Whenever you're ready, walk me through how you'd approach this.
How would you handle service failure in a distributed system?
Discuss the answer in the context of an ML-powered distributed service, covering dependency failures, degraded operation, recovery, and observability. Clarify assumptions before proposing an architecture, and explain how you would preserve correctness, availability, and safe model behavior during incidents.