If you run metrics-server from its Helm chart and scrape it with Prometheus, you've probably set metrics.enabled: true. The chart needs it before it will create a ServiceMonitor. The README says it will "allow unauthenticated access to /metrics", and it does.
It also makes every health check metrics-server answers depend on the Kubernetes apiserver.
Someone reported this in October, already traced to a chart change from January 2022, with the one-line fix. I sent the PR, #1904, with before and after numbers from a kind cluster, and the chart maintainer merged it on 12 November. It isn't in a chart release yet, so the workaround below is what you can use today.
One flag, one value
metrics-server is built on k8s.io/apiserver. Requests to its HTTPS port are authorized by asking the real apiserver, with a SubjectAccessReview. A short list of paths skips that check, and the default list is /healthz, /readyz and /livez.
With metrics.enabled: true the chart renders this:
- --authorization-always-allow-paths=/metrics The flag is a string slice. The first time it's set, the default is thrown away and only your value is kept. So the list is now just /metrics, and the three health paths need a review again.
What the probes do now
The kubelet sends probes without credentials, so metrics-server asks the apiserver whether system:anonymous may read /livez. On a stock cluster the built-in system:public-info-viewer role says yes. The probe passes and nothing looks different.
In an apiserver audit log I counted 10 of these reviews in 90 seconds with the chart's setting, and none with the fix. That load is tiny. The problem is the dependency.
When the apiserver goes away
I stopped the apiserver on the kind node for 60 seconds and polled /livez every 5 seconds.
On chart 3.14.0, /livez started returning 500 straight away:
Internal Server Error: "/livez": Post "https://10.96.0.1:443/apis/authorization.k8s.io/v1/subjectaccessreviews?timeout=10s": dial tcp 10.96.0.1:443: connect: connection refused After about 25 seconds the kubelet killed the container. Each restart panicked at startup, because metrics-server also reads a ConfigMap from the apiserver when it boots. When the apiserver came back, metrics-server was still in CrashLoopBackOff, and it took another 69 to 76 seconds to be Ready across four runs, with 5 restarts. No kubectl top and no HPA resource metrics in that time.
With the fixed chart, every poll returned 200 and nothing restarted.
A 30 second stop gave 500s but no restart, so short blips are fine. Longer ones, like a slow control plane upgrade, now take metrics-server down with them.
Check yours
Charts 3.8.0 to 3.14.0 are affected when metrics.enabled is true. The plain components.yaml manifest isn't.
kubectl -n kube-system get deploy -l app.kubernetes.io/name=metrics-server \ -o yaml | grep always-allow If the only match is --authorization-always-allow-paths=/metrics, add the defaults back through the chart's args:
args: - --authorization-always-allow-paths=/healthz,/readyz,/livez The flag ends up set twice, and later values are appended, so you get all four paths. I tested this on 3.14.0: zero reviews, 200 during the apiserver stop, no restarts.
The full post has the rest of the numbers, the hardened-RBAC case where probes get 403 even with a healthy apiserver, and a table of every configuration I ran.


