We Keep Coding

Castle Windsor Abstract Factory

I'm trying to understand how to use TypedFactoryFacility to create an abstract factory, and I have it working at a basic level, however I don't fully understand how to scale it with runtime dependencies
Suppose I have a service that needs to be created at runtime:
public interface IRuntimeService {
void DoThing();
}
with the following implementation
public class RuntimeService : IRuntimeService {
public void DoThing() {
// Do some work
}
}
To create my IRuntimeService, I've created an abstract factory
public interface IRuntimeServiceFactory {
IRuntimeService CreateService();
}
In my Castle installer, I'm using the TypedFactoryFacility to register my class and abstract factory.
public class TypeInstaller : IWindsorInstaller {
public void Install(IWindsorContainer container, IConfigurationStore store) {
container.AddFacility<TypedFactoryFacility>();
container.Register(Component.For<IRuntimeService>().ImplementedBy<RuntimeService>());
container.Register(Component.For<IRuntimeServiceFactory>().AsFactory());
}
Then in my class that will be using the service, I can use the factory to create new service instances at runtime.
var myService = m_ServiceFactory.CreateService();
Everything above works perfectly, however I'm running into a problem when my RuntimeService class needs to be injected with a dependency chain itself that include runtime parameters.
To expand the example above, suppose I have a new runtime dependency
public interface IRuntimeDependency {
void DoWork();
}
implemented by a class that takes a runtime string value through the constructor
public class RuntimeDependency : IRuntimeDependency {
private readonly string m_Param;
public RuntimeDependency(string param) {
m_Param = param;
}
public void DoWork() {
// Do work involving the param
}
}
And the previously defined service class now needs a reference to the dependency
public class RuntimeService : IRuntimeService {
private readonly IRuntimeDependency m_Dep;
public RuntimeService(IRuntimeDependency dep) {
m_Dep = dep;
}
public void DoThing() {
// Do some work involving the dependency
m_Dep.DoWork();
}
}
How do I now I create instances of my service using the TypedFactoryFacility?
I would expect do just be able to change my factory method to look like
IRuntimeService CreateService(string param);
but Windsor throws an error 'Could not resolve non-optional dependency for parameter 'param' type 'System.String'.
Windsor knows how to create an IRuntimeDependency if I give it a string, and it knows how to create a IRuntimeService if I give it the dependency, so why can't it directly create a IRuntimeService with the string param?
I can make it work by having two distinct factory methods
IRuntimeService CreateService(IRuntimeDependency dep);
IRuntimeDependency CreateDependency(string param);
and creating the dependency, manually myself
var dep = m_ServiceFactory.CreateDependency(param);
var myService = m_ServiceFactory.CreateService(dep );
^^^This works, but the whole point of using a container is so that it will take care of assembling new objects for me. This is a relatively simple example involving only one dependency, but it would easily grow out of control with a more complex object graph.
I could of course create my own factory implementations, but that also nullifies the benefit of using the TypedFactoryFacility which is supposed to create the abstract factory implementations for you. I have a hard time believing there's not an existing solution to this problem but the Windsor examples don't contain any chained run-time dependencies.
I don't think using a FactoryComponentSelector is the correct approach because there's only one possible path to create the RuntimeService instance. It should be able to auto-resolve.

In many or most cases, an object resolved by the container depends on implementations of other interfaces which are also resolved by the container. So as long as all of the interfaces have registered implementations, the container can resolve the entire dependency chain.
But in this case RuntimeDependency depends on a string, which isn't something the container can resolve.
public RuntimeDependency(string param) {
m_Param = param;
}
In this case you can use the DependsOn method to explicitly provide a value to fulfill that dependency.
container.Register(Component.For<IRuntimeDependency, RuntimeDependency>()
.DependsOn(Dependency.OnValue("param","whatEverTheValueIs")));
That value can, of course, come from configuration or wherever else. I use this a lot with SQL connection strings.

