Feature Stores

A centralized data management layer that streamlines the curation, storage, and serving of machine learning features for training and real-time inference.

Feature stores like Tecton, Feast, and Hopsworks eliminate the 'data silo' problem in ML pipelines by providing a unified source of truth for model inputs. These systems manage the entire lifecycle of a feature: they orchestrate point-in-time joins to prevent data leakage during training and maintain low-latency key-value stores (often backed by Redis or DynamoDB) for sub-millisecond retrieval during production inference. By automating the transformation of raw data into curated signals, teams reduce engineering overhead and ensure that the exact same logic used in development is replicated in the live environment.

https://www.featurestore.org