I am about to switch from Windsor to Structuremap for an existing project with ~100 registered components (mostly singletons).
All components inherit from a common base class that provides logging and health tracking and for this reason, contains a "Name" property used to identify component instances.
With Windsor, it was possible to set the component's Name property to the name that was used to register the component in the IOC container (We used a Facility for this).
My question: Is something like this possible with Structuremap?
(I dream of a call to c.For<IFoo>.Use<Bar>.Named("Doe") that magically results in instanceOfBar.Name = "Doe" somewhere.)
Here is what I tried:
using System;
using StructureMap;
using StructureMap.Interceptors;
using System.Diagnostics;
namespace ConsoleApplication1
{
interface IServiceA { }
interface IServiceB { }
class Base
{
public string Name { get; set; }
}
class ComponentA : Base, IServiceA { }
class ComponentB : Base, IServiceB
{
public ComponentB(IServiceA serviceA)
{
this.ServiceA = serviceA;
}
public IServiceA ServiceA { get; private set; }
}
class SetNameInterceptor : TypeInterceptor
{
public bool MatchesType(Type type) { return true; }
public object Process(object target, IContext context)
{
// *** Any other way? This does not work...
string name = context.BuildStack.Current != null ? context.BuildStack.Current.Name : context.RequestedName;
((Base)target).Name = name;
return target;
}
}
class Program
{
static void Main(string[] args)
{
Container container = new Container(c =>
{
c.RegisterInterceptor(new SetNameInterceptor());
c.For<IServiceA>().Use<ComponentA>().Named("A");
c.For<IServiceB>().Use<ComponentB>().Named("B");
});
var b = container.GetInstance<IServiceB>();
// both Fail:
Debug.Assert(((ComponentB)b).Name == "B");
Debug.Assert(((ComponentA)((ComponentB)b).ServiceA).Name == "A");
}
}
}
The above obviously does not work, I tried several variations but had no luck. The registered name of the target object does not seem to be consistently reachable via IContext.
My second best approach would be to define a new "NamedComponent(...)" extension method that resolves to Named(name).WithProperty(x => x.Name).EqualTo(name), but I wonder if this can be avoided to keep component registration as "structuremap-like" as possible?
Am I missing something?
I've never used WithProperty before but if it works the way I'd expect it should do the trick for you.
I think I would favor using EnrichWith though. Something like:
c.For<IFoo>().Use<Foo>().Named(name).EnrichWith(f => f.Name = name);
EnrichWith is a bit more explicit about what it's doing IMO, and lets you call any code on your instance before returning it to the caller. I like that this lets you do a straightforward assignment as well.
There is also a more complex handler you can use with EnrichWith that gives access to the context of the request - this would allow you to do something like this:
c.For<IFoo>().Use<Foo>().Named(name)
.EnrichWith((c, i) => {
i.Name = c.RequestedName;
return i;
});
This may be overkill for your situation but the contextual awareness can be pretty useful.
Related
[This is for a Windows 10 IoT UWP app on a Raspberry Pi 2]:
Concretely, I'm trying to create two serial ports and link each one to a device driver for device that has a serial connection (I have two identical devices and a serial port to talk to each). I have a class (call it DeviceDriver) that implements an IDeviceDriver interface for this type of device. My hardware configuration also includes an external chip that has multiple serial ports. I have a class SerialPort for those and they implement an ISerialPort interface.
I need two instances of DeviceDriver as I have two devices, and two instances of SerialPort - one for each device. I can get Ninject to make one serialport and pass the ISerialPort object to the DeviceDriver constructor. Where I am stuck is that I want two DeviceDriver objects; one that gets linked to a SerialPort (COM1) and the other gets linked to a separate SerialPort (COM2). Examples for Ninject show you can bind multiple different classes to one instance (a Shuriken and Sword can both be bound to IWeapon), but I don't see how to bind a COM1 serialport and a COM2 serialport to ISerialPort - it gives me a compilation error. So how do I create two instances of the same SerialPort class (with different constructor arguments to say one is COM1 and the other is COM2, I know how to specify constructor arguments already) and then tell Ninject which SerialPort to pass to two instances of a DeviceDriver class, where one needs COM1 and one needs COM2?
