You need tests...just not yet post

January 7th, 2013

Every once in a while I come across a blog post or a presentation that helps me figure out something that has been bothering but I was having problems articulating. Lately I've been struggling to come up with strategies to help developers figure out when they should REALLY be writing tests for their code.

If you believe in the power of TDD, then the answer is "you write tests from the beginning of the project and forever after." I am a believer that TDD can be a very powerful design tool for building things at the unit / module / component level.

Then some people asked me my thoughts on the comments by the creator of the Rails project about the usefulness of dependency injection. As happens when I am feeling frustrated, I started firing out the tweets on my feelings on testing and the lonely Crusade that I am on.

Then I saw a talk by Dan North called "Decisions, Decisions" at the recommendation of Ian Barber that made me realize that someone else has articulated my feelings on when you should be doing testing. Slides for the presentation can be found here.

You should watch the whole presentation. So many decisions about programming are made without understanding, REALLY understanding the trade-offs involved when making one decision vs. another.

Let's look at TDD vs. Test whenever. The trade-off being made here is not about quality of code or guarding against regressions. It's about opportunity cost. This had occurred to me but I had dismissed it as being "anti-testing".

But I think I was wrong, and here's why.

In that awesome presentation he talks about "patterns" he observed on the team he was working with at the time. He was amazed at how much work this team was getting done, so he spent time analyzing what they were doing and trying to turn some of their activities into patterns of behaviour. Repeatable processes are good, right?

The one that he saw being related to writing was a pattern he called "Spike and Stabilize." Basically, you write code without any tests until it becomes a solid part of your application's architecture. Once this code has proven it's usefulness, you start writing comprehensive tests in order to verify that it is in fact doing the job and doesn't break going forward.

This is so obvious to me now. Clearly, you are leaving yourself open to the potential of writing not just application code but testing code that gets tossed away if the idea you are implementing is unable to prove it's value.

Opportunity cost is huge and often dismissed by developers too. If you have some prototype code, or even functionality with a half-life measured in 3-month-increments, what else could be worked on while tests for this code are being written?

Could this hybrid approach be an easier sell to skeptical managers or other stakeholders? That's hard to say. The "too busy to write tests" camp still has many members, and is still seems that it takes catastrophic failures before many people join the Testing Crusade.

To be perfectly clear: I find great value in prototyping code and then committing to tests once the prototype is ready to move to a more stable environment.

I try and do prototyping before I commit to writing code, because I often feel that way. My prototypes are usually command-line scripts that try and accomplish the task I've been assigned. That way, it's easier to actually write some tests once I know what I'm doing and guard against regressions later.

The key to all this is being able to identify at what stage in this particular pattern your code is at. Is it still a spike, meaning you are working out implementation details and trying to figure out if it will even have the desired result? Or is it stable, providing solid value to the application as a whole and ready to be wrapped in tests to protect against regressions?

Next time you are thinking about "I need some tests for this", consider the Spike and Stabilize pattern. The answer might be "you need tests...but not yet."