MVVM Backlash

I and many others have pushed hard for Model-View-ViewModel (MVVM) design over the last few years. No wonder people are pushing back. Two recent conversations have stirred me to reexamine my affection for MVVM screen design. What surprises me is that I’m feeling we slightly oversold it.

In the end, I am unrepentant: I continue to practice MVVM and recommend it to others.

But today I am going to take it down from its pedestal and give it a good thrashing. I’ll admit begrudgingly that you can build a large, maintainable application without it. I will discover that MVVM has only two truly distinctive technical advantages: automated testing of the view logic and integrated design time data. Lots of developers don’t care about either. MVVM may not be for those people.

I’m not going to call them bad programmers. They could be good programmers. They could be great programmers. Are you being bad when you don’t write view-logic automated tests? I’m inclined to say so but not prepared to hang you for it.

Disquieting conversations

I have had a long running conversation with sage and MVVM skeptic, Billy Hollis about MVVM. Billy finds a way to talk trash about MVVM just about every chance he gets. His oeuvre on the subject consists largely of twitter bon mots, barbed asides in lectures, and assorted podcast witticisms. He may have penned his opinions on MVVM somewhere but I don’t know where. Therefore, at some risk of misrepresenting his position, I summarize his position as follows:

1. MVVM is a way for code-addicted developers to deflect attention from the horrible designs they inflict on users,
2. MVVM complicates development with little or no compensating benefit,
3. there is too much strident, jargon-filled, punishing MVVM rhetoric that alienates the mass of developers.

I agree completely with #1. MVVM does nothing to improve UX; our obsession with it robs us of time better spent enriching our design education and delivering more satisfying applications. Nonetheless, I expect to continue that trend with this post and more MVVM commentary to come. I do not apologize; while it is one of my obsessions, it’s your fault if you make it yours.

I ruefully acknowledge his 3rd objection although I believe we’ve largely outgrown it.

Objection #2 – that MVVM adds complexity without benefit – is the controversial claim that I address today. That leads me to my second conversation, the one precipitating this post.

John Papa is developing an outline for his forthcoming video series on MVVM. The series targets developers who know nothing about MVVM. It will explain what MVVM is and explore MVVM techniques.

But why bother with MVVM in the first place?  Why not do the most obvious thing and write all code in the View’s code-behind? What is wrong with that? John wants to address those questions in his video in a direct, matter-of-fact way that any developer can appreciate. That isn’t going to be as easy as he first thought 

MVVM is second nature to John as it is for many of use. We’ve forgotten how we got here. We can recite the academic arguments. We have the PowerPoint slides we’ve used over the years. But if you were susceptible to academic argument, you’d be using MVVM already.

What if we couldn’t get by with the old bromides like “Separation of Concerns” and we couldn’t thump Martin Fowler’s pattern book from 2003 like a bible at a revival meeting? How would we demonstrate that MVVM is better than code-behind?

John kindly called me to puzzle this through with him. We had some fun watching the wheels come off. This post is my account of what we learned. It represents my conclusions, not John’s … although I hope he agrees with me.

The usual arguments for MVVM

Here are bullets from a hypothetical MVVM benefits slide; I’ve drawn them from several presentations, blogs, and discussions.

• Separation of concerns (SoC)
• Loose coupling
• Data Binding
• Designer / developer symbiosis
• Reusability
• Avoid spaghetti code
• Consistent, maintainable, scalable
• Support for design time data
• Automated view-logic testing
• Multiple views over the same view logic

John and I kicked these around. It turns out that many of them don’t hold water.

We ruled out “separation of concerns” and “loose coupling” right away because they are just jargon to many people.

[update: "ruled out" was a poor verb choice. What I meant was that we didn't want to chant "SoC" and "Loose Coupling". We wanted to get behind the familiar rhetoric. These principles remain fundamental to us ... as I explained in the next paragraph.]

To me they are a convenient short hand for guidance that has always served me well. But for many they’re just slogans, shouted from a little red book.

John and I aren’t sending anyone to re-education camp. We wanted to show concretely how code-behind necessarily betrays you and ViewModel saves you. When we really looked, we realized that ViewModel isn’t intrinsically more SoC or more loosely coupled than code-behind (stay tuned).

Data binding? The practical implementation of the MVVM pattern requires a powerful data binding infrastructure. But you don’t have to use MVVM to use data binding. You can bind the view to the code-behind.

