CruiseControlOperation to manage Cruise Control

Koperator version 0.22 introduces the CruiseControlOperation custom resource. Koperator executes the Cruise Control related task based on the state of the CruiseControlOperation custom resource. This gives you better control over Cruise Control, improving reliability, configurability, and observability.

Overview

When a broker is added or removed from the Kafka cluster or when new storage is added for a broker, Koperator creates a CruiseControlOperation custom resource. This custom resource describes a task that Cruise Control executes to move the partitions. Koperator watches the created CruiseControlOperation custom resource and updates its state based on the result of the Cruise Control task. Koperator can re-execute the task if it fails.

Cruise Control can execute only one task at a time, so the priority of the tasks depends on the type of the operation:

  • Upscale operations are executed first, then
  • downscale operations, then
  • rebalance operations.

The following Cruise Control tasks are supported:

You can follow the progress of the operation through the KafkaCluster custom resource’s status and through the CruiseControlOperation custom resource’s status.
The following example shows the steps of an add_broker (GracefulUpscale*) operation, but the same applies for the Kafka cluster remove_broker (GracefulDownScale*) and rebalance (when the volumeState is GracefulDiskRebalance*) operations.

  1. Upscale the Kafka cluster by adding a new broker with id “3” into the KafkaCluster CR:

    spec:
    ...  
      brokers:
        - id: 0
          brokerConfigGroup: "default"
        - id: 1
          brokerConfigGroup: "default"
        - id: 2
          brokerConfigGroup: "default"
        - id: 3
          brokerConfigGroup: "default"
    ...
    
  2. A new broker pod is created and the cruiseControlOperationReference is added to the KafkaCluster status.
    This is the reference of the created CruiseControlOperation custom resource.
    The cruiseControlState shows the CruiseControlOperation state: GracefulUpscaleScheduled, meaning that CruiseControlOperation has been created and is waiting for the add_broker task to be finished.

    status:
    ...
      brokersState:
        "3":
         ...
          gracefulActionState:
            cruiseControlOperationReference:
              name: kafka-addbroker-mhh72
            cruiseControlState: GracefulUpscaleScheduled
            volumeStates:
              /kafka-logs:
                cruiseControlOperationReference:
                  name: kafka-rebalance-h6ntt
                cruiseControlVolumeState: GracefulDiskRebalanceScheduled
              /kafka-logs2:
                cruiseControlOperationReference:
                  name: kafka-rebalance-h6ntt
                cruiseControlVolumeState: GracefulDiskRebalanceScheduled
        ...
    
  3. The add_broker Cruise Control task is in progress:

    status:
    ...
      brokersState:
        "3":
         ...
          gracefulActionState:
            cruiseControlOperationReference:
              name: kafka-addbroker-mhh72
            cruiseControlState: GracefulUpscaleRunning
        ...
    
  4. When the add_broker Cruise Control task is completed:

    status:
    ...
      brokersState:
        "3":
         ...
          gracefulActionState:
            cruiseControlOperationReference:
              name: kafka-addbroker-mhh72
            cruiseControlState: GracefulUpscaleSucceeded
        ...
    

There are two other possible states of cruiseControlState, GracefulUpscaleCompletedWithError and GracefulUpscalePaused.

  • GracefulUpscalePaused is a special state. For details, see Control the created CruiseControlOperation.

  • The GracefulUpscaleCompletedWithError occurs when the Cruise Control task fails. If the cruiseControlOperation.spec.errorPolicy is set to retry (which is the default value), Koperator re-executes the failed task every 30s until it succeeds. During the re-execution the cruiseControlState returns to GracefulUpscaleRunning.

    status:
    ...
      brokersState:
        "3":
         ...
          gracefulActionState:
            cruiseControlOperationReference:
              name: kafka-addbroker-mhh72
            cruiseControlState: GracefulUpscaleCompletedWithError
        ...
    

CruiseControlOperation CR overview

The kafka-addbroker-mhh72 CruiseControlOperation custom resource from the previous example looks like:

kind: CruiseControlOperation
metadata:
...
  name: kafka-addbroker-mhh72
...
spec:
...
status:
  currentTask:
    finished: "2022-11-18T09:31:40Z"
    httpRequest: http://kafka-cruisecontrol-svc.kafka.svc.cluster.local:8090/kafkacruisecontrol/add_broker?allow_capacity_estimation=true&brokerid=3&data_from=VALID_WINDOWS&dryrun=false&exclude_recently_demoted_brokers=true&exclude_recently_removed_brokers=true&json=true&use_ready_default_goals=true
    httpResponseCode: 200
    id: 222e30f0-1e7a-4c87-901c-bed2854d69b7
    operation: add_broker
    parameters:
      brokerid: "3"
      exclude_recently_demoted_brokers: "true"
      exclude_recently_removed_brokers: "true"
    started: "2022-11-18T09:30:48Z"
    state: Completed
    summary:
      Data to move: "0"
      Intra broker data to move: "0"
      Number of intra broker replica movements: "0"
      Number of leader movements: "0"
      Number of replica movements: "36"
      Provision recommendation: '[ReplicaDistributionGoal] Remove at least 4 brokers.'
      Recent windows: "1" 
  errorPolicy: retry
  retryCount: 0
  • The status.currentTask describes the Cruise Control task.
  • The httpRequest field contains the whole POST HTTP request that has been executed.
  • The id is the Cruise Control task identifier number.
  • The state shows the progress of the request.
  • The summary is Cruise Control’s optimization proposal. It shows the scope of the changes that Cruise Control will apply through the operation.
  • The retryCount field shows the number of retries when a task has failed and cruiseControlOperation.spec.errorPolicy is set to retry. In this case, the status.failedTask field shows the history of the failed tasks (including their error messages).
    For further information on the fields, see the source code.

