How do you manage test data so tests stay reliable across runs and environments?

Make each test create the data it needs and clean up afterward, with unique identifiers so parallel runs cannot collide. Prefer building data through the API or factories over a shared seeded database, since shared fixtures rot and create hidden coupling between tests. Never depend on data someone loaded manually, and never rely on unmasked production data containing personal information.

The goal is that any test can run alone, in any order, in parallel with the others, and still pass. That rules out a shared seeded dataset, which always starts tidy and is a swamp within two months because nobody knows which test depends on which row. So each test builds what it needs, ideally through API calls or a factory helper rather than the UI, which is faster and far less brittle. Every record gets a unique identifier, usually a run ID plus a timestamp in the email or the name, so ten parallel workers cannot fight over the same account. Cleanup happens in teardown, but I design so leftover data is harmless rather than assuming cleanup always runs, because it will not when a test crashes. So there is also a scheduled job that clears anything older than a day matching the test naming pattern. On production data I push back hard on unmasked copies. If we need realistic volume and shape, I want it generated or properly anonymized, because a test environment full of real customer records is a breach waiting to happen and usually has weaker access controls than production.