Metrics for monitoring Amazon SageMaker with Amazon CloudWatch

You can monitor Amazon SageMaker using Amazon CloudWatch, which collects raw data and processes it into readable, near real-time metrics. These statistics are kept for 15 months. With them, you can access historical information and gain a better perspective on how your web application or service is performing. However, the Amazon CloudWatch console limits the search to metrics that were updated in the last 2 weeks. This limitation ensures that the most current jobs are shown in your namespace.

To graph metrics without using a search, specify its exact name in the source view. You can also set alarms that watch for certain thresholds, and send notifications or take actions when those thresholds are met. For more information, see the Amazon CloudWatch User Guide.

SageMaker endpoint invocation metrics

The AWS/SageMaker namespace includes the following request metrics from calls to InvokeEndpoint.

Metrics are available at a 1-minute frequency.

The following illustration shows how a SageMaker endpoint interacts with the Amazon SageMaker Runtime API. The overall time between sending a request to an endpoint and receiving a response depends on the following three components.

• Network latency – the time that it takes between making a request to and receiving a response back from the SageMaker Runtime Runtime API.

• Overhead latency – the time that it takes to transport a request to the model container from and transport the response back to the SageMaker Runtime Runtime API.

• Model latency – the time that it takes the model container to process the request and return a response.

For more information about total latency, see Best practices for load testing Amazon SageMaker real-time inference endpoints. For information about how long CloudWatch metrics are retained for, see GetMetricStatistics in the Amazon CloudWatch API Reference.

Endpoint Invocation Metrics

Metric	Description
ConcurrentRequestsPerCopy	The number of concurrent requests being received by the inference component, normalized by each copy of an inference component. Valid statistics: Min, Max
ConcurrentRequestsPerModel	The number of concurrent requests being received by the model. Valid statistics: Min, Max
Invocation4XXErrors	The number of InvokeEndpoint requests where the model returned a 4xx HTTP response code. For each 4xx response, 1 is sent; otherwise, 0 is sent. Units: None Valid statistics: Average, Sum
Invocation5XXErrors	The number of InvokeEndpoint requests where the model returned a 5xx HTTP response code. For each 5xx response, 1 is sent; otherwise, 0 is sent. Units: None Valid statistics: Average, Sum
InvocationModelErrors	The number of model invocation requests that did not result in 2XX HTTP response. This includes 4XX/5XX status codes, low-level socket errors, malformed HTTP responses, and request timeouts. For each error response, 1 is sent; otherwise, 0 is sent. Units: None Valid statistics: Average, Sum
Invocations	The number of InvokeEndpoint requests sent to a model endpoint. To get the total number of requests sent to a model endpoint, use the Sum statistic. Units: None Valid statistics: Sum
InvocationsPerCopy	The number of invocations normalized by each copy of an inference component. Valid statistics: Sum
InvocationsPerInstance	The number of invocations sent to a model, normalized by InstanceCount in each ProductionVariant. 1/numberOfInstances is sent as the value on each request. numberOfInstances is the number of active instances for the ProductionVariant behind the endpoint at the time of the request. Units: None Valid statistics: Sum
ModelLatency	The interval of time taken by a model to respond to a SageMaker Runtime API request. This interval includes the local communication times taken to send the request and to fetch the response from the model container. It also includes the time taken to complete the inference in the container. Units: Microseconds Valid statistics: Average, Sum, Min, Max, Sample Count
ModelSetupTime	The time it takes to launch new compute resources for a serverless endpoint. The time can vary depending on the model size, how long it takes to download the model, and the start-up time of the container. Units: Microseconds Valid statistics: Average, Min, Max, Sample Count, Percentiles
OverheadLatency	The interval of time added to the time taken to respond to a client request by SageMaker overheads. This interval is measured from the time SageMaker receives the request until it returns a response to the client, minus the ModelLatency. Overhead latency can vary depending on multiple factors, including request and response payload sizes, request frequency, and authentication/authorization of the request. Units: Microseconds Valid statistics: Average, Sum, Min, Max, Sample Count

Dimensions for Endpoint Invocation Metrics

Dimension	Description
EndpointName, VariantName	Filters endpoint invocation metrics for a ProductionVariant of the specified endpoint and variant.
InferenceComponentName	Filters inference component invocation metrics.

