# Feature Stores

> Systems that define, retrieve, and serve machine-learning features consistently for training and inference.

- HTML version: https://robbiepalmer.me/ideas/feature-stores
- Source: https://docs.feast.dev/getting-started/architecture/overview

A feature store gives named, versioned feature definitions a shared route into model training and
production inference. Historical retrieval builds training sets as the features looked at a chosen
time. Online retrieval returns the latest values for low-latency predictions.

That split exists because the two workloads differ. Training scans large time ranges and needs
point-in-time-correct joins. Online inference usually needs a small vector for one entity within
milliseconds. Materialisation moves selected values from the historical path into an online store,
while a registry records definitions, ownership, and data sources.

Consistency and reuse are the reason to introduce this infrastructure. A feature store cannot fix
a weak definition, and many leave transformation to declarative queries, stream processors, or
other compute engines. It earns its cost when several models need the same governed features or
when training-serving skew has become a recurring failure.

## Questions it prompts

* Can training retrieve the value that was known at prediction time?
* Which definition and source produced this feature value?
* How fresh must the online value be, and what happens when materialisation lags?
* Is shared infrastructure justified, or would a versioned pipeline be simpler?

## Related ideas

- [Feature Engineering](https://robbiepalmer.me/ideas/feature-engineering.md): The work of turning raw observations into model inputs that represent the signal needed for a prediction.
- [Stream/Table Duality](https://robbiepalmer.me/ideas/stream-table-duality.md): A stream can describe the changes to a table, while a table represents the latest state produced by those changes.

## Where it appears

- Technology: [Google BigQuery](https://robbiepalmer.me/technologies/google-bigquery.md)

---

Markdown index of this site: https://robbiepalmer.me/llms.txt