Performance Tuning

Most FastAPI slowness comes from blocking calls in async code, N+1 queries, missing indexes, chatty external calls and large payloads. Profile (py-spy, pyinstrument, traces), then: use async drivers end to end, eager-load relationships, add indexes, cache hot reads in Redis, run independent I/O concurrently, paginate, compress responses (GZip), and consider faster JSON (ORJSONResponse) for very large payloads. Load test with Locust or k6 to verify.

from fastapi.middleware.gzip import GZipMiddleware
import asyncio

app.add_middleware(GZipMiddleware, minimum_size=1024)

sem = asyncio.Semaphore(20)


async def fetch_price(http, sku: str) -> int:
    async with sem:                                   # bounded concurrency
        r = await http.get(f"https://pricing/skus/{sku}")
        r.raise_for_status()
        return r.json()["price_cents"]


async def prices(http, skus: list[str]) -> dict[str, int]:
    results = await asyncio.gather(*(fetch_price(http, s) for s in skus))
    return dict(zip(skus, results))

py-spy top --pid $(pgrep -f "uvicorn")      # live CPU profile of a worker

Common pitfalls

• Optimizing JSON serialization while the database dominates latency.
• Unbounded concurrency overwhelming downstream services.
• Load testing with unrealistic data volumes.

Key points

• Profile first
• Async end to end; bounded concurrency
• Fix queries, cache hot reads, paginate

Further reading: FastAPI: Custom responses