pytest-fastapi-pattern

To test FastAPI lifespan startup with pytest, use httpx AsyncClient configured with ASGITransport so that lifespan startup and shutdown events execute exactly like production, unlike sync TestClient which skips them entirely.

Database state leaks between pytest tests when isolation is absent. Using in-memory SQLite with StaticPool alongside a per-test create_all and drop_all fixture flow ensures deterministic, isolated database state for each individual test.

Upstream LLM API calls in FastAPI tests are mocked using respx. It intercepts HTTP requests to gateways like vLLM or OpenAI and uses assert_all_called to guarantee that the defined mocks are actually exercised during the test.

SSE streaming output in pytest is verified using async streaming iteration to capture byte-level chunks. This allows asserting specific SSE markers like [DONE] and validating the exact chunk sequence returned by the streaming endpoint.

FastAPI sync TestClient skips lifespan startup code, leading to misleading integration test results. Using httpx AsyncClient with ASGITransport resolves this by ensuring lifespan startup and shutdown events run properly within the test.

The best way to test FastAPI gateway applications forwarding to vLLM is using deterministic async pytest patterns. This involves httpx ASGITransport for lifespan, in-memory SQLite isolation, and respx to mock the upstream vLLM HTTP calls.