Choose a project shape

Three questions decide most of a new Databricks project. Answer them before databricks bundle init; each answer links to the page that owns the details. The closest NYC taxi example is named where one exists.

Single model, competition or model set?

You have	Choose	Closest example	Details
One target and one model you already picked	single_model	taxi_fare	Training configuration
One target and 2-8 candidate algorithms to compare	model_competition: CV picks one winner	taxi_high_fare	Training configuration
2-8 targets approved and scored together	multi_target: one versioned model set	taxi_trips	Model sets
One candidate that combines several learners	A voting or stacking model in any layout	none	Ensembles

A competition winner that fails its holdout gate does not promote the runner-up. The layout shapes the generated train job, so choose it at bundle init; see layouts and setup modes.

Local batch, Spark batch, REST or ai_query?

Who reads the predictions	Choose	Closest example	Details
A scheduled job writing a Delta table within max_rows/max_input_mb	score job, batch_compute: single_node (default)	taxi_high_fare	Scoring
A scheduled job on a table too large for one Python process	score job, batch_compute: spark	taxi_fare	Spark scoring
An application sending raw feature values	REST serving endpoint	taxi_fare_endpoint	Serving
SQL users, SQL jobs and dashboards	SQL function calling the endpoint with ai_query	taxi_fare_endpoint	SQL function
Requests that send only an entity key	Endpoint with online feature lookup	taxi_fare_online	Online features

The batch job never calls an endpoint, and an endpoint never writes the prediction table. Models packaged with feature lookup score in batch only with batch_compute: spark. Runtime and cost per path are compared in how predictions are produced.

For rows excluded by preprocessing at scoring, start with taxi_fare_robust. Its Spark and raw-input endpoint examples retain row identity and explain excluded status, null predictions and the resulting monitoring coverage.

For endpoint delivery without a batch prediction table, copy taxi_fare_endpoint. Its serving job follows champion and enrolls the endpoint's telemetry logs for monitoring. Its guide includes a SQL request and the Online serving dashboard.

Which steps run on Spark workers and endpoints?

Spark workers and serving endpoints both run the saved pandas pipeline on independent batches, so the same admission applies to both.

Part	On Spark workers and endpoints	Details
Training, tuning and CV	Never distributed: one bounded pandas or Polars process	Spark scoring
Built-in models	Reviewed linear/logistic/tree models and XGBRegressor; additional models and voting/stacking with matching training probe evidence	Support table
Built-in preprocessing steps	Reviewed steps (including median imputers), plus eligible row-preserving recipes with matching training probe evidence	Recipe boundaries
Your column_step and fitted_step functions, saved scoring rules	Admitted when training lints and probes them	When the probe fails
Steps that need other rows (group, window, global), custom_step classes	Rejected; keep batch_compute: single_node	Custom steps

A single-model or competition Spark project rejects an unsupported model or step before registration, so the failure appears in the train job, not in the first score run. A model set is checked when it is activated and again at scoring.