Vertex AI Metadata Store

A metadata store records ML lineage: artifacts, executions, and the relationships between them, billed per GB of metadata. Cost is minor, but the store documents how other expensive resources relate.

Lineage records billed by the gigabyte

A metadata store meters on stored metadata volume: the artifacts, executions, and contexts that ML Metadata accumulates as pipelines run, charged per GB. Set against endpoints and training jobs, the sums are trivial; even a busy pipeline history amounts to modest storage. The significance of this row is informational rather than financial.

Inventory plus an encryption check

ZopNight inventories metadata stores via Cloud Asset Inventory for complete ML topology and checks each one for customer-managed encryption keys (RC-1347). It does not read the lineage inside the store. Lineage answers questions cost tooling otherwise cannot: which pipeline produced the model an endpoint is serving, which datasets fed a given training run, and whether an expensive resource has any living downstream consumer, so check it in the console before acting on an idle-endpoint recommendation.

Metadata accumulation without ceremony

Even a cheap meter benefits from hygiene. Pipeline-heavy projects pile up execution records for experiments concluded long ago. Default stores collect lineage from every run in the project without anyone owning retention. And deleted pipelines leave artifact rows pointing at Cloud Storage objects that may themselves deserve cleanup, often the more valuable find, since those referenced artifacts bill at real storage rates while the metadata about them costs next to nothing. Pruning the graph is safe: it informs no running system. Worth knowing: a default store appears per project and region the first time a pipeline runs, so most Vertex projects carry one whether or not anyone chose to create it.

Metadata stores in the console

Google Cloud console → Vertex AI → Metadata opens the ML Metadata browser for the region’s store, where artifacts and executions can be traced through their relationships.