The easiest way to reduce waste is to reduce the amount of work you may have to throw away. This means breaking your work down into its smallest possible units and verifying them separately.
Sometimes while debugging, I see multiple problems and their solutions at once. Shotgun debugging is tempting, but if I try several different solutions simultaneously and fix the bug, I may not know which solution actually worked. This also usually leaves a mess behind. Incremental change is a better approach. I make one well-reasoned change, observe and verify its effects, and decide whether to commit to the change or revert it. I learn more and come up with better—and cleaner—solutions.
This may sound like taking baby steps, and it is. Though I can work for 10 or 15 minutes on a feature and get it mostly right, the quality of my code improves immensely when I focus on a very small part and spend time perfecting that one tiny piece before continuing. These short, quick steps build on each other; I rarely have to revert any changes.
If the step doesn’t work, I’ve spent a minute or two learning something and can backtrack a few moments to position myself to make further progress. These frequent course corrections help me get where I really want to go. Baby steps reduce the scope of possible errors to only the most recent changes, which are small and fresh in my mind.
The desire to solve big, hairy problems is common in developers. Pair programming helps us encourage each other to take small steps to avoid unnecessary embellishments. Further, the navigator concentrates on the big picture so that both developers can maintain perspective of the system as a whole.
Test-driven development provides a natural rhythm with its think-test-design-code-refactor cycle. The successful end of every cycle produces a stable checkpoint at which the entire system works as designed; it’s a solid foothold from which to continue. If you go wrong, you can quickly revert to the prior known-good state.
At a higher level, stories limit the total amount of work required for any one pairing session. The maximum size of a step cannot exceed a few days. As well, continuous integration spreads working code throughout the whole team. The project makes continual, always-releasable progress at a reliable pace.
Finally, refactoring enables incremental design. The design of the system proceeds in small steps as needed. As developers add features, their understanding of the sufficient and necessary design will evolve; refactoring allows them to refine the system to meet its optimal current design.
Last summer, I introduced a friend to pair programming. She wanted to automate a family history project, and we agreed to write a parser for some sample genealogical data. The file format was complex, with some interesting rules and fields neither of us understood, but she knew which data we needed to process and which data we could safely ignore.
We started by writing a simple skeleton driven by tests. Could we load a file by name effectively? Would we get reasonable errors for exceptional conditions?
Then the fun began. I copied the first few records out of the sample file for test data and wrote a single test: could our parser identify the first record type? Then I pushed the keyboard to her and said, “Make it pass.”
“What good is being able to read one record, and just the type?” she wondered, but she added two lines of code and the test passed. I asked her to write the next test. She wrote one line to check if we could identify the person’s name from that record and pushed the keyboard back my way.
I wrote three lines of code. The test passed. Then I wrote a test to identify the next record type. Of course it failed. As I passed back the keyboard, we discussed ways to make it pass. I suggested hardcoding the second type in the parsing method. She looked doubtful but did it anyway, and the tests all passed.
“It’s time to refactor,” I said, and we generalized the method by reducing its code. With her next test, I had to parse another piece of data from both record types. This took one line.
We continued that way for two hours, adding more and more of the sample file to our test data as we passed more tests. Each time we encountered a new feature of the file format, we nibbled away at it with tiny tests. By the end of that time, we hadn’t finished, but we had a small amount of code and a comprehensive test suite that would serve her well for further development.