Amazon SageMaker Notebook Instance
Does ZopNight manage Amazon SageMaker Notebook Instance?
SageMaker notebook instances bill per hour from start to stop, and unlike endpoints they have a real stop verb that preserves the instance and its EBS volume. ZopNight discovers notebooks on the 6-hour cycle, stops them overnight via sagemaker:StopNotebookInstance, and restarts them with sagemaker:StartNotebookInstance for working hours.
Rules that fire on Amazon SageMaker Notebook Instance
At a glance
| Field | Value |
|---|---|
| Stop verb | sagemaker:StopNotebookInstance |
| Scheduling notes | notebook stopped via sagemaker:StopNotebookInstance and started via sagemaker:StartNotebookInstance on schedule. |
A SageMaker notebook instance is a managed Jupyter environment on a dedicated ML instance, billed per hour while running. Notebooks left running overnight and on weekends are one of the most common ML cost leaks.
A workstation meter with an off switch
Notebook instances bill per hour at their instance type’s rate while in service, plus per-GB-month for the attached EBS volume. Crucially, and unlike real-time endpoints, the stop verb exists and works: a stopped notebook stops its compute meter entirely, keeps its volume (and the small storage charge), and restarts with all files intact. That makes the notebook the most fixable resource in the SageMaker family. The waste is purely behavioural: an interactive tool used during working hours, billing around the clock because closing the browser does not stop the instance.
Scheduling the working week
ZopNight discovers notebook instances automatically on the 6-hour cycle, collects hourly CloudWatch metrics with a 90-day lookback, and attributes per-notebook cost from Cost Explorer or CUR 2.0. Notebooks are fully schedulable: sagemaker:StopNotebookInstance shuts them down on the configured schedule and sagemaker:StartNotebookInstance brings them back for working hours, converting a 168-hour billing week into roughly a 45-hour one without anyone losing work. Idle-notebook recommendations catch the deeper cases: instances whose usage metrics show nobody has opened them in weeks, where the fix is deletion rather than a schedule.
How notebooks burn nights
The failure mode needs no malice: a notebook is started for a project, the project’s active phase ends, and the instance keeps billing 24/7 through the review period, the vacation, and the reorg. GPU notebook instances sharpen it. An accelerated instance chosen for one training experiment continues at GPU rates for months of light pandas work. Volume sizes creep too, since enlarging is easy and shrinking requires migration.
The notebook fleet at a glance
The SageMaker console’s Notebook instances view shows each instance’s status, type, and volume size. Status InService outside business hours is the audit in one filter, and everything green at midnight is either a scheduled training run or money.