An end-to-end Databricks Asset Bundle for a data-science forecasting demonstration using
samples.tpcds_sf1.store_sales:
- exploratory analysis and governed training-table preparation;
- time-based comparison of naive, seasonal-naive, ETS, and ARIMA forecasts;
- MLflow parent/child runs with metrics, artifacts, and backtest predictions;
- champion registration in the Unity Catalog model registry with a movable alias;
- independently rerunnable batch scoring into a governed Delta table.
All DAB job tasks use serverless compute. Databricks' separate Forecasting (serverless) with
AutoML Public Preview is currently started from the UI, so the automated bundle compares the same
family of open-source statistical models directly. The UC model is registered by MLflow during the training task; the bundle provider's
registered_models resource targets the legacy workspace registry and cannot accept a three-part UC
model name.
Prerequisites: Databricks CLI authentication to the customer's workspace, read access to the
samples catalog, and create privileges in a Unity Catalog catalog chosen for the demo.
databricks bundle validate --strict -t dev --var="catalog=<customer_catalog>"
databricks bundle deploy -t dev --var="catalog=<customer_catalog>"
databricks bundle run forecasting_lifecycle -t dev --var="catalog=<customer_catalog>"Run only the deployed batch inference job after a champion exists:
databricks bundle run forecasting_batch_inference -t dev --var="catalog=<customer_catalog>"The schema defaults to forecasting_demo_<deployer> and can be overridden along with the
forecast horizon or source dataset:
databricks bundle deploy -t dev \
--var="catalog=<customer_catalog>,schema=<customer_schema>,forecast_horizon=30"No notebook widgets are used. Jobs receive DAB variables through a small serverless configuration task. When a notebook is run directly, it uses the current UC catalog and a schema derived from the signed-in user.
The tested workspace runs, model metrics, and batch-output checks are recorded in
docs/VALIDATION.md.