Action Log Test Double
So how do you test-drive an algorithm like this? At first it might seem simple. For each test, create an small array of doubles, predict the results, and then write the test that compares the predicted results with the output of the algorithm.
There are several problems with this approach. The first is data-overload. Even the smallest array is a 3X3 with 9 different values to check. The next smallest is 5X5 with 25 values. Then 9X9 with 81 values. Trying to write a comprehensive set of tests, even for the 3X3 case, would be tedious at best; and very difficult for someone else to understand.
The second problem is that test-driving from the raw results forces us to write the whole algorithm very early in the testing process. There’s no way to proceed incrementally.
[…]
This is a rather different slant an a spy test double. Rather than giving us booleans and flags to inspect about the calls of individual functions, this spy is going to load the actions string with the sequence of things that happened as the algorithm proceeded.
