Wednesday, November 08, 2006

Design for Testing

Good testing influences your design. Too many developers (and managers) are stuck in the waterfall mentality of design, code, test, and deploy. Forces across the industry are pushing to move the testing phase to the beginning of this line, and many just don't understand how that works. Testing should be the first consideration, and thus is influential to all other aspects of the development process.

The problem with testing as a secondary consideration is the design and architecture of the software never lends itself to proper testing when you don't plan way ahead. The consideration of testing can drive your design to be easier to test, but also can encourage generally good programming practices and well-made designs in the architecture. We can use a persisted class as a good example, because this is a use-case where testing is very important, but we have to consider the burden of a full database tied into the class we are testing.

When we develop our database item class before any testing is considered, we create it fairly in a straight-forward manner. After everything is coded, we decide to do some testing but we have a couple problems to face. The first thing to concern us is that our class inherits some ItemSchema super-class, and instances of it must exist in context to some database, which creates a large dependency on the test and thus leads to the test being unreliable. Secondly, we have many functions not easily testable (perhaps they can only be confirmed by locals within the function, which we can not access). We need to redesign everything to solve these issues, but we could have avoided this by using a more testable design in the first place.

To solve the first problem, we have to consider what we are actually testing. We are not testing the persistence framework our application utilizes, but just one ItemSchema sub-class we had to write. Obviously, separation is key. We only care about the functionality we wrote into the type, and we can extract all this into a mixin class, which our original ItemSchema can inherit. However, in our tests, a special TestItem class may also inherit it and perform the testing we need, without bringing a database into the picture.

Tackling the second problem of individual functions, the solutions can vary. If any internal data is important, perhaps it is too internal, and the code generating it could be extracted into its own method, which we can test independently. If our method does not return anything, perhaps this is something it could return and thus we could test for it. However, don't return things you do not need in production, as this bloats the interface and inevitably some code will come to depend on this contract that was only intended as a testing mechanism. It can even be acceptable to wrap a returning method with a public API non-returning version, simply to push a more stable API (it is easier to add returns than remove them) to the consumers of the API.

We have to stop looking at the testing as a second, or third class citizen in the steps we take in development. Our design, development time-lines, and architecture should all be done with testing and quality assurance first in mind. We must focus on how to ensure the stability and accuracy of our code before we can ever trust it, and if we can not trust it, all time developing it is a waste. It is this fact that offsets all arguments by opponents of proper testing (yes, they exist) who are afraid of the time wasted by testing. Wasted time is illusionary, as we only see the time it takes but we don't see the time it save us. However, isn't that the point?

No comments:

I write here about programming, how to program better, things I think are neat and are related to programming. I might write other things at my personal website.

I am happily employed by the excellent Caktus Group, located in beautiful and friendly Carrboro, NC, where I work with Python, Django, and Javascript.

Blog Archive