Ad ranking inference latency

Design the real-time ad ranking service that scores candidate ads for each ad slot on a social feed. For every impression it must fetch dozens of features (user features, ad features, context) from multiple stores, run a ranking model over ~500 candidate ads, and return the winner within a 30ms budget — at 1M requests/second. A slow or failed feature store call cannot be allowed to blow the latency budget or drop the whole request. Design the inference service with its feature-fetch fan-out and degradation strategy.