Review quiet SageMaker capacity before cutting it
On this page
Review quiet SageMaker capacity before requesting changes, because stopping a needed endpoint or notebook can interrupt work. Use low activity as a review signal, not approval to stop a resource.
What you need first
- Tool
- The AWS Command Line Interface (AWS CLI) runs AWS commands from a terminal. Start in a terminal configured for the intended account. For endpoint metrics, use the Monitoring tab on the SageMaker AI console endpoint details page.
- Access
- Ask an authorized AWS colleague for read-only endpoint and notebook lists for the target account and Regions, plus endpoint metrics for the review dates. Ask finance for billing records covering the same scope.
- If you do not use that tool
- Send your AWS colleague the account, Regions and review dates. Request the command results below and Monitoring results with metric values, units, Period and Statistic. Ask resource owners to confirm notebook use.
Why this is worth a look
Skipping owner confirmation can interrupt required work, while low CPU alone can miss request activity or service problems. CloudWatch provides endpoint utilization, invocation, latency and health metrics, and notebook inventory does not prove human use. Compare endpoint signals and ask notebook owners before seeking changes.
Run these commands
BASHRun these read-only AWS CLI commands once per intended Region. They return current inventory, not historical usage, instance-hours or costs.
#!/usr/bin/env bash
set -euo pipefail
# Replace both placeholders with the intended AWS Region.
# Use a terminal configured for the intended account.
aws sagemaker list-endpoints --region "REPLACE_WITH_REGION" --status-equals InService
aws sagemaker list-notebook-instances --region "REPLACE_WITH_REGION"How to confirm it
- 01
Collect the scoped inventory
Record the account and Regions, then run the commands once per Region or request the results from your AWS colleague. Keep endpoint names and notebook entries with status InService for review. Leave pagination enabled and do not add --max-items, which can limit returned inventory.
- 02
Compare demand and health
Choose dates that cover the expected operating schedule. Open each endpoint's Monitoring tab and compare request activity with CPU Utilization, Memory Utilization, Model Latency, Invocation Model Errors and Invocation Failures where available. Record dates, values, units, Period and Statistic using the same settings.
- 03
Confirm use and charges
Ask notebook owners to confirm use during those dates. Ask finance to compare reviewed resources with billing records for the same scope and window, and record which charges are included. Do not turn low metric values into a savings estimate.
- 04
Get an owner decision
Mark low-use resources as review candidates. Ask owners whether availability, failover, latency, dependencies or recovery needs require the current capacity. Request changes separately only after those needs are resolved.
Before making changes
Missing metrics do not prove inactivity. Available metrics depend on endpoint type, and historical detail depends on metric resolution and retention. This review assumes the dates represent normal demand and owners can confirm notebook use. It covers InService endpoints and notebook instances selected for review, not all SageMaker resources, and does not calculate idle hours or savings.