Windsor - pulling Transient objects from the container

Question

How can I pull objects from the container that are transient in nature? Do I have to register them with the container and inject in the constructor of the needing class? Injecting everything into the constructor doesn't feel good. Also just for one class I don't want to create a TypedFactory and inject the factory into the needing class.

Another thought that came to me was "new" them up on need basis. But I am also injecting a Logger component (through property) into all my classes. So if I new them up, I will have to manually instantiate the Logger in those classes. How can I continue to use the container for ALL of my classes?

Logger injection:

Answer 1

From the samples you give it is hard to be very specific, but in general, when you inject ILogger instances into most services, you should ask yourself two things:

1. Do I log too much?
2. Do I violate the SOLID principles?

1. Do I log too much

You are logging too much, when you have a lot of code like this:

try
{
   // some operations here.
}
catch (Exception ex)
{
    this.logger.Log(ex);
    throw;
}

Writing code like this comes from the concern of losing error information. Duplicating these kinds of try-catch blocks all over the place, however, doesn't help. Even worse, I often see developers log and continue (they remove the last throw statement). This is really bad (and smells like the old VB ON ERROR RESUME NEXT behavior), because in most situations you simply have not enough information to determine whether it is safe continue. Often there is a bug in the code or a hiccup in an external resource like a database that caused the operation to fail. To continue means that the user often gets the idea that the operation succeeded, while it hasn't. Ask yourself: what is worse, showing users a generic error message saying that there something gone wrong and ask them to try again, or silently skipping the error and letting users think their request was successfully processed? Think about how users will feel if they found out two weeks later that their order was never shipped. You’d probably lose a customer. Or worse, a patient’s MRSA registration silently fails, causing the patient not to be quarantined by nursing and resulting in the contamination of other patients, causing high costs or perhaps even death.

Most of these kinds of try-catch-log lines should be removed and you should simply let the exception bubble up the call stack.

Shouldn't you log? You absolutely should! But if you can, define one try-catch block at the top of the application. With ASP.NET, you can implement the Application_Error event, register an HttpModule or define a custom error page that does the logging. With Win Forms the solution is different, but the concept stays the same: Define one single top most catch-all.

Sometimes, however, you still want to catch and log a certain type of exception. A system I worked on in the past let the business layer throw ValidationExceptions, which would be caught by the presentation layer. Those exceptions contained validation information for display to the user. Since those exceptions would get caught and processed in the presentation layer, they would not bubble up to the top most part of the application and didn't end up in the application's catch-all code. Still I wanted to log this information, just to find out how often the user entered invalid information and to find out whether the validations were triggered for the right reason. So this was no error logging; just logging. I wrote the following code to do this:

try
{
   // some operations here.
}
catch (ValidationException ex)
{
    this.logger.Log(ex);
    throw;
}

Looks familiar? Yes, looks exactly the same as the previous code snippet, with the difference that I only caught ValidationException exceptions. However, there was another difference that can't be seen by just looking at this snippet. There was only one place in the application that contained that code! It was a decorator, which brings me to the next question you should ask yourself:

2. Do I violate the SOLID principles?

Things like logging, auditing, and security, are called cross-cutting concerns (or aspects). They are called cross cutting, because they can cut across many parts of your application and must often be applied to many classes in the system. However, when you find you're writing code for their use in many classes in the system, you are most likely violating the SOLID principles. Take for instance the following example:

public void MoveCustomer(int customerId, Address newAddress)
{
    var watch = Stopwatch.StartNew();

    // Real operation
    
    this.logger.Log("MoveCustomer executed in " +
        watch.ElapsedMiliseconds + " ms.");
}

Here you measure the time it takes to execute the MoveCustomer operation and you log that information. It is very likely that other operations in the system need this same cross-cutting concern. You start adding code like this for your ShipOrder, CancelOrder, CancelShipping, and other use cases, and this leads to a lot of code duplication and eventually a maintenance nightmare (I've been there.)

The problem with this code can be traced back to a violation of the SOLID principles. The SOLID principles are a set of object-oriented design principles that help you in defining flexible and maintainable (object-oriented) software. The MoveCustomer example violated at least two of those rules:

1. The Single Responsibility Principle—classes should have a single responsibility. The class holding the MoveCustomer method, however, does not only contain the core business logic, but also measures the time it takes to do the operation. In other words, it has multiple responsibilities.
2. The Open-Closed principle (OCP)—it prescribes an application design that prevents you from having to make sweeping changes throughout the code base; or, in the vocabulary of the OCP, a class should be open for extension, but closed for modification. In case you need to add exception handling (a third responsibility) to the MoveCustomer use case, you (again) have to alter the MoveCustomer method. But not only do you have to alter the MoveCustomer method, but many other methods as well, as they will typically require that same exception handling, making this a sweeping change.

