Container Insights with OpenTelemetry metrics for Amazon EKS - Amazon CloudWatch
How OTel Container Insights compares to the Container Insights (enhanced)How metrics are labeledSupported metrics

Container Insights with OpenTelemetry metrics for Amazon EKS

Preview

Container Insights with OpenTelemetry metrics provides visibility into the operational health of your Amazon EKS cluster infrastructure. It is available in public preview at no additional charge in US East (N. Virginia), US West (Oregon), Europe (Ireland), Asia Pacific (Singapore), and Asia Pacific (Sydney).

The Amazon CloudWatch Observability EKS add-on collects open source metrics from your Amazon EKS clusters and sends them to CloudWatch using the OpenTelemetry Protocol (OTLP) at 30 second granularity. These metrics use metric names from their original sources, including cAdvisor, Prometheus Node Exporter, NVIDIA DCGM, Kube State Metrics, and AWS Neuron Monitor. You can query these metrics using PromQL in CloudWatch Query Studio or through the Prometheus compatible query API.

Each metric is automatically enriched with up to 150 labels, including OpenTelemetry semantic convention attributes and Kubernetes pod and node labels. PromQL handles aggregation at query time, so each metric is published once per resource rather than at multiple aggregation levels. The add-on also correlates accelerator metrics from AWS Neuron and AWS Elastic Fabric Adapter with the specific pods and containers using them, providing visibility that is not available from the metric sources alone.

To enable OTel Container Insights on an Amazon EKS cluster, install the Amazon CloudWatch Observability EKS add-on version v6.0.1-eksbuild.1 or later through the Amazon EKS console or through infrastructure as code.

For more information about setting up OTel Container Insights, see Setting up Container Insights.

For more information about querying these metrics with PromQL, see PromQL querying.

How OTel Container Insights compares to the Container Insights (enhanced)

The following table summarizes the differences between Container Insights (enhanced) and OTel Container Insights.

Feature Container Insights (enhanced) OTel Container Insights
Metric names CloudWatch-format metrics (for example, pod_cpu_utilization) Open-source native (for example, container_cpu_usage_seconds_total)
Labels per metric 3–6 predefined dimensions per metric Up to 150 labels, including all Kubernetes pod and node labels
Aggregation Pre-aggregated at multiple levels (cluster, namespace, workload, pod) Raw per-resource metrics; aggregate at query time with PromQL
Query language CloudWatch Metrics API PromQL (Prometheus-compatible)
Metric ingestion CloudWatch Logs in EMF format OTLP endpoint

How metrics are labeled

Each metric collected by OTel Container Insights carries labels from three sources.

Telemetry source native labels

Labels from the original metric source (for example, cAdvisor provides labels such as pod, namespace, and container). These are preserved as datapoint attributes.

OpenTelemetry resource attributes

The add-on appends resource attributes following OpenTelemetry semantic conventions for Kubernetes, Host, and Cloud, such as k8s.pod.name, k8s.namespace.name, k8s.node.name, host.name, and cloud.region. These attributes are consistent across all metric sources.

Kubernetes pod and node labels

All pod labels and node labels discovered from the Kubernetes API are appended as resource attributes with the prefixes k8s.pod.label and k8s.node.label.

For more information about how to query these attributes using PromQL, see PromQL querying.

Supported metrics

The following table lists the metric sources and categories collected by OTel Container Insights.

Metric source Metric category Prerequisites
cAdvisor CPU metrics -
cAdvisor Memory metrics -
cAdvisor Network metrics -
cAdvisor Disk and filesystem metrics -
Prometheus Node Exporter CPU metrics -
Prometheus Node Exporter Memory metrics -
Prometheus Node Exporter Disk metrics -
Prometheus Node Exporter Filesystem metrics -
Prometheus Node Exporter Network metrics -
Prometheus Node Exporter System metrics -
Prometheus Node Exporter VMStat metrics -
Prometheus Node Exporter Netstat and socket metrics -
NVIDIA DCGM GPU utilization and performance metrics NVIDIA device plugin and NVIDIA container toolkit must be installed.
NVIDIA DCGM GPU memory metrics NVIDIA device plugin and NVIDIA container toolkit must be installed.
NVIDIA DCGM GPU power and thermal metrics NVIDIA device plugin and NVIDIA container toolkit must be installed.
NVIDIA DCGM GPU throttling metrics NVIDIA device plugin and NVIDIA container toolkit must be installed.
NVIDIA DCGM GPU error and reliability metrics NVIDIA device plugin and NVIDIA container toolkit must be installed.
NVIDIA DCGM GPU NVLink metrics NVIDIA device plugin and NVIDIA container toolkit must be installed.
NVIDIA DCGM GPU informational metrics NVIDIA device plugin and NVIDIA container toolkit must be installed.
AWS Neuron Monitor NeuronCore metrics Neuron driver and Neuron device plugin must be installed.
AWS Neuron Monitor NeuronDevice metrics Neuron driver and Neuron device plugin must be installed.
AWS Neuron Monitor Neuron system metrics Neuron driver and Neuron device plugin must be installed.
AWS Elastic Fabric Adapter EFA metrics EFA device plugin must be installed.
NVMe NVMe SMART metrics -
Kube State Metrics Pod, node, Deployment, DaemonSet, StatefulSet, ReplicaSet, Job, CronJob, Service, Namespace, PersistentVolume, PersistentVolumeClaim metrics -
Kubernetes API server API server and etcd metrics -