Machine Learning compute instance with no automatic stop

Why it happens

A compute instance is a managed workstation and bills the underlying VM rate for every hour it is up, whether or not anybody is using it: Microsoft says you pay for the instance whenever it is running and to stop it when you are not using it. Two mechanisms can stop one on its own, an idle shutdown time and a start/stop schedule, and this rule fires only when neither will: no idle time is set, and there is no schedule whose action is Stop and whose status is not Disabled. A schedule that only ever starts the instance, or a stop schedule that is switched off, stops nothing and is treated as nothing. The rule reads the resource record rather than a utilisation metric because no Azure Monitor metric exists for this resource type at all; the CPU and GPU metrics that appear to fit are scoped to online endpoint deployments, which is inference serving rather than a workstation. That is also how Microsoft defines idleness for its own shutdown feature, which counts Jupyter kernels, terminals, runs, VS Code connections and custom applications, and never looks at CPU. The cost shown is what the instance costs to run continuously, priced from the ordinary virtual machine meter because Azure Machine Learning adds no surcharge to it. No saving is claimed, because how much of that time went unused is precisely what Azure publishes no metric for, and inventing it would be inventing the only number that matters. Stopping keeps the data: the workspace file share is mounted and persisted, the operating system disk survives a stop and start, and quota is not released. It does not stop the whole bill either, since disk, public IP and load balancer keep charging while the instance is stopped. Two documented conditions make idle shutdown silently fail even when the portal shows it enabled, a workspace managed identity without contributor access and a full disk, so an instance that appears protected may not be.

What the agent does about it

A running Azure Machine Learning compute instance with neither an idle shutdown time nor an enabled stop schedule, so it bills the full VM rate until somebody stops it by hand.