The solution to this problem is to extract the logging into its own class and allow that class to wrap the original class:

// The real thing
public class MoveCustomerService : IMoveCustomerService
{
    public virtual void MoveCustomer(int customerId, Address newAddress)
    {
        // Real operation
    }
}

// The decorator
public class MeasuringMoveCustomerDecorator : IMoveCustomerService
{
    private readonly IMoveCustomerService decorated;
    private readonly ILogger logger;

    public MeasuringMoveCustomerDecorator(
        IMoveCustomerService decorated, ILogger logger)
    {
        this.decorated = decorated;
        this.logger = logger;
    }

    public void MoveCustomer(int customerId, Address newAddress)
    {
        var watch = Stopwatch.StartNew();

        this.decorated.MoveCustomer(customerId, newAddress);
    
        this.logger.Log("MoveCustomer executed in " +
            watch.ElapsedMiliseconds + " ms.");
    }
}

By wrapping the decorator around the real instance, you can now add this measuring behavior to the class, without any other part of the system to change:

IMoveCustomerService command =
    new MeasuringMoveCustomerDecorator(
        new MoveCustomerService(),
        new DatabaseLogger());

The previous example did however just solve part of the problem (only the SRP part). When writing the code as shown above, you will have to define separate decorators for all operations in the system, and you'll end up with decorators like MeasuringShipOrderDecorator, MeasuringCancelOrderDecorator, and MeasuringCancelShippingDecorator. This lead again to a lot of duplicate code (a violation of the OCP principle), and still needing to write code for every operations in the system. What's missing here is a common abstraction over use cases in the system.

What's missing is an ICommandHandler<TCommand> interface.

Let's define this interface:

public interface ICommandHandler<TCommand>
{
    void Execute(TCommand command);
}

And let's store the method arguments of the MoveCustomer method into its own (Parameter Object) class called MoveCustomerCommand:

public class MoveCustomerCommand
{
    public int CustomerId { get; set; }
    public Address NewAddress { get; set; }
}

And let's put the behavior of the MoveCustomer method in a class that implements ICommandHandler<MoveCustomerCommand>:

public class MoveCustomerCommandHandler : ICommandHandler<MoveCustomerCommand>
{
    public void Execute(MoveCustomerCommand command)
    {
        int customerId = command.CustomerId;
        Address newAddress = command.NewAddress;
        // Real operation
    }
}

This might look weird at first, but because you now have a general abstraction for use cases, you can rewrite your decorator to the following:

public class MeasuringCommandHandlerDecorator<TCommand>
    : ICommandHandler<TCommand>
{
    private ILogger logger;
    private ICommandHandler<TCommand> decorated;

    public MeasuringCommandHandlerDecorator(
        ILogger logger,
        ICommandHandler<TCommand> decorated)
    {
        this.decorated = decorated;
        this.logger = logger;
    }

    public void Execute(TCommand command)
    {
        var watch = Stopwatch.StartNew();

        this.decorated.Execute(command);
    
        this.logger.Log(typeof(TCommand).Name + " executed in " +
            watch.ElapsedMiliseconds + " ms.");
    }
}

This new MeasuringCommandHandlerDecorator<T> looks much like the MeasuringMoveCustomerDecorator, but this class can be reused for all command handlers in the system:

ICommandHandler<MoveCustomerCommand> handler1 =
    new MeasuringCommandHandlerDecorator<MoveCustomerCommand>(
        new MoveCustomerCommandHandler(),
        new DatabaseLogger());

ICommandHandler<ShipOrderCommand> handler2 =
    new MeasuringCommandHandlerDecorator<ShipOrderCommand>(
        new ShipOrderCommandHandler(),
        new DatabaseLogger());

This way it will be much, much easier to add cross-cutting concerns to your system. It's quite easy to create a convenient method in your Composition Root that can wrap any created command handler with the applicable command handlers in the system. For instance:

private static ICommandHandler<T> Decorate<T>(ICommandHandler<T> decoratee)
{
    return
        new MeasuringCommandHandlerDecorator<T>(
            new DatabaseLogger(),
            new ValidationCommandHandlerDecorator<T>(
                new ValidationProvider(),
                new AuthorizationCommandHandlerDecorator<T>(
                    new AuthorizationChecker(
                        new AspNetUserProvider()),
                    new TransactionCommandHandlerDecorator<T>(
                        decoratee))));
}

This method can be used as follows:

ICommandHandler<MoveCustomerCommand> handler1 = 
    Decorate(new MoveCustomerCommandHandler());

ICommandHandler<ShipOrderCommand> handler2 =
    Decorate(new ShipOrderCommandHandler());

If your application starts to grow, however, it can get useful to bootstrap this with a DI Container, because a DI Container can support Auto-Registration. This prevents you from having to make changes to your Composition Root for every new command/handler pair you add to the system. Especially when your decorators have generic type constraints, a DI Container be extremely useful.

Most modern DI Containers for .NET have fairly decent support for decorators nowadays, and especially Autofac (example) and Simple Injector (example) make it easy to register open-generic decorators. Both libraries even allow decorators to be applied conditionally based on a given predicate, and Simple Injector even allows them to be applied conditionally based on generic type constraints, allowing the decorated class to be injected as a factory, and allowing a contextual context to be injected into decorators.

Unity and Castle, on the other hand, have Dynamic Interception facilities (as Autofac does to btw). Dynamic Interception has a lot in common with decoration, but it uses dynamic proxy generation under the covers. This can be more flexible than working with generic decorators, but you pay the price when it comes to maintainability, because you often loose type safety and interceptors always force you to take a dependency on the interception library, while decorators are type safe and can be written without taking a dependency on an external library.

Read this article if you want to learn more about this way of designing your application: Meanwhile... on the command side of my architecture.

UPDATE [2019]: I coauthored a book called Dependency Injection Principles, Practices, and Patterns, which goes into much more detail on this SOLID programming style and the design described above (see chapter 10).