That's weird...

Rules for writing tests

· Apostolis Bessas

When writing tests for a (Python) codebase:

  • We use pytest fixtures. They are way more flexible than an inheritance-based system and allow for fine-grained dependencies.
  • Every test case should test for a single behavior. This usually means having a single assert statement in a test case. This makes it easier to know why a test failed and what behavior is tested.
  • The structure of the test suite on the filesystem should mirror that of the codebase, so as to be quick to find the tests for a piece of code.
  • Any package should be free to define its own fixtures in a local conftest.py. Not all fixtures make sense to go to the top-level conftest.py file.
  • A test can override fixtures defined up in the hierarchy, but it should prefer to use a new name that makes sense locally. Otherwise, there could be confusion about which fixture is used.
  • A test case should test one specific thing and test it right at the edge of where it is defined and not at higher levels. For instance, if you want to test the behavior of a function, do not do it via the API that happens to use that function deep in the call hierarchy. This way you are guaranteed that the test does not fail for any other reason than what you are testing for.
  • A test case must avoid any network calls. Otherwise, tests are slow and non-deterministic (e.g. the remote service can go down at any time).
  • A test case should avoid any database calls, if possible. Any I/O is slow.
  • A test case should avoid mocking, when possible. Prefer to use real fixtures or refactor the code to not have as many dependencies. Mocking is fragile and the whole concept relies on people remembering to update the mocks as they refactor the code.
  • Related to mocking: it is mostly useful when you want to avoid side-effects (e.g. kicking off a task). It does not always make sense to do dependency injection just for the sake of testing. In those cases, mocking is your preferable option.
  • There should be switches that are turned off when running the test suite and completely kill unnecessary side-effects and 3rd party integrations entirely (e.g. any integration with a 3rd-party service). This way it is easier to ensure that you avoid any extra I/O and minimize the dependencies for your test.
  • Tests should never cross systems: background tasks should not be allowed to be called, application-level events should not be handled. For instance, (in the case of tasks) write 2 test cases: one that verifies that a task would be called with the proper arguments and one that tests how the task works with those arguments. Otherwise, your tests also test the message bus.
  • Define the input of a test case explicitly. This way you minimize the dependencies of the test case and you are certain what the test does.
  • Functional/end-to-end tests are useful, but they should be a very small part of the test suite and indicated as such. These tests are slow, freagile and thus difficult-to-maintain. They can be useful, e.g. to actually test that making an API call results in a webhook being sent, but (i) that behavior should have been covered by multiple tests already and (b) it makes more sense to have that test against a staging environment and not in the test suite.
  • If a test case takes more than a few hundred milliseconds to run, it is not a unit-test. This is a good indication that a test case does too many things.
  • The default case for a test would be to use pytest.mark.parametrize: a test should have certain inputs that lead to certain outputs. Many functions should be written in a way that take a certain input and that causes a deterministic output. Using parametrize helps you write functions in such a way and it is a very quick way to test for more inputs and outputs.
  • There should be tests not just for the happy path. Code coverage is an indicator for how good a test suite is but a meaningless metric. Code should be tested as much as possible. The goal should be to cover all possible combinations of input/output instead of whether every line has been hit by a single test.

Bonus:

  • The longer a function is, the more combinations there are to test. Write small functions that are as pure as possible.