My DeviceDriver basically looks like this:
public class DeviceDriver :IDeviceDriver
{
ISerialPort m_localPort;
public DeviceDriver(ISerialPort port)
{
m_localPort = port;
}
// Other stuff
// ...
}
Anybody have any thoughts how I can do this? The following link is the only thing I've found, but they are talking about Unity and XML configuration files and it seems overly complex for what I'm trying to do.
Initialising configurable objects with dependency injection container
Thanks!
Let's say we've got the following implementation:
public class SerialPortAddress
{
public SerialPortAddress(string address)
{
this.Address = address;
}
public string Address { get; }
}
public interface ISerialPort
{
SerialPortAddress Address { get; }
}
public class SerialPort : ISerialPort
{
public SerialPort(SerialPortAddress address)
{
this.Address = address;
}
public SerialPortAddress Address { get; }
}
public interface IDeviceDriver
{
ISerialPort SerialPort { get; }
}
public class DeviceDriver : IDeviceDriver
{
public DeviceDriver(ISerialPort serialPort)
{
SerialPort = serialPort;
}
public ISerialPort SerialPort { get; }
}
Multi-Injection
We can then create the bindings as follows and retrieve a list of IDeviceDrivers with their serial ports as follows:
public class Test
{
[Fact]
public void Bla()
{
var com1 = new SerialPortAddress("COM1");
var com2 = new SerialPortAddress("COM2");
var kernel = new StandardKernel();
kernel.Bind<ISerialPort>().To<SerialPort>();
kernel.Bind<IDeviceDriver>().To<DeviceDriver>()
.WithParameter(new TypeMatchingConstructorArgument(
typeof(SerialPortAddress),
(ctx, target) => com1, true));
kernel.Bind<IDeviceDriver>().To<DeviceDriver>()
.WithParameter(new TypeMatchingConstructorArgument(
typeof(SerialPortAddress),
(ctx, target) => com2, true));
var deviceDrivers = kernel.Get<List<IDeviceDriver>>();
deviceDrivers.Should().HaveCount(2)
.And.Contain(x => x.SerialPort.Address == com1)
.And.Contain(x => x.SerialPort.Address == com2);
}
}
Also see Multi Injection
Named Bindings
Alternatively, if you need to know which IDeviceDrivers is which, you can also use named bindings:
[Fact]
public void NamedBindings()
{
const string DeviceDriver1 = "DeviceDriver1";
const string DeviceDriver2 = "DeviceDriver2";
var com1 = new SerialPortAddress("COM1");
var com2 = new SerialPortAddress("COM2");
var kernel = new StandardKernel();
kernel.Bind<ISerialPort>().To<SerialPort>();
kernel.Bind<IDeviceDriver>().To<DeviceDriver>()
.Named(DeviceDriver1)
.WithParameter(new TypeMatchingConstructorArgument(
typeof(SerialPortAddress),
(ctx, target) => com1, true));
kernel.Bind<IDeviceDriver>().To<DeviceDriver>()
.Named(DeviceDriver2)
.WithParameter(new TypeMatchingConstructorArgument(
typeof(SerialPortAddress),
(ctx, target) => com2, true));
kernel.Get<IDeviceDriver>(DeviceDriver1).SerialPort.Address.Should().Be(com1);
kernel.Get<IDeviceDriver>(DeviceDriver2).SerialPort.Address.Should().Be(com2);
}
Factory
Finally, you could also create the components by factory, which requires a factory interface to begin with:
public interface IDeviceDriverFactory
{
IDeviceDriver Create(SerialPortAddress address);
}
using Ninject.Extensions.Factory we can now do the following:
[Fact]
public void Factory()
{
var com1 = new SerialPortAddress("COM1");
var com2 = new SerialPortAddress("COM2");
var kernel = new StandardKernel();
kernel.Bind<ISerialPort>().To<SerialPort>();
kernel.Bind<IDeviceDriver>().To<DeviceDriver>();
kernel.Bind<IDeviceDriverFactory>()
.ToFactory(() => new TypeMatchingArgumentInheritanceInstanceProvider());
var factory = kernel.Get<IDeviceDriverFactory>();
factory.Create(com1).SerialPort.Address.Should().Be(com1);
factory.Create(com2).SerialPort.Address.Should().Be(com2);
}
EDIT: Ninject.Extension.Factory may not run on the raspberry pi.
