Testable Navigation with MVVM Fabric

One of the biggest problems I encountered while writing WPF apps with MVVM is navigation.  The traditional approach to navigation would be to have one view create the view being navigated to.  Now, there’s a couple of issues with that approach which make it less than desirable.  The first issue is that having one view create another view heavily couples the two views.  The second issue is that this approach quickly becomes very difficult (if not impossible) to unit test.

Things get more complicated when view models are introduced.  Since the functionality that initiates the navigation is going to be in the view model, that would mean the view model would have to know about the new view.  When I structure an MVVM solution for WPF, I create a separate class library for the view models.  The client project (which contains the views) needs to have a reference to the view models project in order to use them.  That means there would be a circular dependency if the view model project had a reference to the view project as well.  Doing the traditional approach with this project structure is no longer even an option.

MVVM Fabric solves the navigation problem with some convention, some configuration, a pinch of dependency injection and a dash of the message bus.

10,000 ft View

MVVM Fabric funnels all navigation through one source, the ViewController.  The ViewController subscribes to the message bus to listen for ShowViewMessages requesting views.  When a view request comes through, the ViewController coordinates the resolution and creation of the view and its view model.  Then the ViewController takes the new view and passes it along to the view placer to be placed in the application.

By leaning on the message bus to request a view, writing unit tests for the view model making the request becomes a breeze.  By letting all navigation flow through the ViewController, it de-clutters your codebase by eliminating the need for code creating views throughout the application.  This decoupling will greatly simplify your application and each view model can focus on its own responsibilities and not have to worry about the specifics of any other view or view model.


As I mentioned, there is an element of convention required to leverage the navigation model provided by MVVM Fabric.  The conventions are as follows:

  1. All view models must have a default constructor.
  2. View models have the option of providing a Load method as a way of allowing itself to be loaded during the dynamic creation process provided by MVVM Fabric.  The Load method may take 0 or 1 arguments.
    1. If there is an argument, it can be whatever is required by the view model to load itself.
    2. If more than one thing is required to load the view model, I recommend creating some sort of criteria class that contains any required data.
  3. A class inheriting from ViewTargets must be provided with unique definitions for all navigable views in the application.

The reason that all view models must have a default constructor is that they are dynamically instantiated when a view is loaded.  Since there are many scenarios where a view model will need some information to load itself (picture a detail view, for example), MVVM Fabric looks for a public Load method on the view model and calls it if found.  That Load method may be a no argument method which simply kick-starts the loading needed for the view model or it may take a single parameter.  If the Load method takes a parameter, a parameter must be provided when the view is requested.  That’s where the ShowViewMessage comes into play.

The ShowViewMessage has a ViewTargets parameter (which tells the ViewController which view is being requested) and an optional parameter called LoadArgs.  LoadArgs can be anything required by the view model, but if something is provided a Load method which takes that object type as a parameter must exist on the view model or an exception will be thrown.

If you saw my presentation WPF with MVVM: From the Trenches, you heard me talking about how ViewTargets is an enumeration.  While refactoring my navigation code into a stand-alone library for MVVM Fabric, one of the issues I had to overcome is the fact that enumerations cannot be inherited from and extended from an external object.  The reason I liked using an enumeration to articulate all the navigable views in the first place is because it provided a level of safety at compile time when views are requested.  I could have used strings instead to identify views, but that would not make refactoring easy.  To overcome this limitation of enums, I leveraged a concept introduced by Hugh Ang to create an Enumeration class which acts like an enum, but can be extended.  It’s not a perfect solution, but it does help address the compile-time and refactoring issues.  The new ViewTargets inherits from this Enumeration class and must be further inherited from in your implementation.


In order to leverage the navigation model provided by MVVM Fabric, you must provide implementations to an interface or two: IViewPlacer and optionally IViewAuthorizer.

IViewPlacer is what will actually place the views when the navigation model has loaded the view and it’s view model.  I leave that implementation up to you because each application is different in how views are presented.  The sample application provides an implementation that works with the tabbed interface used by Movies.  MVVM Fabric does contain an implementation of IViewPlacer for views displayed as modal dialogs because this will likely be the same across applications.  The sample application shows the usage of ModalViewPlacer in conjunction with it’s custom IViewPlacer to illustrate how you could implement a combination of the two.

IViewAuthorizer is an interface which can optionally be implemented to give you control over whether the user is authorized to see each view before it is shown.  This allows you to put restrictions on your views from one location.

Enough talk, time for code!

Here’s the part you are probably waiting for, how do I make it work?  Here are the steps that you need to take to get up and running.