It is possible using DynamicParameters.
container.Register(Component.For<IRuntimeService>()
.ImplementedBy<RuntimeService>()
.LifestyleTransient()
.DynamicParameters((k, d) => {
d["dep"] = new RuntimeDependency((string)d["param"]);
}));
Keep in mind that the dictionary keys have to match the parameter names in the CreateService method and RuntimeService constructor.
Edit: You should also make it LifestyleTransient if you intend to create a new instance each time the factory method is called. (The default is singleton)

It seems that what I am asking for is not possible by design.
See this other SO answer.
https://stackoverflow.com/a/3905496/2029835

Pass a dependency instance to a factory method parameter to make Ninject use it within the resolution

I have an abstract factory which creates some service represented by IService interface. In the factory I have two Create methods, because at one of them I allow the consumer to pass an existing IServiceLogger instance to be used by the constructed service tree.
public interface IMyServiceFactory {
IMyService Create(IServiceLogger loggerInstance);
IMyService Create();
}
Because an IServiceLogger should be shared among the service tree, I use the InCallScope when binding it to a concrete implementation.
How can I implement this scenario with Ninject? I've tried the following approaches.
1. Manually create a factory implementation
internal class MyServiceFactory : IMyServiceFactory {
private IResolutionRoot _kernel;
public MyServiceFactory
public IMyService Create(IServiceLogger loggerInstance) {
// what should go here? how can I pass the existing instance to Ninject Get method and make Ninject to use it for the whole resolution tree, just as it were created by Ninject and used as InCallScope?
}
// this one is trivial...
pulbic IMyService Create() {
return _kernel.Get<IMyService>();
}
}
UPDATE
Actually I've found a messy and not too safe way for this. I can get the current bindings via GetBindings, then Rebind IServiceLogger ToConstant, then Get the IMyService instance, and finally restore the original bindings with AddBinding. I don't like it, it feels stinky and what's worse, it's not thread-safe, because another thread can request for a IMyService in the middle of this code and hence use the local temporary binding.
2. Use Ninject.Extensions.Factory
Just use the ToFactory binding, but that's not working, because it just tries to use the parameter as a simple constructor argument (if applicable), and not as an object for the whole resolution tree.

I would give more control to the Kernel of Ninject and do not create a class for the factory at all.
And use Func binding in Ninject like this:
Bind<Func<IMyService>>().ToMethod(s => CreateService);
By binding of the ILoggerService or not binding this you can controll centrally whether you have logger or not in your service.(try by just comment it out)
Here implementation of the Bootstrapper:
public class Bootstrapper
{
private IKernel _kernel = new StandardKernel();
public Bootstrapper()
{
_kernel.Bind<MyStuff>().ToSelf();
_kernel.Bind<IServiceLogger>().To<ServiceLogger>();
_kernel.Bind<IMyService>().To<MyService>();
_kernel.Bind<Func<IMyService>>().ToMethod(s => CreateService);
}
public IKernel Kernel
{
get
{
return _kernel;
}
set
{
_kernel = value;
}
}
private IMyService CreateService()
{
if(_kernel.GetBindings(typeof(IServiceLogger)).Any())
{
return _kernel.Get<IMyService>(new ConstructorArgument("logger", _kernel.Get<IServiceLogger>()));
}
return _kernel.Get<IMyService>();
}
}
Implementation of consumer class for the factory:
internal class MyStuff
{
private readonly Func<IMyService> _myServiceFactory;
public MyStuff(Func<IMyService> myServiceFactory)
{
_myServiceFactory = myServiceFactory;
_myServiceFactory.Invoke();
}
}
Simple implementation of MyService:
internal class MyService
:IMyService
{
public MyService()
{
Console.WriteLine("with no parameters");
}
public MyService(IServiceLogger logger)
{
Console.WriteLine("with logger parameters");
}
}
Simple ServiceLogger:
internal class ServiceLogger
:IServiceLogger
{
public ServiceLogger()
{
}
}
internal interface IServiceLogger
{
}