If that's the case you might need to implement the factory yourself:
public class DeviceDriverFactory : IDeviceDriverFactory
{
private readonly IResolutionRoot resolutionRoot;
public DeviceDriverFactory(IResolutionRoot resolutionRoot)
{
this.resolutionRoot = resolutionRoot;
}
public IDeviceDriver Create(SerialPortAddress address)
{
var serialPortAddressParameter = new TypeMatchingConstructorArgument(
typeof(SerialPortAddress),
(ctx, t) => address)
this.resolutionRoot.Get<IDeviceDriver>(serialPortAddressParameter);
}
}
Bind<IDeviceDriverFactory>().To<DeviceDriverFactory>();
I'm not familiar Ninject or Unity, but Castle Windsor has a lifestyle called Pooled, which will create up to a specified number of instances and then return those instances to the pool of instances after they've been released. When using this lifestyle, Windsor will create as many objects as are requested up to the limit specified, and then either recycle the instance (if you've derived from IRecyclable and implemented the Recycle() method) or dispose of it normally.
You can have your components created using a simple factory method that provides the correct constructor arguments, and then when they are returned to the pool they will be correctly configured.
EDIT:
If you're set on using Ninject, then I would solve this problem by injecting an ISerialPortFactory into the constructor of DeviceDriver and using that to create your ISerialPort objects. Since your DeviceDriver class doesn't care which ISerialPort it's using, the factory can be used to manage the instances that you need.
Your factory would look something like this:
public interface ISerialPortFactory
{
ISerialPort CreateNext();
}
public class SerialPortFactory : ISerialPortFactory
{
public ISerialPort CreateNext()
{
var serialPortConfiguration = GetNextConfiguration();
return new SerialPort(serialPortConfiguration);
}
private GetNextConfiguration()
{
// you could manage some kind of internal registry of COMx configurations here
}
}
And your client DeviceDriver class would look like this:
public class DeviceDriver : IDeviceDriver
{
public DeviceDriver(ISerialPortFactory factory)
{
m_localPort = factory.CreateNext();
}
}
The abstract factory method is sort of a heavy-handed way of getting what you want, but it's a surefire way to get exactly what you need since you have complete control over it. Its main use case is to resolve dependencies where you don't necessarily know the exact implementation you want until runtime.
Someone told me that before dependency injection frameworks came around there developers would use a factory to implement DI. Can anyone provide an example how a factory pattern can be used for DI. I mean just by thinking about it a factory is a depenendency injector but i cant find any examples on the web.
This off the top of my head, untested code (with C#)
public class CarFactory : ICarFactory{
private static CarFactory instance = null;
public static ICarFactory SingletonInstance {
get {
if (this.instance == null){
this.instance = new CarFactory();
return this.instance;
}
},
set {
this.instance = value;
}
}
public ICar CreateCar(string make){
switch(make)
{
case "Toyota": return new Toyota();
case "Honda" : return new Honda();
default: throw new Exception();
}
}
}
public interface ICarFactory {
ICar CreateCar(string make);
}
public class Toyota : ICar
{
}
public class Honda : ICar
{
}
And usage would be something like :
ICar car = CarFactory.SingletonInstance.CreateCar("Toyota");
Exposing the singleton instance of the CarFactory publically enables you to mock the CarFactory for your unit tests and you can have your mocked CarFactory return mocked ICars when calling CreateCar.