1.  Inherit from ViewTargets and provide values for your views.

2.   Provide an implementation for IViewPlacer (unless you strictly use ModalViewPlacer)

Note: Your custom ViewPlacer would also be a good place to subscribe to CloseViewMessages from the message bus.  Since views have the ability to request that they be closed, it only makes sense to keep the showing and closing functionality in the same place.  You can see this in action with the sample application.

3.  Provide a way to wire up the IViewConfigurationResolver when the application starts up.

Note: The sample application calls ContainerConfiguration.InitContainer() in the application’s OnStartup method.  I also do some configuration for Windsor Container in this method.

4 (Optional).  Provide an implementation for IViewAuthorizer.  This one is up to you since application security varies on a case-by-case basis.

5.  Instantiate the ViewController and all the pieces it needs.  This will probably be done in the code-behind of your main window.  The sample application uses Castle Windsor, so I will leave it up to you how you get all the pieces standing.

You’re There!

The steps described above may seem like a lot of work, but the work is almost all front-loaded.  When a new view is added, only steps 1 and 3 must be revised to accommodate the new view.

Now, let’s look at how to initiate navigation.

That’s it!  This example shows how to pass a parameter to the Detail view.  You could remove the LoadArgs parameter from the ShowViewMessage for those cases where there is either no need for a Load method on the view model or the Load method has no parameters.

From a testing perspective, the above example is extremely simple to write unit tests for.  You would simply assert that the message bus received a ShowViewMessage with the correct arguments.


The navigation model provided by MVVM Fabric takes a pain point with the MVVM pattern and smooths it over in a testable manner.  With all the navigation being handled in one place, it reduces the surface area that you must test for in your application.  A testable application is a maintainable application!

MVVM Fabric: Using MessageBus

While working with MVVM and trying to keep clean separation of view models, it becomes clear that there is a need for some sort of communication mechanism to allow view models to interact with each other without knowing about each other.  A message bus works perfectly to fit this need.

Message Bus Overview

The message bus acts as a communication switchboard for your application.  All messages route through the message bus and the message bus gets them to any interested parties.

The message bus is an example of the publisher/subscriber pattern, similar to events in .NET.  However, events require at least one object to know about the other in order to subscribe to the event.  With the message bus, objects only need to know about the message bus to subscribe to or publish messages.  All messages go through a single instance of the message bus and that single instance must be available throughout the application or the pattern quickly breaks down.

With the message bus, there is no longer the need to bubble events from one object to the next, eventually getting them to an object that cares.  Objects subscribe to specific message types.  When that type of message is published through the message bus, the message bus passes the message along to any subscribers.  If there are no subscribers, the message quietly goes away.

A message is simply a class.  It can be an empty class, which acts more as a signal, or a sort of data transfer object, which contains relevant information.  The message bus doesn’t care what the message is, but it does care about the message’s type.  The message’s type is how the message bus knows who is subscribed.

A while back I wrote a post describing how to develop a simple message bus.  The message bus I wrote for MVVM Fabric uses the same interface I described in that post, but has a more robust implementation.  More specifically, I added thread-safety and the use of WeakReferences so that it doesn’t promote memory leaks.  The message bus provides a clean interface which can be used to send and receive messages throughout the application.  Let’s take a look at an example.

Message Bus In Action

First of all, here is a simple example of a message.

Notice that the message doesn’t derive from anything (though it very well could).  In this case, I am using the message to pass keywords used for a search.  Next up is the class which will publish the message.

I only included the important parts for brevity. Notice that we will be using the IMessageBus interface, so that this can be mocked out in testing.  The Search method creates a new instance of the SearchMessage and calls Publish on the message bus passing the message as a parameter and using the message’s type as the generic argument.

To make sure that the message doesn’t fall on deaf ears, we need a subscriber.

Notice the method called HandleSearch, which takes a SearchMessage as a parameter.  The constructor subscribes the HandleSearch method to messages of type SearchMessage and is all ready to search.


The message bus is something that I have used frequently in my travels with WPF and MVVM to keep my view models decoupled.  It provides a simple way to provide fire and forget functionality in a testable manner.  Using the message bus in your application will allow you to keep your view models nicely de-coupled.

Roll Your Own Simple Message Bus / Event Aggregator

I was recently working on a project that had a need for some sort of centralized messaging system.  After doing some research into the matter, I decided to move forward with the Event Aggregator from the Prism team.  Prism’s Event Aggregator worked fine, but there were a few things about it that I found cumbersome.  While preparing for a presentation I gave recently, called “WPF with MVVM: From the Trenches,” I began to think about what it would take to create my own event aggregator that would smooth over some of the bumps.

