Comparison¶

Picking a circuit breaker is mostly about which failure model, runtime and coordination story you need. This page compares interlock-cb with the established Python circuit breakers — honestly. interlock-cb is the youngest of the five (first released in 2026); pybreaker, circuitbreaker, aiobreaker and purgatory have carried production traffic for years, and for many projects one of them is exactly the right choice.

Feature table¶

_{Compared against pybreaker 1.4.1, circuitbreaker 2.1.3, aiobreaker 1.2.0 and purgatory 3.0.1, as documented in July 2026. Something out of date or unfair? Please open a PR.}

The four established libraries, honestly¶

pybreaker is the original Python circuit breaker and the most proven one: a small, stable sync breaker with listeners and optional Redis-backed state, maintained for well over a decade. Its async support targets Tornado, not asyncio, and it trips on a consecutive-failure count rather than a failure rate. If you run a synchronous stack and want the most battle-tested option, start here.

circuitbreaker has the smallest API of the five: one @circuit decorator that also handles async functions, plus the only built-in fallback function in this table. There are no listeners, no shared state and no rate window — which is precisely its appeal when you need a guard on a handful of call sites and nothing else.

aiobreaker is pybreaker ported to native asyncio, with the same listener and Redis-storage features. Its last release was in 2021; for new projects, prefer an actively maintained alternative.

purgatory brings a fully typed sync + async breaker with Redis storage and event hooks. Sync and async live in separate factory classes (SyncCircuitBreakerFactory / AsyncCircuitBreakerFactory), and tripping is a consecutive-failure threshold with a TTL on the open state.

Where interlock-cb differs¶

• Rate over a window, not a streak. A consecutive-failure counter resets on any single success, so a dependency failing 90% of requests under load can keep a breaker closed indefinitely. interlock trips on the failure rate across a count- or time-based sliding window (configuration).
• Slow calls are failures too. A dependency that answers in 30 s can be worse than one that errors fast. slow_call_duration_threshold + slow_call_rate_threshold trip the breaker on latency degradation alone.
• One class, both runtimes. The same CircuitBreaker instance guards sync and async callables — decorator, context manager, or call() — and the decorator preserves the wrapped signature for type checkers.
• Coordination built for fleets. With the Redis storage, tripping is atomic across racing instances and half-open probes are budgeted globally, with graceful degradation to local state when Redis is down (Redis integration).
• Meets your stack where it is. Per-host breakers ship for httpx2, httpx, aiohttp and requests; tenacity glue composes retries correctly; FastAPI and Litestar map rejections to 503 + Retry-After (integrations overview).

The honest trade-off: interlock-cb requires Python ≥ 3.11 and has not had years in production. Reach for an established library if that maturity matters more than the feature gap; choose interlock-cb when you want rate-based windows, slow-call detection, coordinated state and a fully typed API.