How should I architect logging within my application?

So I've done a lot of research on this and haven't found any answers where I said, "yes, THAT". I'm hoping the ever-erudite StackOverflow crowd can help me out.

I've run into this issue in a couple of different scenarios. Say I have a C# app, and there are important things that I want to log.

public class MyClass
{
    ... 

    public void ImportantMethod()
    {
        DoInterestingThing();

        var result = SomethingElseImportant();
        if (result == null)
        {
            logger.Log("I wasn't expecting that. No biggie.");
            return;
        }

        MoreInterestingStuff(); 
}

The thing I'm interested in is where do I get logger from.

As I see it I have a few options.

1. Have it injected into MyClass in the constructor.
2. Retrieve it using a globally available service locator.
3. Use method decorators and AOP to have my logging done for me.

None of these seem like great options. #3 looks to be impossible since I'm logging right in the middle of my business logic and not just doing a simple trace of my method calls, input parameters and/or exceptions thrown. #2, though simple seems like it would be really difficult to unit test. Of course, I would want to unit test everything. #1, though it would work fine, clutters up all my business logic with logging objects that have nothing to do with the business objects themselves.

Any alternative ideas, or thoughts on one of the options above? Much thanks!

EDIT: just to be clear, I already know how to do DI (I use Unity), and I already know a good logging framework (I use log4net). Just wondering a how to use logging in an architecture sense across an application in the smartest way.

* EDIT *

I marked Mark Seeman's answer as the solution. I went through my application and found that most of my logging calls were doing the same thing a decorator could do. Namely, log the entry to the method, any exceptions thrown, and exit return values.

Some cases I still needed to log directly inside a method. An example would be where I want to fail fast in a method which doesn't return anything but not throw an Exception. In those cases I have a singleton which holds a reference a LogProvider which will in turn retrieve a named log instance. The code looks similar to this:

private ILog logger = LogProviderFactory.Instance.GetLogger(typeof(Foo));

LogProviderFactory has a method SetProvider which allows you to swap out the singleton. So in unit testing I can do:

// LogProviderFactory.Instance now is our mock
LogProviderFactory.SetProvider(MockLogProvider);

The logging decorator uses the same LogProvider as the singleton (which it obtains through injection), so logging is unified throughout the system.

So really the end solution was mostly option #3, and a hybrid #2 (where it's the service locator pattern but the service is 'injected' into the locator).

AOP

As far as "aspect oriented programming" goes, I was a bit disappointed in the limitations of the language. Hoping AOP will be treated as a first-class citizen in future releases.

• I tried PostSharp but couldn'开发者_StackOverflow中文版t get it running on my machine correctly. Additionally it was a big restriction that you had to have PostSharp installed on your system to use it (as opposed to just calling a dll which comes with the solution or something similar).
• I used LinFu and was able to get it partially working. It blew up in a few instances however. The new 2.0 release is barely documented, so that was a hurdle.
• Interface interception with Unity however seems to work well out of the box. I got lucky that most things I wanted to log were in classes implementing interfaces.

Use a logging Decorator.

Two bits:

(1) - A pre-built logging framework.

Some people like Log4Net but I'm a EntLibs fan. This does the heavy lifting in terms of actual logging. Tools like EntLibs will let you log to different types of logging repositories (database, message queue, rolling text file, etc). They'll also let you log to different instances based on categories and so forth. They are usually highly configurable.

(2) - Custom class(es) that wrap the logging framework.

So "logger" is something you write, and it calls the logging framework to do the actual logging.

I like this approach for several reasons:

• You decouple the logging framework (#1) from the rest of your application as the custom wrappers go into a separate assembly.
• By writing your own Logging API you can define method signatures that suit your needs, and you can expand on them.
• If your working on a team, you can make the method signatures very easy to use so that no-one has grounds to say that using the logging was too hard.
• It keeps the logging consistent. It also makes it easy to do code searches for "illegal" code that writes to files, console or event log easy as there won't be any as part of your logging (it's all in the framework).
• By writing specific custom classes for each tier you can pre-populate a lot data behind the scenes making life easier for whoever's writing the actual application code. You can set severity, priority, default event Ids, categories and more.
• It scales well in terms of application complexity and growth; it might seem heavy handed for smaller apps but you'd have plenty of head-room if it starts to grow on you over time.

Here's an example of a Informational logging class in a project I've worked on. It has a bunch of easy to call public methods, and one private method that calls the framework (ConcreteLogInformation).

public static void LogInformation(string title, string message)

public static void LogInformation(string title, Dictionary<string, object> extendedProperties)

public static void LogInformation(string title, int eventId, Dictionary<string, object> extendedProperties)

public static void LogInformation(string title, string message, Dictionary<string, object> extendedProperties)

public static void LogInformation(string title, string message, int eventId)

public static void LogInformation(string title, string message, int eventId, Dictionary<string, object> extendedProperties)

public static void LogInformation(string title, string message, int eventId, string category)

public static void LogInformation(string title, string message, int eventId, Dictionary<string, object> extendedProperties, string category)

private static void ConcreteLogInformation(string title, string message, int eventId, Dictionary<string, object> extendedProperties, string category)

In the Using the request context in a Factory or Factory Method to do contextual binding section of the Ninject Contetual Binding docs I have an example of leveraging your container to inject an appropriate logger for your class by doing (in Ninjectese):

Bind<ILog>().ToMethod( context => LogFactory.CreateLog( context.Request.Target.Type ) );

For tracing type stuff, Mark's interception article describes the best approach.

And can I ask again that you read @Mark Seemann's cited articles in depth before just discarding them without an upvote.