Saturday, September 17, 2011

IObservable<T> and Delegate Equality

Equality and identity are often intermingled in the .NET framework. .NET delegates correctly implement a straightforward structural equality, where if two delegates are equal if they designate the same method and the same object, regardless of whether the delegate is a unique instance:
class Bar
{
public void Foo(int i)
{
}
}
var bar = new Bar();
Action<int> a1 = bar.Foo;
Action<int> a2 = bar.Foo;
Console.WriteLine("Delegate Equality: {0}", a1 == a2);
// prints:
// Delegate Equality: True

However, an IObservable<T> created from delegates does not respect structural equality, and reverts to an identity criterion:
// use a1 and a2 from previous code sample
var o1 = Observer.Create<int>(a1);
var o2 = Observer.Create<int>(a2);
Console.WriteLine("Observable Equality: {0}", o1 == o2 || o1.Equals(o2)
|| EqualityComparer<IObserver<int>>.Default.Equals(o1, o2));
//prints:
// Observable Equality: False

This is unfortunate, because it means that by default we cannot implement previous event handling patterns using IObservable without significantly restructuring the code to propagate the IDisposable returned by IObservable.Subscribe. By this, I mean code that properly managed event registration and unregistration has no easy transition to using IObservable, it must be completely rewritten.

Like event equality, the equality of IObservers created from delegates should be structural, not based on identity. Thus, manually managing subscriptions would be possible via an explicit "Unsubscribe" operation.

This decision has a real consequence that I just hit: I can implement IObservable given an object implementing INotifyPropertyChanged, but could not do the reverse using the built-in abstractions. You'd either have to define your own IObservers that implement structural equality, or you'd have to store the actual event locally and trigger it manually when a new value is available, as I have done with NamedProperty<T> in Sasa.

On a slightly related note, I've switched Sasa over to use Mercurial on Sourceforge, and have closed the old subversion repo.

No comments: