Circuit Breaker
Failing fast when a dependency is unhealthy, the closed/open/half-open state machine, tuning thresholds, fallbacks and Resilience4j configuration.
Intent
Popularized by Michael Nygard's Release It!, a circuit breaker wraps calls to a remote dependency. When failures exceed a threshold, it opens and fails calls immediately for a cool-down period, protecting both the caller (threads, latency) and the struggling dependency (load). After the cool-down it half-opens to test recovery.
stateDiagram-v2 [*] --> Closed Closed --> Open: failure rate or slow-call rate above threshold Open --> HalfOpen: wait duration elapsed HalfOpen --> Closed: trial calls succeed HalfOpen --> Open: trial calls fail
Java example (Resilience4j, programmatic)
CircuitBreakerConfig config = CircuitBreakerConfig.custom()
.slidingWindowType(SlidingWindowType.COUNT_BASED)
.slidingWindowSize(50)
.minimumNumberOfCalls(20)
.failureRateThreshold(50) // % of failed calls
.slowCallDurationThreshold(Duration.ofMillis(500))
.slowCallRateThreshold(60) // % of slow calls also opens the breaker
.waitDurationInOpenState(Duration.ofSeconds(10))
.permittedNumberOfCallsInHalfOpenState(5)
.recordExceptions(IOException.class, TimeoutException.class)
.ignoreExceptions(BusinessValidationException.class) // 4xx-like errors are not dependency failures
.build();
CircuitBreaker cb = CircuitBreaker.of("recommendations", config);
Supplier<List<Product>> guarded = CircuitBreaker.decorateSupplier(cb, () -> recoClient.top(userId));
List<Product> result = Try.ofSupplier(guarded).getOrElse(List.of()); // fallback: empty carouselTuning guidance
- Count only dependency failures (timeouts, 5xx, connection errors), not client errors.
- Include slow calls: a dependency answering in 10 s is as harmful as one that fails.
- Minimum call volume avoids opening on a handful of requests at night.
- Expose breaker state as metrics and alert on prolonged open state.
Pros and cons
| Pros | Cons |
|---|---|
| Prevents cascading failures and thread exhaustion | Wrong thresholds cause flapping or late reaction |
| Gives dependencies room to recover | Fallbacks may hide outages if unmonitored |
| Fast failure improves user-perceived latency | Per-instance breakers see only local traffic |
Interview questions
Where should circuit breakers live: library or service mesh?
Mesh-level outlier detection and ejection protect at the connection level without code changes; library breakers know business context and can choose meaningful fallbacks. Many systems use both, with the application owning fallback behaviour.
What makes a good fallback?
Cached or stale data, a default value, a reduced feature (hide the widget), or a queued request for later processing. Never a fallback that silently produces wrong financial results.