Feature Store
A feature store is a centralized data system that stores, manages, and serves the engineered features used to train and run machine learning models, ensuring consistency between the features used during training and those used during live…
Definition
A feature store is a centralized data system that stores, manages, and serves the engineered features used to train and run machine learning models, ensuring consistency between the features used during training and those used during live inference.
Overview
As organizations scale from one-off machine learning projects to many models running in production, a recurring problem emerges: the same underlying signal — say, 'average order value in the last 30 days' — often gets recomputed slightly differently by different teams, or is computed one way during training and a subtly different way during live serving. That mismatch, known as training-serving skew, silently degrades model accuracy in ways that are hard to diagnose. A feature store is the infrastructure built to solve this problem directly. A feature store typically has two components: an offline store, optimized for computing and retrieving large batches of historical feature values used to train models, and an online store, optimized for low-latency lookups of the latest feature values needed to serve real-time predictions. Feature stores also handle versioning, so teams can track which feature definitions were used by which model version, and discovery, so teams can find and reuse features that other teams have already built rather than recomputing them from scratch. Feature stores sit downstream of feature engineering — engineers still design and compute the features — but they centralize storage, governance, and delivery of those features across the organization. They are commonly paired with a model registry, which tracks model versions and their approval status, as part of a broader MLOps stack; both are frequently covered together in courses like MLOps & Model Deployment.
Key Concepts
- Centralizes storage and serving of engineered features across an organization
- Splits into an offline store for training data and an online store for low-latency serving
- Prevents training-serving skew by ensuring consistent feature computation
- Supports feature versioning to track which definitions were used by which model
- Enables feature discovery and reuse across multiple teams and models
- Commonly paired with a model registry as part of an MLOps platform
Use Cases
Frequently Asked Questions
From the Blog
How to A/B Test an LLM Feature With Real Users
Judge model changes on behaviour, not on offline scores. Pick one primary behavioural metric, define guardrails that stop the experiment automatically, randomise at the unit users actually experience, and hold the test long enough for the slow signals — retention and follow-up rate — to arrive.
Read More AI & TechnologyHow to Store and Query Metadata Alongside Embeddings
Design the payload before you index anything: flat, typed, low-cardinality fields for the things you will filter on, with tenant and permission keys applied server-side on every query. Then decide deliberately between pre-filtering and post-filtering, because that choice determines whether restrictive filters return empty results.
Read More AI & TechnologyAgent memory: what to keep in context and what to store outside it
Agent memory is two problems, not one. Separate working state from durable knowledge, and learn summarisation, retrieval and eviction that keep both usable.
Read More AI & TechnologyBuilding a golden evaluation set for an LLM feature
Your golden set makes every later decision measurable. Learn to mine real inputs and past failures, size and stratify it, and keep it honest as the product changes.
Read More