One of my goals for the new year was to follow TDD on a real project at work. I actually got my chance very early this year with a fairly basic Silverlight project. The project was short and simple, basically a fancy list of links and resources managed in sharepoint and exposed in a silverlight interface allowing a variety of queries and searches. It was large enough to be more than just a toy project, but small enough that I didn’t worry about doing much damage by trying out TDD for the first time.
I learned alot, and I think the work I did makes a good case study for someone interested in getting started with TDD. In my next few blog posts. I plan to walk readers through my development environment, the specifics of the techniques I followed, and the lessons I learned.
As I said at the start, the project was written in Silverlight. For my testing I used the Silverlight Unit Test Framework, which allows for asynchronous testing, which is vitally important for any webservices based integration testing. On top of that I used a fantastic continuous test runner named Statlight. Statlight is a small console application that automatically runs your unit tests every time your test project’s .xap file changes. This means that running your tests is as easy as hitting Ctrl + Shift + B to build the project and Statlight does the rest. I quickly got in the habit of building after every code change so that I was getting instant feedback on what I was doing.
Since this was an experiment, I tried to stick as close to the rules of TDD as possible. This meant I never wrote a line of code until I had already written a test covering it, and that my tests were extremely granular. Even simple tasks like parsing a single line of XML returned from a webservice had a test devoted to it. I also tried not to overthink some of the details of my design, instead trying to put of design decisions until I had already written the test necessitating them.
Overall, my experience was hugely positive. I’m convinced that TDD definitely makes me more effective and productive and I want to leverage it wherever I can in the future. In general I found there were 3 major benefits to TDD, and I learned 3 lessons about how to do TDD better next time. Let’s start with the good
Flow – It was shocking how good it felt to be able to code without stopping. With TDD my brain stayed in code mode for hours at a time. Usually, I slip in and out of this mode out the day, especially when I’m manually testing code I’ve just written. With TDD, that never happened, and it made my concentration and focus 20x better. When I’m manually testing, there are all sorts of interruptions and opportunities for distraction. Waiting for the page I’m testing to load? I’ll just go browse google reader for a bit. Stepping through a tedious bit of code so I can examine the value of one variable? Let me just skim this email while I do that. With TDD though, my brain never gets an opportunity to slip away from the task at hand. Throughout the day I was laser focused on whatever I was doing.
Similarly, if I did have to step away for an interruption (meetings, lunch, help another dev, etc.) it was easy to get back into the flow and figure out where I was. Just hit Ctrl + Shift + B and see what test failed. Since each test was so small and covered such a small area, I didn’t have a ton of details about what I was doing slip away when I got distracted.
Design – I didn’t totally abandon upfront design, but I did do less design than I usually do. I mostly sketched out the layers at the boundaries of the application, the pieces that interacted with the user and the pieces that interacted with the data source, SharePoint, since both of those were external pieces that I couldn’t exercise complete control over. Once I had those layers designed though, I let TDD evolve the internal architecture of the application, which actually led to a couple of neat design decisions I don’t think I would have come up with otherwise. The coolest of these was how I handled loading up a given model for a given page. In our application the same view could be wired up to a variety of different models. The specific model depended on the url the user used. I ended up with two separate objects which handled this process, the Model Locator which parsed the incoming URL, and the Model Map, which tied each model to a path-like-string which represented how the data was maintained in the data store. The Model Locator would use the URL to extract the key elements to identify the right model, and then pass those into the Model Map, which would use those elements to find the right model by building the path representation for the model. The end result was a nice decoupling between the path structure the user used to browse to a model, and the way it was actually handled by the data layer. If I had been designing up front, I am almost positive I would have missed this approach, and put too much of the logic into the Model Locator itself, tightly coupling the data structure and the navigation structure. Instead, I put off making any decisions about how the Model Locator interacted with the data until the last minute, and by then it was clear that a new class would improve the design significantly.
Refactoring Ease of Mind – Not everything about this project was perfect. In fact, towards the middle there were some significant pain points because I had to be temporarily put on another higher priority project. To keep things moving another developer was assigned to the project. There wasn’t enough time invested in communication and as a result, he ended up taking a different approach in some key ares, and duplicating some work I’d already done. By the time I came back, his code was wired up to the UI, and it didn’t make sense to try and reincorporate the pieces of my code that were performing some of the same functions. Unfortunately, there were a number of pieces that handled things like search and model location that were still expecting the classes defined in my code. All of those had to be modified to work with his architecture instead.
This would have been a really scary refactoring to do in the timeline we had, except for the the automated tests I already had covering all of my code. With only a few minor tweaks, that test suite was modified to test my search services using his new classes, and we had extremely detailed information about where my code was now broken. After less than a day of work, we’d switched everything over without a hitch. And because of the tests, we had confidence that everything would work fine.
I won’t say much more in summary, because I think the benefits speak for themselves. Next post, I’ll talk about what I’d do differently next time, and how I plan to get better at TDD in the future.