Entity

A one-dimensional array of numerical values representing the aggregated features for a specific entity at a specific point in time. The entity ID acts as the primary key that links disparate feature values—such as user demographics and real-time clickstream data—into a single, coherent input for a machine learning model.

A data engineering guarantee that feature values are reconstructed exactly as they existed historically for a specific entity. When building a training dataset, the system joins features based on the entity key and the precise timestamp, preventing future data from leaking into the past and ensuring the model learns from realistic, time-consistent examples.

The low-latency database component designed to serve pre-computed feature vectors at inference time. It is typically a key-value store optimized for point lookups, where the entity ID (e.g., user_id or product_sku) is the key used to retrieve all associated features in milliseconds to avoid delaying a real-time prediction.

A logical collection of related features sharing a common entity key and update cadence. For example, a 'user_profile' feature group might contain age, membership_tier, and lifetime_value, all keyed by user_id. This grouping simplifies management, materialization, and consistent joining logic within the feature store.

The practice of discovering and consuming existing feature definitions from a shared registry. By standardizing on a common entity model (e.g., a canonical customer_id), teams can avoid duplicating engineering work. A fraud model can reuse the same customer aggregation features originally built for a marketing propensity model, accelerating development.

Feature values computed incrementally on real-time event data and keyed by entity. Unlike static batch features, streaming features capture immediate intent—such as items viewed in the current session by a specific visitor_id. These are materialized into the online store with sub-second latency to power dynamic personalization.