Collaborate within and across teams to manage data and models
The collaboration surfaces
Data. BigQuery datasets with IAM at project/dataset/table level; authorized views and Analytics Hub for sharing without copying; policy tags (Dataplex/Data Catalog taxonomies) for column-level security; row-level security for filtering. Dataplex provides lakes/zones, discovery, quality and lineage across BigQuery, GCS and other sources.
Models. The Vertex AI Model Registry holds model versions, aliases (e.g. default, production) and parent/child lineage to the pipeline run and artifacts that produced them. Vertex AI Experiments and ML Metadata record parameters, metrics and artifacts. Feature Registry / Feature Store shares feature definitions so two teams do not recompute the same feature differently.
Environments. Teams typically separate dev/test/prod projects, connected by a host project with Shared VPC, cross-project IAM, and shared notebooks (Colab Enterprise, Vertex AI Workbench) or repositories.
Traps
- Model Registry ≠ Model Garden. Registry stores your models; Garden is a catalogue of third-party and foundation models.
- Aliases vs versions. Deployments and CI reference an alias so promotion does not require redeploying a new endpoint; questions often hinge on this.
- Lineage is automatic only inside Vertex AI Pipelines. Notebook training leaves no lineage unless you log it to Experiments.
- Policy tags are not IAM roles - they are fine-grained access control applied through Dataplex on BigQuery columns.
- Authorized view vs granting dataset access: the view exposes only selected rows/columns, the grant exposes everything.
Rule of thumb
Metadata should flow one way - artifact → model version → endpoint - and the only way to get that for free is to run every step as a Vertex AI Pipelines component.
Worked example: Team A trains in project ml-dev, Team B serves in ml-prod. Promote by registering the model in the production project's registry (or importing it), tagging the version with the production alias, granting Team B roles/aiplatform.user, and deploying by alias.