31 lines
1.6 KiB
Markdown
31 lines
1.6 KiB
Markdown
# Worker CPU rebalancing
|
|
|
|
The upstream Descheduler chart runs every five minutes. `values.yaml` is the
|
|
policy; the bootstrap Application selects its chart version.
|
|
|
|
The policy uses Metrics Server CPU usage: worker nodes below 25% are possible
|
|
receivers and nodes above 50% are possible eviction sources. The control-plane
|
|
node is excluded from eviction. Kubernetes chooses replacement placement; a
|
|
replacement can land on any eligible worker, including the original node.
|
|
No workload is pinned to the Xeon or an Odroid by this policy.
|
|
|
|
Each cycle permits at most two evictions overall, one per node and namespace.
|
|
Pods must be at least one hour old, belong to a controller with at least two
|
|
replicas, have a PodDisruptionBudget, have no PVC, and fit another eligible
|
|
node. Storage, networking, monitoring, certificate management, and GitOps
|
|
namespaces are excluded. EmptyDir alone is not a protection for these
|
|
replicated stateless workloads.
|
|
|
|
To test a policy change, set `cmdOptions.dry-run: true` and inspect Job logs
|
|
before enabling evictions. To stop balancing, set `suspend: true` in values and
|
|
let Argo sync. Existing running Jobs must finish before suspension takes full
|
|
effect.
|
|
|
|
Review after changes to workload mix. CPU requests determine whether a
|
|
replacement can fit even when observed utilization is low. The Descheduler
|
|
cannot guarantee that a particular worker gains CPU usage; scheduling rules on
|
|
the workload and available memory still govern placement.
|
|
|
|
Live evictions were enabled on 2026-11-01 after successful dry-run cycles.
|
|
The 2026-11-22 dry run selected one Traefik replica and respected the caps.
|