The measure of a test suite is not coverage percentage. It is whether a green run gives the team enough confidence to release on a Friday afternoon, and whether a red run makes anyone stop what they are doing.
Flakiness is the fatal defect
A suite that fails randomly trains everyone to re-run it until it passes. From that moment the suite is decorative: it will catch a genuine regression and be dismissed as noise. Fixing flakiness therefore outranks adding coverage, always.
The moment a team starts re-running a red build, the suite has stopped working.
The three causes of flakiness
Almost every intermittent failure traces to timing assumptions, shared state between tests, or non-deterministic data. Waiting for a condition rather than a duration, isolating data per test, and freezing time and randomness fix the overwhelming majority. Adding a retry hides the symptom and preserves the bug.
Test at the level that catches real bugs
Unit tests around pure logic are fast and precise. Integration tests catch the things unit tests mock away — and mocked-away boundaries are exactly where bugs live. A modest number of tests against a real database usually finds more genuine defects than an exhaustive suite of mocked unit tests.
End-to-end: few, critical, reliable
Browser tests are slow and the most prone to flake, so spend them on the handful of journeys that cost money when broken — signup, checkout, the core workflow. Ten reliable end-to-end tests beat two hundred that half the team ignores.
Speed is a feature of the suite
A suite that takes forty minutes will be skipped under deadline pressure. Parallelise aggressively, run a fast subset on every commit and the full set before merge. Under ten minutes for the pre-merge suite is a target worth engineering for, because it changes behaviour.
Let failures write your tests
Every production bug should leave behind a test that would have caught it. This grows the suite along the axis of what actually breaks in your system, which no coverage target can direct you towards. Over a year it becomes a record of the ways your product fails, and it is worth more than any percentage.
