[Operational] cluster-scheduler
Deploys the cluster-scheduler binary on the management cluster. The
kcp-side API objects it needs live in the
poc-kcm-kcp chart;
install that one into the workspace owning the exports first.
Install
kubectl create namespace edge-connect-system
# The kubeconfig pointing at the kcp front proxy, created out of band.
kubectl -n edge-connect-system create secret generic kcp-kubeconfig \
--from-file=kubeconfig=/path/to/kcp.kubeconfig
helm upgrade --install cluster-scheduler deploy/charts/cluster-scheduler \
--namespace edge-connect-system \
--set kcp.connection.kubeconfigSecret.name=kcp-kubeconfig
The chart refuses to render without a kcp connection: pass
kcp.connection.kubeconfigSecret.name for a Secret you manage, or
kcp.connection.inlineKubeconfig to have the chart create one.
Values
| Key | Default | Notes |
|---|---|---|
replicaCount | 1 | Safe above 1 while leaderElection.enabled is true |
image.repository | edp.buildth.ing/devfw-cicd/cluster-scheduler | |
image.tag | "" | Empty means the chart appVersion, so a chart upgrade is the only thing that changes which bytes run |
endpointSlice | cluster-tenant | Must match the APIExport name the kcp chart deploys |
catalogWorkspacePath | root:providers | Workspace holding the offerings and machine types |
unschedulableRequeue | 5m | Re-check interval for an unschedulable cluster |
leaderElection.enabled | true | Lease lives in the release namespace on the management cluster |
serviceAccount.name | cluster-scheduler | kcp maps this to mgmt:system:serviceaccount:<namespace>:<name>, which the kcp chart's export subjects are written against — keep the two in step |
observability.networkPolicy.from | the monitoring namespace | The metrics endpoint is plaintext and unauthenticated, so reachability is the control; an empty list denies all metrics ingress |
observability.serviceMonitor.enabled | false | Requires the Prometheus Operator CRDs |
Probes
/healthz is a liveness ping. /readyz is the caches-synced gate: a replica
whose kcp provider has not connected reports not-ready, so a rollout does not
shift traffic onto a replica that can serve no reconcile.