“Designer / developer symbiosis” refers to the ability of a designer and a developer to work on the view independently without risk of corrupting the other’s work.  In MVVM, no matter what the designer does to the View, the ViewModel code is untouched. Nothing the developer does to the ViewModel code affects the View XAML. Upon closer examination (see below), I feel that MVVM doesn’t have a strong advantage here … even when (as is extremely rare) the project is blessed with a crew of designer people to assist the developer people.

The “spaghetti code” argument is a canard. MVMV in not inherently more “consistent, maintainable, and scalable.” Bad code is bad code whether it’s in the ViewModel or the code-behind. The view logic is not magically transformed by refactoring to a separate class. If you are inclined to reduce repetition and externalize dependencies into supporting services, you will do so in code-behind as surely as you will do it in a ViewModel. If not … well we’ve all seen train-wreck ViewModels haven’t we?

Design time data support? Check! John and I believe developers and designers are more effective when they design, maintain and evolve the view in Blend (or Cider) using design time data. A ViewModel can easily deliver design data to the tooling and “real” data to the runtime application. The code-behind cannot provide design data to the tooling because the design tools don’t execute the code behind.  I expand on this point below.

Automated view-logic testing? Check! We agreed that automated testing of code-behind is so difficult as to be impractical. Billy agrees too. Score one for MVVM. Of course if you don’t do automated testing of view logic …   I talk more about this below.

Multiple views over the same view logic? Check! ViewModel is for you if you think you’ll need different views for different users – perhaps as part of a localization requirement or perhaps because the presentation should vary based on role. You can build such variation into a single xaml view but that gets messy quickly.

Definitive benefits

Three claims survive close analysis; ViewModel has a decisive advantage when you care about

• design time data
• automated view testing
• multiple views over the same view logic

Do you really care?

I do care … as I said at the top. I care about view-logic testing and design time data. Multiple views are usually just a nice-to-have; but I care a great deal about the first two.

The unfortunate reality is that many (most?) developers don’t seem to care about these things. I can want them to care … but I haven’t figured out how to make them care.

I believe the “design time data” argument may become attractive. Not many people design with data yet. I think that’s because they don’t know about it. Once they discover it and see how handy it is … and see that ViewModel is easy … that could shift the balance.

People have heard the argument for testing for years. I don’t think there is anything new to say, nothing more persuasive to say, that hasn’t been said. The testing adoption statistics are not encouraging. My anecdotal evidence is not encouraging. You know how many ViewModels I’ve seen tested in the wild? You know how many ViewModels I’ve seen tested in a book, article, or demo that cites “testability” as a reason for MVVM? I want to say “none.”

Actually I’ve done it and I’ve got a BookClub demo that shows how. ViewModel automated testing is the way IdeaBlade professional services rolls.

Please, please, please prove me wrong and tell us all where we can see some ViewModel testing.

I might do it if it’s easy

I fear that the unequivocal arguments for MVVM will have only a tenuous grasp on the hearts of many developers. I will argue until the cows come home that MVVM inclines you toward good practices (e.g., SoC, loose coupling). I must admit that if you have good discipline you’ll be fine without it.

I remind you, pace Billy Hollis, that MVVM won’t do a thing for your application’s user experience. Don’t get caught up in the code at the expense of the design! That’s unbalanced. That’s not healthy.

I suppose what I am saying (and lot of people are saying ) is “take this MVVM medicine … it’s good for you … even though you may not feel different right away. ”

That means MVVM must be easy to do. If it were hard you would be mad to follow our lead on the strength of these assurances. We’re telling you to create an extra class – the ViewModel class – for almost every View in your application. You might write a hundred “extra” classes in a big application. That’s a tough sell right there.

Fortunately, MVVM is easy to do. There’s always someone around to make a good thing harder than it needs to be. Laurent Bugnion’s MVVM Light proves it doesn’t have to be hard. I hope you’ll find – as we have – that you’ll quickly get the hang of it and it’s smooth sailing once you do. Design data and automated testing will rapidly reveal their value.

I have a good feeling about John’s video series. I think it’s going to help you feel good about MVVM.

Appendix: Observation Details

The remainder of this post explores a few of the earlier observations in a bit more detail. Indulge yourself if you will.

