Review low-use SageMaker endpoints and notebook instances

Why this is worth a look

The SageMaker endpoint Monitoring tab presents operational, invocation and health metrics. These include resource utilization, InvokeEndpoint activity, latency and endpoint failures, so review more than CPU alone. CloudWatch provides historical context, while notebook-instance pricing is based on the selected instance type and duration of use. A notebook LastModifiedTime records modification, not confirmed user activity.

A console-based inventory and monitoring checklist for existing endpoints and notebook instances. It makes no resource changes and uses existing billing records as a comparison input.

Execution surface: the SageMaker AI console, endpoint Monitoring tabs, CloudWatch metrics, organization records, and existing billing records. Assumptions: you can view SageMaker inventory, CloudWatch metrics, relevant tags or internal ownership records, and the billing data used for comparison. Access requirements are organization-specific; this checklist does not define one AWS permission set.

Use one governed Region at a time. Record the Region and choose a representative UTC review window. In the CloudWatch console, metric search returns metrics updated within the last two weeks; available historical data also depends on metric resolution and retention.

1. List endpoints with status InService. Record each endpoint name, creation time, last-modified time and status. List notebook instances with status InService and record each name, creation time, last-modified time and status. Record lifecycle configuration and code repository details for notebooks where shown. Add owner, environment and operating schedule from tags or internal records where available.
2. For each endpoint, open the Monitoring tab. Review the available operational metrics, including CPU Utilization and Memory Utilization, plus invocation metrics such as request activity, Model Latency and Invocation Model Errors, and health metrics such as Invocation Failures. Available metrics depend on endpoint type. Record displayed values, units and timestamps.
3. Compare the same Region and UTC window with existing billing data. State which billing records are included in the review. Do not infer user activity from a notebook LastModifiedTime.
4. Mark a resource as a low-use candidate for owner review only. Document availability, failover, latency, dependencies and recovery requirements before a separate change request.

How to confirm it

1. 01

   Inventory one Region at a time

   Use the SageMaker AI console in each governed Region. List endpoints and notebook instances with status InService, recording the supported names, statuses, creation times and last-modified times. Add owner, environment and schedule from tags or internal records where available.

2. 02

   Set a representative review window

   Choose and record a representative UTC window for the workload. CloudWatch console search finds metrics updated within the last two weeks, so confirm that the metric is discoverable and that the selected history is available.

3. 03

   Review endpoint monitoring signals

   Open each endpoint Monitoring tab and review the available operational, invocation and health metrics. CPU and memory are documented examples. Also review request activity, Model Latency, Invocation Model Errors and Invocation Failures where shown. Metric availability depends on endpoint type, so record displayed values, units and timestamps rather than assuming every metric exists.

4. 04

   Compare cost and escalate carefully

   Compare the same Region and UTC window with existing billing data, stating which billing records are included. Treat notebook LastModifiedTime as modification metadata, not user activity. Mark low-use resources for owner review only, then document availability, dependencies and recovery needs before any separate change request.

Primary sources