Feature store with point-in-time correctness
Design a feature store for a fraud-detection model where the same feature definitions (e.g. 'card transaction count in last 24h', 'avg basket value last 7d') are used both to train offline on years of historical data and to score transactions online at 5k TPS with a 20ms feature-fetch budget. The bank's biggest recurring problem is training/serving skew: models look great offline and underperform in production. Design the store to make online and offline features provably consistent, and explain how you guarantee point-in-time correctness when generating training data.
Clarify scale and constraints first. Propose a clean component breakdown, then go deep on the hard parts (data model, bottlenecks, consistency, failure modes) and name the trade-offs you are making.