The default 2-worker gunicorn could only serve 2 concurrent tablet requests,
queueing the rest, and the rate limiter saw every tablet as the same Nginx
container IP, so 20 users would have collectively burned through the
100 req/min general bucket.
- gunicorn: 5 workers x 4 gthread, --forwarded-allow-ips=*, access log
- uvicorn: 4 workers, --proxy-headers, --forwarded-allow-ips=*
- RateLimitMiddleware: resolve real client IP from
X-Forwarded-For -> X-Real-IP -> request.client.host
- Bump rate_limit_general 100 -> 300 req/min/IP (per tablet now)
- Flask: ProxyFix(x_for=1, x_proto=1, x_host=1) so request.remote_addr
is the tablet IP, not the Nginx IP
- APIClient: forward X-Forwarded-For + X-Real-IP to FastAPI for both
JSON and multipart/files calls; safe no-op outside request context
- 12 new tests (7 server + 5 client) covering header precedence,
forwarding behavior and ProxyFix install
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adriano
86df67f2e5