slowapi limit like @limiter.limit('60/minute') lets ~30x more through under gunicorn/uvicorn workers: default in-memory storage is per process; set storage_uri to Redis
Fix: give the slowapi Limiter a shared backend, e.g. limiter = Limiter(key_func=get_remote_address, storage_uri="redis://redis:6379/0"). Without storage_uri, slowapi (via the limits package) keeps counters in process memory, so every gunicorn worker enforces its own separate copy of the limit.
What we saw: FastAPI 0.115 + slowapi 0.1.9 under gunicorn 23 with worker_class = "uvicorn.workers.UvicornWorker" and workers = cpu_count()*2 + 1 (25 workers on that box). The route was decorated @limiter.limit("60/minute"). A 30-second run at 10 req/s with oha (oha -z 30s -c 5 -q 10) got 180 x 200 and 120 x 429. A shared 60/minute limit allows about 60 in that window. A 10-second burst (-c 10, no rate cap) got 1,321 x 200 before the 429s took over. Which worker takes a request decides which counter it hits, so you see 429s, just far too few.
How to verify: run Redis (the app was already using it for caching), pass the same URL as storage_uri, restart, then repeat a fixed-rate load test. The count of 200s per minute should match the decorator no matter how many workers run. redis-cli KEYS 'LIMITER*' shows the counters being written.
Also: get_remote_address returns the proxy's IP when the app sits behind a reverse proxy or a container network. Then all clients share one bucket. Run uvicorn/gunicorn with forwarded-allow-ips and key on the real client address when that applies.