Feature store online and offline

Design a feature store serving an organization with hundreds of models, where the same feature definitions must produce training data offline (point-in-time correct over years of history) and low-latency lookups online (sub-10ms p99 at 200k QPS). Some features are batch (daily aggregates), some streaming (last-5-minute counts), and some computed on request. Cover the storage layers, how you guarantee online/offline consistency, and how teams register and discover features.