I've written a series of post on AOP lately (
here,
here and
here), and in the last part I promised to tackle mixins and introductions in a future post. When I was doing my research for just that, I came cross a Java framework (just humor me :p) called
Qi4j (that's 'chee for jay'), written by Swedish
Richard Öberg, pioneering the idea of
Composite Oriented Programming, which instantly put a spell on me. Essentially, it takes the concepts from Aspect Oriented Programming to the extreme, and for the past week I’ve dug into it with a passion. This post is the first fruits of my labor.
OOP is Not Object Oriented!
One of the things that Richard Öberg argues, is that OOP is not really object oriented at all, but rather class oriented. As the Qi4j website proclaims, "class is the first class citizen that objects are derived from. Not objects being the first-class citizen to which one or many classes are assigned". Composite oriented programming (COP) then, tries to work around this limitation by building on a set of core principles; that behavior depends on context, that decoupling is a virtue, and that business rules matter more. For a short and abstract explanation of COP,
see this page. In the rest of this post I'll try and explain some of its easily graspable benefits through a set of code examples, and then in a future post we'll look at how I've morphed the AOP framework I started developing in the previous posts in this series into a lightweight COP framework that can actually make it compile and run.
Lead by Example
Lets pause for a short aside: obviously the examples presented here are going to be architectured beyond any rational sense, but the interesting part lies in seeing the bigger picture; imagine the principles presented here applied on a much larger scale and I'm sure you can see the benefits quite clearly when we reach the end.
Imagine that we have a class Division, which knows how to divide one number by another:
public class Division
{
public Int64 Dividend { get; set; }
private long _divisor = 1;
public Int64 Divisor
{
get { return _divisor; }
set
{
if(value == 0)
{
throw new ArgumentException("Cannot set the divisor to 0; division by 0 is not allowed.");
}
_divisor = value;
}
}
public Int64 Calculate()
{
Trace.WriteLine("Calculating the division of " + this.Dividend + " by " + this.Divisor);
Int64 result = this.Dividend/this.Divisor;
Trace.WriteLine("Returning result: " + result);
return result;
}
}
Consider the code presented above. Do you like it? If you've followed the discussion on AOP in the previous posts, then you should immediately be able to identify that there are several aspects tangled together in the above class. We've got data storage (the Dividend and Divisor properties), data validation (the argument check on the Divisor setter), business logic (the actual calculation in the Calculate method) and diagnostics (the Trace calls), all intertwined. To what extent is this class reusable if I wanted to implement addition, subtraction or multiplication calculations? Not very, at least not unless we refactored it. We could make the Calculate method and the properties virtual, and thus use inheritance to modify the logic of the calculation - and since this is a tiny example, it would probably look OK. But again, think bigger - how would this apply to a huge API? It would easily become quite difficult to manage as things got more and more complex.
Design by Composition
With a COP framework, we can implement each aspect as a separate object and then treat them as
mixins which blend together into a meaningful
composite. Sounds confusing? Lets refactor the above example using an as of yet imaginary COP framework for .NET (which I’m currently developing and will post the source code for in a follow-up post), and it'll all make sense (hopefully!).
Above, we identified the four different aspects in the Division class - so let's implement each of them. First, we have the data storage:
public interface ICalculationDataAspect // aspect contract
{
long Number1 { get; set; }
long Number2 { get; set; }
}
public class CalculationDataAspect : ICalculationDataAspect // aspect implementation
{
public long Number1 { get; set; }
public long Number2 { get; set; }
}
In this example, the data storage is super easy – we just provide a set of properties (using the C# 3.0 automatic properties notation) that can hold the values in-memory. The second aspect we found, was the business logic – the actual calculation:
public interface ICalculationLogicAspect
{
long Calculate();
}
public class DivisionLogicAspect : ICalculationLogicAspect
{
[AspectRef] ICalculationDataAspect _data;
public long Calculate()
{
return _data.Number1 / _data.Number2;
}
}
Here we follow the same structure again, by defining the aspect as an interface and providing an implementation of it. In order to perform the calculation however, we need access to the data storage aspect so that we can read out the numbers we should perform the calculation on. Using attributes, we can tell the COP framework that we require this reference, and it will provide it for us at runtime using some dependency injection trickery behind the scenes. It is important to notice that we’ve now placed a
constraint on any possible composition of these aspects – the DivisionLogicAspect now requires an ICalculationDataAspect to be present in any composition it is part of (our COP framework will be able to validate such constraints, and tell us up front should we break any). It is still loosely coupled however, because we only hold a constraint on the
contract of that aspect, not any specific implementation of it. We'll see the benefit of that distinction later.
