DataTerminology.comNEW
Also known as: Online Store (Feature Store)
A low-latency data store that serves precomputed feature values for real-time model inference.
DESCRIPTION
An online feature store provides millisecond-level access to feature values for serving predictions in production systems. It ensures consistency with offline training data by synchronizing computed features from pipelines. Typical implementations use high-performance key-value or vector stores to meet strict SLA requirements. It supports time-travel-safe lookups and often enforces point-in-time correctness via feature timestamps. Access is usually through a feature registry and online serving APIs. Robust online stores minimize training–serving skew and enable consistent, repeatable inference. They also integrate with streaming ingestion for continuous updates. Security, caching, and monitoring are critical to maintain performance and correctness at scale.
ALTERNATIVE DEFINITIONS
AWS
The online store is a low latency, high throughput database for real-time ML inference.
Stresses low latency and high throughput for production inference.
Feast
The online store is used for low-latency access to feature values for online serving.
Emphasizes serving features at low latency through an online store.
Tecton
An online store provides low-latency access to feature values for serving predictions.
Highlights prediction serving and low-latency access.
RELATED TERMS
Feature Store Feature Engineering Feature Pipeline Real-time Inference Model Serving Feature Registry Streaming Ingestion Feature Materialization Low-Latency Storage
COMPETING TERMS
Offline Feature Store On-demand Feature Computation
RELATIONS
- Broader: Feature Store
KNOWLEDGE AREAS
MLOps Data Engineering AI Real-time & Streaming Data Data Services Data Storage & Operations
TARGET AUDIENCE
MLOps Engineer AI Engineer Data Engineer Data Scientist Cloud Architect Data Architect Software Developer
EXTERNAL SOURCES
- Wikipedia - Feature store
- AWS - SageMaker Feature Store overview
- Feast - Feast concepts and architecture
- Tecton - Online vs. offline store concepts
- Databricks - Databricks Feature Store guide
TOP QUESTIONS
Why is an Online Feature Store critical for real-time inference?
It guarantees low-latency, consistent access to features, reducing training–serving skew. This consistency directly improves prediction accuracy and reliability in production.
Category: Relevancy
How does an Online Feature Store differ from an Offline Feature Store?
The online store prioritizes low-latency serving for inference, while the offline store focuses on historical data for training and batch scoring. They synchronize to maintain consistency.
Category: Differences
What SLAs matter for an Online Feature Store?
Typical SLAs include single-digit millisecond p99 latency, high availability, and strict freshness windows. Throughput and multi-region replication may also be required.
Category: Requirements
What are common mistakes with Online Feature Stores?
Not enforcing point-in-time correctness, poor key design, and inconsistent update policies cause skew and stale features. Missing observability hides latency and freshness regressions.
Category: Mistakes
How is an Online Feature Store secured?
Use strong IAM, per-feature or per-entity access controls, encryption in transit/at rest, audit logging, and network isolation such as VPC peering or private endpoints.
Category: Security
What data models are typical in Online Feature Stores?
Key-value schemas keyed by entity IDs with timestamped feature values are common. Some stores also support vector embeddings and approximate nearest neighbor indexes.
Category: Architecture
How do streaming updates integrate with an Online Feature Store?
Streaming pipelines compute and materialize feature updates continuously, using CDC or event streams to minimize latency and maintain freshness for online serving.
Category: Application
What are the risks of relying on an Online Feature Store?
Operational failures can cause elevated latency or stale features, degrading predictions. Inconsistent schemas or missing backfills can introduce bias and break compatibility.
Category: Risks
What tools commonly power Online Feature Stores?
Backends often include Redis, DynamoDB, Cassandra, or specialized vector databases. Feature frameworks integrate these with a registry and serving APIs.
Category: Tools