12

I'm building a WPF application using the MVVM pattern. Right now, my viewmodels calls the service layer to retrieve models (how is not relevant to the viewmodel) and convert them to viewmodels. I'm using constructor injection to pass the service required to the viewmodel.

It's easily testable and works well for viewmodels with few dependencies, but as soon as I try to create viewModels for complex models, I have a constructor with a LOT of services injected in it (one to retrieve each dependencies and a list of all available values to bind to an itemsSource for example). I'm wondering how to handle multiple services like that and still have a viewmodel that I can unit test easily.

I'm thinking of a few solutions:

  1. Creating a services singleton (IServices) containing all the available services as interfaces. Example: Services.Current.XXXService.Retrieve(), Services.Current.YYYService.Retrieve(). That way, I don't have a huge constructor with a ton of services parameters in them.

  2. Creating a facade for the services used by the viewModel and passing this object in the ctor of my viewmodel. But then, I'll have to create a facade for each of my complexe viewmodels, and it might be a bit much...

What do you think is the "right" way to implement this kind of architecture ?

  • I think the "right" way to do it is to create a separate layer that calls the services and does whatever casting is necessary to create the ViewModel. Your ViewModels should not be responsible for creating themselves. – Amy Blankenship Nov 4 '13 at 18:29
  • @AmyBlankenship: View Models shouldn't have to (or even necessarily be able to) create themselves, but inevitably will sometimes be responsible for creating other view models. An IoC container with automatic-factory support is a huge help here. – Aaronaught Nov 4 '13 at 18:52
  • "Will sometimes" and "should" are two different animals ;) – Amy Blankenship Nov 4 '13 at 19:34
  • @AmyBlankenship: Are you suggesting that view models shouldn't create other view models? That's a tough pill to swallow. I can understand saying that view models shouldn't use new to create other view models, but think of something as simple as an MDI application where clicking on a "new document" button or menu will add a new tab or open a new window. The shell/conductor must be able to create new instances of something, even if it's hidden behind one or a few layers of indirection. – Aaronaught Nov 4 '13 at 21:15
  • Well, certainly it must have the ability to request that a View be made somewhere. But to make it itself? Not in my world :). But then again, in the world I live in, we call VM "Presentation Model." – Amy Blankenship Nov 4 '13 at 23:15
21

In fact, both of these solutions are bad.

Creating a services singleton (IServices) containing all the available services as interfaces. Example: Services.Current.XXXService.Retrieve(), Services.Current.YYYService.Retrieve(). That way, I don't have a huge constructor with a ton of services parameters in them.

This is essentially the Service Locator Pattern, which is an anti-pattern. If you do this, you will no longer be able to understand what the view model actually depends on without looking at its private implementation, which will make it very difficult to test or refactor.

Creating a facade for the services used by the viewModel and passing this object in the ctor of my viewmodel. But then, I'll have to create a facade for each of my complexe viewmodels, and it might be a bit much...

This isn't so much an anti-pattern but it is a code smell. Essentially you're creating a parameter object, but the point of the PO refactoring pattern is to deal with parameter sets that are used frequently and in a lot of different places, whereas this parameter would only ever be used once. As you mention, it would create a lot of code bloat for no real benefit, and wouldn't play nice with a lot of IoC containers.

In fact, both of the above strategies are overlooking the overall issue, which is that coupling is too high between view models and services. Simply hiding these dependencies in a service locator or parameter object does not actually change how many other objects the view model depends on.

Think of how you would unit-test one of these view models. How big is your setup code going to be? How many things need to be initialized in order for it to work?

A lot of people starting out with MVVM try to create view models for an entire screen, which is fundamentally the wrong approach. MVVM is all about composition, and a screen with many functions should be composed of several different view models, each of which depends on only one or a few internal models/services. If they need to communicate with each other, you do so via pub/sub (message broker, event bus, etc.)

What you actually need to do is refactor your view models so that they have fewer dependencies. Then, if you need to have an aggregate "screen", you create another view model to aggregate the smaller view models. This aggregate view model doesn't have to do very much by itself, so it in turn is also fairly easy to understand and test.