SageMaker inference component metrics

The /aws/sagemaker/InferenceComponents namespace includes the following metrics from calls to InvokeEndpoint for endpoints that host inference components.

Metrics are available at a 1-minute frequency.

Metric	Description
CPUUtilizationNormalized	The value of the CPUUtilizationNormalized metric reported by each copy of the inference component. The value ranges between 0%–100%. If you set the NumberOfCpuCoresRequired parameter in the settings for the inference component copy, the metric presents the utilization over the reservation. Otherwise, the metric presents the utilization over the limit.
GPUMemoryUtilizationNormalized	The value of the GPUMemoryUtilizationNormalized metric reported by each copy of the inference component.
GPUUtilizationNormalized	The value of the GPUUtilizationNormalized metric reported by each copy of the inference component. If you set the NumberOfAcceleratorDevicesRequired parameter in the settings for the inference component copy, the metric presents the utilization over the reservation. Otherwise, the metric presents the utilization over the limit.
MemoryUtilizationNormalized	The value of MemoryUtilizationNormalized reported by each copy of the inference component. If you set the MinMemoryRequiredInMb parameter in the settings for the inference component copy, the metrics present the utilization over the reservation. Otherwise, the metrics present the utilization over the limit.

Dimensions for Inference Component Metrics

Dimension	Description
InferenceComponentName	Filters inference component metrics.

SageMaker multi-model endpoint metrics

The AWS/SageMaker namespace includes the following model loading metrics from calls to InvokeEndpoint.

Metrics are available at a 1-minute frequency.

For information about how long CloudWatch metrics are retained for, see GetMetricStatistics in the Amazon CloudWatch API Reference.

Multi-Model Endpoint Model Loading Metrics

Metric	Description
ModelLoadingWaitTime	The interval of time that an invocation request has waited for the target model to be downloaded, loaded, or both in order to run inference. Units: Microseconds Valid statistics: Average, Sum, Min, Max, Sample Count
ModelUnloadingTime	The interval of time that it took to unload the model through the container's UnloadModel API call. Units: Microseconds Valid statistics: Average, Sum, Min, Max, Sample Count
ModelDownloadingTime	The interval of time that it took to download the model from Amazon Simple Storage Service (Amazon S3). Units: Microseconds Valid statistics: Average, Sum, Min, Max, Sample Count
ModelLoadingTime	The interval of time that it took to load the model through the container's LoadModel API call. Units: Microseconds Valid statistics: Average, Sum, Min, Max, Sample Count
ModelCacheHit	The number of InvokeEndpoint requests sent to the multi-model endpoint for which the model was already loaded. The Average statistic shows the ratio of requests for which the model was already loaded. Units: None Valid statistics: Average, Sum, Sample Count

Dimensions for Multi-Model Endpoint Model Loading Metrics

Dimension	Description
EndpointName, VariantName	Filters endpoint invocation metrics for a ProductionVariant of the specified endpoint and variant.

The /aws/sagemaker/Endpoints namespaces include the following instance metrics from calls to InvokeEndpoint.

Metrics are available at a 1-minute frequency.

For information about how long CloudWatch metrics are retained for, see GetMetricStatistics in the Amazon CloudWatch API Reference.

Multi-Model Endpoint Model Instance Metrics

Metric

Description

LoadedModelCount

The number of models loaded in the containers of the multi-model endpoint. This metric is emitted per instance. The Average statistic with a period of 1 minute tells you the average number of models loaded per instance. The Sum statistic tells you the total number of models loaded across all instances in the endpoint. The models that this metric tracks are not necessarily unique because a model might be loaded in multiple containers at the endpoint. Units: None Valid statistics: Average, Sum, Min, Max, Sample Count

Dimensions for Multi-Model Endpoint Model Loading Metrics

Dimension	Description
EndpointName, VariantName	Filters endpoint invocation metrics for a ProductionVariant of the specified endpoint and variant.

