Vertex AI Featurestore (Legacy)

The legacy Vertex AI Featurestore provides managed feature storage and serving billed for online serving nodes and storage. Legacy featurestores forgotten during migrations to the newer architecture keep billing.

Two meters left running by migrations

A legacy featurestore charges on two axes: provisioned online-serving nodes, billed for every hour they stand regardless of read volume, and the feature data stored inside the service. Neither meter cares that the team’s attention has moved on. The type is superseded by feature groups and feature online stores, the current-generation Feature Store built on BigQuery. That is exactly why forgotten legacy stores are so common: the migration succeeds, the new stack takes traffic, and the old store keeps its nodes warm for nobody.

Flagging superseded feature infrastructure

ZopDev inventories legacy featurestores via Cloud Asset Inventory and highlights them as migration and cleanup candidates. The framing is deliberate: for a deprecated-generation service the useful recommendation is rarely rightsizing; it is confirming the migration actually finished and retiring the old store outright. The type carries no stop verb; teardown is the whole play.

Migration leftovers with familiar shapes

The archetypal finding is a featurestore whose serving nodes still run for an application that switched to the BigQuery-backed store months earlier. Close behind: feature data retained in full for entity types no model reads anymore, and dev or staging featurestores cloned from production sizing that nobody revisited. Because serving nodes are provisioned rather than demand-driven, none of these shrink on their own. Every one persists until a human deletes it.

Locating legacy stores in the console

Google Cloud console → Vertex AI → Feature Store separates the legacy featurestores from the newer feature groups and online stores. Each legacy store shows its online-serving node configuration, the number to check when confirming a migration really ended the spend. Deleting the store retires both of its meters at once.