Open-data wants a multi-task job of notebooks where profiling finishes before transform starts. What should they create?

Notebook-based civic ETL can be orchestrated as a multi-task job with precedence constraints. Open-data workloads that require profiling to finish before transform should create a multi-task job where the profile task is an upstream dependency of the transform task. That gives a durable schedule, clear stage boundaries, and job UI visibility—while still allowing notebooks as the task implementations.

A multi-task job from notebooks with shared precedence so the profile task runs before the transform task fails by concept: shared precedence is not a real ordering construct—shared only describes reusing cluster resources across tasks instead of enforcing task order, so there is no dependency to ensure transform waits for profile to finish. A multi-task job from notebooks with arrival precedence so the profile task runs before the transform task likewise provides no guaranteed precedence: adding arrival to the name does not create a dependency—the arrival concept only starts the job when files arrive, so without an upstream dependency transform may still start before profile finishes. A multi-task job from notebooks with retry precedence so the profile task runs before the transform task confuses fault tolerance with ordering: adding retry to the name does not create a dependency—a retry concept only controls retries on failure and does not enforce that profile completes before transform starts.

Exam caveat: multi-task precedence orders tasks; it does not automatically version profile outputs—persist profile results to a table or volume if transform must read them. Operational check: run the job and confirm profile completes before transform starts in the run timeline; fail profile and confirm transform does not start; review open-data runbooks that cite the job DAG rather than shared precedence configuration.