Take a card-fraud model. Its inputs might include “number of purchases on this card in the last hour” and “average purchase size over 30 days.” During training, those values are calculated from months of history. In production, they must be calculated again in milliseconds for each new transaction. If the two calculations differ even slightly, say one counts refunds and the other does not, the model sees different inputs live than it learned from and quietly performs worse.
A feature store gives each feature one shared definition. The offline store keeps historical values, with timestamps, so a training set can be rebuilt as it would have looked on a past date without leaking future information. The online store keeps the latest value for each customer or card so a live service can fetch it fast. Teams can also reuse features someone else already built instead of rewriting them.
Feature stores belong mainly to classic predictive machine learning, the fraud, recommendation and pricing models built on tables of business data. Most generative AI applications built on large language models feed the model text and documents instead, so they rarely need one, though the same consistency problem returns whenever a prompt includes computed customer data. Feature stores are usually run as part of MLOps.
In episode 17 at 16:00, Galileo CTO and co-host Atindriyo Sanyal recalls the period “when feature stores were created and evangelized,” along with model monitoring, lineage and versioning, as the data-management side of AI. That remark is the co-host’s, not a guest’s; episode 17 was produced while Galileo sponsored the show (intro at 0:20).
Sources
- Feast documentation: Introduction — The open-source Feast project describes a feature store as helping teams define, manage, validate and serve features for production ML, built from an offline store for historical training data and an online store for low-latency serving; checked October 24, 2026.
- Google Cloud: About Feature Store — Official documentation for Google Cloud's managed feature store, which serves features for online prediction from data managed in BigQuery; checked October 24, 2026.
