Skip to main content
Version: v1.12

Workload Rebalancer

Objectives

In general, once workload replicas are scheduled, the scheduling result remains fixed, and the replica propagation does not change. Now, assuming that you want to actively trigger a fresh rescheduling in a special case, you can achieve it by WorkloadRebalancer.

This guide will cover how to trigger a rescheduling using WorkloadRebalancer.

Prerequisites

Multi-cluster Karmada has been installed

Run the commands:

git clone https://github.com/karmada-io/karmada
cd karmada
hack/local-up-karmada.sh
export KUBECONFIG=~/.kube/karmada.config:~/.kube/members.config

Note:

Before following this guide, you should have created at least three kubernetes clusters, one will be used to host Karmada control plane, and the rest will be member clusters. For convenience, we use hack/local-up-karmada.sh script to quickly prepare the clusters mentioned.

Once the script is executed, you will see Karmada control plane installed with multi member clusters.

Tutorial

Step 1: Create a Deployment

First, create a Deployment named foo, filling a new file deployment.yaml with the following content:

deployment.yaml
apiVersion: apps/v1
kind: Deployment
metadata:
name: foo
labels:
app: test
spec:
replicas: 3
selector:
matchLabels:
app: foo
template:
metadata:
labels:
app: foo
spec:
terminationGracePeriodSeconds: 0
containers:
- image: nginx
name: foo
resources:
limits:
cpu: 10m
memory: 10Mi
---
apiVersion: policy.karmada.io/v1alpha1
kind: PropagationPolicy
metadata:
name: default-pp
spec:
placement:
clusterTolerations:
- effect: NoExecute
key: workload-rebalancer-test
operator: Exists
tolerationSeconds: 0
clusterAffinity:
clusterNames:
- member1
- member2
replicaScheduling:
replicaDivisionPreference: Weighted
replicaSchedulingType: Divided
weightPreference:
dynamicWeight: AvailableReplicas
resourceSelectors:
- apiVersion: apps/v1
kind: Deployment
name: foo
namespace: default

Then run the following command to create those resources:

kubectl --context karmada-apiserver apply -f deployment.yaml

You can check whether this step succeeded like this:

$ karmadactl --karmada-context karmada-apiserver get deploy foo
NAME CLUSTER READY UP-TO-DATE AVAILABLE AGE ADOPTION
foo member1 2/2 2 2 20s Y
foo member2 1/1 1 1 20s Y

Thus, 2 replicas were propagated to member1 cluster and 1 replica was propagated to member2 cluster.

Step 2: Add NoExecute taint to member1 cluster to mock cluster failover

  • Run the following command to add NoExecute taint to member1 cluster:
$ karmadactl --karmada-context=karmada-apiserver taint clusters member1 workload-rebalancer-test:NoExecute
cluster/member1 tainted

Then, rescheduling will be triggered due to cluster failover, and all replicas will be propagated to member2 cluster, as you can see:

$ karmadactl --karmada-context karmada-apiserver get deploy foo
NAME CLUSTER READY UP-TO-DATE AVAILABLE AGE ADOPTION
foo member2 3/3 3 3 57s Y
  • Run the following command to remove the previous NoExecute taint from member1 cluster:
$ karmadactl --karmada-context=karmada-apiserver taint clusters member1 workload-rebalancer-test:NoExecute-
cluster/member1 untainted

Removing the taint will not cause replicas to repropagate, because the scheduling result stays fixed, so all replicas will keep being in member2 cluster.

Step 3. Apply a WorkloadRebalancer to trigger rescheduling

In order to trigger the rescheduling of the above resources, you can create a new file workload-rebalancer.yaml with the following content:

apiVersion: apps.karmada.io/v1alpha1
kind: WorkloadRebalancer
metadata:
name: demo
spec:
workloads:
- apiVersion: apps/v1
kind: Deployment
name: foo
namespace: default

Then run the following command to apply it:

kubectl --context karmada-apiserver apply -f workload-rebalancer.yaml

You will get a workloadrebalancer.apps.karmada.io/demo created result, which means the API was created successfully.

Step 4: Check the status of WorkloadRebalancer.

Run the following command:

$ kubectl --context karmada-apiserver get workloadrebalancer demo -o yaml
apiVersion: apps.karmada.io/v1alpha1
kind: WorkloadRebalancer
metadata:
creationTimestamp: "2024-05-25T09:49:51Z"
generation: 1
name: demo
spec:
workloads:
- apiVersion: apps/v1
kind: Deployment
name: foo
namespace: default
status:
finishTime: "2024-05-25T09:49:51Z"
observedGeneration: 1
observedWorkloads:
- result: Successful
workload:
apiVersion: apps/v1
kind: Deployment
name: foo
namespace: default

Thus, you can observe the rescheduling result at status.observedWorkloads field of workloadrebalancer/demo. As you can see, deployment/foo was rescheduled successfully.

Step 5: Observe the real effect of WorkloadRebalancer

You can observe the propagation status of deployment/foo replicas:

$ karmadactl --karmada-context karmada-apiserver get deploy foo
NAME CLUSTER READY UP-TO-DATE AVAILABLE AGE ADOPTION
foo member1 2/2 2 2 3m14s Y
foo member2 1/1 1 1 4m37s Y

As you see, rescheduling was triggered and 2 replicas migrated back to member1 cluster, while 1 replica in member2 cluster remains unchanged.

Besides, you can observe that a schedule event was emitted by default-scheduler:

$ kubectl --context karmada-apiserver describe deployment foo
...
Events:
Type Reason Age From Message
---- ------ ---- ---- -------
...
Normal ScheduleBindingSucceed 3m34s (x2 over 4m57s) default-scheduler Binding has been scheduled successfully. Result: {member1:2, member2:1}
Normal AggregateStatusSucceed 3m20s (x20 over 4m57s) resource-binding-status-controller Update resourceBinding(default/foo-deployment) with AggregatedStatus successfully.
...

Step 6: Update and auto-clean WorkloadRebalancer

Assuming you want the WorkloadRebalancer resource to be auto cleaned in the future, you can just edit it and set spec.ttlSecondsAfterFinished field to 300, just like:

apiVersion: apps.karmada.io/v1alpha1
kind: WorkloadRebalancer
metadata:
name: demo
spec:
ttlSecondsAfterFinished: 300
workloads:
- apiVersion: apps/v1
kind: Deployment
name: foo
namespace: default

Once this modification has been applied, this WorkloadRebalancer resource will be auto deleted after 300 seconds.