FastAPI Interview Questions for Senior (7+ Years)

Use StreamingResponse with an async generator. The generator yields tokens as they arrive from the LLM API: async def stream(): async for chunk in openai_client.stream(): yield chunk.text. Return StreamingResponse(stream(), media_type="text/event-stream"). On the frontend, read with EventSource or fetch with streaming. This reduces time-to-first-token from 5s+ to under 500ms perceived latency.

Use FastAPI BackgroundTasks for lightweight tasks (email notifications after AI processing). For heavy AI workloads (embedding generation, model inference), use a proper task queue: Celery + Redis or ARQ. The pattern: receive request → validate → enqueue task → return task_id immediately → client polls /tasks/{task_id} for results. This prevents request timeouts on long-running AI operations.

Use slowapi (FastAPI wrapper for limits library) for per-IP rate limiting, or implement custom middleware using Redis for per-user limits. For AI APIs specifically: limit by user tier (free: 10 req/min, paid: 100 req/min), by model (expensive GPT-4: stricter limits), and use token counting for OpenAI costs. Decorate with @limiter.limit("10/minute") and return 429 with Retry-After header.

Use multi-stage Docker build: Stage 1 (builder) installs dependencies, Stage 2 (production) copies only necessary files. Run with Uvicorn + Gunicorn: CMD ["gunicorn", "main:app", "-w", "4", "-k", "uvicorn.workers.UvicornWorker"]. Use --workers = (2 × CPU cores) + 1. For AI APIs with heavy model loading, use 1 worker per GPU and load models at startup in lifespan events, not per-request.