Skip to content

Security Model

Authentication

Request filtering funnel diagram Request filtering funnel diagram

Every request except public endpoints must include:

X-API-Key: <your-api-key>
  • Missing key → 401 Unauthorized
  • Unknown key → 401 Unauthorized
  • Insufficient scope → 403 Forbidden

API key comparison uses hmac.compare_digest — constant-time, preventing timing attacks.

Scopes

Scope Grants access to
predict /predict, /predict/batch, /predict/async*
read_models /models
admin /debug/*, /admin/*

/health, /ready, /metrics — no auth required (public endpoints). /jobs/{id} — auth required, no specific scope.

Rate limiting

Per-tenant sliding-window rate limits keyed on tenant_id.

Endpoint Limit
/predict 10 req / 1s
/models 2 req / 1s

Exceeded → 429 Too Many Requests

In-process (REDIS_URL not set): per-process counters only. Accurate for single-process deployments.

Redis-backed (REDIS_URL set): atomic Lua script with sorted set. Accurate across multiple processes.

Every response includes X-RateLimit-Mode: local or X-RateLimit-Mode: distributed.

Payload guard

Requests with a body larger than 1 MB are rejected before reaching any route handler → 413 Request Entity Too Large.

Production checklist

  • [ ] Set ENV=production
  • [ ] Set API_KEYS with strong randomly generated keys
  • [ ] Run behind TLS-terminating reverse proxy
  • [ ] Restrict /metrics and /debug/* to internal networks
  • [ ] Use Redis-backed rate limiting for multi-process deployments

See Auth Configuration.