Design an ML Feature Store

Requirements

• •Register and version feature definitions (feature groups) with schema and transformation logic
• •Compute features from raw data via batch pipelines (daily) and streaming pipelines (real-time)
• •Serve features online at < 10ms latency for inference requests
• •Generate point-in-time correct feature snapshots for model training (no future data leakage)
• •Support backfilling historical feature values when a new feature is added

Non-Functional Requirements

• •Online feature lookup < 10ms (p99)
• •Training dataset generation for 100M entities < 4 hours
• •Support 1,000 distinct feature definitions
• •Streaming features must be < 60 seconds stale in the online store

Constraints

• •Features are numerical and categorical, up to 1,000 features per entity
• •Entities are users, products, and transactions
• •Training jobs run weekly; online inference runs continuously

Evaluation Criteria