Take Two

November 10, 2012

On the apprentice team's project, this week's iteration was dedicated to implementing a particular new feature. An implementation was discussed and approved between us. At the start of today, I began working on it. I was close to finishing, perhaps two tests away from completion, when it was realized that a separate effort to implement a persistence layer was revealing conflicts with some existing elements of our application.

We reconvened and discussed the matter. After thoroughly evaluating our needs and our options, we reached a firm conclusion on how to resolve it. A consequence of this meeting was that all of my work to that point was unusable. It was well-tested and well-constructed, but did not support the changes which needed to be made. Sadly, this happens to everyone at one time or another.

After this experience, which is certainly not the first and likely won't be the last, I feel the need to understand what went wrong and how to help prevent these circumstances from arising again.

What went wrong?

A previous design decision involved an assumption of compatibility with a not-yet-chosen database. Implementation details, however few, preceded the choice of database. Essentially, we were building a part of our application on an expectation that would not be met.

The next problem, which is beginning to reveal itself, is limited abstraction separating the core of our application from the library governing the persistence layer. Too much of our application is using direct calls to external library functions.

What could have prevented this?

Both issues stem from the same root cause: a lapse in discipline. Following acceptable programming practices should have clued us in to the mistakes we were making, at a time when correcting or prevent such mistakes would be trivial. Couplings and open implementations are substantial gaffes. The distinct isolation of components should be of paramount importance, for reasons even greater than what was revealed with this experience.

Implementation details should be decided only when they can be tested, and not a moment sooner. Describing application logic in terms not native to the application is a clue that dependencies are creeping in. When making these decisions, and implementing application logic, it should remain as high-level as possible.

Using a stand-in for a library before that library is available is a challenge. It can be done, provided a layer of abstraction simplifies the flow of information from one side of this layer to the other. In our case, this decision backfired. A non-trivial structural change is now required to both complete the current story and also to accommodate future design decisions. Going forward, observing better practices (especially the Dependency Inversion Principle) will help us avoid similar challenges.