Feature Store

Full definition

A feature store provides two interfaces: an offline store (columnar warehouse) for training and batch scoring, and an online store (low-latency key-value DB) for real-time inference. Both are populated from the same feature pipelines, eliminating training-serving skew — a leading cause of production model failure. Uber introduced the pattern with Michelangelo (2017); modern open-source implementations include Feast and Tecton. AIDOLS uses feature stores as the backbone of any production ML platform we deliver through services.

Why it matters

Without a feature store, every team rebuilds feature logic in Python notebooks, then again in production code, then again in monitoring. Bugs emerge silently as offline and online definitions diverge. A feature store collapses that fan-out into one source of truth and is correlated with materially shorter time-to-production for new models.

Example

A fintech defines "user_avg_transaction_value_30d" once in the feature store. Training jobs pull historical values; the fraud model in production reads the same feature at sub-10ms latency through the online store.

Related terms

• MLOps
  MLOps is the discipline of operating machine-learning and AI systems reliably in production — covering data pipelines, model training, deployment, monitoring, drift detection, governance, and incident response — analogous to DevOps for traditional software.
• Model Serving
  Model serving is the runtime infrastructure that hosts a trained model and exposes it as an API for low-latency online inference at scale, handling batching, autoscaling, GPU sharing, versioning, and routing.
• Online Inference
  Online inference returns model predictions synchronously in response to a live user or system request, typically under 100ms end-to-end for tabular models and under 2s for streaming LLM completions.
• Model Monitoring
  Model monitoring is the continuous measurement of a deployed model's inputs, outputs, and performance signals — including data drift, prediction drift, label drift, latency, and ground-truth accuracy — to detect degradation before it harms business outcomes.