IMPORTANT UPDATE
While my original answer gave me a working solution, by an accidental InteliSense navigation I've just found that there is a built-in tool for exactly this issue. I just have to use the built-in TypeMatchingArgumentInheritanceInstanceProvider which does this, and even more, because there are no more needs for naming conventions due to the parameter type matching.
It would be good to have a more detailed documentation about these options, or maybe it's just me who can't find it currently.
ORIGINAL ANSWER
I tried a few ways, and ended up with a slightly different, kind of a convention based approach utilizing Ninject's context parameter inheritance.
The convention is used at constructor argument naming through the dependency tree. For example whenever an IServiceLogger instance is injected to a service class, the argument should be called serviceLogger.
With the above convention in mind, I've tested the following approach. Firstly I've implemented a custom instance provider for the factory extension. This custom provider overrides the mechanism for creating constructor parameters for the context to let the developer specify several named arguments which should be set as inherited. This way all the parameters with the specified names will inherit through the whole request graph during the get operation.
public class ParameterInheritingInstanceProvider : StandardInstanceProvider
{
private readonly List<string> _parametersToInherit = new List<string>();
public ParameterInheritingInstanceProvider(params string[] parametersToInherit)
{
_parametersToInherit.AddRange(parametersToInherit);
}
protected override IConstructorArgument[] GetConstructorArguments(MethodInfo methodInfo, object[] arguments)
{
var parameters = methodInfo.GetParameters();
var constructorArgumentArray = new IConstructorArgument[parameters.Length];
for (var i = 0; i < parameters.Length; ++i)
constructorArgumentArray[i] = new ConstructorArgument(parameters[i].Name, arguments[i], _parametersToInherit.Contains(parameters[i].Name));
return constructorArgumentArray;
}
}
Then after at binding configuration I just threw it in with the corresponding parameter name.
kernel.Bind<IMyServiceFactory>().ToFactory(() => new ParameterInheritingInstanceProvider("serviceLogger"));
Finally I reviewed parameter naming, and for exampled changed loggerInstance in the factory interface to serviceLogger to match the convention.
This solution is still not the nicest one as it has several limitations.
It is error prone. One can make bugs which are hard to track by not keeping the naming convention, because currently it silently fails if the convention does not match. This could be improved probably, I'll think about it later.
It handles only constructor injection, however this should not be a big issue as that's the suggested technique. For example I almost never do other kind of injections.

I realise this was asked a long time ago but I was looking to do the same sort of thing myself and finally worked out that you can use the IParameter array passed to the Get() method to specify a ContructorArgument to use only for the current Get() call. This allowed me to use a specific constructor value when creating a Hangfire Job allowing the Hangfire job to use a different database connection on each invocation if required.
EnvironmentName forcedEnv = new EnvironmentName() { Name = dbName };
// For this instantiation, set the 'envName' parameter to be the one we've specified for this job
var instance = ResolutionExtensions.Get((IResolutionRoot) _kernel, jobType,
new IParameter[] {new ConstructorArgument("envName", forcedEnv, true)});
return instance;
By setting the shouldInherit value to true you can ensure the value gets passed down the resolution chain. So it get's passed to any objects in the dependency tree that use that argument (but only for this particular instantiation).

StructureMap - Circular Dependencies and Lazy Initialization of Setter Properties

So we have ran into what seems like a very common issue with StructureMap and IoC containers in general I assume. Bidirectiona/Circuar dependencies.
Given the following code, it is currently causing a circular dependency since we have it 'autowiring' properties.
public class ServiceA:IServiceA
{
public IServiceB ServiceBDependency {get;set;}
}
public class ServiceB:IServiceB
{
public IServiceA ServiceADependency {get;set;}
}
I see the 'dependency' of each of these on eachother, however, I feel that as a property, they are not true dependencies which is what separates them from using constructor injection.
It seems that there should be a way for these services to be resolved...and then have the properties injected after the objects have been created?
I know of various ways to get around this, including the true clean way of rearchitecting my services, but i am curious as to what options I have with instantiation and service registration with StructureMap. It seems like a fairly common issue that would need a solution.

