Tests Should FailThu 02 August 2018 by Moshe Zadka
"eyes have they, but they see not" -- Psalms, 135:16
Eyes are expensive to maintain. They require protection from the elements, constant lubrication, behavioral adaptations to protect them and more. However, they give us a benefit. They allow us to see: to detect differences in the environment. Eyes register different signals when looking at an unripe fruit and when looking at a ripe fruit. This allows us to eat the ripe fruit, and wait for the unripe fruit to ripen: to behave differently, in a way that ultimately furthers our goals (eat yummy fruits).
If our eyes did not get different signals that influenced our behavior, they would not be cost effective. Evolution is a harsh mistress, and the eyes would be quickly gone if the signals from them were not valuable.
Writing tests is expensive. It takes time to write them, time to review them, time to modify them as code evolves. A test that never fails is like an eye that cannot see: it always sends the same signal, "eat that fruit!". In order to be valuable, a test must be able to fail, and that failure must modify our behavior.
The only way to be sure that a test can fail is to see it fail. Test-driven-development does it by writing tests that fail before modifying the code. But even when not using TDD, making sure that tests fail is important. Before checking in, break your code. Best of all is to break the code in a way that would be realistic for a maintenance programmer to do. Then run the tests. See them fail. Check it in to the branch, and watch CI fail. Make sure that this CI failure is clearly communicated: something big must be red, and merging should be impossible, or at least require using a clearly visible "override switch".
If there is no code modification that makes the test fail, of if such code modification is weird or unrealistic, it is not a good test. If a test failure does not halt the CI with a visible message, it is not a good CI. These are false gods, with eyes that do not see, and mouths that do not speak.
Real tests have failures.