Vertex AI Feature Online Store

A feature online store serves feature values at low latency for online predictions, billed for provisioned serving nodes. Like any always-on serving tier, it bills whether or not models query it.

Feature serving that bills through quiet hours

An online store holds provisioned serving capacity so predictions can fetch fresh feature values at request time. That capacity bills for every hour it stands, and lookup volume never factors in: a store feeding a production recommender and a store nobody queries pay the same for equal provisioning. Data synced in from feature groups adds a storage component alongside the node charges.

Catching stores nothing reads

ZopDev discovers online stores via Cloud Asset Inventory and flags stores with no serving traffic as idle spend. The no-traffic signal is decisive for this type because provisioning is a manual, sticky choice. A store stays exactly as large as someone once made it until someone else notices. There is no stop verb here; reclaiming the spend means downsizing the store or deleting it once its readers are confirmed gone.

Provisioning inherited from launch day

The recurring shapes: an online store stood up for an experiment whose model never got promoted to production; capacity sized for a seasonal or launch peak that has long passed; and a store still syncing feature views for an endpoint that was undeployed months ago. Each keeps paying node-hours for readiness nothing requires, and the sync jobs compound it by continuing to refresh values no prediction will ever read.

Online stores in the console

Google Cloud console → Vertex AI → Feature Store, in the online stores view, shows each store’s serving configuration and the feature views synced into it. An online store whose feature views have no consuming models left is the cue to reclaim it. Downsizing loses no data, because the source of truth for every feature value stays in BigQuery.