When you are changing an interface or API that is widely used, it can be tempting to do the change in one go and break a lot of tests at once, assuming you can just fix them. You risk ending up chasing failing tests around for hours, with lack of good feedback why they fail. Furthermore, you risk breaking something along the way, because you have created holes in your safety net!
When you make sure that there is only one failing test at any point in time, there is no doubt about what to work on next. If you have many tests failing at the same time, you have to decide where to start, which takes extra mental effort (decision fatigue). Worse, when a substantial part of your tests are failing, you don’t know what state you are in at all, and you cannot work on the code base with any degree of confidence.
Tests fail for a reason
If your next step breaks many tests at the same time, you probably have one of these issues:
- your design has too many dependencies on one part
- your tests have too much duplication
- your tests share setup code where they should not
The art of Test Driven Development is being able to take a small next step in any circumstance.
Taking smaller steps
The solution for having multiple failing tests is to take a smaller step. If more than one test fails, we look at the failures to understand what is happening, revert our change, and think of a smaller step to take.
So how do you reduce step size?
Refactor first before adding the next test. Change the code while keeping the behaviour, so that all tests stay green. Move the code in the direction you need, so that the intended change just fits in. Some useful techniques are:
- Test cases follow common setup - we group tests that have a common setup, which will reduce test setup duplication. We will post more about this heuristic later on.
- Create example objects, test fixtures, or test data builders which encapsulate specific dependencies and allow reuse across tests.
- Extract a domain specific language (DSL) that encapsulates steps in your tests, to isolate the effect of changes.
Let’s look an example from the Agile Fluency® Diagnostic application that we wrote about before. The backend component, written in Python, has a Facilitator class with a function to extend the facilitator’s license by a year. This function looks like this:
1 2 3 4 5 6 7 8 9 class Facilitator: ... def extend_license(self): current_expiry_date = self.license_valid_until try: new_license_valid_until = current_expiry_date.replace(year=current_expiry_date.year + 1) except ValueError: new_license_valid_until = current_expiry_date + timedelta(days=366) return replace(self, license_valid_until=new_license_valid_until)
Some of the tests are:
1 2 3 4 5 6 7 8 9 class TestExtendLicense: def test_extends_the_license_for_a_year(self): facilitator = aValidFacilitator(license_valid_until=date(2020,1,29)) assert_that(facilitator.extend_license().license_valid_until, equal_to(date(2021,1,29))) def test_handles_29th_of_feb_and_makes_if_march_first(self): facilitator = aValidFacilitator(license_valid_until=date(2020,2,29)) assert_that(facilitator.extend_license().license_valid_until, equal_to(date(2021,3,1))) ...
The tests are using test data builders to set up a facilitator. A facilitator
license_valid_until date property that we check.
We would like to change the signature of the
extend_license function to have a
days parameter, representing the number of days to extend the license. This
could also be months or an absolute date, but this doens’t matter for now.
Just making the change would break all existing tests that use
and possibly other tests that indirectly touch this function. What other options
do we have here?
- Introduce the parameter using a default value. The tests should remain green. You can then start adding parameter values to the tests, one by one, and remove the default value when all invocations have been changed. In some programming languages like Java, you would introduce an overload for the function instead of using a default value.
- Create a helper function in the test that wraps the call. Use this helper function everywhere, and make sure the tests remain green. You have isolated the dependency on the function in this helper function, so it is relatively straightforward to make the change. Strictly speaking you would still have multiple failing tests if you make the change, as all tests are affected through the helper function. They fail, however, because of that single invocation and you are much more in control.
- Perhaps the tests contain too many duplicated calls to the function anyway.
Does the invocation belong to the ‘Given’
part of the test? Should we move it
to the setup code? In the tests shown above, this is not the case. The
extend_licenseare the behaviour under test - they are part of the ‘When’, not the ‘Given’.
The main effect of applying this heuristic is being more in control. When there is just one test failing, you know the state of the code and you know how close you are to working software. You are always either in a state to commit & share your changes or one step away from commitable code.
A single failure provides fast, specific, actionable feedback. You know what to do next. Multiple failures produce noise, leaving you in a state of confusion.
Keeping test failures low keeps your code continuously in a working, and ultimately deployable state. This facilitates continuous integration & continuous delivery (CI/CD).
- Mike Hill wrote a series of posts about the value of taking baby steps: Many More Much Smaller Steps
This is a post in our series on Test Driven Development.