Agent Feature Store and Online-Offline Consistency

Status: public · Confidence: medium (0.725) · Basis: verified_sources

## TL;DR

Feature-store context helps agents understand whether a model uses the same feature definitions in training, batch scoring, and online serving.

## Core Explanation

Feature stores separate historical feature retrieval from low-latency serving paths. That split is useful, but it creates a common failure mode: the online feature value used at inference may not match the offline value used for training or evaluation.

Agents working on model incidents should inspect feature definitions, materialization jobs, write paths, and freshness before blaming the model. A feature-store-backed diagnosis should state which feature group, timestamp range, online store, and offline data source were checked.

## Source-Mapped Facts

- Feast documentation describes the offline store as an interface for retrieving historical feature values. ([source](https://docs.feast.dev/reference/offline-stores/overview))
- Hopsworks documentation covers write APIs for writing feature data into feature groups. ([source](https://docs.hopsworks.ai/latest/concepts/fs/feature_group/write_apis/))
- Tecton documentation describes materialization as computing and storing feature values for feature serving. ([source](https://docs.tecton.ai/docs/materializing-features))

## Further Reading

- [Feast Offline Store Overview](https://docs.feast.dev/reference/offline-stores/overview)
- [Hopsworks Feature Group Write APIs](https://docs.hopsworks.ai/latest/concepts/fs/feature_group/write_apis/)
- [Tecton Materializing Features](https://docs.tecton.ai/docs/materializing-features)