To power several machine learning use cases like route planning and pricing, Lyft needs to compute features from historical data (through batch queries) and in real-time (through event streams) for both training and low-latency online inference. As such, Lyft developed a Feature Service that includes feature definitions, feature ingestion and processing, and retrieval components (and also guarantees online-offline consistency). Feature definitions are specified in SQL with metadata in JSON. Feature ingestion is handled by Flyte or Flink and feature retrieval is supported by GRPC and REST endpoints (with DynamoDB, Hive, Elasticsearch, and Redis used for data management).