Feature degradation keeps a system useful under partial failure by shedding non-essential features instead of erroring on everything. The user keeps the critical path; optional extras pause until the system recovers. It is the concrete tactic behind the graceful degradation quality.

It works only when features carry an explicit criticality ranking and the core path has no hard dependency on the parts being shed.

How It Works

  • Rank features by criticality: a core set that must stay up, and optional features that can pause.
  • Watch health signals per dependency: error rate, latency, queue depth, or remaining capacity.
  • When a signal crosses its threshold, disable or simplify the optional features bound to that dependency.
  • Serve a defined fallback for each shed feature: a cached value, a sensible default, or a clear “temporarily unavailable” state.
  • Re-enable features once the signal recovers and holds for a stabilization window.

Failure Modes

  • The core path itself depends on a shed component, so degradation takes down the function it meant to protect.
  • Fallbacks return stale or wrong data silently, and users act on it.
  • Thresholds sit too close to normal load, so features flap on and off and confuse users.
  • No re-enable path exists, so the system stays degraded long after the dependency recovers.

Verification

  • Inject a dependency failure and assert the core path keeps a success rate above its target while optional features report unavailable.
  • Measure that core-path response time stays within budget in degraded mode.
  • Confirm features re-enable within the stabilization window after health is restored.
  • Rate Limiting and load shedding drop requests by priority; feature degradation drops whole features.
  • Circuit Breaker cuts a single failing dependency; feature degradation decides what the user still sees once it is cut.
  • Bulkheads isolate resource pools so degradation stays contained to one feature’s callers.
  • Fail-Safe Defaults define the safe values a degraded feature falls back to.

References