SageMaker endpoints averaging at most one invocation a day that could run on a smaller instance

A trickle of requests on a full-size instance

A real-time endpoint keeps its instances running whether it receives a thousand requests a minute or one a day. The Invocations metric in the AWS/SageMaker namespace counts InvokeEndpoint requests, and the metrics reference says to read it with the Sum statistic. An endpoint whose daily sum is in single digits is running an instance sized for traffic that is not there.

Counting invocations per day

aws cloudwatch get-metric-statistics --namespace AWS/SageMaker --metric-name Invocations \
  --dimensions Name=EndpointName,Value=my-endpoint Name=VariantName,Value=AllTraffic \
  --start-time 2026-08-26T00:00:00Z --end-time 2026-09-25T00:00:00Z \
  --period 86400 --statistics Sum

aws sagemaker describe-endpoint-config --endpoint-config-name my-endpoint-config \
  --query 'ProductionVariants[].[VariantName,InstanceType,InitialInstanceCount]'

The low-but-not-zero band

1. The endpoint is InService and not covered by a Savings Plan or reservation, since on-demand rate differences overstate savings on committed capacity.
2. At least 7 days of invocation data exist in the metric history ZopNight holds.
3. Total invocations over the last 7 days, divided by 7, is above 0 and at most 1.
4. A concrete one-size-smaller ml.* instance exists, and ZopNight has hourly rates for both types.
5. The endpoint has a positive monthly cost.

If invocation data is missing or covers fewer than 7 days, ZopNight accepts an invocations_low=true tag on the endpoint in place of steps 2 and 3.

Endpoints outside the band

An endpoint with exactly zero invocations is not a downsize case; it belongs to Idle SageMaker Endpoint, No Invocations, and the lower bound keeps the two rules from giving opposite advice on one endpoint. GPU instance types, the smallest size of a family, and types outside ml.* naming produce nothing. The same applies when a rate is missing or the smaller type is not cheaper.

Rate-ratio saving

saving = monthly endpoint cost x (1 - smaller type rate / current type rate)

There is no fallback percentage: without both rates there is no figure and no finding.

Options for a rarely used endpoint

1. Resize the variant to the smaller instance by creating a new endpoint configuration and calling aws sagemaker update-endpoint. ZopNight can also apply this resize from the recommendation.
2. For traffic this sparse, consider Serverless Inference, which AWS describes as suited to workloads with idle periods between traffic spurts that can tolerate cold starts.
3. Check ModelLatency after the change to confirm the smaller instance still meets its target.