diff --git a/packages/system/vertical-pod-autoscaler/values.yaml b/packages/system/vertical-pod-autoscaler/values.yaml index 46678903e..b870049be 100644 --- a/packages/system/vertical-pod-autoscaler/values.yaml +++ b/packages/system/vertical-pod-autoscaler/values.yaml @@ -37,6 +37,15 @@ vertical-pod-autoscaler: prometheus-cadvisor-job-name: cadvisor admissionController: + extraArgs: + # The admission controller registers its mutating webhook for Pod CREATE in every + # namespace with failurePolicy Ignore and, by default, a 30s timeout. When its pod is + # down and the Service IP blackholes (e.g. a cold start after a full cluster shutdown, + # before the CNI is back), every Pod creation waits the full 30s here on top of the + # other webhooks and fails with "context deadline exceeded" - including the CNI's own + # DaemonSet pods, so the cluster cannot recover by itself (#3125). A mutation takes + # milliseconds; 5s keeps a slow response working and lets Ignore actually ignore. + webhook-timeout-seconds: 5 resources: limits: memory: 256Mi