Fitting Retrieval and Ranking Into a 200 ms Budget

Your marketplace has 5 million listings and a p99 latency target of 200 ms for the search response. A "Paris, next weekend, 2 guests" query leaves about 60,000 listings after geo/date/capacity filters. Your heavy neural ranker takes about 0.3 ms per listing on CPU when batched.

1. Show with numbers why scoring all candidates with the heavy ranker is infeasible, and design the stage funnel that makes it fit.
2. Allocate a rough per-stage latency budget and name one fallback per stage.
3. Where would caching help and where would it be dangerous?