The third aspect we have, is validation. We want to ensure that the divisor is never set to 0, because trying to divide by zero is not a pleasant experience. Validation is a type of advice, which was introduced at length earlier in my AOP series. We've seen it implemented using the IAdvice interface of my AOP framework, allowing us to dynamically hook up to a method invocation. However, the advice we’re implementing here is specific to the data aspect, so with our COP framework we can define it as
concern for that particular aspect, which gives us a much nicer implementation than an AOP framework could - in particular because of its type safety. Just look at this:
public abstract class DivisionValidationConcern : ICalculationDataAspect
{
[ConcernFor] protected ICalculationDataAspect _proceed;
public abstract long Number1 { get; set; }
public long Number2
{
get { return _proceed.Number2; }
set
{
if (value == 0)
{
throw new ArgumentException("Cannot set the Divisor to 0 - division by zero not allowed.");
}
_proceed.Number2 = value; // here, we tell the framework to proceed with the call to the *real* Number2 property
}
}
}
I just love that, it's so friggin' elegant ;). Remember that an advice is allowed to control the actual method invocation by telling the target when to proceed – we’re doing the exact same thing above, only instead of dealing with a generic method invocation we're actually using the interface of the aspect we're advising to control the specific invocation directly. In our validation, we validate the value passed into the Divisor setter, and if we find it valid then we tell the target (represented by a field annotated with an attribute which tells the COP framework to inject the reference into it for us, much like we did with aspects earlier) to proceed with the invocation; otherwise we throw an exception. This particular concern is abstract, because we only wanted to advise a subset of the methods in the interface. That's merely a convenience offered us by the framework - under the covers it will automatically complete our implementation of the members we left abstract.
Only one aspect remains now, and that is the logging:
public class LoggingAdvice :
IAdvice {
public object Execute(AdviceTarget target)
{
Trace.WriteLine("Invoking method " + target.TargetInfo.Name + " on " + target.TargetInfo.DeclaringType.FullName);
object retValue;
try
{
retValue = target.Proceed();
}
catch(Exception ex)
{
Trace.WriteLine("Method threw exception: " + ex.Message);
throw;
}
Trace.WriteLine("Method returned " + retValue);
return retValue;
}
}
We’ve implement it as a regular advice, like we've seen earlier in AOP, because it lends itself to much wider reuse than the validation concern did.
Having defined all our aspects separately, it is now time to put them back together again into something that can actually
do something. We call this the
composite, and it is defined as follows:
[Mixin(typeof(ICalculationDataAspect), typeof(CalculationDataAspect))]
[Mixin(typeof(ICalculationLogicAspect), typeof(DivisionLogicAspect))]
[Concern(typeof(DivisionValidationConcern))]
[Concern(typeof(LoggingAdvice))]
public interface IDivision : ICalculationDataAspect, ICalculationLogicAspect
{ }
Basically, we’ve just defined the implementation of an interface IDivision as a composition of the data and logic aspects, and sprinkled it with the two concerns (the validation concern and the logging advice). We can now use it to perform divisions:
IDivision division = Composer.Compose<IDivision>().Instantiate();
division.Number1 = 10;
division.Number2 = 2;
Int64 sum = division.Calculate();
That’s pretty cool, no? Take a moment to just think about what doors this opens. To what extent do you think our code is reusable
now, if we wanted to implement addition, subtraction and so forth? That’s right – all we’d need to do is substitute the implementation of the calculation aspect with one that performs the required calculation instead of division, and we're done. Let’s do subtraction, for example:
public class SubtractionLogicAspect : ICalculationLogicAspect
{
[AspectRef] ICalculationDataAspect _data;
public long Calculate()
{
return _data.Number1 - _data.Number2;
}
}
That’s it! The rest we can reuse as is, building a new composite:
[Mixin(typeof(ICalculationDataAspect), typeof(CalculationDataAspect))]
[Mixin(typeof(ICalculationLogicAspect), typeof(SubtractionLogicAspect))]
[Pointcut(typeof(LoggingAdvice))]
public interface ISubtraction : ICalculationDataAspect, ICalculationLogicAspect
{ }
Notice that we just left out the validation concern in this composite, as it is no longer needed. What if we wanted our subtraction to only ever return positive numbers? Easy! We’ll just implement an absolute number concern:
public class AbsoluteNumberConcern : ICalculationLogicAspect
{
[ConcernFor] protected ICalculationLogicAspect _proceed;
public long Calculate()
{
long result = _proceed.Calculate();
return Math.Abs(result);
}
}
And then update the composition to include it:
[Mixin(typeof(ICalculationDataAspect), typeof(CalculationDataAspect))]
[Mixin(typeof(ICalculationLogicAspect), typeof(SubtractionLogicAspect))]
[Concern(typeof(AbsoluteNumberConcern))]
[Pointcut(typeof(LoggingAdvice))]
public interface ISubtraction : ICalculationDataAspect, ICalculationLogicAspect
{ }
To Be Continued…
I hope this post has whet your appetite for more on this subject, as I will certainly pursue it further in future posts. I’ve already implemented a prototype framework that supports the above examples, which builds on my
previously posted AOP framework, and I’ll post the source code for that soon. If you want to dig deeper right now (and don’t mind a bit of Java), then I suggest you head over to the
Qi4j website and poke about there.
Richard Öbergs blog also provides great insight.
27. February 2008 18:18