Now replace the Cars in the factory by actual dependencies such as classes that implement services. And voila, you have a Factory that contains all your dependencies. You can now use the Factory to "resolve" your dependencies. You can take the example and push it further by using generic types and a dictionary (hashtable) where the key is the type name and the value the implementation instance such as:
public T Create<T>(){
return mydictionary.get(typeof(T).Name);
}
Something like that... You get the drift...
Hope it helps!
Following on from this question I would like autofac to inject the type of the declaring object into the constructor of my NLog service, so that it can correctly log which type is logging entries.
My NLogService class looks like this...
public class NLogService : ILogService
{
private readonly Logger _logger;
public NLogService(Type t)
{
var consumerType = t.DeclaringType.FullName;
_logger = LogManager.GetLogger(consumerType);
}
However it fails on app startup because it obviously cannot work out what to inject into the constructor of the NLogService with the following error...
None of the constructors found with
'Public binding flags' on type
'MyProduct.Domain.Services.Logging.NLogService'
can be invoked with the available
services and parameters: Cannot
resolve parameter 'System.Type t' of
constructor 'Void .ctor(System.Type)'.
So, my question is - how do i instruct autofac to inject the type of the calling class?
I tried this...
public NLogService(Type t)
{
var method = MethodBase.GetCurrentMethod();
Type consumingType = method.DeclaringType;
var consumerType = consumingType.FullName;
var consumerType = t.DeclaringType.FullName;
_logger = LogManager.GetLogger(consumerType);
}
But i just end up with MyProduct.Domain.Services.Logging.NLogService
What i want is the type of the class that is doing the actual logging.
i have already tried this suggestion and it didnt work for me either.
Could make your NLogService generic, i.e. NLogService<T> and use Autofac's open generics support?
Then you could do this:
public class NLogService<T> : ILogger<T>
{
private readonly Logger _logger;
public NLogService()
{
_logger = LogManager.GetLogger(typeof(T).FullName);
}
}
There is no real good way to do this with Autofac, because does not have support for 'context based injection' (which is what you are trying to do). There is a workaround, but it aint pretty...
What you can do is revert to property injection and define a base class or interface for that ILogService property. For instance, you can define the following interface:
public interface ILoggerContainer
{
public ILogService Logger { get; set; }
}
Now you can implement this interface on all types that need a logger:
public class Consumer : IConsumer, ILoggerContainer
{
public ILogService Logger { get; set; }
}
With this in place you can configure Autofac as follows:
builder.RegisterType<ILoggerContainer>()
.OnActivating(e =>
{
var type = typeof(LogService<>)
.MakeGenericType(e.Instance.GetType());
e.Instance.Logger = e.Context.Resolve(type);
});
Another workaround, that you may find cleaner is to inject an ILogger<T> with the same type as the type of the parent type:
public class Consumer : IConsumer
{
public Consumer(ILogger<Consumer> logger) { }
}
This makes the configuration much easier and prevents you from having to have a base class. Which one is most appropriate is up to you.
As I said, these are workarounds, but to be honest, you might need to reconsider your logging strategy in your application. Perhaps you are logging at too many places. In the applications I write there is hardly ever a need to log, and when I do, I write an logging message that is expressive enough so that there is no need to communicate the type that triggered the event. And when you log exception, you will always have a complete stack trace (and exception logging should almost only happen in the outer layer of your application and not within services anyway).
