice for a project on a timeline for budget, but this is a learning experience where I can really explore these types of questions. My instinct tells me the right answer is some manner of balance between the loosely-coupled option and the event-driven option.

So does it make sense to inject the GUI with an instance of the TransportControl? Or does it make more sense for components wishing to control playback to broadcast a message according to their intent? Thoughts?

UPDATE: After reading in more detail on the Wikipedia entries I linked to above, I think the best solution for my example case is a complementary application of both event-based and service-based architectures. Specifically, changes in state throughout the application should be expressed as events, but components that need to invoke certain behavior should do so on a service object or proxy, even if that proxy merely relays an event under the hood.

Using GWT’s Application Framework Judiciously

I am very excited to see the inclusion of a proper application framework in the upcoming Google Web Toolkit 2.1 release. The addition of this framework is a clear signal that GWT is a serious library for enterprise scale applications.

As a seasoned developer, I am very familiar with the traditional Model-View-Controller approach to building web applications. The MVC approach works great for server-based frameworks like Spring MVC or Struts. However, ever since I started working with GWT way back at version 1.4, I’ve had a hard time mapping MVC into a truly client-side application where the entire GUI is built in GWT.

The MVC approach is tried and true in the desktop software world. But on the web, there are a few extra quirks that make traditional MVC a little tricky. Here are some of the challenges I’ve run into.

History Management and Support

Because it is a web application running in a browser, there is an implied expectation that the forward and back buttons work, regardless of whether your web application is mostly client- or server-side. This presents a unique challenge to MVC because most desktop applications don’t have forward or back buttons.

Bookmark Support

Another important feature to any good web application is the ability to bookmark or share the link you’re currently viewing. It should be possible to bookmark a page or paste the current URL into an email, for example, and restore the state of the application based solely on that URL.

Composite View Structure

When working with traditional MVC frameworks like Spring MVC, the process is pretty straight forward. For any given URL, only a single Controller responds to the request. The Controller prepares a single Model and that Model is handed off to a single View. Common elements of the UI can be rendered with include statements, but the View generally has a very direct relationship with the URL.

Conversely, with a more desktop-like application built in GWT, the UI is frequently built as a Composite. There are several components in play at any given time and there isn’t necessarily a straight forward relationship between the current URL and the current set of components viewable to the user. This makes intra-application messaging and navigation a bit more challenging; you can’t just rely on a DispatcherServlet, for example, to coordinate user interaction.

Enter the GWT Application Framework

Clearly, the new App module in GWT is designed to address many of these concerns. It’s great to see History Management and the notion modular composite structure being supported as first class features.

The application framework in GWT is all new to me, though, and it isn’t really MVC. Instead, it’s more of a Model-View-Activity (or -Presenter) approach. In this case, the Activity takes the place of the Controller and is responsible for only the slice of the UI represented by its View. This MVA approach handles application navigation very well, assuming your application is, like a traditional web application, a series of screens or “activities” through which the user navigates linearly.

So what happens when the UI dictates that two or more of these activities be present at once? Well GWT can handle that as well by using multiple ActivityManagers. The ActivityManager is responsible for coordinating all of the Activities for a designated Display – the area of UI over which control can be delegated to an Activity.

This is all well and good, but Project: Apollo presented me with a unique challenge in this new MVA architecture. What do you do with UI components that are always present? Case in point: Apollo has a control bar that is always present in the UI, similar to the button on the front face of a DVD player.

The control bar allows the user to control playback and displays pertinent playback information such as the elapsed time and name of the current track. The control bar, by nature of its omnipresence, does not participate in the navigational functions of the ActivityManager, nor is there any other Activity that would fill its slot.

I’ve also gleaned through GWT code in trunk that the usual way of handling Activities is through an ActivityMapper; the ActivityMapper is responsible for resolving and creating instances of Activities and their corresponding Views. However, control bar isn’t managed as such. I have a self-imposed rule in development of Project: Apollo that UI components should be UiBinder compatible; I should be able to include the control bar in my UI declaratively through the constructs afforded by UiBinder. By contrast, Activities depend on the current state and history of the application, so their lifecycle would not be appropriate to manage through UiBinder, hence the Display slot concept.

Bring Back the Controller

The Java docs for Activity state:

Implemented by objects that control a piece of user interface, with a life cycle managed by an ActivityManager, in response to PlaceChangeEvent events as the user navigates through the app.

It’s easy to see that my control bar doesn’t strictly qualify as an Activity. In this case, MVA isn’t a good fit; I need a different approach. One option, for simpler components, is to let the UiBinder template be the “view” and the backing Widget code be the “controller”. For more complex components, though, I still want to break out application logic to a separate class and leave UI-centric logic in the View-backing code.

Taking a hint from previous iterations of GWT’s application framework, I am choosing to call my pseudo-controllers “delegates”. The Delegate represents a code structure to which the actual View will delegate user interaction for handling. Delegates contain familiar Activity-esque functions to facilitate lifecycle events where necessary.

public interface Delegate<W extends IsWidget> {
    void shutdown();
    void startup(W widget, EventBus eventBus);
}

For my control bar, then, I have broken the code into a ControlBarWidget (the View), a ControlBarDelegate (the Delegate), and a ControlBarModel (the Model).

To keep with my requirements that the control bar be UiBinder compatible, I am allowing the View to handle creation of a default implementation of the Delegate. I’m using GWT.create() to instantiate it so the actual implementation can always be easily swapped out later via deferred binding, but for now, the need to swap out implementations is low to non-existent.

The Judicious Part

The lesson I am driving at here is that the new application framework in GWT may be great for some portions of an application, but it probably isn’t appropriate for everything. Parts of the application that do not participate in history or navigation (such as global menus or controls) are not good candidates for Activities.

I think GWT and the new application framework represent a fantastic toolset for building user interfaces and entire applications. You just have to be smart about when and how you use them.