nopanic-redis

Redis-backed circuit breakers and rate limits for nopanic, shared by every process and host that uses them.

pip install nopanic-redis

A nopanic.circuit_breaker lives in one process. Behind a load balancer with 40 workers, each worker has to discover a dead dependency on its own, and a rate_limit(100) lets 4,000 calls through. These policies keep that state in Redis, so the fleet trips together and shares one quota:

import redis
from nopanic_redis import redis_circuit_breaker, redis_rate_limit

r = redis.Redis()   # redis.asyncio.Redis() for async functions

payments = redis_circuit_breaker(r, name="payments", min_calls=20, reset_timeout=15.0)
openai_quota = redis_rate_limit(r, name="openai", rate=500, per=60.0, max_wait=5.0)

@openai_quota
@payments
def charge(order): ...

Semantics

• Same model as the in-process policies. redis_circuit_breaker takes the parameters of nopanic.CircuitBreaker (failure-rate threshold over a bucketed sliding window, min_calls, reset_timeout, half_open_max_calls, on). redis_rate_limit is nopanic.rate_limit (token bucket, burst, waiting by default, RateLimited(retry_after=...) with max_wait, ambient deadline respected, cancelled waiters refunded).
• Identity is the name. Every instance with the same name and prefix is the same breaker or bucket, whichever process created it. Keys use a Redis Cluster hash tag, so each policy's keys stay in one slot.
• Atomic and cheap. Each decision is one Lua script, one round trip: a breaker costs two per call (admit, record), a rate limit one.
• Clocks. Time comes from the callers' wall clocks (time.time, injectable via clock=); keep hosts NTP-synced. Skew only shifts window boundaries slightly.
• Redis down is not your outage. With fail_open=True (default) calls go through unguarded and a breaker.store_error / ratelimit.store_error event is emitted; fail_open=False rejects with CircuitOpen / RateLimited instead.
• Sync or async by client. A redis.Redis client serves sync functions, a redis.asyncio.Redis client serves async ones (use two instances with the same name to guard both; they share state).
• breaker.get_state() and breaker.reset() act fleet-wide (awaitable with an async client). Streams are not supported.

Events are the usual nopanic events, so nopanic-otel and events.log_to work unchanged.

License

MIT.