Feature Store

A feature store is the centralized infrastructure that computes, stores, and serves the features used by machine-learning models — both during training (where features must match the model's expected distribution) and at inference (where features must be available with low latency in production). Feature stores solve the chronic mismatch between training pipelines (typically batch, offline, in a data warehouse) and serving pipelines (real-time, low-latency, in production), ensuring that "the value of feature X for user Y at time T" is computed the same way in both places — eliminating training-serving skew. Leading offerings include Feast (open-source, CNCF), Tecton (commercial, originally from Uber Michelangelo), Databricks Feature Store, Hopsworks, AWS SageMaker Feature Store, Vertex AI Feature Store, and Snowflake's feature engineering primitives. For LLM applications, feature stores serve a related but distinct role: they manage the user-context and session-state features that enrich prompts at inference time — user profile, recent activity, entitlements, locale, account tier — feeding into a personalized system prompt or RAG filter. A practical pattern: define features once in a feature definition (Feast feature views, Tecton feature definitions), batch-compute them daily into an offline store (S3, Snowflake, BigQuery), stream low-latency features into an online store (Redis, DynamoDB, Cassandra), and serve them via a feature-store SDK at training and inference. For LLM-personalized applications, the feature store delivers "this user's last 5 interactions, current entitlements, and active subscription tier" into the prompt context at single-digit-millisecond latency. AI governance teams use feature stores to enforce consistent feature definitions across teams, audit feature usage by model, and apply access controls so sensitive features (income, health metrics) are not exposed to models that should not see them.

Feature serving from 25 years of personalized content delivery: Centralpoint has computed and served user-context features — audience, entitlement, locale, preference — for 25 years to drive personalized content delivery. Extending those features to enrich AI prompts is the same engineering with a new consumer. Features stay on-premise, tokens meter per skill, and feature-personalized chatbots deploy through one line of JavaScript.


Related Keywords:
Feature Store,Feature Store,Oxcyon, AI, AI Governance, Generative AI, Inference, Inference, Inferencing, RAG, Prompts, Skills Manager,