Your product calls an external LLM API and you are on a 99.9% uptime SLA. The provider goes dark for 4 hours. Design the failover.

The first real signal is whether the candidate does the error-budget math: 99.9% is ~43 min/month, so 4 hours is roughly 5.6x the monthly budget and the question is not 'wait it out' but 'fail over or degrade'. Strong answers put every provider behind one interface, add a circuit breaker so you stop hammering a dead provider, and define a fallback ladder: secondary provider, then smaller/cheaper model, then cache, then queue, then a graceful 'reduced features' mode. The reserve probe is the consistency one: what do you cache (idempotent reads, embeddings) versus what must you queue and replay (writes, anything with side effects), and how do you avoid double-execution on replay. Watch for candidates who name a backup provider but never address prompt portability or output-format drift between providers.