Skip to content

Latest commit

 

History

History
126 lines (107 loc) · 5.42 KB

user-guide.md

File metadata and controls

126 lines (107 loc) · 5.42 KB
Raw

User Guide

Starting with descheduler release v0.10.0 container images are available in the official k8s container registry.

Descheduler Version Container Image Architectures
v0.32.0 registry.k8s.io/descheduler/descheduler:v0.32.0 AMD64
ARM64
ARMv7
v0.31.0 registry.k8s.io/descheduler/descheduler:v0.31.0 AMD64
ARM64
ARMv7
v0.30.1 registry.k8s.io/descheduler/descheduler:v0.30.1 AMD64
ARM64
ARMv7
v0.30.0 registry.k8s.io/descheduler/descheduler:v0.30.0 AMD64
ARM64
ARMv7
v0.29.0 registry.k8s.io/descheduler/descheduler:v0.29.0 AMD64
ARM64
ARMv7

Note that multi-arch container images cannot be pulled by kind from a registry. Therefore starting with descheduler release v0.20.0 use the below process to download the official descheduler image into a kind cluster.

kind create cluster
docker pull registry.k8s.io/descheduler/descheduler:v0.20.0
kind load docker-image registry.k8s.io/descheduler/descheduler:v0.20.0

Policy Configuration Examples

The examples directory has descheduler policy configuration examples.

CLI Options

The descheduler has many CLI options that can be used to override its default behavior. Please check the CLI Options documentation for details

Production Use Cases

This section contains descriptions of real world production use cases.

Balance Cluster By Pod Age

When initially migrating applications from a static virtual machine infrastructure to a cloud native k8s infrastructure there can be a tendency to treat application pods like static virtual machines. One approach to help prevent developers and operators from treating pods like virtual machines is to ensure that pods only run for a fixed amount of time.

The PodLifeTime strategy can be used to ensure that old pods are evicted. It is recommended to create a pod disruption budget for each application to ensure application availability.

descheduler -v=3 --evict-local-storage-pods --policy-config-file=pod-life-time.yml

This policy configuration file ensures that pods created more than 7 days ago are evicted.

---
apiVersion: "descheduler/v1alpha2"
kind: "DeschedulerPolicy"
profiles:
  - name: ProfileName
    pluginConfig:
    - name: "PodLifeTime"
      args:
        maxPodLifeTimeSeconds: 604800
    plugins:
      deschedule:
        enabled:
          - "PodLifeTime"

Balance Cluster By Node Memory Utilization

If your cluster has been running for a long period of time, you may find that the resource utilization is not very balanced. The following two strategies can be used to rebalance your cluster based on cpu, memory or number of pods.

Balance high utilization nodes

Using LowNodeUtilization, descheduler will rebalance the cluster based on memory by evicting pods from nodes with memory utilization over 70% to nodes with memory utilization below 20%.

apiVersion: "descheduler/v1alpha2"
kind: "DeschedulerPolicy"
profiles:
  - name: ProfileName
    pluginConfig:
    - name: "LowNodeUtilization"
      args:
        thresholds:
          "memory": 20
        targetThresholds:
          "memory": 70
    plugins:
      balance:
        enabled:
          - "LowNodeUtilization"

Balance low utilization nodes

Using HighNodeUtilization, descheduler will rebalance the cluster based on memory by evicting pods from nodes with memory utilization lower than 20%. This should be use NodeResourcesFit with the MostAllocated scoring strategy based on these doc. The evicted pods will be compacted into minimal set of nodes.

apiVersion: "descheduler/v1alpha2"
kind: "DeschedulerPolicy"
profiles:
  - name: ProfileName
    pluginConfig:
    - name: "HighNodeUtilization"
      args:
        thresholds:
          "memory": 20
    plugins:
      balance:
        enabled:
          - "HighNodeUtilization"

Autoheal Node Problems

Descheduler's RemovePodsViolatingNodeTaints strategy can be combined with Node Problem Detector and Cluster Autoscaler to automatically remove Nodes which have problems. Node Problem Detector can detect specific Node problems and report them to the API server. There is a feature called TaintNodeByCondition of the node controller that takes some conditions and turns them into taints. Currently, this only works for the default node conditions: PIDPressure, MemoryPressure, DiskPressure, Ready, and some cloud provider specific conditions. The Descheduler will then deschedule workloads from those Nodes. Finally, if the descheduled Node's resource allocation falls below the Cluster Autoscaler's scale down threshold, the Node will become a scale down candidate and can be removed by Cluster Autoscaler. These three components form an autohealing cycle for Node problems.

NOTE

Once kubernetes/node-problem-detector#565 is available in NPD, we need to update this section.