dce_001 · Deployment E2E: real GCS model + real BQ + real execute_live_refresh

Status: ✓ Passing · Last run: 2026-06-09T06:39:15 · Pipeline: deployment_e2e · History: 1/4 runs passed (25%)

What was tested

The actual production deployment path, exercised end-to-end:

1. Train a real sklearn LogisticRegression on synthetic history. 2. joblib.dump the trained model to a real GCS blob in gs://auto-insight-api/validation_e2e/<case>/model.pkl. 3. Upload new-data rows to a real BigQuery table. 4. Build a real DeploymentSpec with one Python section whose transformed_code downloads the GCS model, queries BQ, calls predict_proba, returns a scored DataFrame. 5. Invoke deployment_composer.skills.meta.execute_live_refresh. run_sections_capture_outputs — the same function the Cloud Scheduler hits at /projects/<id>/deployment/render. 6. Assert the production runtime returned a scored DataFrame with no section errors and predictions in [0, 1].

This is the only validation lane that proves the GCS round-trip works, that exec() with the curated globals dict actually loads a pickle, and that the BQ client integration is intact end-to-end.

What we planted in the data

• 500 entities
• Signal strength: 0.50 (sigmoid slope multiplier)
• n_new_rows = 50

What the system did

• Training rows built: 500
• Features used in model: 3
• Model selected: DeploymentE2ERunner

What we checked — all assertions passed

How this could have gone wrong

(no assertion descriptions provided)

Why this case matters

This is the hand-on-the-stove test. If the production runtime's pickle.load path breaks, or the bq global is no longer injected, or the GCS download is misconfigured, THIS case fires — before Yifan's daily refresh stops working.

Reproducing

# from auto_insight_api/
python -m validation.v2 run dce_001 --pipeline deployment_e2e -v

• Case config: validation/v2/cases/dce_001_e2e_happy.yaml
• Data shape: deployment_e2e_happy
• Analytics type: deployment_e2e

← back to validation report