Thursday, March 26, 2009

View Models: POCOs versus DependencyObjects

This post is part of a short series I am doing on MVVM infrastructure. In this series, I will share some thinking and code that has helped to produce cleaner, simpler MVVM code in my applications.

Related posts:

  1. POCOs versus DependencyObjects
  2. ViewModel
  3. DelegateCommand
  4. ActiveAwareCommand

If you're leveraging the MVVM pattern in your WPF/Silverlight development, you will quickly be faced with a decision regarding the implementation of your view models: should they be DependencyObjects, or POCOs (Plain Old CLR Objects)? I've seen and worked with applications that use both of these options.

This post is intended to discuss these two options. It certainly won't touch on all issues and nuances of either option, but it will cover the main ones I have found to be problematic. I will order them from least significant to most.


I was hesitant to even mention this one because I haven't done any measuring and haven’t found it to be a problem at all. There is a theoretical performance benefit to using DependencyObjects for view models for a couple of reasons:

  1. Lower memory usage.
  2. Faster binding performance.

The former assumes your view models have a lot of properties, and that those properties tend to take on their default values. WPF’s dependency object system is optimised for this case. Your typical WPF control has dozens or even hundreds of properties, most of which are set to their default value. Extra memory is only used by properties if they take on a non-default value. If your view models follow a similar pattern then you might get some memory usage benefits from using DependencyObjects. But I'd also be questioning your view model design if that was the case.

The second point is more relevant, since the primary job of a view model is to provide properties to which the view can bind. WPF's binding system supposedly performs better when binding to DependencyObjects than to CLR properties. That may well be the case (again, I haven't measured), but the difference must be negligible because this has not proven to be a problem for me thus far.

Property Change Notification in a Related View Model

Suppose you have two view models that are related. For example, a ParentViewModel and ChildViewModel. Let’s say that the ChildViewModel has a reference to a ParentViewModel. When the Savings property on the ParentViewModel changes, we want to update the Inheritance property on the ChildViewModel. Using DependencyObjects as view models, we can simply use data binding to achieve this:

//this code is in the ChildViewModel constructor
var binding = new Binding("Savings") { Source = parent};
BindingOperations.SetBinding(this, ChildViewModel.InheritanceProperty, binding);Using POCOs we need to do a little more work. Typically your POCO view models will implement INotifyPropertyChanged, so the code would look more like this: 

//this code is in the ChildViewModel constructor
parent.PropertyChanged += delegate(object sender, PropertyChangedEventArgs e)
    if (e.PropertyName == "Savings")
        Inheritance = parent.Savings;

Neither of these options is particularly appealing to me, but the binding approach with DependencyObjects is more flexible and less typing. Both suffer from magic strings ("Savings", in this case).

I recognised this as a problem with POCO view models a while back and have been busy working on a library that solves it. To whet your appetite, using my library would allow you to do this with any POCOs:

//this code is in the ChildViewModel constructor
var binding = new TypedBinding<ParentViewModel, ChildViewModel>(parent, x => x.Savings, this, x => x.Inheritance);

Note the strongly-typed binding using lambda expressions. Note also that I'm planning an even terser, fluent interface. Something like:

bindingManager.Bind(parent).Property(x => x.Savings).To(this).Property(x => x.Inheritance);

Serialization of View Models

Sometimes you will want to serialize your view model. For example, you might want to implement the IEditableObject interface so that changes to your view model can be rolled back. An easy way to do this is to serialize a snapshot of your view model and then roll back if necessary:

public class MyViewModel : IEditableObject
    private object[] _copy; 

    public MyViewModel()
        Name = string.Empty;

    public int Age { get; set; } 

    public string Name { get; set; } 

    public void BeginEdit()
        //take a copy of current state
        var members = FormatterServices.GetSerializableMembers(GetType());
        _copy = FormatterServices.GetObjectData(this, members);

    public void CancelEdit()
        //roll back to copy
        var members = FormatterServices.GetSerializableMembers(GetType());
        FormatterServices.PopulateObjectMembers(this, members, _copy);

    public void EndEdit()
        //discard copy
        _copy = null;

This works fine for POCO view models, but not so for DependencyObjects. Recall that for an object to be serializable (and I'm talking strictly about IFormatter-based serialization here), it and all its subclasses must be marked with the Serializable attribute.

DependencyObjects are not marked as serializable.

Incidentally, you might instead follow the pattern where you wrap your data inside a serializable struct and just serialize/restore that struct instead of the view model itself. That works great for POCOs, but – yet again – you will run into headaches if you use DependencyObjects.

And there are other reasons you might want to serialize your view model. Perhaps you want to clone it. Or perhaps you want to save certain view model objects over application restarts. DependencyObject-based view models will cause you grief when this need arises. Basically, your only option is to implement serialization surrogates, or use a non-IFormatter-based serialization mechanism.

