Descheduler for Kubernetes

Introduction

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 first time for scheduling. As Kubernetes clusters are very dynamic and their state change over time, there may be desired 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.

Build and Run

Build descheduler:

$ make

and run descheduler:

$ ./_output/bin/descheduler --kubeconfig <path to kubeconfig> --policy-config-file <path-to-policy-file>

For more information about available options run:

$ ./_output/bin/descheduler --help

Running Descheduler as a Job Inside of a Pod

Descheduler can be run as a job inside of a pod. It has the advantage of being able to be run multiple times without needing user intervention. Descheduler pod is run as a critical pod to avoid being evicted by itself, or by kubelet due to an eviction event. Since critical pods are created in kube-system namespace, descheduler job and its pod will also be created in kube-system namespace.

Create a container image

First we create a simple Docker image utilizing the Dockerfile found in the root directory:

$ make image

Create a cluster role

To give necessary permissions for the descheduler to work in a pod, create a cluster role:

$ cat << EOF| kubectl create -f -
kind: ClusterRole
apiVersion: rbac.authorization.k8s.io/v1beta1
metadata:
  name: descheduler-cluster-role
rules:
- apiGroups: [""]
  resources: ["nodes"]
  verbs: ["get", "watch", "list"]
- apiGroups: [""]
  resources: ["pods"]
  verbs: ["get", "watch", "list", "delete"]
EOF

Create the service account which will be used to run the job:

$ kubectl create sa descheduler-sa -n kube-system

Bind the cluster role to the service account:

$ kubectl create clusterrolebinding descheduler-cluster-role-binding \
    --clusterrole=descheduler-cluster-role \
    --serviceaccount=kube-system:descheduler-sa

Create a configmap to store descheduler policy

Descheduler policy is created as a ConfigMap in kube-system namespace so that it can be mounted as a volume inside pod.

$ kubectl create configmap descheduler-policy-configmap \
     -n kube-system --from-file=<path-to-policy-dir/policy.yaml>

Create the job specification (descheduler-job.yaml)

apiVersion: batch/v1
kind: Job
metadata:
  name: descheduler-job
  namespace: kube-system
spec:
  parallelism: 1
  completions: 1
  template:
    metadata:
      name: descheduler-pod
      annotations:
        scheduler.alpha.kubernetes.io/critical-pod: "true"
    spec:
        containers:
        - name: descheduler
          image: descheduler
          volumeMounts:
          - mountPath: /policy-dir
            name: policy-volume
          command:
          - "/bin/sh"
          - "-ec"
          - |
            /bin/descheduler --policy-config-file /policy-dir/policy.yaml
        restartPolicy: "Never"
        serviceAccountName: descheduler-sa
        volumes:
        - name: policy-volume
          configMap:
            name: descheduler-policy-configmap

Please note that the pod template is configured with critical pod annotation, and the policy policy-file is mounted as a volume from the config map.

Run the descheduler as a job in a pod:

$ kubectl create -f descheduler-job.yaml

Policy and Strategies

Descheduler's policy is configurable and includes strategies to be enabled or disabled. Two strategies, RemoveDuplicates and LowNodeUtilization are currently implemented. As part of the policy, the parameters associated with the strategies can be configured too. By default, all strategies are enabled.

RemoveDuplicates

This strategy makes sure that there is only one pod associated with a Replica Set (RS), Replication Controller (RC), Deployment, or Job running on 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 RS or RC, for example, running on same node. Once the failed nodes are ready again, this strategy could be enabled to evict those duplicate pods. Currently, there are no parameters associated with this strategy. To disable this strategy, the policy would look like:

apiVersion: "descheduler/v1alpha1"
kind: "DeschedulerPolicy"
strategies:
  "RemoveDuplicates":
     enabled: false

LowNodeUtilization

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, and number of pods in terms of percentage. If a node's usage is below threshold for all (cpu, memory, and number of pods), 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. 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. An example of the policy for this strategy would look like:

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

There is another parameter associated with LowNodeUtilization strategy, called numberOfNodes. This parameter can be configured to activate the strategy only when 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.

Pod Evictions

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

• Critical pods (with annotations scheduler.alpha.kubernetes.io/critical-pod) are never evicted.
• Pods (static or mirrored pods or stand alone pods) not part of an RC, RS, Deployment or Jobs are never evicted because these pods won't be recreated.
• Pods associated with DaemonSets are never evicted.
• Pods with local storage are never evicted.
• Best efforts pods are evicted before Burstable and Guaranteed pods.

Pod disruption Budget (PDB)

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

Roadmap

This roadmap is not in any particular order.

• Addition of test cases (unit and end-to-end)
• Ability to run inside a pod as a job
• Strategy to consider taints and tolerations
• Consideration of pod affinity and anti-affinity
• Strategy to consider pod life time
• 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

Note

This project is under active development, and is not intended for production use. Any api could be changed any time with out any notice. That said, your feedback is very important and appreciated to make this project more stable and useful.