I'd like to show you my approach. I am using only the setter injection. And we often have many objcets referencing each other, in our web app. E.g. IUserFacade requires IUserState on user creation while IUserState requires IUserFacade on userState deletion (to check for constraint).
e.g.:
public interface IUserFacade
{
... // User facade Add, Update, Delete
IUserStateFacade { set; }
}
public interface IUserStateFacade
{
...
IUserFacade { set; }
}
In reality, we have many objects with cross references, and even more complicated. And it would really be very costy, if all the referenced objects should be created each time, even if not used during the request. We need the "lazy", the proxy objects to be placed in the setters.
The way how to do it, is a mix of the: 1) StructureMap (IoC top) and 2) Castle (proxying top) libraries. Below I will show some snippets of objects needed for this solution. More could be found inside open source project Catharsis
Wrapper. This object would be injected into each Property by SM (StructureMap) instead of real implementor. It is the sleeping implementation. It is waiting for a first call. If it will never happen (IUserFacade is deleting user, no need to access referenced IUserStateFacade during such request) this wrapper will sleep for ever (request). Once touched, SM will create the real object and Wrapper will pass all calls to that.
The Castle interceptor:
public class Wrapper : IInterceptor
{
object _lazy;
protected readonly Type Type;
public Wrapper(Type type)
{
Type = type;
}
public void Intercept(IInvocation invocation)
{
if (_lazy.IsNull()) // lazily instantiate the instance
{
_lazy = ObjectFactory.GetInstance(Type);
}
try
{
var method = invocation.Method;
if (method.ContainsGenericParameters)
{
method = method.MakeGenericMethod(invocation.GenericArguments);
}
invocation.ReturnValue = method.Invoke(_lazy, invocation.Arguments);
}
catch (TargetInvocationException ex)
{
// PublishingManager.Publish(.... // publish exception
throw;
}
}
}
ProxyInstance. Now, we need an object, clear and understandable to SM. That object will be mapped to all interfaces (IUserFacade...) Instead of returning implementation of the UserFacade.
We can also use our custom AOP filters here.
This ProxyInstance will be provided with the real implementor type, and building up the Wrapper.
The StructureMap Instance:
public class ProxyInstance : Instance
{
protected readonly ProxyGenerator Factory = new ProxyGenerator();
protected readonly Type ConcreteType;
public ProxyInstance(Type type)
{
ConcreteType = type; // the type for our Wrapper, the real implementation
}
protected override object build(Type pluginType, BuildSession session)
{
var aopFilters =
// my custom way how to inject more AOP filters
AopFilterManager.GetFilters()
// the core for us, one of the interceptors is our Wrapper
.Union(new[] { new Wrapper(ConcreteType) })
.ToArray();
// Castle will emit a proxy for us, but the Wrapper will do the job
var proxy = Factory
.CreateClassProxy(ConcreteType, AopFilterManager.AopOptions, aopFilters);
return proxy;
}
And now just use some standard way how to map it in the SM (I am using custom ProxyConvention but it is out of the scope here). Let's use simplified explicit mapping:
registry
.For<IUserFacade>()
.HybridHttpOrThreadLocalScoped()
.Use(new ProxyInstance(typeof(UserFacade)));
...
Also, each of our objects created via SM, implements IService. So, the default setter injection could be set like this:
registry.SetAllProperties
(
set => set.TypeMatches
(
type => type
.GetInterfaces()
.Any(i => i.IsEquivalentTo(typeof(IService)))
)
);
From that moment, when we need to work with a IUserFacade (the direct ObjectFactory call, or accessed via Wrapper), we recieve the real implementor. All its properties (setter injection) will be pre-populated with our ProxyInstance / Wrapper.
If any of these properties is accessed, e.g. IUserStateFacade the same (discribed above for IUserFacade) will happen again.
Because the Lifecycle is Thread or Request based, we have only one implementor in runtime/web request
Because we do inject the Wrappers while using setter injection, no issues with circular infinite loops. Only the first level is injected each time

How to avoid Service Locator Anti-Pattern?

I'm trying to remove a Service Locator from an abstract base class, but I'm not sure what to replace it with. Here is a psuedo-example of what I've got:
public abstract class MyController : Controller
{
protected IKernel kernel;
public MyController(IKernel kernel) { this.kernel = kernel); }
protected void DoActions(Type[] types)
{
MySpecialResolver resolver = new MySpecialResolver(kernel);
foreach(var type in types)
{
IMyServiceInterface instance = resolver.Get(type);
instance.DoAction();
}
}
}
The problem with this is that the instanciator of a derived class doesn't know what bindings the kernel must have in order to keep MySpecialResolver from throwing an exception.
This might be intrinsicly intractable because I don't know from here which types I'll have to resolve. The derived classes are responsible for creating the types parameter, but they aren't hardcoded anywhere. (The types are based on the presence of attributes deep within the derived class's composition hierarchy.)
I've trying to fix this with lazy loading delegates, but so far I haven't come up with a clean solution.
Update
There are really two issues here, one is that the IoC container is passed to the controller, acting as a service locator. This is easy to remove--you can move the location up or down the call stack using all sorts of techniques.
The second issue is the difficult one, how can you ensure that the controller has the necessary services when the requirements aren't exposed until runtime. It should have been obvious from the start: you can't! You will always be dependent upon either the state of the service locator or contents of a collection. In this particular case no amount of fiddling will ever resolve the problem described in this article with staticly typed dependencies. I think that what I'm going to end up doing is passing a Lazy array into the controller constructor and throwing an exception if a required dependency is missing.

