Scheduling in Kubernetes is the process of binding pending pods to nodes, and is performed by a component of Kubernetes called kube-scheduler. The scheduler's decisions, whether or where a pod can or can not be scheduled, are guided by its configurable policy which comprises of set of rules, called predicates and priorities. The scheduler's decisions are influenced by its view of a Kubernetes cluster at that point of time when a new pod appears for scheduling. As Kubernetes clusters are very dynamic and their state changes over time, there may be desire to move already running pods to some other nodes for various reasons:

• Some nodes are under or over utilized.
• The original scheduling decision does not hold true any more, as taints or labels are added to or removed from nodes, pod/node affinity requirements are not satisfied any more.
• Some nodes failed and their pods moved to other nodes.
• New nodes are added to clusters.

Consequently, there might be several pods scheduled on less desired nodes in a cluster. Descheduler, based on its policy, finds pods that can be moved and evicts them. Please note, in current implementation, descheduler does not schedule replacement of evicted pods but relies on the default scheduler for that.

• Quick Start
  • Run As A Job
  • Run As A CronJob
  • Run As A Deployment
  • Install Using Helm
  • Install Using Kustomize
• User Guide
• Policy and Strategies
  • RemoveDuplicates
  • LowNodeUtilization
  • HighNodeUtilization
  • RemovePodsViolatingInterPodAntiAffinity
  • RemovePodsViolatingNodeAffinity
  • RemovePodsViolatingNodeTaints
  • RemovePodsViolatingTopologySpreadConstraint
  • RemovePodsHavingTooManyRestarts
  • PodLifeTime
  • RemoveFailedPods
• Filter Pods
  • Namespace filtering
  • Priority filtering
  • Label filtering
  • Node Fit filtering
• Pod Evictions
  • Pod Disruption Budget (PDB)
• High Availability
  • Configure HA Mode
• Metrics
• Compatibility Matrix
• Getting Involved and Contributing
  • Communicating With Contributors
• Roadmap
  • Code of conduct

The descheduler can be run as a Job, CronJob, or Deployment inside of a k8s cluster. It has the advantage of being able to be run multiple times without needing user intervention. The descheduler pod is run as a critical pod in the kube-system namespace to avoid being evicted by itself or by the kubelet.

kubectl create -f kubernetes/base/rbac.yaml
kubectl create -f kubernetes/base/configmap.yaml
kubectl create -f kubernetes/job/job.yaml

kubectl create -f kubernetes/base/rbac.yaml
kubectl create -f kubernetes/base/configmap.yaml
kubectl create -f kubernetes/cronjob/cronjob.yaml

kubectl create -f kubernetes/base/rbac.yaml
kubectl create -f kubernetes/base/configmap.yaml
kubectl create -f kubernetes/deployment/deployment.yaml

Starting with release v0.18.0 there is an official helm chart that can be used to install the descheduler. See the helm chart README for detailed instructions.

The descheduler helm chart is also listed on the artifact hub.

You can use kustomize to install descheduler. See the resources | Kustomize for detailed instructions.

Run As A Job

kustomize build 'github.com/kubernetes-sigs/descheduler/kubernetes/job?ref=v0.25.0' | kubectl apply -f -

Run As A CronJob

kustomize build 'github.com/kubernetes-sigs/descheduler/kubernetes/cronjob?ref=v0.25.0' | kubectl apply -f -

Run As A Deployment

kustomize build 'github.com/kubernetes-sigs/descheduler/kubernetes/deployment?ref=v0.25.0' | kubectl apply -f -

See the user guide in the /docs directory.

Descheduler's policy is configurable and includes strategies that can be enabled or disabled. By default, all strategies are enabled.

The policy includes a common configuration that applies to all the strategies:

As part of the policy, the parameters associated with each strategy can be configured. See each strategy for details on available parameters.

Policy:

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
nodeSelector: prod=dev
evictFailedBarePods: false
evictLocalStoragePods: true
evictSystemCriticalPods: true
maxNoOfPodsToEvictPerNode: 40
ignorePvcPods: false
strategies:
  ...

The following diagram provides a visualization of most of the strategies to help categorize how strategies fit together.

This strategy makes sure that there is only one pod associated with a ReplicaSet (RS), ReplicationController (RC), StatefulSet, or Job running on the same node. If there are more, those duplicate pods are evicted for better spreading of pods in a cluster. This issue could happen if some nodes went down due to whatever reasons, and pods on them were moved to other nodes leading to more than one pod associated with a RS or RC, for example, running on the same node. Once the failed nodes are ready again, this strategy could be enabled to evict those duplicate pods.

It provides one optional parameter, excludeOwnerKinds, which is a list of OwnerRef Kinds. If a pod has any of these Kinds listed as an OwnerRef, that pod will not be considered for eviction. Note that pods created by Deployments are considered for eviction by this strategy. The excludeOwnerKinds parameter should include ReplicaSet to have pods created by Deployments excluded.

Parameters:

Example:

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
strategies:
  "RemoveDuplicates":
     enabled: true
     params:
       removeDuplicates:
         excludeOwnerKinds:
         - "ReplicaSet"

This strategy finds nodes that are under utilized and evicts pods, if possible, from other nodes in the hope that recreation of evicted pods will be scheduled on these underutilized nodes. The parameters of this strategy are configured under nodeResourceUtilizationThresholds.

The under utilization of nodes is determined by a configurable threshold thresholds. The threshold thresholds can be configured for cpu, memory, number of pods, and extended resources in terms of percentage (the percentage is calculated as the current resources requested on the node vs total allocatable. For pods, this means the number of pods on the node as a fraction of the pod capacity set for that node).

If a node's usage is below threshold for all (cpu, memory, number of pods and extended resources), the node is considered underutilized. Currently, pods request resource requirements are considered for computing node resource utilization.

There is another configurable threshold, targetThresholds, that is used to compute those potential nodes from where pods could be evicted. If a node's usage is above targetThreshold for any (cpu, memory, number of pods, or extended resources), the node is considered over utilized. Any node between the thresholds, thresholds and targetThresholds is considered appropriately utilized and is not considered for eviction. The threshold, targetThresholds, can be configured for cpu, memory, and number of pods too in terms of percentage.

These thresholds, thresholds and targetThresholds, could be tuned as per your cluster requirements. Note that this strategy evicts pods from overutilized nodes (those with usage above targetThresholds) to underutilized nodes (those with usage below thresholds), it will abort if any number of underutilized nodes or overutilized nodes is zero.

Additionally, the strategy accepts a useDeviationThresholds parameter. If that parameter is set to true, the thresholds are considered as percentage deviations from mean resource usage. thresholds will be deducted from the mean among all nodes and targetThresholds will be added to the mean. A resource consumption above (resp. below) this window is considered as overutilization (resp. underutilization).

NOTE: Node resource consumption is determined by the requests and limits of pods, not actual usage. This approach is chosen in order to maintain consistency with the kube-scheduler, which follows the same design for scheduling pods onto nodes. This means that resource usage as reported by Kubelet (or commands like kubectl top) may differ from the calculated consumption, due to these components reporting actual usage metrics. Implementing metrics-based descheduling is currently TODO for the project.

Parameters:

Example:

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
strategies:
  "LowNodeUtilization":
     enabled: true
     params:
       nodeResourceUtilizationThresholds:
         thresholds:
           "cpu" : 20
           "memory": 20
           "pods": 20
         targetThresholds:
           "cpu" : 50
           "memory": 50
           "pods": 50

Policy should pass the following validation checks:

• Three basic native types of resources are supported: cpu, memory and pods. If any of these resource types is not specified, all its thresholds default to 100% to avoid nodes going from underutilized to overutilized.
• Extended resources are supported. For example, resource type nvidia.com/gpu is specified for GPU node utilization. Extended resources are optional, and will not be used to compute node's usage if it's not specified in thresholds and targetThresholds explicitly.
• thresholds or targetThresholds can not be nil and they must configure exactly the same types of resources.
• The valid range of the resource's percentage value is [0, 100]
• Percentage value of thresholds can not be greater than targetThresholds for the same resource.

There is another parameter associated with the LowNodeUtilization strategy, called numberOfNodes. This parameter can be configured to activate the strategy only when the number of under utilized nodes are above the configured value. This could be helpful in large clusters where a few nodes could go under utilized frequently or for a short period of time. By default, numberOfNodes is set to zero.

This strategy finds nodes that are under utilized and evicts pods from the nodes in the hope that these pods will be scheduled compactly into fewer nodes. Used in conjunction with node auto-scaling, this strategy is intended to help trigger down scaling of under utilized nodes. This strategy must be used with the scheduler scoring strategy MostAllocated. The parameters of this strategy are configured under nodeResourceUtilizationThresholds.

The under utilization of nodes is determined by a configurable threshold thresholds. The threshold thresholds can be configured for cpu, memory, number of pods, and extended resources in terms of percentage. The percentage is calculated as the current resources requested on the node vs total allocatable. For pods, this means the number of pods on the node as a fraction of the pod capacity set for that node.

If a node's usage is below threshold for all (cpu, memory, number of pods and extended resources), the node is considered underutilized. Currently, pods request resource requirements are considered for computing node resource utilization. Any node above thresholds is considered appropriately utilized and is not considered for eviction.

The thresholds param could be tuned as per your cluster requirements. Note that this strategy evicts pods from underutilized nodes (those with usage below thresholds) so that they can be recreated in appropriately utilized nodes. The strategy will abort if any number of underutilized nodes or appropriately utilized nodes is zero.

NOTE: Node resource consumption is determined by the requests and limits of pods, not actual usage. This approach is chosen in order to maintain consistency with the kube-scheduler, which follows the same design for scheduling pods onto nodes. This means that resource usage as reported by Kubelet (or commands like kubectl top) may differ from the calculated consumption, due to these components reporting actual usage metrics. Implementing metrics-based descheduling is currently TODO for the project.

Parameters:

Example:

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
strategies:
  "HighNodeUtilization":
     enabled: true
     params:
       nodeResourceUtilizationThresholds:
         thresholds:
           "cpu" : 20
           "memory": 20
           "pods": 20

Policy should pass the following validation checks:

• Three basic native types of resources are supported: cpu, memory and pods. If any of these resource types is not specified, all its thresholds default to 100%.
• Extended resources are supported. For example, resource type nvidia.com/gpu is specified for GPU node utilization. Extended resources are optional, and will not be used to compute node's usage if it's not specified in thresholds explicitly.
• thresholds can not be nil.
• The valid range of the resource's percentage value is [0, 100]

There is another parameter associated with the HighNodeUtilization strategy, called numberOfNodes. This parameter can be configured to activate the strategy only when the number of under utilized nodes is above the configured value. This could be helpful in large clusters where a few nodes could go under utilized frequently or for a short period of time. By default, numberOfNodes is set to zero.

This strategy makes sure that pods violating interpod anti-affinity are removed from nodes. For example, if there is podA on a node and podB and podC (running on the same node) have anti-affinity rules which prohibit them to run on the same node, then podA will be evicted from the node so that podB and podC could run. This issue could happen, when the anti-affinity rules for podB and podC are created when they are already running on node.

Parameters:

Example:

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
strategies:
  "RemovePodsViolatingInterPodAntiAffinity":
     enabled: true

This strategy makes sure all pods violating node affinity are eventually removed from nodes. Node affinity rules allow a pod to specify requiredDuringSchedulingIgnoredDuringExecution type, which tells the scheduler to respect node affinity when scheduling the pod but kubelet to ignore in case node changes over time and no longer respects the affinity. When enabled, the strategy serves as a temporary implementation of requiredDuringSchedulingRequiredDuringExecution and evicts pod for kubelet that no longer respects node affinity.

For example, there is podA scheduled on nodeA which satisfies the node affinity rule requiredDuringSchedulingIgnoredDuringExecution at the time of scheduling. Over time nodeA stops to satisfy the rule. When the strategy gets executed and there is another node available that satisfies the node affinity rule, podA gets evicted from nodeA.

