Workload Rebalance

In general, once workload replicas are scheduled, the scheduling result remains fixed, and the replica propagation does not change. Even if rescheduling is triggered by modifying replicas or placement configuration, the system will try to maintain the existing replica propagation as closely as possible, making only minimal adjustments when necessary. This approach minimizes disruptions and preserves balance across clusters.

However, in some scenarios, users may wish to actively trigger a fresh rescheduling, disregarding the previous propagation entirely to establish a new replica propagation across clusters.

Applicable Scenarios

Scenario 1

In a cluster failover scenario, replicas are propagated across two clusters, member1 and member2. If the member1 cluster fails, all replicas would migrate to the member2 cluster.

As a cluster administrator, I would like the replicas to be repropagated across both clusters when member1 recovers, allowing the resources of the member1 cluster to be re-utilized and enhancing high availability.

Scenario 2

In application-level failover, low-priority applications may be preempted, resulting in shrinking from multi clusters to single cluster due to limited cluster resources (refer to Application-level Failover).

As a user, I would like the replicas of low-priority applications to be repropagated across multiple clusters when resources are sufficient, ensuring high availability for the application.

Scenario 3

In the Aggregated scheduling type, replicas may still be propagated across multiple clusters due to resource constraints.

As a user, I would like the replicas to be propagated in an aggregated strategy when any cluster has sufficient resource to accommodate all replicas, so that the application better meets actual business requirements.

Scenario 4

In a disaster-recovery scenario, replicas migrate from a primary cluster to a backup cluster in the event of a primary cluster failure.

As a cluster administrator, I would like replicas to migrate back once the primary cluster is restored, so that:

  1. disaster-recovery mode is reestablished, to ensure the reliability and stability of the cluster federation.
  2. backup cluster costs are reduced.

Feature of WorkloadRebalancer

Reschedule a workload

Assume that you want to trigger the rescheduling of a deployment named demo-deploy. You can just apply the following WorkloadRebalancer.

  1. apiVersion: apps.karmada.io/v1alpha1
  2. kind: WorkloadRebalancer
  3. metadata:
  4.   name: demo
  5. spec:
  6.   workloads:
  7.     - apiVersion: apps/v1
  8.       kind: Deployment
  9.       name: demo-deploy
  10.       namespace: default

Then, scheduler will reschedule this deployment, which disregards the previous propagation entirely and seeks to establish an entirely new replica propagation across clusters.

  • If it succeeds, you will get the following result:
  1. apiVersion: apps.karmada.io/v1alpha1
  2. kind: WorkloadRebalancer
  3. metadata:
  4.   name: demo
  5.   generation: 1
  6.   creationTimestamp: "2024-05-22T11:16:10Z"
  7. spec:
  8.   ...
  9. status:
  10.   finishTime: "2024-05-22T11:16:10Z"
  11.   observedGeneration: 1
  12.   observedWorkloads:
  13.     - result: Successful
  14.       workload:
  15.         apiVersion: apps/v1
  16.         kind: Deployment
  17.         name: demo-deploy
  18.         namespace: default
  • If the resource binding of deployments/demo-deploy does not exist, you will get the following result:
  1. apiVersion: apps.karmada.io/v1alpha1
  2. kind: WorkloadRebalancer
  3. metadata:
  4.   name: demo
  5.   generation: 1
  6.   creationTimestamp: "2024-05-22T11:16:10Z"
  7. spec:
  8.   ...
  9. status:
  10.   finishTime: "2024-05-22T11:16:10Z"
  11.   observedGeneration: 1
  12.   observedWorkloads:
  13.     - reason: ReferencedBindingNotFound
  14.       result: Failed
  15.       workload:
  16.         apiVersion: apps/v1
  17.         kind: Deployment
  18.         name: demo-deploy
  19.         namespace: default
  • If an abnormal failure occurs during processing, such as a network problem or a throttling issue, the WorkloadRebalancer will continue to retry and you will get the following result:
  1. apiVersion: apps.karmada.io/v1alpha1
  2. kind: WorkloadRebalancer
  3. metadata:
  4.   name: demo
  5.   generation: 1
  6.   creationTimestamp: "2024-05-22T11:26:10Z"
  7. spec:
  8.   ...
  9. status:
  10.   observedGeneration: 1
  11.   observedWorkloads:
  12.     - workload:
  13.         apiVersion: apps/v1
  14.         kind: Deployment
  15.         name: demo-deploy
  16.         namespace: default

tip: there will be no finishTime in status field, nor result/reason field in each observed workload in this case since it still retrying.