I agree with #chrisichris and #Mark Seemann.
Ditch the kernel from the controller. I'd switch your resolver composition a little bit so that your controller can remove the dependency on the IoC container and allow the resolver to be the only item that worries about the IoC container.
Then I would let the resolver get passed into the constructor of the controller. This will allow your controller to be far more testable.
For example:
public interface IMyServiceResolver
{
List<IMyServiceInterface> Resolve(Type[] types);
}
public class NinjectMyServiceResolver : IMyServiceResolver
{
private IKernal container = null;
public NinjectMyServiceResolver(IKernal container)
{
this.container = container;
}
public List<IMyServiceInterface> Resolve(Type[] types)
{
List<IMyServiceInterface> services = new List<IMyServiceInterface>();
foreach(var type in types)
{
IMyServiceInterface instance = container.Get(type);
services.Add(instance);
}
return services;
}
}
public abstract class MyController : Controller
{
private IMyServiceResolver resolver = null;
public MyController(IMyServiceResolver resolver)
{
this.resolver = resolver;
}
protected void DoActions(Type[] types)
{
var services = resolver.Resolve(types);
foreach(var service in services)
{
service.DoAction();
}
}
}
Now your controller isn't coupled to a specific IoC container. Also your controller is much more testable since you can mock the resolvers and not require an IoC container at all for your tests.
Alternatively, if you don't get to control when a controller is instantiated, you can modify it slightly:
public abstract class MyController : Controller
{
private static IMyServiceResolver resolver = null;
public static InitializeResolver(IMyServiceResolver resolver)
{
MyController.resolver = resolver;
}
public MyController()
{
// Now we support a default constructor
// since maybe someone else is instantiating this type
// that we don't control.
}
protected void DoActions(Type[] types)
{
var services = resolver.Resolve(types);
foreach(var service in services)
{
service.DoAction();
}
}
}
You would then call this at your application start up to initialize the resolver:
MyController.InitializeResolver(new NinjectMyServiceResolver(kernal));
We did this to handle elements created in XAML who require dependencies resolved but we wanted to remove Service Locator like requests.
Please excuse any syntactical errors :)
I'm writing a blog post series on the topic of refactoring an MVVM application with Service Locator calls in the view models you might find interesting. Part 2 is coming soon :)
http://kellabyte.com/2011/07/24/refactoring-to-improve-maintainability-and-blendability-using-ioc-part-1-view-models/

Maybe you should just do away the Kernel, Types and MySpecialResolver and let the subclasses call DoActions with the IMyServiceInterface instances they need as argument directly. And let the subclasses decide how they get to these instances - they should know best (or in case they don't know which exactly the one who ever decides which instances of IMyServiceInterface are needed)

I would have liked to have a bit more information before posting this answer, but Kelly put me on the spot. :) Telling me to put my code where my mouth is, so to speak.
Like I said in my comment to Kelly, I disagree with moving the resolver/locator from a static implementation to an injected implementation. I agree with ChrisChris that the dependencies the derived type needs should be resolved in that class and not delegated to the base class.
That said, here is how I would remove the service location...
Create Command Interface
First of all I would create a command interface for the specific implementation. In this case the types sent with the DoActions method are generated from attributes, so I would create an IAttributeCommand. I am adding a Matches method to the command in order to declare the command for use by certain types.
public interface IAttributeCommand
{
bool Matches(Type type);
void Execute();
}
Add Command Implementations
To implement the interface, I pass in the specific dependencies I need to execute my command (to be resolved by my container). I add a predicate to my Matches method, and define my Execute behavior.
public class MyTypeAttributeCommand : IAttributeCommand
{
MyDependency dependency;
SomeOtherDependency otherDependency;
public MyTypeAttributeCommand (MyDependency dependency, ISomeOtherDependency otherDependency)
{
this.dependency = dependency;
this.otherDependency = otherDependency
}
public bool Matches(Type type)
{
return type==typeof(MyType)
}
public void Execute()
{
// do action using dependency/dependencies
}
}
Register Commands with Container
In StructureMap (use your favorite container), I would register the array like so:
Scan(s=>
{
s.AssembliesFromApplicationBaseDirectory();
s.AddAllTypesOf<IAttributeCommand>();
s.WithDefaultConventions();
}
Select and Execute Commands Based on Type
Finally, on the base class, I define an IAttributeCommand array in my constructor arguments to be injected by the IOC container. When the derived type passes in the types array, I will execute the correct command based on the predicate.
public abstract class MyController : Controller
{
protected IAttributeCommand[] commands;
public MyController(IAttributeCommand[] commands) { this.commands = commands); }
protected void DoActions(Type[] types)
{
foreach(var type in types)
{
var command = commands.FirstOrDefault(x=>x.Matches(type));
if (command==null) continue;
command.Execute();
}
}
}
If you multiple commands can handle one type, you can change the implementation: commands.Where(x=>x.Matches(type)).ToList().ForEach(Execute);
The effect is the same, but there is a subtle difference in how the class is constructed. The class has no coupling to an IOC container and there is no service location. The implementation is more testable as the class can be constructed with its real dependencies, with no need to wire up a container/resolver.

