What is jobs compute?

The lifecycle is the point: the cluster exists because a job needs it, and disappears when the job ends. No idle window, no one forgetting to shut it down, and a DBU rate priced accordingly.

The costly anti-pattern is the reverse: scheduled production work running on an all-purpose cluster, paying an interactive premium of roughly 3–4× for a workload no human is interacting with. It usually starts innocently (a notebook developed interactively gets scheduled as-is) and persists because nothing breaks. The bill is the only symptom.

Try this: list your scheduled jobs and confirm each one runs on jobs compute (or serverless jobs). Every one that doesn’t is paying the all-purpose rate for batch work.