Normally the software projects I worked in never consisted of a proper unit testing phase. What we did all these years was not UT but IT ( integration testing ). Some people still oppose when I say you can’t unit test a code without completely isolating the method under test. This is probably due to their lack of knowledge or something related to the rocket science that I still don’t know. In either case, I love to direct them to this one book called “The art of Unit Testing by Roy Oshervoe”. In one project, I got an opportunity to study about unit testing and submit a report on whether it is feasible under the current workflow.The knowledge gained from that particular assignment, I consider precious.
What I’m going to talk about is not about unit testing in general. I suggest you to take a look at that book I mentioned to get a good picture of current unit testing methods etc; Ignore if you already know. During my assignment, while trying out different techniques like dependency injection in my project scenario, I come across a difficult situation. I was testing a C++/CLI wrapper for a module in C++. Its basically an adapter that marshals the native C++ types in to corresponding C# types and vice versa. Wrappers normally have little or less logic. But, there are some ‘if’s and ‘while’s involved even though that counts as the conversion logic only. To ensure nearly a complete code coverage, I need to test those minute details too. Dependency to the module that is being wrapped is obvious. I need to get rid of it in order to test it.
A normal dependency breaking scenario consists of two classes that are coupled together that we use inversion of control(IoC) to loosen the coupling. IoC is a process by which the assembler binds the object coupling at run-time. Normally we achieve this through Dependency Injection, where the interface or base class of the dependent class is injected via the constructor or setter functions.
We can refactor the code so that the class under test uses an interface that is added to the top of the dependent class in the hierarchy. During testing, the interface can be replaced with a stub class in order to isolate the function of the class to be tested from the dependency. But, what if such a re-factoring is not possible from the dependent side? What if we are using a class from a library that cannot be modified? In the case of our CLI wrapper the dependency was from a c++ library which should not be modified. So introducing an interface or a base class above the hierarchy is not easy.
After some (re)search I found out a solution that is appropriate while considering the clarity and definitely not, the complexity. It suggested the introduction of yet another wrapper to the low level dependency that delegates the functions that the high level code uses. So in our case if the class Engine is a low level class where we cannot introduce an interface above it, we can add the class
EngineWrapper that wraps the functions of
Engine and we can add an interface
IEngineWrapper above it. So, the
EngineCLIWrapper must keep only the interface and accept the instance via a constructor or a setter function. The unit testing class
EngineCLIWrapperTest breaks the dependency by deriving a stub class from the wrapper interface and “injecting” it to the instance of class under test.
The previous solution looks pretty good, but if I am to implement this, I will certainly feel a guilt of over-decoupling. We created an
IEngineWrapper interface only to be implemented by a wrapper that points to the actual dependency. Was all these necessary? The answer would be YES if the dependency is towards non-virtual methods. As the only objective was to test the CLI wrapper, I had this question in my mind about the inclusion of an adapter to an adapter for sake of unit testing. Luckily the dependency was towards some virtual functions of the low-level class. After spending some thought a I could think of a simple solution to this problem to be specific. The Idea is to subclass the original “low-level” dependent class, stub or mock it by overriding the dependent virtual function so that it poses no menace while unit testing. Now, even though the c++ class
Engine does not have any virtual functions, we can make the wrapper functions of EngineWrapper virtual and derive a stub directly from it instead of using
IEngineWrapper as shown below.
I’m not sure whether this is a correct approach or not, but I believe I spent some seat time in sorting this out. Do chip in your valuable suggestions/opinions. In other news, when I started this post, I was thinking that I invented a new dependency breaking method but ended up wall bashing when I realized that I forgot about the ‘virtual’ keyword. Silly me!