Parameters:

Example:

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
strategies:
  "RemovePodsViolatingNodeAffinity":
    enabled: true
    params:
      nodeAffinityType:
      - "requiredDuringSchedulingIgnoredDuringExecution"

This strategy makes sure that pods violating NoSchedule taints on nodes are removed. For example there is a pod "podA" with a toleration to tolerate a taint key=value:NoSchedule scheduled and running on the tainted node. If the node's taint is subsequently updated/removed, taint is no longer satisfied by its pods' tolerations and will be evicted.

Node taints can be excluded from consideration by specifying a list of excludedTaints. If a node taint key or key=value matches an excludedTaints entry, the taint will be ignored.

For example, excludedTaints entry "dedicated" would match all taints with key "dedicated", regardless of value. excludedTaints entry "dedicated=special-user" would match taints with key "dedicated" and value "special-user".

Parameters:

Example:

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
strategies:
  "RemovePodsViolatingNodeTaints":
    enabled: true
    params:
      excludedTaints:
      - dedicated=special-user # exclude taints with key "dedicated" and value "special-user"
      - reserved # exclude all taints with key "reserved"

This strategy makes sure that pods violating topology spread constraints are evicted from nodes. Specifically, it tries to evict the minimum number of pods required to balance topology domains to within each constraint's maxSkew. This strategy requires k8s version 1.18 at a minimum.

By default, this strategy only deals with hard constraints, setting parameter includeSoftConstraints to true will include soft constraints.

Strategy parameter labelSelector is not utilized when balancing topology domains and is only applied during eviction to determine if the pod can be evicted.

Parameters:

Example:

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
strategies:
  "RemovePodsViolatingTopologySpreadConstraint":
     enabled: true
     params:
       includeSoftConstraints: false

This strategy makes sure that pods having too many restarts are removed from nodes. For example a pod with EBS/PD that can't get the volume/disk attached to the instance, then the pod should be re-scheduled to other nodes. Its parameters include podRestartThreshold, which is the number of restarts (summed over all eligible containers) at which a pod should be evicted, and includingInitContainers, which determines whether init container restarts should be factored into that calculation.

Parameters:

Example:

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
strategies:
  "RemovePodsHavingTooManyRestarts":
     enabled: true
     params:
       podsHavingTooManyRestarts:
         podRestartThreshold: 100
         includingInitContainers: true

This strategy evicts pods that are older than maxPodLifeTimeSeconds.

You can also specify states parameter to only evict pods matching the following conditions:

• Pod Phase status of: Running, Pending
• Container State Waiting condition of: PodInitializing, ContainerCreating

If a value for states or podStatusPhases is not specified, Pods in any state (even Running) are considered for eviction.

Parameters:

Example:

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
strategies:
  "PodLifeTime":
     enabled: true
     params:
       podLifeTime:
         maxPodLifeTimeSeconds: 86400
         states:
         - "Pending"
         - "PodInitializing"

This strategy evicts pods that are in failed status phase. You can provide an optional parameter to filter by failed reasons. reasons can be expanded to include reasons of InitContainers as well by setting the optional parameter includingInitContainers to true. You can specify an optional parameter minPodLifetimeSeconds to evict pods that are older than specified seconds. Lastly, you can specify the optional parameter excludeOwnerKinds and if a pod has any of these Kinds listed as an OwnerRef, that pod will not be considered for eviction.

Parameters:

Example:

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
strategies:
  "RemoveFailedPods":
     enabled: true
     params:
       failedPods:
         reasons:
         - "NodeAffinity"
         includingInitContainers: true
         excludeOwnerKinds:
         - "Job"
         minPodLifetimeSeconds: 3600

The following strategies accept a namespaces parameter which allows to specify a list of including, resp. excluding namespaces:

• PodLifeTime
• RemovePodsHavingTooManyRestarts
• RemovePodsViolatingNodeTaints
• RemovePodsViolatingNodeAffinity
• RemovePodsViolatingInterPodAntiAffinity
• RemoveDuplicates
• RemovePodsViolatingTopologySpreadConstraint
• RemoveFailedPods

For example:

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
strategies:
  "PodLifeTime":
     enabled: true
     params:
        podLifeTime:
          maxPodLifeTimeSeconds: 86400
        namespaces:
          include:
          - "namespace1"
          - "namespace2"

In the examples PodLifeTime gets executed only over namespace1 and namespace2. The similar holds for exclude field:

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
strategies:
  "PodLifeTime":
     enabled: true
     params:
        podLifeTime:
          maxPodLifeTimeSeconds: 86400
        namespaces:
          exclude:
          - "namespace1"
          - "namespace2"

The strategy gets executed over all namespaces but namespace1 and namespace2.

It's not allowed to compute include with exclude field.

All strategies are able to configure a priority threshold, only pods under the threshold can be evicted. You can specify this threshold by setting thresholdPriorityClassName(setting the threshold to the value of the given priority class) or thresholdPriority(directly setting the threshold) parameters. By default, this threshold is set to the value of system-cluster-critical priority class.

Note: Setting evictSystemCriticalPods to true disables priority filtering entirely.

E.g.

Setting thresholdPriority

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
strategies:
  "PodLifeTime":
     enabled: true
     params:
        podLifeTime:
          maxPodLifeTimeSeconds: 86400
        thresholdPriority: 10000

Setting thresholdPriorityClassName

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
strategies:
  "PodLifeTime":
     enabled: true
     params:
        podLifeTime:
          maxPodLifeTimeSeconds: 86400
        thresholdPriorityClassName: "priorityclass1"

Note that you can't configure both thresholdPriority and thresholdPriorityClassName, if the given priority class does not exist, descheduler won't create it and will throw an error.

The following strategies can configure a standard kubernetes labelSelector to filter pods by their labels:

• PodLifeTime
• RemovePodsHavingTooManyRestarts
• RemovePodsViolatingNodeTaints
• RemovePodsViolatingNodeAffinity
• RemovePodsViolatingInterPodAntiAffinity
• RemovePodsViolatingTopologySpreadConstraint
• RemoveFailedPods

This allows running strategies among pods the descheduler is interested in.

For example:

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
strategies:
  "PodLifeTime":
    enabled: true
    params:
      podLifeTime:
        maxPodLifeTimeSeconds: 86400
      labelSelector:
        matchLabels:
          component: redis
        matchExpressions:
          - {key: tier, operator: In, values: [cache]}
          - {key: environment, operator: NotIn, values: [dev]}

The following strategies accept a nodeFit boolean parameter which can optimize descheduling:

• RemoveDuplicates
• LowNodeUtilization
• HighNodeUtilization
• RemovePodsViolatingInterPodAntiAffinity
• RemovePodsViolatingNodeAffinity
• RemovePodsViolatingNodeTaints
• RemovePodsViolatingTopologySpreadConstraint
• RemovePodsHavingTooManyRestarts
• RemoveFailedPods

If set to true the descheduler will consider whether or not the pods that meet eviction criteria will fit on other nodes before evicting them. If a pod cannot be rescheduled to another node, it will not be evicted. Currently the following criteria are considered when setting nodeFit to true:

• A nodeSelector on the pod
• Any tolerations on the pod and any taints on the other nodes
• nodeAffinity on the pod
• Resource requests made by the pod and the resources available on other nodes
• Whether any of the other nodes are marked as unschedulable

E.g.

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
strategies:
  "LowNodeUtilization":
    enabled: true
    params:
      nodeFit: true
      nodeResourceUtilizationThresholds:
        thresholds:
          "cpu": 20
          "memory": 20
          "pods": 20
        targetThresholds:
          "cpu": 50
          "memory": 50
          "pods": 50

Note that node fit filtering references the current pod spec, and not that of it's owner. Thus, if the pod is owned by a ReplicationController (and that ReplicationController was modified recently), the pod may be running with an outdated spec, which the descheduler will reference when determining node fit. This is expected behavior as the descheduler is a "best-effort" mechanism.

Using Deployments instead of ReplicationControllers provides an automated rollout of pod spec changes, therefore ensuring that the descheduler has an up-to-date view of the cluster state.

When the descheduler decides to evict pods from a node, it employs the following general mechanism:

• Critical pods (with priorityClassName set to system-cluster-critical or system-node-critical) are never evicted (unless evictSystemCriticalPods: true is set).
• Pods (static or mirrored pods or standalone pods) not part of an ReplicationController, ReplicaSet(Deployment), StatefulSet, or Job are never evicted because these pods won't be recreated. (Standalone pods in failed status phase can be evicted by setting evictFailedBarePods: true)
• Pods associated with DaemonSets are never evicted.
• Pods with local storage are never evicted (unless evictLocalStoragePods: true is set).
• Pods with PVCs are evicted (unless ignorePvcPods: true is set).
• In LowNodeUtilization and RemovePodsViolatingInterPodAntiAffinity, pods are evicted by their priority from low to high, and if they have same priority, best effort pods are evicted before burstable and guaranteed pods.
• All types of pods with the annotation descheduler.alpha.kubernetes.io/evict are eligible for eviction. This annotation is used to override checks which prevent eviction and users can select which pod is evicted. Users should know how and if the pod will be recreated.
• Pods with a non-nil DeletionTimestamp are not evicted by default.

Setting --v=4 or greater on the Descheduler will log all reasons why any pod is not evictable.

Pods subject to a Pod Disruption Budget(PDB) are not evicted if descheduling violates its PDB. The pods are evicted by using the eviction subresource to handle PDB.

In High Availability mode, Descheduler starts leader election process in Kubernetes. You can activate HA mode if you choose to deploy your application as Deployment.

Deployment starts with 1 replica by default. If you want to use more than 1 replica, you must consider enable High Availability mode since we don't want to run descheduler pods simultaneously.

The leader election process can be enabled by setting --leader-elect in the CLI. You can also set --set=leaderElection.enabled=true flag if you are using Helm.

To get best results from HA mode some additional configurations might require:

• Configure a podAntiAffinity rule if you want to schedule onto a node only if that node is in the same zone as at least one already-running descheduler
• Set the replica count greater than 1

The metrics are served through https://localhost:10258/metrics by default. The address and port can be changed by setting --binding-address and --secure-port flags.

The below compatibility matrix shows the k8s client package(client-go, apimachinery, etc) versions that descheduler is compiled with. At this time descheduler does not have a hard dependency to a specific k8s release. However a particular descheduler release is only tested against the three latest k8s minor versions. For example descheduler v0.18 should work with k8s v1.18, v1.17, and v1.16.

Starting with descheduler release v0.18 the minor version of descheduler matches the minor version of the k8s client packages that it is compiled with.

Are you interested in contributing to descheduler? We, the maintainers and community, would love your suggestions, contributions, and help! Also, the maintainers can be contacted at any time to learn more about how to get involved.

To get started writing code see the contributor guide in the /docs directory.

In the interest of getting more new people involved we tag issues with [good first issue][good_first_issue]. These are typically issues that have smaller scope but are good ways to start to get acquainted with the codebase.

We also encourage ALL active community participants to act as if they are maintainers, even if you don't have "official" write permissions. This is a community effort, we are here to serve the Kubernetes community. If you have an active interest and you want to get involved, you have real power! Don't assume that the only people who can get things done around here are the "maintainers".

We also would love to add more "official" maintainers, so show us what you can do!

This repository uses the Kubernetes bots. See a full list of the commands [here][prow].

You can reach the contributors of this project at:

• Slack channel
• Mailing list

Learn how to engage with the Kubernetes community on the community page.

This roadmap is not in any particular order.

• Consideration of pod affinity
• Strategy to consider number of pending pods
• Integration with cluster autoscaler
• Integration with metrics providers for obtaining real load metrics
• Consideration of Kubernetes's scheduler's predicates

Participation in the Kubernetes community is governed by the Kubernetes Code of Conduct.