The following technique works well in our experience:
Create an attribute like below, which can be applied at class level or at the injection site:
[AttributeUsage(AttributeTargets.Parameter | AttributeTargets.Class)]
public class LoggerAttribute : Attribute
{
public readonly string Name;
public LoggerAttribute(string name)
{
Name = name;
}
}
Create an Autofac module that you register with the ContainerBuilder:
public class LogInjectionModule : Module
{
protected override void AttachToComponentRegistration(IComponentRegistry registry, IComponentRegistration registration)
{
registration.Preparing += OnComponentPreparing;
}
static void OnComponentPreparing(object sender, PreparingEventArgs e)
{
var typePreparing = e.Component.Activator.LimitType;
// By default, the name supplied to the logging instance is the name of the type in which it is being injected into.
string loggerName = typePreparing.FullName;
//If there is a class-level logger attribute, then promote its supplied name value instead as the logger name to use.
var loggerAttribute = (LoggerAttribute)typePreparing.GetCustomAttributes(typeof(LoggerAttribute), true).FirstOrDefault();
if (loggerAttribute != null)
{
loggerName = loggerAttribute.Name;
}
e.Parameters = e.Parameters.Union(new Parameter[]
{
new ResolvedParameter(
(p, i) => p.ParameterType == typeof (Logger),
(p, i) =>
{
// If the parameter being injected has its own logger attribute, then promote its name value instead as the logger name to use.
loggerAttribute = (LoggerAttribute)
p.GetCustomAttributes(typeof(LoggerAttribute),true).FirstOrDefault();
if (loggerAttribute != null)
{
loggerName = loggerAttribute.Name;
}
// Return a new Logger instance for injection, parameterised with the most appropriate name which we have determined above.
return LogManager.GetLogger(loggerName);
}),
// Always make an unamed instance of Logger available for use in delegate-based registration e.g.: Register((c,p) => new Foo(p.TypedAs<Logger>())
new TypedParameter(typeof(Logger), LogManager.GetLogger(loggerName))
});
}
}
You can now inject a named Logger in any one of these ways depending on individual scenarios:
By default, the injected logger name will be given the full type name of the class it is injected into:
public class Foo
{
public Foo(Logger logger)
{
}
}
Use a constructor parameter [Logger] attribute to override the logger name:
public class Foo
{
public Foo([Logger("Meaningful Name")]Logger logger)
{
}
}
Use a class-level [Logger] attribute to set the same logger name override for all constructor overloads:
[Logger("Meaningful Name")]
public class Foo
{
public Foo(Logger logger, int something)
{
}
public Foo(Logger logger, int something, DateTime somethingElse)
{
}
}
Use constructor parameter [Logger] attributes on each constructor overload to set different logger names depending on the context of how you were constructed:
public class Foo
{
public Foo(Logger("Meaningful Name")]Logger logger, int something)
{
}
public Foo(Logger("Different Name")]Logger logger, int something, DateTime somethingElse)
{
}
}
IMPORTANT NOTE: If you register types to be resolved with logger constructor injection using Autofac's delegate registration, you MUST use the two parameter overload like so: Register((c,p) => new Foo(p.TypedAs<Logger>()).
Hope this helps!
It is possible to do this without generics.
However, please note that in Autofac 6.x, the resolution process has changed to use a resolve pipeline. This doesn't matter for most scenarios, but it does when you want to use the lifetime events like OnPreparing, etc. Most of the answers here on SO around overriding the Preparing event are very old and are now outdated. You can't override Preparing directly anymore.
There is an example on the Autofac documentation site doing this for log4net, and it works with NLog with only minor changes. Here is the basic idea:
public class Log4NetMiddleware : IResolveMiddleware
{
public PipelinePhase Phase => PipelinePhase.ParameterSelection;
public void Execute(ResolveRequestContext context, Action<ResolveRequestContext> next)
{
// Add our parameters.
context.ChangeParameters(context.Parameters.Union(
new[]
{
new ResolvedParameter(
(p, i) => p.ParameterType == typeof(ILog),
(p, i) => LogManager.GetLogger(p.Member.DeclaringType)
),
}));
// Continue the resolve.
next(context);
// Has an instance been activated?
if (context.NewInstanceActivated)
{
var instanceType = context.Instance.GetType();
// Get all the injectable properties to set.
