test-data-strategies

Use fixtures for expensive immutable values shared read-only, factories as the default for domain objects needing fresh instances per test, and builders when objects have many fields or ordered construction steps. Prefer established libraries like factory_boy or fishery over hand-rolled factories.

Order-dependent tests usually come from mutable fixtures with broad scope. Default to function scope, escalate scope only for immutable values like frozen dataclasses or MappingProxyType, and give external state like database rows explicit teardown via transaction rollback.

Seed random and Faker centrally in an autouse fixture with an overridable TEST_SEED environment variable. Also eliminate hidden nondeterminism from uuid4(), datetime.now() in factory defaults, set iteration order, and random default values.

Pseudonymized production data is not anonymized; quasi-identifiers like ZIP code, birth date, and sex allow re-identification by linkage. Prefer synthetic data by default, and require irreversible transformation, production-grade access controls, and expiry for any real data.

Random values for unique columns collide far sooner than intuition suggests, producing intermittent uniqueness violations. Use a monotonic per-run sequence such as itertools.count to guarantee uniqueness deterministically.