Why it matters
Notebook instances are often left running after the work is done, and GPU instances in particular are expensive to keep on overnight.How nOps finds it
nOps reads Google Cloud billing for Workbench instance CPU, memory and GPU. Billing does not name the instance, so nOps groups by project, region and machine family. A group is flagged when it was billed in at least 95% of the hours of the last 14 complete days. The saving is its lowest typical hourly cost (the 10th-percentile hour) times 108 off-hours a week (nights and weekends), scaled to 30 days and capped at its cost. Disk is not counted, and savings under $5 a month are not shown.Before you start
Permission to update Workbench instances in the project.How to implement
- Ask the instance owners whether any long-running notebook or job needs it overnight or at weekends.
- In the Google Cloud console, open Vertex AI → Workbench → Instances, edit the instance, and turn on Idle shutdown with a timeout (for example 60 minutes).
- If you need a fixed schedule instead, stop instances with a Cloud Scheduler job.
- Set the same option in Terraform or your instance templates.
Risks and considerations
- Unsaved work in memory is lost when the instance stops. The boot and data disks are kept and still billed.
- The estimate assumes nobody uses the instances nights and weekends (108 of 168 hours).
- Because billing groups instances by machine family, check which instances in the project make up the cost.
- To roll back, turn idle shutdown off.
FAQs
Why does the recommendation not name a single instance?
Why does the recommendation not name a single instance?
Google Cloud billing for Workbench does not include an instance id, so nOps groups by project, region and machine family. The recommendation lists instance names when billing labels carry them.