awesome-python/website/data/category_intros/testing.md

Use pytest as your Python testing framework. Add Hypothesis to find the edge cases you missed, and Playwright to test in a real browser.

How to choose:

• Unit and integration tests: pytest
• Property-based tests: Hypothesis
• Running the suite across Python versions or dependency sets: tox, or nox if you'd rather configure it in Python
• End-to-end browser tests: Playwright
• An existing Selenium suite, or tests spread across many machines with Selenium Grid: Selenium, or SeleniumBase for a pytest-ready framework on top of it
• Acceptance tests in a readable keyword syntax: Robot Framework
• Load testing: Locust
• An API with an OpenAPI or GraphQL schema: Schemathesis
• Faking HTTP: responses for Requests, RESPX for HTTPX, VCR.py to record and replay real traffic
• Freezing the clock: freezegun
• Test data: factory_boy for ORM models, Polyfactory for dataclasses and Pydantic models, Faker or Mimesis for single fake values
• Coverage: coverage.py

With pytest, write tests as plain functions with assert, and share setup through fixtures, which pytest says offer dramatic improvements over xUnit-style setup and teardown. For a new project, its good practices recommend the importlib import mode and a src layout. You don't have to rewrite an old unittest suite first: pytest runs it as is, so you can move it over one file at a time.

A Hypothesis test is a pytest test with @given on top. You describe the inputs with strategies, and Hypothesis picks the values, including edge cases you might not have thought about. It still works with fixtures and parametrize.

For the browser, Playwright calls its pytest plugin, pytest-playwright, the recommended way to write end-to-end tests. Find elements by what the user sees, starting with get_by_role(), and assert with expect(), which keeps retrying until the condition is met or it times out.

Run tox or nox locally and in CI. pytest's own docs point to tox because it tests the installed package, not your checkout, which catches packaging mistakes.

For coverage, run coverage run --branch -m pytest. You don't need a pytest plugin for it: coverage.py calls one unnecessary for most purposes.

When you mock with unittest.mock, patch where an object is looked up, not where it's defined. Pass autospec=True too, so a call with the wrong signature raises a TypeError.