Settings
Controller flags
| Flag | Default | Meaning |
|---|---|---|
--bestbefore-resync-period |
5m |
How often every NodeClaim is re-evaluated, on top of its exact maxAge deadline. Minimum 10s. |
--metrics-bind-address |
0 (disabled) |
Metrics endpoint. The chart sets :8443. |
--metrics-secure |
true |
Require authn/authz on the metrics endpoint. |
--health-probe-bind-address |
:8081 |
Health and readiness probes. |
--leader-elect |
false |
Leader election. The chart sets it. |
The resync period bounds how late a rotation can start if a deadline is missed; it does not make k-belt poll the API server, because reconciles read from the controller’s cache. At 2,000 NodeClaims the default works out at roughly seven cached reads a second. Each NodeClaim’s resync is jittered by ±10% so they don’t all land together after a restart.
Chart values
The ones most installs touch. Everything else is in values.yaml.
| Value | Default | Meaning |
|---|---|---|
image.tag |
chart appVersion |
Controller image tag |
replicaCount |
1 |
Extra replicas stand by via leader election |
controller.resyncPeriod |
5m |
Sets --bestbefore-resync-period |
metrics.enabled |
true |
Serve metrics and create the Service |
metrics.secure |
true |
Protect the metrics endpoint |
metrics.serviceMonitor.enabled |
false |
Create a Prometheus Operator ServiceMonitor |
resources |
10m/64Mi requests | Controller pod resources |
priorityClassName |
"" |
Worth setting so k-belt isn’t evicted by the rotations it starts |
Upgrading
Helm does not upgrade CRDs. After helm upgrade, apply the CRD yourself:
kubectl apply -f https://raw.githubusercontent.com/rjbez17/k-belt/main/charts/k-belt/crds/bestbefore.k-belt.io_bestbefores.yaml
Permissions
k-belt runs with a ClusterRole that can get, list, watch and patch NodeClaims and Nodes,
read NodePools and BestBefores, write its own status, and create events. It cannot create or delete
nodes: every replacement is Karpenter’s doing.