Featurewhere

Featurewhere is a conceptual framework and software approach for managing the "where" of features in machine learning pipelines. It emphasizes provenance, locality, and deployment context for features, aiming to improve reproducibility, governance, and explainability by documenting where features are computed, stored, and consumed.

Origin and scope: The term emerged in discussions of increasingly complex feature engineering workflows, where data sources migrate across environments and feature definitions drift. Featurewhere builds on ideas from feature stores and data provenance tools, focusing on the spatial and operational location of features within pipelines.

Architecture and components: A featurewhere system typically includes a feature catalog, a provenance ledger, locality-aware compute orchestration, and deployment adapters. Users declare feature definitions with metadata about data sources, computation steps, storage location, and access controls, enabling automatic lineage tracking.

Capabilities: It enables lineage auditing, reproducibility across environments, and locality optimization by scheduling feature computations near the data source or consumer. It supports versioning, rollback, access governance, and integration with machine learning platforms.

Applications: Use cases include regulated domains such as finance or healthcare, where traceability and compliance matter, as well as real-time scoring systems and experimentation pipelines that require consistent feature states across models.

Limitations: Critics note added complexity and performance overhead, as well as the need for standard schemas and interoperability with existing data catalogs and feature stores.

See also: feature store, data lineage, data provenance.