Reschedule a batch of resources

In actual scenarios, you need to trigger rescheduling in the application dimension, that is, you need to trigger rescheduling of a batch of resources. Assuming the resources are deployment/demo-deploy, configmap/demo-config and clusterrole/demo-role, you can define the WorkloadRebalancer like this:

  1. apiVersion: apps.karmada.io/v1alpha1
  2. kind: WorkloadRebalancer
  3. metadata:
  4.   name: demo
  5. spec:
  6.   workloads:
  7.     - apiVersion: apps/v1
  8.       kind: Deployment
  9.       name: demo-deploy
  10.       namespace: default
  11.     - apiVersion: v1
  12.       kind: ConfigMap
  13.       name: demo-config
  14.       namespace: default
  15.     - apiVersion: rbac.authorization.k8s.io/v1
  16.       kind: ClusterRole
  17.       name: demo-role

You will get the following result:

  1. apiVersion: apps.karmada.io/v1alpha1
  2. kind: WorkloadRebalancer
  3. metadata:
  4.   name: demo
  5.   generation: 1
  6.   creationTimestamp: "2024-05-22T11:36:10Z"
  7. spec:
  8.   ...
  9. status:
  10.   finishTime: "2024-05-22T11:36:10Z"
  11.   observedGeneration: 1
  12.   observedWorkloads:
  13.     - result: Successful
  14.       workload:
  15.         apiVersion: apps/v1
  16.         kind: Deployment
  17.         name: demo-deploy
  18.         namespace: default
  19.     - result: Successful
  20.       workload:
  21.         apiVersion: rbac.authorization.k8s.io/v1
  22.         kind: ClusterRole
  23.         name: demo-role
  24.     - result: Successful
  25.       workload:
  26.         apiVersion: v1
  27.         kind: ConfigMap
  28.         name: demo-config
  29.         namespace: default

tip: the observedWorkloads is ordered according to the dictionary order of apiVersion, kind, namespace, and name.

Modify WorkloadRebalancer

The rebalancer also supports updates, the guideline is:

  • if a new workload is added to spec list, just add it into the status list too and do the rebalance.
  • if a workload is deleted from the previous spec list, keep it in status list if it had already succeeded, and remove it if not.
  • if a workload is modified, just regard it as a deletion of the old one and an insertion of a new one.

Assume the current WorkloadRebalancer is as follows:

  1. apiVersion: apps.karmada.io/v1alpha1
  2. kind: WorkloadRebalancer
  3. metadata:
  4.   name: demo
  5.   generation: 1
  6.   creationTimestamp: "2024-05-22T11:36:10Z"
  7. spec:
  8.   workloads:
  9.     - apiVersion: apps/v1
  10.       kind: Deployment
  11.       name: demo-deploy-1
  12.       namespace: default
  13.     - apiVersion: apps/v1
  14.       kind: Deployment
  15.       name: demo-deploy-2
  16.       namespace: default
  17. status:
  18.   finishTime: "2024-05-22T11:36:10Z"
  19.   observedGeneration: 1
  20.   observedWorkloads:
  21.     - result: Successful
  22.       workload:
  23.         apiVersion: apps/v1
  24.         kind: Deployment
  25.         name: demo-deploy-1
  26.         namespace: default
  27.     - reason: ReferencedBindingNotFound
  28.       result: Failed
  29.       workload:
  30.         apiVersion: apps/v1
  31.         kind: Deployment
  32.         name: demo-deploy-2
  33.         namespace: default

Now, if I edit target workloads from demo-deploy-1 + demo-deploy-2 to only demo-deploy-3, the result will be:

  1. apiVersion: apps.karmada.io/v1alpha1
  2. kind: WorkloadRebalancer
  3. metadata:
  4.   name: demo
  5.   generation: 2
  6.   creationTimestamp: "2024-05-22T11:36:10Z"
  7. spec:
  8.   workloads:
  9.     - apiVersion: apps/v1
  10.       kind: Deployment
  11.       name: demo-deploy-3
  12.       namespace: default
  13. status:
  14.   finishTime: "2024-05-22T11:40:10Z"
  15.   observedGeneration: 2
  16.   observedWorkloads:
  17.     - result: Successful
  18.       workload:
  19.         apiVersion: apps/v1
  20.         kind: Deployment
  21.         name: demo-deploy-1
  22.         namespace: default
  23.     - result: Successful
  24.       workload:
  25.         apiVersion: apps/v1
  26.         kind: Deployment
  27.         name: demo-deploy-3
  28.         namespace: default

You can see in status.observedWorkloads:

  • demo-deploy-1 is not specified in latest spec, but it had already succeeded, so it keep existing in status.
  • demo-deploy-2 is not specified in latest spec and it once failed, so it is removed from status.
  • demo-deploy-3 is newly added in latest spec, so it is added to status.

Auto clean WorkloadRebalancer

You can use spec.ttlSecondsAfterFinished to specify when the object will be automatically deleted after it finished execution (finished execution means each target workload is finished with result of Successful or Failed).

The guideline is:

  • If this field is set, ttlSecondsAfterFinished seconds after the WorkloadRebalancer finishes, it will be automatically deleted.
  • If this field is unset, the WorkloadRebalancer won’t be automatically deleted.
  • If this field is set to zero, the WorkloadRebalancer will be deleted immediately right after it finishes.

Here is an example:

  1. apiVersion: apps.karmada.io/v1alpha1
  2. kind: WorkloadRebalancer
  3. metadata:
  4.   name: demo
  5. spec:
  6.   ttlSecondsAfterFinished: 60
  7.   workloads:
  8.     - apiVersion: apps/v1
  9.       kind: Deployment
  10.       name: demo-deploy
  11.       namespace: default

Then, it will be deleted 60 seconds after the WorkloadRebalancer finished execution.

What’s next

For a detailed demo of the workload rebalancer you can refer to the tutorial Workload Rebalancer