First up, lets talk about the name, “Event Aggregator.”  During the course of our project, we used the event aggregator to publish data (or messages) in addition to events.  During a discussion about the event aggregator, Jason Bock pointed out that it is really more of a “Message Bus” than an “Event Aggregator.”  I agree.  So, for the remainder of this post, I will refer to it as a “Message Bus.”

The next bump that I wanted to smooth over was the fact that, with Prism, we had to create both an Event class which derived from CompositePresentationEvent and an EventArgs class for every message we wanted to publish.  With Prism, I wrote an extension method, called GetEventViaArgs(), which would get the CompositePresentationEvent based on the type of the args.  That eliminated the need for an explicit Event class, but that also made me question the need for an Event at all.  What if the message type was all that we needed for subscribing and publishing?  Eliminating the need for an Event would simplify testing as well because we wouldn’t have to mock out both getting the event and subscribing/publishing the message, which so happens to be another bump that I wanted to smooth over.

A Starting Point

So, lets take a look at the method signatures that I had in mind.  A Message Bus should provide functionality to Subscribe to a message, Unsubscribe from a message and Publish a message.

This looks straight-forward enough.  Whatever implementation I come up with, it should leverage generics to make sure all interested parties are notified of messages they care about.  The method signature I would like to see for the callbacks would be something like this:

Since I am a big fan of dependency injection and inversion of control, I wanted to define an interface, called IMessageBus, to be used instead of a concrete implementation.  Keeping in mind how I wanted to use this bad boy, here is the interface I came up with:

Let me break this down a bit…  TMessage is the Type of the message to be sent.  Action is a way for me to take, as a parameter, a method which has a void return value and takes a parameter of type TMessage.  Also note that TMessage is just a class that contains whatever data we need to convey the message.

So far, so good.  We now have a contract in place for our message bus.  Next up is implementation.

Making It Happen

First, we need a data structure that will have some sort of key/value association.  The key will be the Type of TMessage and the value will be the collection of Action which will be called when a message of type TMessage is published.  Since I plan to use generics at the method level, I decided to use Object in place of Action, then cast it before using it.  So, here is our MessageBus class so far:

Now that we have a data structure in place to keep track of subscribers, time to implement the Subscribe method.  Now, when a new subscriber is added, we need to first find out if there is already a subscriber list for that message type.  If there is, use it.  Otherwise, create a new one.

Next up is Unsubscribe.  When a subscriber makes a call to Unsubscribe, we need to make sure they are actually a subscriber in the first place and remove them if they are.  Also, if they are the last subscriber, we can go ahead and remove the subscriber list for that message type.

Almost there.  Now we just need to be able to Publish.  As you probably noticed above, I thought it would be nice to both explicitly define what type of message to publish and let the message type be determined dynamically.  For either case, we need to get the list of subscribers for the type of message being published and invoke each of their callback methods with the message as a parameter.  First up is the explicitly define flavor:

Pretty straight-forward here too.  If there are subscribers for that message type, iterate over them (casting them in the process) and invoke them.

The dynamic version of Publish is a little more involved, only because we have to use reflection in place of generics due to the dynamic aspect:

This one is essentially doing the same thing as the other Publish method.  If there are subscribers for that message type, iterate over them (finding their Invoke method in the process) and invoke them.  Don’t be confused by the two different Invokes that you see…  The first Invoke is the name of the method on the Action type and the second Invoke is what reflection uses to execute a method.

There we have it!  A simple message bus that is easy to use.  But how about testing with it?  Since testing was one of the bumps I wanted to smooth over, how did we do?  Let’s find out!

Testing Our Shiny New Message Bus

I use RhinoMocks, so here is how I would verify that some code is publishing the correct message:

I generate a stub for IMessageBus which will set my local searchMessage variable to the actual SearchMessage that is published.  I then could use searchMessage to verify that the message that is published has the correct data.

As an alternative, if all I wanted to do was verify that Publish was called on my Message Bus, I could do the following:

The key changes here are that I am using a Mock instead of a Stub and I am setting up an expectation.  That way I can verify that Publish was called without caring about the details:

Testing for messages being published is now much cleaner with our new MessageBus than it would be with Prism’s Event Aggregator.  But what about testing the handling of these messages that get published?  Well, the best way that I could come up with for that is the same way I did it with Prism’s Event Aggregator, which is to use the concrete implementation of the MessageBus and actually subscribe and publish with it in the unit test.  Oh well, at least we were able to clean up some of the testing aspect!


So, here we are at the end.  My goal here was to show that it isn’t all that difficult to implement your own message bus system or, at the very least, show the general concept behind how it works.  I hope that this has been useful.  If you want play with the code that we have written, as well as the tests that I wrote to beat on it a little, you can download the source here: