The module lifecycle stage: General Availability
Configuring the pod redistribution interval
To set how often the descheduler module runs the pod redistribution cycle, use the deschedulingInterval parameter.
For example, to run descheduler every 5 minutes, set deschedulingInterval: Frequent:
apiVersion: deckhouse.io/v1alpha1
kind: ModuleConfig
metadata:
name: descheduler
spec:
enabled: true
settings:
deschedulingInterval: Frequent
Supported values:
Frequent: Runs every 5 minutes. Suitable for clusters where faster pod redistribution is important.Moderate: Runs every 15 minutes (default).Rare: Runs every 30 minutes. Suitable for clusters where it’s important to minimize the number of pod redistributions.
Example LowNodeUtilization strategy
---
apiVersion: deckhouse.io/v1alpha2
kind: Descheduler
metadata:
name: low-node-utilization
spec:
strategies:
lowNodeUtilization:
enabled: true
thresholds:
cpu: 20
targetThresholds:
cpu: 50
Example LowNodeUtilization strategy with thresholds relative to the average
The strategy keeps the CPU utilization of the nodes within the corridor of ±5% around the average utilization of the cluster nodes and evicts at most two pods from one node per descheduling cycle. A virtual machine that is still migrating from the node takes one of the two places:
---
apiVersion: deckhouse.io/v1alpha2
kind: Descheduler
metadata:
name: low-node-utilization-deviation
spec:
strategies:
lowNodeUtilization:
enabled: true
useDeviationThresholds: true
thresholds:
cpu: 5
targetThresholds:
cpu: 5
evictionLimits:
node: 2
Example LowNodeUtilization strategy on the 15-minute CPU average
The strategy keeps the nodes within the corridor of ±5% around the cluster average of the CPU utilization averaged over 15 minutes, so short load spikes do not trigger evictions. It evicts one pod from each over utilized node per descheduling cycle; keep evictionLimits.node unset in this mode, since it raises that number. Use the default Moderate descheduling interval (15 minutes). The prometheus module must be enabled:
---
apiVersion: deckhouse.io/v1alpha2
kind: Descheduler
metadata:
name: low-node-utilization-prometheus
spec:
strategies:
lowNodeUtilization:
enabled: true
metricsSource: Prometheus
useDeviationThresholds: true
thresholds:
cpu: 5
targetThresholds:
cpu: 5
Example HighNodeUtilization strategy
---
apiVersion: deckhouse.io/v1alpha2
kind: Descheduler
metadata:
name: high-node-utilization
spec:
strategies:
highNodeUtilization:
enabled: true
thresholds:
cpu: 50
memory: 50
Protecting pods by StorageClass
Use the protectedStorageClasses parameter to forbid evicting pods that use PersistentVolumeClaims provisioned by specific StorageClasses. This is useful for storage that is bound to a node (for example, local-path), where evicting a pod either loses the data locality or leaves the pod unschedulable.
---
apiVersion: deckhouse.io/v1alpha2
kind: Descheduler
metadata:
name: protect-local-storage
spec:
protectedStorageClasses:
- local-path
strategies:
lowNodeUtilization:
enabled: true
thresholds:
cpu: 20
targetThresholds:
cpu: 50
With this configuration, the descheduler can still evict every other pod, but a pod with a PersistentVolumeClaim from the local-path StorageClass is never evicted.
Keep the following in mind:
- If the parameter is omitted or empty, the behavior does not change: pods with PersistentVolumeClaims are evicted as usual.
- A pod is also considered protected when its PersistentVolumeClaim cannot be resolved, that is, the PVC does not exist or its
storageClassNameis empty. Such pods are never evicted while the parameter is set.