SageMaker jobs and endpoint metrics

The /aws/sagemaker/ProcessingJobs, /aws/sagemaker/TrainingJobs, /aws/sagemaker/TransformJobs, and /aws/sagemaker/Endpoints namespaces include the following metrics for training jobs and endpoint instances.

Metrics are available at a 1-minute frequency.

Note

Amazon CloudWatch supports high-resolution custom metrics and its finest resolution is 1 second. However, the finer the resolution, the shorter the lifespan of the CloudWatch metrics. For the 1-second frequency resolution, the CloudWatch metrics are available for 3 hours. For more information about the resolution and the lifespan of the CloudWatch metrics, see GetMetricStatistics in the Amazon CloudWatch API Reference.

Tip

To profile your training job with a finer resolution down to 100-millisecond (0.1 second) granularity and store the training metrics indefinitely in Amazon S3 for custom analysis at any time, consider using Amazon SageMaker Debugger. SageMaker Debugger provides built-in rules to automatically detect common training issues. It detects hardware resource utilization issues (such as CPU, GPU, and I/O bottlenecks). It also detects non-converging model issues (such as overfit, vanishing gradients, and exploding tensors). SageMaker Debugger also provides visualizations through Studio Classic and its profiling report. To explore the Debugger visualizations, see SageMaker Debugger Insights Dashboard Walkthrough, Debugger Profiling Report Walkthrough, and Analyze Data Using the SMDebug Client Library.

Processing Job, Training Job, Batch Transform Job, and Endpoint Instance Metrics

Metric	Description
CPUReservation	The sum of CPUs reserved by containers on an instance. The value ranges between 0%–100%. In the settings for an inference component, you set the CPU reservation with the NumberOfCpuCoresRequired parameter. For example, if there 4 CPUs, and 2 are reserved, the CPUReservation metric is 50%.
CPUUtilization	The sum of each individual CPU core's utilization. The CPU utilization of each core range is 0–100. For example, if there are four CPUs, the CPUUtilization range is 0%–400%. For processing jobs, the value is the CPU utilization of the processing container on the instance. For training jobs, the value is the CPU utilization of the algorithm container on the instance. For batch transform jobs, the value is the CPU utilization of the transform container on the instance. For endpoint variants, the value is the sum of the CPU utilization of the primary and supplementary containers on the instance. Note For multi-instance jobs, each instance reports CPU utilization metrics. However, the default view in CloudWatch shows the average CPU utilization across all instances. Units: Percent
CPUUtilizationNormalized	The normalized sum of the utilization of each individual CPU core. The value ranges between 0%–100%. For example, if there are four CPUs, and the CPUUtilization metric is 200%, then the CPUUtilizationNormalized metric is 50%.
DiskUtilization	The percentage of disk space used by the containers on an instance uses. This value range is 0%–100%. This metric is not supported for batch transform jobs. For processing jobs, the value is the disk space utilization of the processing container on the instance. For training jobs, the value is the disk space utilization of the algorithm container on the instance. For endpoint variants, the value is the sum of the disk space utilization of the primary and supplementary containers on the instance. Units: Percent Note For multi-instance jobs, each instance reports disk utilization metrics. However, the default view in CloudWatch shows the average disk utilization across all instances.
GPUMemoryUtilization	The percentage of GPU memory used by the containers on an instance. The value range is 0–100 and is multiplied by the number of GPUs. For example, if there are four GPUs, the GPUMemoryUtilization range is 0%–400%. For processing jobs, the value is the GPU memory utilization of the processing container on the instance. For training jobs, the value is the GPU memory utilization of the algorithm container on the instance. For batch transform jobs, the value is the GPU memory utilization of the transform container on the instance. For endpoint variants, the value is the sum of the GPU memory utilization of the primary and supplementary containers on the instance. Note For multi-instance jobs, each instance reports GPU memory utilization metrics. However, the default view in CloudWatch shows the average GPU memory utilization across all instances. Units: Percent
GPUMemoryUtilizationNormalized	The normalized percentage of GPU memory used by the containers on an instance. The value ranges between 0%–100%. For example, if there are four GPUs, and the GPUMemoryUtilization metric is 200%, then the GPUMemoryUtilizationNormalized metric is 50%.
GPUReservation	The sum of GPUs reserved by containers on an instance. The value ranges between 0%–100%. In the settings for an inference component, you set the GPU reservation by NumberOfAcceleratorDevicesRequired. For example, if there are 4 GPUs and 2 are reserved, the GPUReservation metric is 50%.
GPUUtilization	The percentage of GPU units that are used by the containers on an instance. The value can range between 0–100 and is multiplied by the number of GPUs. For example, if there are four GPUs, the GPUUtilization range is 0%–400%. For processing jobs, the value is the GPU utilization of the processing container on the instance. For training jobs, the value is the GPU utilization of the algorithm container on the instance. For batch transform jobs, the value is the GPU utilization of the transform container on the instance. For endpoint variants, the value is the sum of the GPU utilization of the primary and supplementary containers on the instance. Note For multi-instance jobs, each instance reports GPU utilization metrics. However, the default view in CloudWatch shows the average GPU utilization across all instances. Units: Percent
GPUUtilizationNormalized	The normalized percentage of GPU units that are used by the containers on an instance. The value ranges between 0%–100%. For example, if there are four GPUs, and the GPUUtilization metric is 200%, then the GPUUtilizationNormalized metric is 50%.
MemoryReservation	The sum of memory reserved by containers on an instance. The value ranges between 0%–100%. In the settings for an inference component, you set the memory reservation with the MinMemoryRequiredInMb parameter. For example, if a 32 GiB instance reserved 1024 MB, the MemoryReservation metric is 29.8%.
MemoryUtilization	The percentage of memory that is used by the containers on an instance. This value range is 0%–100%. For processing jobs, the value is the memory utilization of the processing container on the instance. For training jobs, the value is the memory utilization of the algorithm container on the instance. For batch transform jobs, the value is the memory utilization of the transform container on the instance. For endpoint variants, the value is the sum of the memory utilization of the primary and supplementary containers on the instance. Units: Percent Note For multi-instance jobs, each instance reports memory utilization metrics. However, the default view in CloudWatch shows the average memory utilization across all instances.

Dimensions for Processing Job, Training Job and Batch Transform Job Instance Metrics

Dimension

Description

Host

For processing jobs, the value for this dimension has the format [processing-job-name]/algo-[instance-number-in-cluster]. Use this dimension to filter instance metrics for the specified processing job and instance. This dimension format is present only in the /aws/sagemaker/ProcessingJobs namespace. For training jobs, the value for this dimension has the format [training-job-name]/algo-[instance-number-in-cluster]. Use this dimension to filter instance metrics for the specified training job and instance. This dimension format is present only in the /aws/sagemaker/TrainingJobs namespace. For batch transform jobs, the value for this dimension has the format [transform-job-name]/[instance-id]. Use this dimension to filter instance metrics for the specified batch transform job and instance. This dimension format is present only in the /aws/sagemaker/TransformJobs namespace.

SageMaker Inference Recommender jobs metrics

The /aws/sagemaker/InferenceRecommendationsJobs namespace includes the following metrics for inference recommendation jobs.

Inference Recommender Metrics

Metric	Description
ClientInvocations	The number of InvokeEndpoint requests sent to a model endpoint, as observed by Inference Recommender. Units: None Valid statistics: Sum
ClientInvocationErrors	The number of InvokeEndpoint requests that failed, as observed by Inference Recommender. Units: None Valid statistics: Sum
ClientLatency	The interval of time taken between sending an InvokeEndpoint call and receiving a response as observed by Inference Recommender. Note that the time is in milliseconds, whereas the ModelLatency endpoint invocation metric is in microseconds. Units: Milliseconds Valid statistics: Average, Sum, Min, Max, Sample Count, Percentiles
NumberOfUsers	The number of concurrent users sending InvokeEndpoint requests to the model endpoint. Units: None Valid statistics: Max, Min, Average

Dimensions for Inference Recommender Job Metrics

Dimension	Description
JobName	Filters Inference Recommender job metrics for the specified Inference Recommender job.
EndpointName	Filters Inference Recommender job metrics for the specified endpoint.

SageMaker Ground Truth metrics

Ground Truth Metrics

Metric	Description
ActiveWorkers	A single active worker on a private work team submitted, released, or declined a task. To get the total number of active workers, use the Sum statistic. Ground Truth tries to deliver each individual ActiveWorkers event once. If this delivery is unsuccessful, this metric may not report the total number of active workers. Units: None Valid statistics: Sum, Sample Count
DatasetObjectsAutoAnnotated	The number of dataset objects auto-annotated in a labeling job. This metric is only emitted when automated labeling is enabled. To view the labeling job progress, use the Max metric. Units: None Valid statistics: Max
DatasetObjectsHumanAnnotated	The number of dataset objects annotated by a human in a labeling job. To view the labeling job progress, use the Max metric. Units: None Valid statistics: Max
DatasetObjectsLabelingFailed	The number of dataset objects that failed labeling in a labeling job. To view the labeling job progress, use the Max metric. Units: None Valid statistics: Max
JobsFailed	A single labeling job failed. To get the total number of labeling jobs that failed, use the Sum statistic. Units: None Valid statistics: Sum, Sample Count
JobsSucceeded	A single labeling job succeeded. To get the total number of labeling jobs that succeeded, use the Sum statistic. Units: None Valid statistics: Sum, Sample Count
JobsStopped	A single labeling jobs was stopped. To get the total number of labeling jobs that were stopped, use the Sum statistic. Units: None Valid statistics: Sum, Sample Count
TasksAccepted	A single task was accepted by a worker. To get the total number of tasks accepted by workers, use the Sum statistic. Ground Truth attempts to deliver each individual TaskAccepted event once. If this delivery is unsuccessful, this metric may not report the total number of tasks accepted. Units: None Valid statistics: Sum, Sample Count
TasksDeclined	A single task was declined by a worker. To get the total number of tasks declined by workers, use the Sum statistic. Ground Truth attempts to deliver each individual TasksDeclined event once. If this delivery is unsuccessful, this metric may not report the total number of tasks declined. Units: None Valid Statistics: Sum, Sample Count
TasksReturned	A single task was returned. To get the total number of tasks returned, use the Sum statistic. Ground Truth attempts to deliver each individual TasksReturned event once. If this delivery is unsuccessful, this metric may not report the total number of tasks returned. Units: None Valid statistics: Sum, Sample Count
TasksSubmitted	A single task was submitted/completed by a private worker. To get the total number of tasks submitted by workers, use the Sum statistic. Ground Truth attempts to deliver each individual TasksSubmitted event once. If this delivery is unsuccessful, this metric may not report the total number of tasks submitted. Units: None Valid statistics: Sum, Sample Count
TimeSpent	Time spent on a task completed by a private worker. This metric does not include time when a worker paused or took a break. Ground Truth attempts to deliver each TimeSpent event once. If this delivery is unsuccessful, this metric may not report the total amount of time spent. Units: Seconds Valid statistics: Sum, Sample Count
TotalDatasetObjectsLabeled	The number of dataset objects labeled successfully in a labeling job. To view the labeling job progress, use the Max metric. Units: None Valid statistics: Max

Dimensions for Dataset Object Metrics

Dimension	Description
LabelingJobName	Filters dataset object count metrics for a labeling job.

