A fallback is an alternative behavior used when a preferred operation or dependency is unavailable. It helps a service keep working from the user’s perspective, sometimes at reduced quality. This is graceful degradation: bending rather than failing completely. section 7.0 section 7.5.4
The right fallback depends on what has failed:
- For an overloaded service, shed load by rejecting or throttling some requests early. Serving a portion of users can be better than allowing every request to time out. section 7.1.2
- For a backed-up request queue, adaptive LIFO keeps FIFO order normally but processes newer requests first under heavy load, when older requests may already have been abandoned. section 7.2
- For an unavailable dependency, distinguish hard from soft dependencies. A hard dependency needs a strategy to keep the service running; a soft one can fail without breaking the critical user journey. section 7.5.1
When more capacity makes overload worse
A REST service has a database connection pool of 100 per instance. During a traffic spike, Kubernetes starts more instances, which add connections until the database slows and requests time out. Rejecting some traffic early could instead preserve service for some users. section 7.1.2
A fallback is only useful if it still works when needed. Test it continuously: a path left unexercised for months may be dead code. If dependency downtime becomes invisible to users, a formerly hard dependency can become soft—but maintaining that status takes ongoing effort. section 7.5.4 section 7.8























