DataTerminology.comNEW
Also known as: Feature Repository, Feature Platform
Centralized system to store, manage, and serve machine learning features consistently for training and low-latency inference.
DESCRIPTION
A feature store provides a governed hub for curated, reusable ML features, separating feature computation from model use. It typically offers an offline store for training data generation and an online store for low-latency serving. Feature definitions and transformation logic are versioned to ensure training–serving consistency and reproducibility. Integrations include data orchestration, model registry, monitoring, and access control. Feature stores reduce duplication, accelerate experimentation, and improve model reliability by standardizing semantics. They can compute features batch, streaming, or on demand, and support point-in-time joins to prevent leakage. Modern platforms integrate real-time pipelines and embedding management for deep learning and LLM applications.
ALTERNATIVE DEFINITIONS
Google Cloud
Vertex AI Feature Store is a fully managed repository for storing, serving, and sharing ML features.
Highlights managed service and collaboration across teams.
AWS
Amazon SageMaker Feature Store is a fully managed, purpose-built repository to store, share, and manage features for machine learning models.
Emphasizes purpose-built nature and ML lifecycle integration.
Databricks
Databricks Feature Store is a centralized repository of features that enables feature sharing, discovery, and serving to models in production and for training.
Stresses discoverability and production serving alongside training.
Feast
Feast is an open source feature store for machine learning that helps teams manage, discover, and serve features for online and offline use cases.
Open-source, multi-engine focus supporting online and offline features.
RELATED TERMS
Feature Engineering Online Feature Store Offline Feature Store Model Serving Model Registry Feature Registry Data Drift MLOps Vector Database Embeddings
COMPETING TERMS
RELATIONS
KNOWLEDGE AREAS
MLOps Data Engineering AI Data Science Data Management
TARGET AUDIENCE
MLOps Engineer AI Engineer Data Scientist Data Engineer AI Product Manager
EXTERNAL SOURCES
- Wikipedia - Feature store
- Google Cloud - Vertex AI Feature Store docs
- AWS - SageMaker Feature Store guide
- Feast - Open-source feature store docs
TOP QUESTIONS
Why use a Feature Store instead of Ad-hoc feature pipelines?
It centralizes definitions, ensures training–serving consistency, and promotes reuse. This reduces duplication and leakage, speeding experiments and improving reliability.
Category: Relevancy
What is the difference between an Online and Offline Feature Store?
Offline stores generate training datasets at scale; online stores serve low-latency features for inference. Both share definitions to maintain consistency.
Category: Distinction
How do Feature Stores prevent training–serving skew?
They version transformation logic and provide point-in-time joins, ensuring identical semantics between training and production serving.
Category: Quality
What is a hot topic around Feature Stores?
Real-time feature computation and managing embeddings for LLM applications are rapidly evolving capabilities.
Category: Topic
How are Feature Stores governed?
They integrate access control, lineage, and metadata catalogs, enabling stewardship and compliance for shared ML assets.
Category: Governance
What are common mistakes when adopting a Feature Store?
Treating it as only a table store, ignoring point-in-time correctness, and lacking ownership models undermine trust and reuse.
Category: Mistakes
What are prerequisites for implementing a Feature Store?
Reliable data pipelines, clear feature ownership, CI/CD for transformations, and a registry to version feature definitions are essential.
Category: Requirements
What competes with a Feature Store for serving ML features?
Traditional data warehouses or custom services can serve features but often lack point-in-time and low-latency guarantees.
Category: Alternatives
How do Feature Stores support monitoring?
They integrate with model monitoring to track data drift, freshness SLAs, and feature value distributions.
Category: Operations