What is the design pattern in this code?

Say I have a singleton-ish, factory-ish, reflection-ish class that receives some input, and spits back a new instance of a concrete implementation of some interface. What kind of design is this? Is there a better way to do what I want?
Here's some code to illustrate the point:
using System;
using System.Collections.Generic;
// static factory class
public static class ArticleFactory
{
// given an SKU, store the Type object for an IArticle object
private static Dictionary<string, Type> articleRegistry = new Dictionary<string, Type>();
// allow public registration of SKU-to-Type object relationships
public static bool Register(string sku, Type typeInfo)
{
if(!articleRegistry.ContainsKey(sku))
{
articleRegistry.Add(sku, typeInfo);
return true;
}
return false;
}
// given a SKU, give me an instance of the related IArticle object
public static IArticle NewArticle(string sku)
{
if(articleRegistry.ContainsKey(sku))
{
// use reflection to invoke the default constructor
return articleRegistry[sku].GetConstructor(Types.EmptyTypes).Invoke(null) as IArticle;
}
return null;
}
}
// example concrete-implementation of an IArticle
public class Jeans : IArticle
{
public decimal GetPrice() { return SomeDecimal(); }
}
// WHERE DO I CALL THIS LINE?
ArticleFactory.Register("0929-291", typeof(Jeans));
// Later on, if another group needs to write the class for Snowboards,
// how can they self-register their class, without changing any "Main()"
// or "Page_Init()" function?

Looks like you've already identified the pattern. It's the Factory Method Pattern. Or rather, a somewhat half-baked implementation of one. A slightly better approach would be to first make it an interface:
public interface IArticleFactory
{
IArticle CreateArticle(string sku);
}
Then implement the factory without any Reflection at all:
public class MyArticleFactory
{
private Dictionary<string, Func<IArticle>> instantiators =
new Dictionary<string, Func<Iarticle>>();
public MyArticleFactory()
{
Register("Jeans", () => new Jeans());
Register("Shirt", () => new Shirt());
// etc.
}
public IArticle CreateArticle(string sku)
{
Func<IArticle> instantiator;
if (creators.TryGetValue(sku, out instantiator))
return instantiator();
throw new UnknownSkuException(sku);
}
protected void Register(string sku, Func<IArticle> instantiator)
{
creators.Add(sku, instantiator);
}
}
A few important differences:
Registration isn't public, nor should it be. Registration usually either resides in a configuration file somewhere or is private.
Does not require the IArticle concrete types to have a default parameterless constructor. This can easily register articles with parameterized constructors (as long as it knows what parameters to use).
Throws an exception on duplicate registrations. I don't like the idea of simply returning false; if you try to register the same factory method twice, that ought to be considered a bug.
It's not static. You can replace this factory with a different factory. You can unit-test it.
Of course, an even better approach would just be to use any of the myriad of existing .NET Dependency Injection/Inversion of Control Frameworks, such as Ninject or AutoFac.

I don't know if it has a "name" as such, but it looks like some kind of manual service resolver. The problem I can see (from experience, sadly) is that it is inflexible in real terms, in that:
the registration only has a single configuration
it is hard to unit test
Personally I'd look at an IoC container if I was doing this in a new system; the IoC can handle this relationship, and provide a lot more capabilities for free (lifetimes, enrichment, extra setup, etc), and solve many associated problems.
BTW, it may be easier to:
return Activator.CreateInstance(articleRegistry[sku]);

I think what you're doing here is basically Dependency Injection (or Inversion of Control is what the cool kids call it). Have a look at these links:
Explanation from Wikipedia: http://en.wikipedia.org/wiki/Dependency_Injection
Two DI .Net frameworks:
StructureMap: http://structuremap.sourceforge.net/QuickStart.htm
Castle Windsor: http://www.castleproject.org/container/index.html

It's just a factory pattern that happens to use reflection in its implementation. Rather than using reflection, though, it would probably be more efficient to simply put instances of factory classes directly in the dictionary, though this might require some boilerplate code.