graceful-degradation

Implement graceful degradation by wrapping external calls with fallbacks: serve stale cache data when sources fail, use circuit breakers to stop calling failing services, and toggle feature flags to degrade non-critical features. The Skill provides async Python implementations of each pattern.

A circuit breaker stops calling a failing dependency after a failure threshold, preventing cascade overload. A fallback defines what to return instead, such as cached data or a default value. They are typically combined: the breaker detects failure, the fallback provides the degraded response.

Store responses in Redis with a TTL extended by a staleness window, including a timestamp. When the live call raises an exception, read the cached entry and return it if its age is within the acceptable stale TTL, logging the degraded serve.

Critical paths like payments, authentication, and data writes must fail loudly rather than degrade silently, since serving stale or default data there causes correctness or security problems. Degradation is appropriate for read-only, non-critical features like recommendations or search.

Test fallback paths by mocking dependencies to raise timeouts or connection errors, then asserting the fallback response is returned. Also verify circuit breaker state transitions after repeated failures and confirm stale cache is served when the source is unavailable.

Expose a health endpoint reporting per-dependency status as ok, degraded, or error, and emit metrics for fallback usage. Configure Prometheus alerts for degradation active beyond a duration threshold and fallback rates exceeding a percentage of total requests.