Google Vertex AI Feature Store system Dag exceeds import timeout

Start with the Vertex AI Feature Store system DAG and the import behavior described in airflow-core/tests/unit/always/test_example_dags.py, running the provided Breeze pytest command first. Trace the module-scope google.cloud.aiplatform_v1beta1 and related heavy imports, then defer them until task execution. Done means the system DAG imports within the default timeout without loading that root package and passes the stated import, database-query, and connection-access checks.

Under which category would you file this issue?

Providers

Apache Airflow version

main (development)

What happened and how to reproduce it?

The Google provider's Vertex AI Feature Store system Dag imports google.cloud.aiplatform_v1beta1 at module scope. Importing that package triggers an expensive third-party import chain while Airflow scans the Dag file.

In a clean Breeze process, importing the unmodified system Dag took 30.505 seconds and loaded google.cloud.aiplatform_v1beta1. This exceeds Airflow's default dagbag_import_timeout of 30 seconds and can surface as a Dag import timeout or a third-party circular-import failure.

Steps to reproduce:

1. Check out main with the Google provider development dependencies installed.

2. Run the Vertex AI Feature Store system Dag import test in Breeze:

   BACKEND=sqlite breeze run pytest airflow-core/tests/unit/always/test_example_dags.py \
       -k 'test_should_be_importable and vertex_ai_feature_store' -xvs
   

3. Observe that parsing the Dag loads google.cloud.aiplatform_v1beta1 at module scope and may exceed the 30-second Dag import timeout.

What you think should happen instead?

Scanning the system Dag should not load the Vertex AI Feature Store client package. Heavy Google clients and protobuf-related imports should be deferred until task execution so the Dag remains importable within the default timeout.

Operating System

macOS host with the Apache Airflow Breeze CI container, Python 3.10.

Deployment

Other Docker-based deployment

Apache Airflow Provider(s)

No response

Versions of Apache Airflow Providers

No response

Official Helm Chart version

Not Applicable

Kubernetes Version

No response

Helm Chart configuration

No response

Docker Image customizations

No response

Anything else?

After removing the parse-time aiplatform_v1beta1 imports, the system Dag imports without loading that root package and passes Airflow's generic system Dag import, database-query, and connection-access checks.

Are you willing to submit PR?

• Yes I am willing to submit a PR!

Code of Conduct

• I agree to follow this project's Code of Conduct