Separation of Concerns

Code-behind is code that supports the view. It sits in its own file, apart from the XAML file that determines what the user sees and touches. It’s purpose is to assist the View in presenting data and responding to user gestures. Well that’s the definition of a ViewModel too.

To be sure the view and its code-behind are the same class. In this respect, the visual design and view logic are technically comingled in a clear violation of SoC.  Practical consequences attend. But are they sufficiently compelling? Or can a good developer achieve practical separation with little effort?

Naïve developers tend to crowd the code-behind with logic that doesn’t belong there. I’ve seen people implement data access in code-behind. I’ve done it myself … in quick demos. That’s a bad idea for a real application. But it’s not a better idea when you relocate that code to a ViewModel. Pesistence logic belongs in a service that you make available to the ViewModel … or the code-behind.

The same argument applies to logging, and navigation, and all other cross-cutting, re-usable tasks that should be treated as external dependencies by the ViewModel or code-behind. You either think this way or you don’t.

I considered a detailed discussion of Robert Martin’s SOLID principles as they apply to using or not using a ViewModel. It struck me as an arid exercise. Only “single responsibility” and “dependency inversion” seemed to have practical relevance; there’s not much use for inheritance or interfaces in view design. I’d make the same argument as I did for SoC.

Culturally ViewModel is more aligned with SOLID thinking but, as I said, you can get the practical benefits in code-behind without much effort.

Loose Coupling

The code-behind approach makes it easy to tightly couple the UI elements to the code. That sounds really bad. I think it sounds bad. Then I try to think of all the awful things that will happen to you … and I don’t come up with any; not unless you care about testability which I do. But bear with me. If I didn’t care about testability – if I merely paid it lip service – would tight coupling matter? I’m having a tough time showing how.

It’s not as if the View and the code-behind are living separate lives. How often (outside of tests) would you instantiate one without the other?

Moreover, if I do care about tight coupling, I can easily loosen it in the code-behind approach. Nothing stops me. It’s no harder (or easier) than coupling to a ViewModel.  I can data bind the xaml to the code-behind and use Behaviors to bind to methods and events in the code-behind. All I have to do is set the view’s DataContext to this in the constructor. Done. The choice is mine.

Automated Testing

Testability is MVVM’s sole uncontested and incontrovertible advantage. I can think of no one who argues that it is practical to test logic in the code-behind.

Update: Bill Kempf took up the challenge in a follow-up blog post that shows you can (a) write a code-behind ViewModel and (b) test it. Thus he reminds us that MVVM is a pattern admitting of many implementations ... including, it seems, a code-behind implementation. p.s: Bill was making a point, not a recommendation. Be amused by it but please don't follow it!

ViewModel testability alone should be sufficient reason to adopt MVVM designs. And MVVM proponents dutifully recite MVVM’s testability advantage.

Too bad so few MVVM proponents actually write ViewModel tests. Find me a demo with automated ViewModel tests. Do they even exist? The hypocrisy is mildly amusing.

One possible explanation: demo ViewModels don’t seem worthy of tests. The first ViewModel you write is typically just a bunch of properties. Property tests are low priority because they generally lack decision logic. Testing simple getters and setters is a complete waste of time.

Even a primitive ViewModel gets its model from somewhere, typically a data service. However, in the beginning, the data service is usually a fake that always delivers the goods in a timely fashion. Nothing goes wrong so you let it slide. Later, if you are inclined toward automated testing, you will test the data service independently of the ViewModel.  VM testing can seem a well-intentioned luxury that you never get around to doing.

In fact, in a mature ViewModel there will be plenty to test. The VM is an integration point for asynchronous services and user interactions. Users do weird things and services fail. How does the ViewModel handle service exceptions? How does it handle null results, empty results, one result, multiple results? Does it signal changes to IsBusy at the right times … as when the service fails? Can the VM run offline? What if the user isn’t authenticated? What if the save fails with a concurrency failure? How does the VM respond to tombstoning if this is a Windows Phone application? Does the VM signal view state transitions? Does it wait for view animations to complete and, if so, how do you test the timing? Does the VM send messages? Does it respond to messages? How does it learn which data to show? What if the requested data item doesn’t exist or is invalid?

