Skip to main content
Version: Next

Metrics Server

This guide explains how to run metrics-server in the cloud so it can scrape resource usage metrics from edge nodes (kubelet) in an OpenYurt cluster.

By default, metrics-server prefers the node InternalIP when talking to kubelet. When the cloud and edge sit in different network domains, the cloud often cannot reach edge InternalIPs directly. You then rely on Raven for cross-domain connectivity, configure metrics-server to use Hostname instead of InternalIP, and resolve node hostnames to the Raven gateway in the cloud. Traffic flows through that gateway and the tunnel to the target edge node.

Prerequisites​

  • A working OpenYurt cluster (if you need to create or convert one, see the installation summary).
  • For edge scenarios, Raven must be installed and Raven Proxy DNS (raven-proxy-dns) must be deployed so node hostnames resolve to the cloud-side gateway. If Raven is not installed yet, complete step 1 below first.

Configuration steps​

1. Install Raven (skip if already installed)​

When edge node IPs are not reachable from the cloud, Raven provides cross-network-domain communication. Follow Manually Setup, section 3.3 Install raven component.

2. Deploy Raven Proxy DNS​

Node hostnames must resolve to the Raven cloud gateway so accesses in the form hostname:port are forwarded to the correct edge node. See Raven L7 proxy preparation for raven-proxy-dns.

3. Change metrics-server flags​

To make metrics-server use the node Hostname (instead of InternalIP) when resolving and contacting kubelet, add the following container argument:

--kubelet-preferred-address-types=Hostname

A typical approach is to edit the metrics-server Deployment in the kube-system namespace and add this flag to the container args. If --kubelet-preferred-address-types is already set, merge the value so Hostname is preferred, according to how metrics-server is deployed in your cluster.

4. Verify​

After the change rolls out, run:

kubectl top pods -A
kubectl top node <edge-node-name>

If CPU and memory metrics appear for pods and the edge node, the setup is working.