Q1What is the must-have feature list for a production AI gateway?
Six items: (1) per-channel circuit breakers, (2) configurable fallback chains, (3) double-entry billing (upstream cost + sale price), (4) sk-level concurrency / QPS / token / cost quotas, (5) async task lifecycle for image / video / music, (6) admin escape hatches (kill switch, cache invalidate, manual breaker).
Q2Does Anima have manual kill switches?
Yes. POST /api/admin/v1/circuits/{channel}/open immediately opens the breaker, PATCH /api/admin/v1/channels/{code} enabled=false disables the channel, PATCH /api/admin/v1/apps/{id}/keys/{kid} enabled=false bans a single sk. All take effect via Redis pub/sub within ~1s.
Q3How are queues handled?
Three modes per sk: REJECT (immediate 429), WAIT (Redis ZSet wait queue with timeout), ASYNC (task_id is returned, a worker drains the queue). The mode is set on the user_api_key row and can be changed live without a deploy.
Q4What metrics does the gateway expose?
Prometheus at /actuator/prometheus: request count + latency histograms per channel / alias / app, circuit-breaker state, queue length, upstream cost vs charge spread. Grafana dashboards JSON ship with the repo.
Q5Is configuration zero-downtime?
Yes. Every config table (channel / alias / price / quota / sk / token-group) has a version_no optimistic lock plus effective_from/to time window. ConfigPublisher writes DB then publishes invalidation events; L1 (Caffeine) + L2 (Redis) caches refresh within 60s without restart.
Q6What is the worst-case data path on failure?
If both Redis and the primary DB go down, gateway returns 503 with a service_unavailable code. Read paths gracefully degrade to L1 in-process cache (30s TTL). Write paths refuse — billing integrity is preferred over availability.
All answers
compare
Best AI gateway for production workloads
Production checklist for picking an AI gateway: fallback chains, circuit breakers, billing transparency, SLA. How Anima scores on each.
By Anima Team· 6/15/2026· locale: en
All answers© 2026 Orux AI