Concurrency, Performance, and the Honest Test Suite

Testing async code

The rule: drive the event loop yourself. asyncio.run(coro()) in a test is fine; what's not fine is time.sleep — it freezes the loop and tests nothing. Use await asyncio.sleep(0) to yield control deterministically, fake clocks for timeouts, and assert on task states (task.done(), task.cancelled()) rather than wall-clock guesses.

Timing-sensitive tests

Never assert elapsed < X on a loaded machine — CI machines lie. The stable alternatives:

• Injected clocks (every module in this course): pass now in, advance it by hand.
• Call counting: assert "at most N attempts" instead of "finished within Y ms".
• Deterministic schedulers: for queue/worker logic, step the scheduler explicitly (the deliver() sweep from the distributed module).

Test doubles for I/O

A network dependency in a unit test is a flaky test with extra steps. Fake the port (the interface you already defined in the architecture module): an in-memory queue, a canned clock, a fake repository. Keep one thin integration suite that exercises the real adapters, and let it be the only place network flakiness can live.

What the suite cannot tell you

• That your performance holds under production data volumes — load test separately.
• That your security holds against a motivated adversary — audit separately.
• That users want the feature — that's not a test suite's job.

Honest scope: the suite verifies behavior you specified, deterministically. Everything else needs a different tool.

The maintenance contract

Tests are code: naming, duplication, dead tests deleted, helpers factored. A test suite is a system under active maintenance — budget for it, review it, refactor it. The pyramid's shape is a policy: when the integration tier grows fat, push logic down into unit-testable cores; when unit tests multiply on a thin layer, the design is telling you the logic belongs elsewhere.