Online vs offline feature store — architecture.
mediumAnswer
- Offline store: columnar format (Parquet / Delta / Iceberg) on object storage.
- Cheap, high throughput, for training + batch inference.
- Online store: low-latency key-value (Redis, DynamoDB, ScyllaDB) for real-time serving (sub-10ms).
- Sync via streaming (Kafka + Flink) or batch push.
- Must guarantee point-in-time correctness (as-of joins) between training data + online reads.
- Feast, Tecton, Hopsworks are canonical implementations.
- Custom KV store often works for smaller teams.
Check yourself — multiple choice
- Random
- Offline: columnar (Parquet/Delta) for training + batch; Online: KV (Redis/DynamoDB) for real-time <10ms; sync via Kafka/Flink; point-in-time correctness required; Feast/Tecton
- Same thing
- Not real
FS: offline columnar + online KV + streaming sync + PIT correctness.
#features#infrastructure
Practise MLOps & Data Quality
215 interview questions in this topic.
Related questions
- How do you compute real-time features (streaming aggregates)?
- Push vs pull materialization for features.
- When is a feature store worth the operational cost?
- What problem does a feature store solve?
- How do you handle missing values in a production pipeline?
- How do you handle high-cardinality categorical features?