Test maintenance

Test maintenance

Test code is production code with a longer memory: it outlives features and gets read far more than it is written. Left untended it accretes copy-pasted setup, tests that assert nothing, and duplicates that all break together on one change. Maintaining it means the same moves you make in application code, plus one nerve most people lack: deleting a test that has stopped paying for its runtime.

Method

1. Extract setup into named builders, not shared mutable state. Replace repeated arrange blocks with a factory or builder (makeUser({suspended: true})) that returns fresh objects. Avoid a shared setUp that mutates instance fields across tests, which couples them and breeds order-dependent flakiness.
2. Dedupe the assertion, keep the intent visible. Pull a repeated multi-line check into a custom matcher or a helper (assertValidInvoice(inv)), but leave the input and the expected outcome in the test body. A test whose meaning now lives entirely in a helper is unreadable at the call site.
3. Delete tests that no longer earn their runtime. Remove tests that assert a deleted feature, duplicate another test's coverage exactly, or only restate the implementation. Confirm the deletion is safe with a coverage or mutation run: if killing the test does not drop either number, it was not pulling weight.
4. Fix or quarantine flaky tests on sight, never retry blindly. A test that fails intermittently is worse than no test: it trains the team to ignore red. Diagnose the nondeterminism (time, ordering, shared state), fix it, or move it to a quarantined lane with an owner and a deadline rather than wrapping it in a retry.
5. Refactor tests and production in separate commits. When you change both, commit the test refactor while behavior is green first, then the behavior change. Mixing them means a red suite cannot tell you whether the refactor or the feature broke.
6. Keep one behavior per test and one reason to fail. Split a test that guards several behaviors so a failure names one cause. Rename tests whose description drifted from what they now check, since a lying test name costs more than no name.

Signals

• Does changing one production behavior break exactly the tests about that behavior, and no unrelated others?
• Can you delete a test and immediately tell from coverage or mutation whether it mattered?
• Is there a flaky test currently failing that everyone has silently agreed to ignore?

Boundaries

This keeps an existing suite healthy; it does not decide the initial mix of tests, which is testing-strategy, or how to design a single test, which is unit-test-design. Deleting a test is a judgment call: when coverage cannot tell you whether it matters, keep it and ask a reviewer.