Azure ML Compute Instance
Does ZopNight manage Azure ML Compute Instance?
Azure ML compute instances bill per hour for their VM size, GPU SKUs especially, from start until someone stops them, and the platform never stops them on its own. ZopNight discovers each instance under its workspace, flags ones running outside working hours, and stops and starts them on schedule.
Rules that fire on Azure ML Compute Instance
No active rule family targets Azure ML Compute Instance today. Rules that used to are retired, and retired rules publish no pages and fire no findings. Scheduling and permissions coverage are unaffected.
At a glance
| Field | Value |
|---|---|
| Scheduling notes | stopped via the ML compute stop operation and started on schedule, halting VM billing while stopped. |
ML compute instances are single-user development VMs for data scientists, billed per hour while running. Notebooks left running after the workday are one of the most common ML cost leaks.
One person, one VM, one always-running meter
A compute instance is a dedicated VM assigned to a single data scientist, and it bills for that VM every hour it is running regardless of whether a notebook cell has executed in days. Because ML development gravitates toward GPU sizes, the hourly rate is often a multiple of an ordinary dev VM’s. Nothing in the platform stops the machine when its owner logs off. The meter simply continues until a stop is issued.
Where these instances surface in ZopNight
Discovered via the AML enricher under the parent workspace. Cost Management billing attributes VM spend to each instance, recommendations flag instances running outside working hours, and schedules stop them automatically. Since each instance belongs to one person, a working-hours schedule maps cleanly onto how the machine is actually used.
Stop semantics for notebook VMs
ZopNight stops an instance through the ML compute stop operation, which halts VM billing while stopped, and starts it again on schedule. State on the OS disk persists across the cycle, so a stopped notebook resumes where its owner left it. The trade is a short start-up wait in the morning against every night and weekend of VM hours recovered.
The after-hours GPU habit
The dominant leak is simple: a GPU instance started for an experiment on Tuesday still running the following Monday. Its close cousins are the instance kept up “so the environment stays warm” and the departed teammate’s machine nobody claimed. A 45-hour working week uses about a quarter of the 168 hours an unscheduled instance bills.
Spotting running instances in ML studio
Azure ML studio → Compute → Compute instances lists each instance with its state, size, and assigned user. Every row showing Running outside office hours is a candidate.