Equality and Hashing

It is often useful to compare view models for equality, or stick them in a dictionary. For example, suppose you have a ReportsViewModel that is responsible for managing and exposing a set of ReportViewModels (note the plural versus singular in those names). Each ReportViewModel contains the name of the report, the parameters to the report, and the results of its execution:


Now suppose you'd like to cache report executions. If the user runs the same report with the same parameters, you'd like to just give them the existing results. To achieve this, you'd naturally attempt to override the Equals() and GetHashCode() methods in the ReportViewModel class (and implement IEquatable<ReportViewModel> if you're thorough). But if the ReportViewModel class inherits from DependencyObject, you'll find that you can't.

The DependencyObject class overrides and seals both the Equals() and GetHashCode() methods.

This inability to redefine equality in your view models is both annoying and limiting in certain scenarios. There may be ways around the problem. For example, you could implement a class that implements IEqualityComparer<ReportViewModel> and use it where appropriate. However, this can quickly lead to a mess and anti-DRY code base.

Of course, if your view model is a POCO, it won't suffer from this problem. You'd just provide the most appropriate implementation of equality inside your view model class.

Thread Affinity of View Models

One of the responsibilities your view models will typically take on is that of doing heavy work on a background thread. For example, suppose you have a refresh button in your UI that causes a heap of widget data to be loaded from the database and displayed in a list. You'd hardly want the UI to hang while the database access is taking place, so you decide to put the work in a background thread:

public class MyViewModel
    //this gets called when the user clicks the refresh button - we'll worry about how that happens in a later post
    public void LoadWidgets()
        //do the heavy lifting in a BG thread
            var widgets = new List<Widget>();
            //pretend I execute a database query here, would you kindly?
            var dataReader = command.ExecuteReader(); 

            while (dataReader.Read())
                widgets.Add(new WidgetViewModel(dataReader["Name"]));

            //now we have all our widgets read from the DB, so assign to the collection that the UI is bound to

All we're doing here is doing as much work on a background thread as we can. Only at the last step do we switch to the UI thread to update the collection that the UI thread is bound to (and with the right infrastructure, even that would be optional).

Again, this is great for POCO view models, but falls flat on its face for DependencyObjects.

A DependencyObject has thread affinity - it can only be accessed on the thread on which it was created.

Since we're creating a bunch of WidgetViewModels on a background thread, those view models will only be accessible on that thread. Therefore, as soon as the UI thread tries to access them (via bindings) an exception will be thrown.

The only option here is to construct and populate each view model on the UI thread. This is ugly, error-prone, and can negate much of the benefit of doing the work on a background thread in the first place. If we need to create many view models, suddenly the UI thread spends much of its time constructing and initialising those view models.

Code Readability

You'll notice that all of these problems can be worked around, and I don't dispute that. However, all those workarounds result in an incomprehensible mess of code. One of the most important qualities of code is its readability. Using DependencyObjects as view models causes an explosion of supporting code and workarounds that obscure the intent of the code. What's more, all these workarounds gain you . . . nothing.

To me, this is the final nail in the coffin for DependencyObjects as view models. I want my view models to be as readable and maintainable as the XAML I get from using MVVM in the first place. Rather than sweep The Ugly from the view to the view model, I'd much prefer to banish it entirely.


I think the advantages of POCO view models over DependencyObject view models are clear and convincing. Moreover, the disadvantages of using POCOs are practically zero. As such, I always use the POCO option in my projects. That said, all the problems I mentioned can be worked around, and you may have a convincing argument to do so (please let me know in the comments if you do). But, for the here and now, I will be sticking with POCOs for my view models.

In the next two or three posts, I will be sharing some infrastructure I have put together for MVVM applications, starting with a base ViewModel class.

† Note that WPF does do some auto-marshalling for us with simple properties. However, it cannot automatically marshal changes to collections, so you will inevitably end up wanting to create non-trivial DispatcherObjects on the UI thread.


Colin Eberhardt said...

A very interesting post ... thanks.

Regards, Colin E.

Kent Boogaart said...

Thanks Colin - glad you liked it.

Max Palmer said...

Really interesting post. Really like the fluent syntax for the binding code.

Kevin Rohling said...

Great post, very informative!

wesaday said...

Cool. Looking forward to reading more.

chris said...

I have been in the process of moving from DOs as VMs to POCOs + what we have called 'property models' for the properties and it gives a feel something like what you have (less magic strings etc).

class ViewModelX
IPropertyModel[int] MyIntProperty {get;set;}

interface IPropertyModel : INotifyPropertyChanged, IComparable
event EventHandler ValueChanged;
bool IsOverridden {get;}
bool HasValue {get;}
AddDependentProperty(IPropertyModel prop);
RemoveDependencyOn(IPropertyModel prop)

We have many impls of IPropertyModel[T], the normal tree, for nullable types, for enumerables, etc.

