Tight coupling between modules is a bad idea, and the worst form of coupling is cyclic dependencies between modules. Fortunately, there are a few techniques we can use to break the cycles. They are Callback, Escalation, and Demotion, and I’m going to walk through some examples that show each of them in action.
Then, once the dependencies are broken, we’ll look at two more techniques that allow us to invert and eliminate the relationship altogether. The code for each of the samples can be found in the edcie project on my Google code repository. Each example includes a build script and a simple test case. To execute them though, you’ll need GraphViz if you want to use JarAnalyzer. To invoke the build scripts without invoking JarAnalyzer, you can simply type:
Keep in mind that each variation of the system has the exact same behavior!
The example we’re going to use to drive the remainder of our discussion is incredibly simple. We’ve got a Customer and a Bill class that we’re going to bundle into two separate modules - cust.jar and bill.jar. There’s also a test case called PaymentTest that serves as a sample client to drive the interactions between the two classes. The test case is bundled into the billtest.jar module. The initial class diagram is seen at right. Note the bi-directional relationship between the two classes.
As we progress, we’ll add more classes and abstractions to the system to help increase the modularity. Additionally, we’re going to use JarAnalyzer to illustrate the relationships between the modules and also help us assess the quality of our design. The module structure is below, as generated by JarAnalyzer. You can see how to use JarAnalyzer in the build by reviewing the build file. Again, our goal is to break the cyclic dependency between cust.jar and bill.jar, and we’re going to look at three different ways to do this before moving on to examine different ways to massage acyclic module relationships.
Initial Module Structure with Cyclic Dependencies
The first technique we’re going to apply is called Escalation. With Escalation, we break the cyclic dependencies by escalating the cause of the dependency to a higher level entity. Before we do that, we need to more fully understand why a cyclic dependency exists in this example. This reason follows:
A Customer has a list of Bill instances. When the pay method on Bill is invoked, the Bill needs to determine if a discount should be applied. The discount is a product of the Customer the bill belongs to, not necessarily the Bill. Therefore, the Bill class calls a method on Customer to determine the appropriate discount amount. Think of it this way…The Customer represents a payee and we negotiate a discount with each payee. The calculation of this discounted amount is encapsulated within the Customer.
To break this dependency, we want to escalate the cause of the dependency up to a higher level class - the CustomerMediator. The mediator now encapsulates calculation of the discount and passes that to the bill class. The best way to see this change is to look at the modified PaymentTest class. Now, I’ve modified the build script and have bundled the mediator into it’s own module, as shown below. If you dig a bit more deeply into the class structure, you’ll wonder why I didn’t just pass the discount amount from Customer into Bill. Don’t worry about that. This example is slightly contrived because escalation isn’t the best way to solve this type of problem. The key takeaway is that we’ve escalated the dependency up to the mediator.jar bundle, breaking the cyclic dependency.
Escalating the Cause of the Cyclic Dependency
A slightly better way to solve this particular type of cyclic dependency (where we have a true composite relationship between Customer and Bill) is to use demotion. With demotion, we push the cause of the dependency to a lower level module. Exactly the opposite of escalation. We do this by introducing a DiscountCalculator class that will be passed into the Bill class. Our modified PaymentTest class will create the calculator and pass it in for us. The Customer class will serve as the factory for the DiscountCalculator, since it’s the Customer that knows the discount that must be applied. The new class structure can be seen at right.
Now we’ll modify our build script to create an additional calc.jar bundle which will contain the DiscountCalculator class. Our resulting module structure is shown below.
Demoting the Cause of the Cyclic Dependency
Already you can see how this is a more natural solution than escalation for this particular type of cyclic dependency problem. What’s the key difference you might ask? With escalation, notice how I would be able to deploy the cust.jar and bill.jar modules independently. While demotion is a more natural solution in this situation, it also means that to deploy bill.jar or cust.jar, I must also deploy calc.jar. The right solution is always going to be contextual and the ideal solution is likely to shift throughout the development lifecycle.
Using a Callback is similar to the Observer pattern. With this approach, we’ll refactor our DiscountCalculator class to an interface, and then modify the Customer class to implement this interface. This new class structure can be seen at right.
As it happens in this specific situation, using a Callback represents a combination of demotion and our initial solution. We’ll go back to passing the Customer into the Bill, but will pass it in as a DiscountCalculator type. Whereas in the Demotion example we bundled the DiscountCalculator in a separate module, we’ll now just include it in our bill.jar module. Note that putting the DiscountCalculator in the cust.jar module would introduce the cyclic dependency we’re trying to get rid of. The new module structure, which resembles our original version minus the cyclic dependency, is shown below.
Using a Callback to Eliminate the Cyclic Dependency
Now we’re going to play around a bit with the module relationships. While Callback seems like the most logical solution, what if we wanted to use the cust.jar module without the bill.jar module? Callback, as it’s implemented, doesn’t allow us to do this. But with a bit of trickery, I can actually invert the relationship between the cust.jar and bill.jar modules.
I start by refactoring the Bill class to an interface. Then, to avoid split packages (where classes in the same package are bundled into separate modules), I move the Bill class into the same package as the Customer class. The new class diagram is shown at right, and the inverted module structure is shown below.
Inverted Module Structure
Inverting the relationships allows us to deploy the cust.jar module independent of the bill.jar module. Again, it’s all about need. But I’d like to explore another option based on another important need - the ability to test modules independently. Before inverting the relationships, I am able to test the bill.jar module independently. After inverting the relationships, I can test the cust.jar module independently. But what if I want to test (or deploy) both modules independently? To do this, I need to completely eliminate the relationship altogether.
As it turns out, because I’ve got a pretty flexible class structure after I inverted the relationships (lot’s of abstract coupling), I can do this by simply bundling the two interfaces - Bill and DiscountCalculator - into a separate module. No other coding changes required. I start by moving them to a new package called base. Then, I modify my build script to bundle these two interfaces into a separate base.jar module, and we have successfully eliminated the relationship between the bill.jar and cust.jar modules, as shown below.
Eliminating Relationships Between Modules
We’ve come a long ways from our original version of the system. Two modules with a cyclic relationship to a module structure where the original modules don’t have any relationship to each other. This means these modules can be tested and deployed independently. If you follow this blog, you know I’ve written plenty about the tradeoffs between flexibility and complexity, use and reuse, and many other architectural and design challenges. I hope this little exercise has helped drive some of those concepts home.
As a final note, to explore some of these design decisions a bit more deeply, and to examine the tradeoffs a bit more objectively, I encourage you to run the builds for each of the projects and examine the dependencies.html file in the stats directory. To do this, you’ll need to make sure JarAnalyzer is executed, which requires GraphViz. As you’ll see when comparing the initial version to the final version, we made considerable progress in improving the quality of the design.