Pod priorities (Priority Classes)
Available Priority Classes
DKP creates a set of predefined PriorityClass objects in the cluster and assigns them to Deckhouse components and user applications. The Kubernetes scheduler takes pod priority into account when allocating resources: if there are not enough resources, pods with lower priorities will be evicted first.
For example, if pods are assigned priorityClassName: production-low and the cluster runs out of resources, the scheduler will first evict pods with priorityClassName: develop, then cluster-low, and so on. Therefore, when selecting a priority class, you should consider the type of application and the environment in which it runs. If a pod has no specified priority, the scheduler treats it as the lowest.
Do not use the system-node-critical, system-cluster-critical, cluster-medium, or cluster-low priority classes, as they are reserved for critical cluster components.
Below is a table of available priority classes.
| Priority Class | Description | Value |
|---|---|---|
| system-node-critical | Cluster components that must be present on the node. Also fully protected from eviction by kubelet. Examples: node-exporter, csi, and others. |
2000001000 |
| system-cluster-critical | Cluster components without which the cluster cannot function properly. This PriorityClass must be assigned to MutatingWebhooks and Extension API servers. Also fully protected from eviction by kubelet. Examples: kube-dns, kube-proxy, cni-flannel, cni-cilium, and others. |
2000000000 |
| production-high | Stateful applications in production environments whose absence leads to complete service unavailability or data loss. Examples: PostgreSQL, Memcached, Redis, MongoDB, and others. |
9000 |
| cluster-medium | Cluster components responsible for monitoring (alerts, diagnostics) and autoscaling. Without monitoring it’s impossible to assess incidents; without autoscaling, applications may lack necessary resources. Examples: deckhouse, node-local-dns, grafana, upmeter, and others. |
7000 |
| production-medium | Main stateless applications in production that serve end-users. | 6000 |
| deployment-machinery | Cluster components used for build and deploy processes. | 5000 |
| production-low | Production applications (cron jobs, admin panels, batch processes) that can be temporarily unavailable. If batch/cron jobs cannot be interrupted, they should be classified as production-medium. |
4000 |
| staging | Staging environments for applications. | 3000 |
| cluster-low | Non-critical but desirable cluster components. Examples: dashboard, cert-manager, prometheus, and others. |
2000 |
| develop (default) | Development environments for applications. Default priority class if none is specified. | 1000 |
| standby | Not intended for applications. Used internally to reserve nodes. | -1 |