I don’t anticipate all of these possibilities at the start. Instead I write a few minimal ViewModel tests to put the structure in place. Once I get the hang of it in a new project, this takes a couple of minutes for each new VM. If it takes longer, I count myself blessed because I’ve discovered trouble in my MVVVM triad before the heat is on. I act immediately to get it pointed in the right direction; then I move on. Let me repeat, I make no attempt to test everything. I make sure I have at least one test. Then I add tests for whatever happens to catch my interest.

Over time I’ll trace a number of bugs to the VM. Over time I’ll add capabilities or refactor portions of the VM to helper classes. When I fix those bugs, add new logic, or refactor, the friction of adding supporting tests is low because I’ve established a base line.

This approach works for me. I’m not a testing fanatic. I skip tests even when I know that I shouldn’t. But let me say this: I never regret writing them. I’ve been saved by tests many times and suffered as many times for lack of them. If we work together on a project, we’re going to write tests together. That’s how I roll.

Design time data

John and I believe strongly in the virtue of designing screens with realistic design data that is shaped exactly like the runtime data. We can’t think of any way to satisfy this requirement in the code behind approach. The design tools don’t run the code-behind so the code-behind can’t deliver design time data.

Some folks recommend Blend’s XML sample data, a secondary data source that doesn’t depend upon a ViewModel. That’s a viable way to drive out the first draft of a view. I feel it is a significantly inferior alternative because:

• You can’t re-use the sample data as fakes for developer, QA, and automated testing.
• Many data types (e.g., DateTime) are represented as strings in sample data. You don’t bind to a string as you would a DateTime. That means you write two bindings – one for design, one for production – or devise some other hideous workaround.
• Complex object graphs are trickier to maintain in XML sample data
• You have to update the sample data structures as your model evolves
• Sample data files tend to be scattered around your projects, complicating maintenance.
• It’s tedious at best to keep the sample data out of the production builds.

The first objection is a big one for me. I’ve always got fake data around during development and testing.

Forget about automated testing for the moment. Everyone tests their applications in some fashion. You make a change to a view, you want to see how it looks … now … before you forget what you changed. You want quick turnaround, not a long delay while the application tries to reach the server … which could be down. View development shouldn’t depend on a data service that’s broken or hasn’t been written yet.

It will take effort to create and maintain fake development and test data. I don’t want to do that twice – once for runtime simulations and again – in a completely different format - for view design.

Designer / developer symbiosis

We can’t count this as a huge plus for MVVM. The separation of the XAML and the code behind is pretty good to begin with. It is not like Windows Forms where the changes made on the design surface produced reams of code-behind. The risk of developer/designer cross-contamination is low and reasonable, easy-to-follow work rules reduce it further.

There are no technical obstacles to a designer and developer working on the view simultaneously because the view xaml and code-behind are in separate physical files.

The greater risk is that a designer or developer will introduce breaking changes that aren’t detected until runtime. MVVM is no help here at all. In fact, the risk of conflicting changes is slightly higher due to View/ViewModel separation. Each party works happily in isolation, unaware of the effects on the other. The compiler is no help … as it might be if the designer’s changes happen to corrupt the code. ViewModel automated tests don’t help because, by nature, they don’t consider the view at all. 

On the plus side, ViewModel discipline prohibits references to View components; it even prohibits references to most types in the presentation technology. If you follow that discipline, you free the designer to explore visuals and interaction patterns without worrying about breaking the code. While technically, you can be as rigorous in code-behind, the temptation to be lax is almost irresistible. Accordingly, I count “Designer / developer symbiosis” among the cultural claims to MVVM advantage.

Caliburn

I am a Caliburn fan but I am positioning it as an afterthought in today’s discussion. Caliburn is an UI framework grounded in MVVM. Caliburn automatically wires visual components to code by applying conventions. It can automatically wire a button named “Save” to the “Save” and “CanSave” methods in the ViewModel. The designer doesn’t add behaviors to the XAML; the developer doesn’t write ICommands. There is less manual work for everyone. And Caliburn offers other convention-based time-savers throughout the UI development workflow. Its contributions could weigh heavily in the balance for MVVM.

On the other hand, it’s one more framework to learn and it’s not in the MVVM mainstream. Developers new to MVVM are usually new to the entire XAML platform; they are overwhelmed by foreign concepts and unprepared to shoulder the additional burden of learning Caliburn. Great as it is, the road to Caliburn is bumpier than it could be. We at IdeaBlade are doing our part to smooth that road.