Since 4.8, prometheus-adapter is now highly available in OpenShift, but only one of the two instances of prometheus-adapter is responding to requests from the apiserver. This is caused by the kube-aggregator not loadbalancing requests between the replicas behind the API service.

This has the following impacts:

• availability problems - if the active prometheus-adapter has a failure it can take some time for apiserver to notice this.
• wastes of resources - requires allocating resources needed to serve API for both instances, but only one is doing the job.

FWIW, I tried adding Service load balancing, but proxying doesn't seem to go through it as I would have expected and queries are still rooted to the same instance, so this shouldn't solve the issue.

Upstream issue in metrics-server: https://github.com/kubernetes-sigs/metrics-server/issues/764

Since we are not responsible for the kube-aggregator, it might be a good idea to discuss this with the apiserver team.