This pattern of property models reduces the property changed notification to the IPropertyModel.Value's property change and internally we generally attach the the IProperty's ValueChanged event. This gives a completely typed feel and removes several burdens from the hosting VM.

... I will blog post about it more completely I guess to show what I mean. In any case, I am in agreement about your general POCO stance. Used in combination with a good PropertyModel, you can have a relatively string free impl..

Miroslav Miroslavov said...

Nice post Kent,
but how do you get the reference to the Dispatcher from a POCO class( MyViewModel in your example ) ?

Best Regards,
Miroslav M.

Kent Boogaart said...

Thanks everyone.

@Miroslav: Ah, you shall have to wait and see my next post ;)

MF said...

if using DependencyObjects you don't need a magic string for binding to a property.
you can go:
binding.Path = new PropertyPath(ParentViewModel.SavingsProperty);
instead of using the string in the constructor of the binding (as long as you don't have a more complex path)

MF said...

I have never done it myself. But i believe that you can access the properties of a DependencyObject on another thread if you have frozen it (has to implement IFreezable). Of course once it is frozen you can't modify it or un-freeze it, so you would have to clone it before you freeze it. WHich is a little annoying, but it is a work around.

Miroslav MIroslavov said...

Thanks Kent,
I'll be glad to read your next post.


Glenn Block said...

Nice job Kent.

I like the approach, and agree on Pocos. Especially when it comes to having code that is more testable.

Actually I've been working on a similar FI approach using Expressions. Not for vm to vm binding per se, rather for setting up the binding between the UI and the VM itself, including, data, actions and UI state. I'll be posting something on it soon.

Kent Boogaart said...

@MF: Good points, both. Regarding the former, I believe you can even do compile-checked complex path bindings via the the PropertyPath class. However, it gets quite verbose and messy.

Also, regarding freezing Freezables, you're right. A frozen Freezable is thread-agnostic. However, I think that's a really limited use case for view models. Maybe it would be useful for reference style data, but that's about it I think.

@Glenn: Thanks. I look forward to seeing your FI approach.

Regarding testing, I left that out because my experience was that there wasn't much difference between testing a DO versus POCO. I'd be interested in hearing where the pain points were for you.

Philipp said...


Your fluent interface to inter-object binding looks *very* promising. I've published a comparable (but quite simple) framework a while ago that also supports binding scenarios based on expressions, but it appears to me you put quite some work in yours. Looking forward to it ;)


Brian said...

Yes, I completely agree. I have been working with MVVM for some time now, and I ONLY use dependency objects on the UserControls if I want the properties to play nicely in the XAML.

As far as ViewModels, I won't let DepencencyObjects touch the code. The reason is simple: Testability and separation. Using NUnit, I can't even consider constructing a DependencyObject... even if I tried.

But I don't try. I have found that making it so my ViewModel code does not depend on any of the WPF/Silverlight namespaces, they are just classes that can compiled and used anywhere. Keeping DependencyObjects away only helps it.

Finally, I have also come up with a similar binding framework. It is not quite as elegant as your lambda method (I like the FI a lot), but it works well... binder.Bind(source, "Prop1", dest, "Prop2"); Again, it Doesn't rely on the WPF/Silverlight namespaces, so it is very easily testable. I also added a binder.HandleChange(source, "Prop1", (value) => DoSomething(value)) to avoid any switch statements that occur when listening to the property notify event.

Anyways, reading your article really helped me solidify my opinions on DOs and I don't feel like such a black sheep anymore.

Brian Genisio

Anonymous said...

Hi Kent

The threading code causes the following error:

"This type of CollectionView does not support changes to its SourceCollection from a thread different from the Dispatcher thread."

Phil Steel said...

Re the Threading code, I think you need a BeginInvoke call instead of Invoke.

Patrick said...

Thanks for this post. It was very helpful.

Anonymous said...

Hello Kent,

To me, there is no question.
For a userControl, you use DP (for every fields used as input-output to the control).
Anywhere else is INotifyPropertyChanged. Dot.
But a problem still exists, how to make a userControl with MVVM and DP while DP should be defined in the userControl ??? Too me that is the question... How to do that ???

Also, too me, you should never serialize a ViewModel. It's an error. You can serialize your dataLayer but never a viewModel.
Also, too me, you could have a very complex ViewModel(many classes) but I can't see why you would have relation between DataModel. Too me, relations are for the dataLayer.

Hope I'm not too wrong...

Eric Ouellet

Kent Boogaart said...

@Eric: to me, you're either writing an application-aware view or an application-agnostic control. The latter would have custom dependency properties but no view model. The former, has a view model and might use the application-agnostic controls in its view. It might bind the dependency properties on those controls to the view model.

I think there is often a relationship between view models, whether it's manifested explicitly in the object model or not. It is often the responsibility of a parent view model to create child view models.

Anonymous said...


Monirul Islam

Manikandan Ganesan said...

Very interesting!