Amazon SageMaker Feature Store metrics

Feature Store Consumption Metrics

Metric	Description
ConsumedReadRequestsUnits	The number of consumed read units over the specified time period. You can retrieve the consumed read units for a feature store runtime operation and its corresponding feature group. Units: None Valid statistics: All
ConsumedWriteRequestsUnits	The number of consumed write units over the specified time period. You can retrieve the consumed write units for a feature store runtime operation and its corresponding feature group. Units: None Valid statistics: All
ConsumedReadCapacityUnits	The number of provisioned read capacity units consumed over the specified time period. You can retrieve the consumed read capacity units for a feature store runtime operation and its corresponding feature group. Units: None Valid statistics: All
ConsumedWriteCapacityUnits	The number of provisioned write capacity units consumed over the specified time period. You can retrieve the consumed write capacity units for a feature store runtime operation and its corresponding feature group. Units: None Valid statistics: All

Dimensions for Feature Store Consumption Metrics

Dimension	Description
FeatureGroupName, OperationName	Filters feature store runtime consumption metrics of the feature group and the operation that you've specified.

Feature Store Operational Metrics

Metric	Description
Invocations	The number of requests made to the feature store runtime operations over the specified time period. Units: None Valid statistics: Sum
Operation4XXErrors	The number of requests made to the Feature Store runtime operations where the operation returned a 4xx HTTP response code. For each 4xx response, 1 is sent; else, 0 is sent. Units: None Valid statistics: Average, Sum
Operation5XXErrors	The number of requests made to the feature store runtime operations where the operation returned a 5xx HTTP response code. For each 5xx response, 1 is sent; else, 0 is sent. Units: None Valid statistics: Average, Sum
ThrottledRequests	The number of requests made to the feature store runtime operations where the request got throttled. For each throttled request, 1 is sent; else, 0 is sent. Units: None Valid statistics: Average, Sum
Latency	The time interval to process requests made to the Feature Store runtime operations. This interval is measured from the time SageMaker receives the request until it returns a response to the client. Units: Microseconds Valid statistics: Average, Sum, Min, Max, Sample Count, Percentiles

Dimensions for Feature Store Operational Metrics

Dimension	Description
FeatureGroupName, OperationName	Filters feature store runtime operational metrics of the feature group and the operation that you've specified. You can use these dimensions for non batch operations, such as GetRecord, PutRecord, and DeleteRecord.
OperationName	Filters feature store runtime operational metrics for the operation that you've specified. You can use this dimension for batch operations such as BatchGetRecord.

SageMaker pipelines metrics

The AWS/Sagemaker/ModelBuildingPipeline namespace includes the following metrics for pipeline executions.

Two categories of Pipelines execution metrics are available:

• Execution Metrics across All Pipelines – Account level pipeline execution metrics (for all pipelines in the current account)

• Execution Metrics by Pipeline – Pipeline execution metrics per pipeline

Metrics are available at a 1-minute frequency.

Pipelines Execution Metrics

Metric	Description
ExecutionStarted	The number of pipeline executions that started. Units: Count Valid statistics: Average, Sum
ExecutionFailed	The number of pipeline executions that failed. Units: Count Valid statistics: Average, Sum
ExecutionSucceeded	The number of pipeline executions that succeeded. Units: Count Valid statistics: Average, Sum
ExecutionStopped	The number of pipeline executions that stopped. Units: Count Valid statistics: Average, Sum
ExecutionDuration	The duration in milliseconds that the pipeline execution ran. Units: Milliseconds Valid statistics: Average, Sum, Min, Max, Sample Count

Dimensions for Execution Metrics by Pipeline

Dimension	Description
PipelineName	Filters pipeline execution metrics for a specified pipeline.

Pipelines Step Metrics

The AWS/Sagemaker/ModelBuildingPipeline namespace includes the following metrics for pipeline steps.

Metrics are available at a 1-minute frequency.

Metric	Description
StepStarted	The number of steps that started. Units: Count Valid statistics: Average, Sum
StepFailed	The number of steps that failed. Units: Count Valid statistics: Average, Sum
StepSucceeded	The number of steps that succeeded. Units: Count Valid statistics: Average, Sum
StepStopped	The number of steps that stopped. Units: Count Valid statistics: Average, Sum
StepDuration	The duration in milliseconds that the step ran. Units: Milliseconds Valid statistics: Average, Sum, Min, Max, Sample Count

Dimensions for Pipelines Step Metrics

Dimension	Description
PipelineName, StepName	Filters step metrics for a specified pipeline and step.