Pytest mastery

Pytest mastery

Fixtures are dependency injection for tests: each test states what it needs, pytest builds it, and scope controls how much is shared.

Method

1. Scope fixtures by cost, not convenience. Default function scope for anything mutable; session scope only for expensive immutable resources (a container, a compiled artifact). A session-scoped mutable fixture is the classic source of order-dependent flakiness.
2. Layer conftest.py by directory. Repo-root conftest holds universal fixtures; each package's tests get their own conftest for local ones. Never import fixtures across test files; discovery through conftest is the mechanism.
3. Parametrize instead of copy-pasting. @pytest.mark.parametrize("raw,expected", CASES, ids=str) turns a table of cases into individual test items with readable names. Stack two parametrize decorators to get the cross product.
4. Use markers as a vocabulary, registered in pyproject. slow, integration, requires_gpu; select with -m "not slow" locally and run everything in CI. Unregistered markers are typos waiting to pass silently; set --strict-markers.
5. Prefer tmp_path, monkeypatch, and capsys over hand-rolled setup. tmp_path gives an isolated directory per test; monkeypatch.setenv and .setattr undo themselves; capsys asserts on output without redirecting streams manually.
6. Assert one behavior per test, plainly. Bare assert with pytest's rewriting gives rich diffs; pytest.raises(ValueError, match="port") pins the error and its message. If a test needs a comment to explain what it checks, its name is wrong.
7. Keep the suite fast enough to run on every save. Measure with --durations=10; push anything over a second behind a slow marker. pytest -x --ff (fail fast, failed first) is the debugging loop.

Boundaries

• Do not mock what you own; restructure so the seam is a real interface. Mock only true externals (network, clock, randomness).
• Fixture chains deeper than three levels hide the arrange step; inline the setup or build an explicit builder function.