If you've done this properly, it should be obvious just from looking at the code, because you'll have short, succinct, specific, and testable view models.

  • Yeah, that's what I'll probably end up doing ! Thanks a lot, sir. – alfa-alfa Nov 4 '13 at 19:20
  • Well, I already assumed he already tried this, but didn't succeed. @alfa-alfa – Euphoric Nov 4 '13 at 19:35
  • @Euphoric: How do you "not succeed" in this? As Yoda would say: Do or do not, there is no try. – Aaronaught Nov 4 '13 at 20:48
  • @Aaronaught For example he really needs all the data in single viewmodel. Maybe he has grid and different columns come from different services. You cannot do that with composition. – Euphoric Nov 4 '13 at 20:57
  • @Euphoric: Actually, you can solve that with composition, but that can be done beneath the view model level. It's simply a matter of creating the right abstractions. In that case, you only need one service to handle the initial query to get a list of IDs, and a sequence/list/array of "enrichers" which annotate with their own information. Make the grid itself its own view model, and you've solved the problem with effectively two dependencies, and it's extremely easy to test. – Aaronaught Nov 4 '13 at 21:13
1

I could write a book about this...in fact I am ;)

First off, there is no universally "right" way to do things. You have to take other factors into consideration.

It's possible that your services are too fine grained. Wrapping the Services with Facades that provide the interface that a specific Viewmodel or even a cluster of related ViewModels would use might be a better solution.

Simpler still would be to wrap the services into a single Facade that all viewmodels use. Of course that can potentially be a very big interface with a lot of unnecessary functions for the average scenario. But I would say it's no different than a message router that handles every message in your system.

In fact, what I've seen a lot of architectures eventually evolve into is a message bus built around something like the Event Aggregator pattern. Testing is easy there because your test class just registers a listener with the EA and fires the appropriate event in response. But that's an advance scenario that takes time to grow to. I say start with the unifying facade and go from there.

  • A massive service façade is very different from a message broker. It's almost at the opposite end of the dependency spectrum. A hallmark of this architecture is that the sender knows nothing about the receiver, and there may be many receivers (pub/sub or multicast). Perhaps you're confusing it with RPC-style "remoting" which is just exposing a traditional service over a remote protocol, and the sender is still coupled to the receive, both physically (endpoint address) and logically (return value). – Aaronaught Nov 4 '13 at 20:48
  • The similarity is that the facade acts like a router, the caller doesn't know which service/services handles the call just like a client sending a message doesn't know who handles the message. – Michael Brown Nov 4 '13 at 20:57
  • Yes, but the façade is then a God Object. It has all of the dependencies that the View Models have, probably more because it's going to be shared by several. In effect you've eliminated the benefits of loose coupling that you worked so hard for in the first place, because now, whenever something touches the mega-façade, you have no clue what functionality it really depends on. Picture writing a unit test for a class that uses the façade. You create a mock for the façade. Now, which methods do you mock? What does your setup code look like? – Aaronaught Nov 4 '13 at 21:04
  • This is very different from a message broker because the broker also knows nothing about the implementation of the message handlers. It uses IoC under the hood. The façade knows everything about the recipients because it has to forward calls to them. The bus has zero coupling; the façade has obscenely high efferent coupling. Almost everything that you change, anywhere, will affect the façade as well. – Aaronaught Nov 4 '13 at 21:07
  • I think part of the confusion here - and I see this quite a lot - is just what a dependency means. If you have a class that depends on one other class, but calls 4 methods of that class, it has 4 dependencies, not 1. Putting it all behind a façade doesn't change the number of dependencies, it just makes them harder to understand. – Aaronaught Nov 4 '13 at 21:08
0

Why not combine both?

Create a facade and put all services your viewmodels use. Then you can have single facade for all yourviewmodels without the bad S word.

Or you can use property injection instead of constructor injection. But then, you need to ensure that those are getting injected properly.

  • This would be a better answer if you provided an example in pseudo-C#. – Robert Harvey Nov 4 '13 at 16:50

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.