metrics-server.md

Using metrics-server instead of prometheus-adapter

By default, kube-prometheus deploys prometheus-adapter to serve the Kubernetes resource metrics API (metrics.k8s.io). This is a full-featured solution that derives CPU and memory metrics from Prometheus queries, and can also serve custom and external metrics APIs.

However, for use cases where only the resource metrics API is needed (for example, when using KEDA for custom metrics scaling), the lightweight upstream metrics-server may be preferable.

Switching to metrics-server

Set the resourceMetricsAPI field in values.common to 'metrics-server':

(import 'kube-prometheus/main.libsonnet') +
{
  values+:: {
    common+: {
      resourceMetricsAPI:: 'metrics-server',
    },
    metricsServer+: {
      kubeletInsecureTLS:: true,
    },
  },
}

This replaces the prometheus-adapter manifests with metrics-server manifests. Only one resource metrics API provider can be active at a time since they both register the v1beta1.metrics.k8s.io APIService.

When to set kubeletInsecureTLS

On clusters that do not use proper kubelet serving certificates (for example, kind, minikube, or clusters without a kubelet-serving CA), set kubeletInsecureTLS:: true. On production clusters with a properly configured kubelet PKI, this should be left at the default (false).

Emitting manifests

If you are writing your own top-level jsonnet (instead of using example.jsonnet), emit the resource metrics manifests using the helper library:

(import 'kube-prometheus/lib/resource-metrics-api.libsonnet')(kp)

This automatically selects the correct component based on the resourceMetricsAPI flag.

Single-node clusters

The default configuration deploys 2 replicas with hard pod anti-affinity, which requires at least 2 nodes. On single-node clusters (minikube, single-node kind, etc.), set replicas:: 1 or switch to soft anti-affinity:

values+:: {
  metricsServer+: {
    replicas:: 1,
  },
},

Available configuration

The following hidden fields can be overridden in values.metricsServer:

Field	Default	Description
replicas	2	Number of metrics-server replicas
resources	{requests: {cpu: '100m', memory: '200Mi'}}	Resource requests/limits
kubeletInsecureTLS	false	Skip kubelet TLS verification
metricResolution	'15s'	How often metrics are scraped from kubelets
securePort	10250	Metrics-server HTTPS port
extraArgs	[]	Additional command-line arguments
podAntiAffinity	'hard'	Pod anti-affinity type ('hard' or 'soft')
podAntiAffinityTopologyKey	'kubernetes.io/hostname'	Topology key for anti-affinity
insecureSkipTLSVerify	true	APIService TLS verification
priorityClassName	'system-cluster-critical'	Pod priority class

Generating and deploying manifests

To generate manifests using the provided example:

make manifests-metrics-server

This builds from examples/metrics-server.jsonnet and outputs YAML manifests to manifests/. Then deploy:

kubectl apply --server-side -f manifests/setup
kubectl apply -f manifests/

Full example

See examples/metrics-server.jsonnet for a complete working example that deploys the full kube-prometheus stack with metrics-server instead of prometheus-adapter.