IssueTracker

This is a feature request to make the metrics-server component of GKE highly available.

Currently, metrics-server runs as a Deployment with a single replica.

Problem
The metrics-server is registered as the backend for the metrics.k8s.io/v1beta1 APIService.

When the metrics-server Pod is unhealthy, disrupted, or rescheduled, it can no longer serve API requests for the metrics endpoint.

This causes disruption to Kubernetes 1st-party & 3rd-party controllers, especially those that use API Discovery to discover which API groups and resources are available.

Two notable error examples:

* The Kubernetes namespace garbage collector fails to fully clean up Namespaces when the metrics-server is unavailable.
* Config Sync fails to sync and/or update resource status when the metrics-server is unavailable.

Possible Solution
The metrics-server Deployment used by GKE does not specify a replica count, which makes it default to 1 replica.

The official component GitHub repo has example YAML for deploying a highly available metrics-server:
https://github.com/kubernetes-sigs/metrics-server/?tab=readme-ov-file#high-availability

In addition to the recommended configuration, it would also be a good idea to define the PodDisruptionBudget to avoid both Pods being disrupted at the same time, and use enable-aggregator-routing to share the traffic load between the instances.

Cost & Node Requirement Concerns
For metrics-server to be made HA requires clusters to have at least 2 nodes. So some tweaks may be required for this solution to work on single-node and zero-node clusters, like Autopilot.

One way to handle this might be to create a simple controller that modifies the metrics-server Deployment config depending on how many Nodes are in the cluster at any given time. This way the config could be changed to be single-replica on one-node clusters, or even scale to zero on zero-node clusters without causing constant errors about the Deployment not having any healthy replicas.