// If you wanted to ensure the properties were only UNSET properties,
// here's where you'd do it.
var properties = instanceType
.GetProperties(BindingFlags.Public | BindingFlags.Instance)
.Where(p => p.PropertyType == typeof(ILog) && p.CanWrite && p.GetIndexParameters().Length == 0);
// Set the properties located.
foreach (var propToSet in properties)
{
propToSet.SetValue(context.Instance, LogManager.GetLogger(instanceType), null);
}
}
}
}
Please also note that you have to understand how middleware works in Autofac. The documentation is a good place to start.
How do I handle classes with static methods with Ninject?
That is, in C# one can not have static methods in an interface, and Ninject works on the basis of using interfaces?
My use case is a class that I would like it to have a static method to create an
unpopulated instance of itself.
EDIT 1
Just to add an example in the TopologyImp class, in the GetRootNodes() method, how would I create some iNode classes to return? Would I construct these with normal code practice or would I somehow use Ninject? But if I use the container to create then haven't I given this library knowledge of the IOC then?
public interface ITopology
{
List<INode> GetRootNodes();
}
public class TopologyImp : ITopology
{
public List<INode> GetRootNodes()
{
List<INode> result = new List<INode>();
// Need code here to create some instances, but how to without knowledge of the container?
// e.g. want to create a few INode instances and add them to the list and then return the list
}
}
public interface INode
{
// Parameters
long Id { get; set; }
string Name { get; set; }
}
class NodeImp : INode
{
public long Id
{
get { throw new NotImplementedException(); }
set { throw new NotImplementedException(); }
}
public string Name
{
get { throw new NotImplementedException(); }
set { throw new NotImplementedException(); }
}
}
// Just background to highlight the fact I'm using Ninject fine to inject ITopology
public partial class Form1 : Form
{
private ITopology _top;
public Form1()
{
IKernel kernal = new StandardKernel(new TopologyModule());
_top = kernal.Get<ITopology>();
InitializeComponent();
}
}
If you're building a singleton or something of that nature and trying to inject dependencies, typically you instead write your code as a normal class, without trying to put in lots of (probably incorrect) code managing the singleton and instead register the object InSingletonScope (v2 - you didnt mention your Ninject version). Each time you do that, you have one less class that doesnt surface its dependencies.
If you're feeling especially bloody-minded and are certain that you want to go against that general flow, the main tools Ninject gives you is Kernel.Inject, which one can use after you (or someone else) has newd up an instance to inject the dependencies. But then to locate one's Kernelm you're typically going to be using a Service Locator, which is likely to cause as much of a mess as it is likely to solve.
EDIT: Thanks for following up - I see what you're after. Here's a hacky way to approximate the autofac automatic factory mechanism :-
/// <summary>
/// Ugly example of a not-very-automatic factory in Ninject
/// </summary>
class AutomaticFactoriesInNinject
{
class Node
{
}
class NodeFactory
{
public NodeFactory( Func<Node> createNode )
{
_createNode = createNode;
}
Func<Node> _createNode;
public Node GenerateTree()
{
return _createNode();
}
}
internal class Module : NinjectModule
{
public override void Load()
{
Bind<Func<Node>>().ToMethod( context => () => Kernel.Get<Node>() );
}
}
[Fact]
public void CanGenerate()
{
var kernel = new StandardKernel( new Module() );
var result = kernel.Get<NodeFactory>().GenerateTree();
Assert.IsType<Node>( result );
}
}
The ToMethod stuff is a specific application of the ToProvider pattern -- here's how you'd do the same thing via that route:-
...
class NodeProvider : IProvider
{
public Type Type
{
get { return typeof(Node); }
}
public object Create( IContext context )
{
return context.Kernel.Get<Node>();
}
}
internal class Module : NinjectModule
{
public override void Load()
{
Bind<Func<Node>>().ToProvider<NodeProvider>();
}
}
...
I have not thought this through though and am not recommending this as A Good Idea - there may be far better ways of structuring something like this. #Mark Seemann? :P
I believe Unity and MEF also support things in this direction (keywords: automatic factory, Func)
EDIT 2: Shorter syntax if you're willing to use container-specific attributes and drop to property injection (even if Ninject allows you to override the specific attributes, I much prefer constructor injection):
class NodeFactory
{
[Inject]
public Func<Node> NodeFactory { private get; set; }
public Node GenerateTree()
{
return NodeFactory();
}
}
EDIT 3: You also need to be aware of this Ninject Module by #Remo Gloor which is slated to be in the 2.4 release
EDIT 4: Also overlapping, but not directly relevant is the fact that in Ninject, you can request an IKernel in your ctor/properties and have that injected (but that doesn't work directly in a static method).
public class TheController : Controller
{
IThe the;
public TheController( IThe the)
{
//when User.IsInRole("r1") The1 should be injected else r2
this.the = the;
}
}
public class The1 : IThe{}
public class The2 : IThe{}
//anybody knows a good way of doing this ?
IHandlerSelector is the way to go. See this post for an example of usage.
Alternatively if you prefer AutoFac-like experience you can use factory for that:
container.Register(Component.For<IThe>().UsingFactoryMethod(
c => HttpContext.Current.User.IsInRole("r1") ?
c.Resolve<IThe>("r1") :
c.Resolve<IThe>("r2"));
Or if you want to use specific IThe just in one context, you can use DynamicParameters:
container.Register(Component.For<TheFactory>().Lifestyle.Transient.DynamicParameters(
(c, d) => HttpContext.Current.User.IsInRole("r1") ?
d["the"] = c.Resolve<IThe>("r1") :
c.Resolve<IThe>("r2"));
However the most correct way of doing this is IHandlerSelector
The container-agnostic approach obviously employs an Abstract Factory:
public interface ITheFactory
{
IThe Create(IPrincipal user);
}
You can take a dependency on ITheFactory instead of IThe:
public class TheController : Controller
{
private readonly IThe the;
public TheController(ITheFactory theFactory)
{
if (theFactory == null)
{
throw new ArgumentNullException("theFactory");
}
this.the = theFactory.Create(this.User);
}
}
I can't really remember if this.User is populated at this time, but if it isn't, you can just keep a reference to the factory and lazily resolve your dependency the first time it's requested.
However, Controller.User is a bit special because it ought to be available as Thread.CurrentPrincipal as well. This means that in this special case you don't actually have to introduce an Abstract Factory. Instead, you can write a Decorator that performs the selection every time it's used:
public class UserSelectingThe : IThe
{
private readonly IThe the1;
private readonly IThe the2;
public UserSelectingThe(IThe the1, IThe the2)
{
if (the1 == null)
{
throw new ArgumentNullException("the1");
}
if (the2 == null)
{
throw new ArgumentNullException("the2");
}
this.the1 = the1;
this.the2 = the2;
}
// Assuming IThe defines the Foo method:
public Baz Foo(string bar)
{
if (Thread.CurrentPrincipal.IsInRole("r1"))
{
return this.the1.Foo(bar);
}
return this.the2.Foo(bar);
}
}
In this case, you would be able to use your original TheController class unchanged.
In Autofac:
var builder = new ContainerBuilder();
// Give the different implementations names
builder.RegisterType<The1>.Named<IThe>("r1");
builder.RegisterType<The2>.Named<IThe>("r2");
// Use a function for the default IThe
builder.Register(
c => HttpContext.Current.User.IsInRole("r1") ?
c.Resolve<IThe>("r1") :
c.Resolve<IThe>("r2"))
.As<IThe>()
.ExternallyOwned();
If you have a lot of roles, you can use a method instead of the inline expression, e.g.:
builder.Register(c => ChooseTheImplementation(c))
(BTW, The "ExternallyOwned" modifier tells the container that the result of the function is disposed elsewhere, e.g. via the concrete components. You can usually leave it out but it makes good documentation :))