Control the created CruiseControlOperation

Stop a task

The task execution can be stopped gracefully when the CruiseControlOperation is deleted. In this case the corresponding cruiseControlState or the cruiseControlVolumeState will transition to Graceful*Succeeded.

Handle failed tasks

cruiseControlOperation.spec.errorPolicy defines how the failed Cruise Control task should be handled. When the errorPolicy is set to retry, Koperator re-executes the failed task every 30 seconds. When it is set to ignore, Koperator treats the failed task as completed, thus the cruiseControlState or the cruiseControlVolumeState transitions to Graceful*Succeeded.

Pause a task

When there is a Cruise Control task which can not be completed without an error and the cruiseControlOperation.spec.errorPolicy is set to retry, Koperator will re-execute the task until it succeeds. You can pause automatic re-execution by adding the following label on the corresponding CruiseControlOperation custom resource. For details see this example. To continue the task, remove the label (or set to any other value than true).

Pausing is useful when the reason of the error can not be fixed any time soon but you want to retry the operation later when the problem is resolved.

A paused CruiseControlOperation tasks are ignored when selecting operations for execution: when a new CruiseControlOperation with the same operation type (status.currentTask.operation) is created, the new one is executed and the paused one is skipped.

kind: CruiseControlOperation
metadata:
...
  name: kafka-addbroker-mhh72
  labels:
    pause: "true"
...

Automatic cleanup

You can set automatic cleanup time for the created CruiseControlOperations in the KafkaCluster custom resource.
In the following example, the finished (completed successfully or completedWithError and errorPolicy: ignore) CruiseControlOperation custom resources are automatically deleted after 300 seconds.

apiVersion: kafka.banzaicloud.io/v1beta1
kind: KafkaCluster
...
spec:
...
  cruiseControlConfig:
    cruiseControlOperationSpec:
      ttlSecondsAfterFinished: 300
...

Example for the ignore and pause use-cases

This example shows how to ignore and pause an operation.

  1. Using the original example with four Kafka brokers from the Overview as the starting point, this example removes two brokers at the same time by editing the KafkaCluster custom resource and deleting broker 2 and broker 3.

     Spec:
     ...  
       brokers:
        - id: 0
          brokerConfigGroup: "default"
        - id: 1
          brokerConfigGroup: "default"
    
  2. The brokers (kafka-removebroker-lg7qm, kafka-removebroker-4plfq) will have separate remove_broker operations. The example shows that the first one is already in running state.

    status:
    ...
      brokersState:
        "2":
         ...
          gracefulActionState:
            cruiseControlOperationReference:
              name: kafka-removebroker-lg7qm
            cruiseControlState: GracefulDownscaleRunning
         ...
         "3":
         gracefulActionState:
            cruiseControlOperationReference:
              name: kafka-removebroker-4plfq
            cruiseControlState: GracefulDownscaleScheduled
        ...
    
  3. Assume that something unexpected happened, so the remove_broker operation enters the GracefulDownscaleCompletedWithError state.

    status:
    ...
      brokersState:
        "2":
         ...
          gracefulActionState:
            cruiseControlOperationReference:
              name: kafka-removebroker-lg7qm
            cruiseControlState: GracefulDownscaleCompletedWithError
         ...
         "3":
         gracefulActionState:
            cruiseControlOperationReference:
              name: kafka-removebroker-4plfq
            cruiseControlState: GracefulDownscaleScheduled
        ...
    
  4. At this point, you can decide how to handle this problem using one of the three possible options: retry it (which is the default behavior), ignore the error, or use the pause label to pause the operation and let Koperator execute the next operation.

    • Ignore use-case: To ignore the error, set the cruiseControlOperation.spec.errorPolicy field to ignore. The operation will be considered as a successful operation, and the broker pod and the persistent volume will be removed from the Kubernetes cluster and from the KafkaCluster status. Koperator will continue to execute the next task: remove_broker for kafka-removebroker-4plfq.

      status:
      ...
        brokersState:
          ...
           "3":
           gracefulActionState:
              cruiseControlOperationReference:
                name: kafka-removebroker-4plfq
              cruiseControlState: GracefulDownscaleRunning
          ...
      
    • Pause use-case: To pause this task, add the pause: true label to the kafka-removebroker-lg7qm CruiseControlOperation. Koperator won’t try to re-execute this task, and moves on to the next remove_broker operation.

      status:
      ...
        brokersState:
          "2":
           ...
            gracefulActionState:
              cruiseControlOperationReference:
                name: kafka-removebroker-lg7qm
              cruiseControlState: GracefulDownscalePaused
           ...
           "3":
           gracefulActionState:
              cruiseControlOperationReference:
                name: kafka-removebroker-4plfq
              cruiseControlState: GracefulDownscaleRunning
          ...
      

      When the second remove_broker operation is finished, only the paused task remains:

      status:
      ...
        brokersState:
          "2":
           ...
            gracefulActionState:
              cruiseControlOperationReference:
                name: kafka-removebroker-lg7qm
              cruiseControlState: GracefulDownscalePaused
           ...
      

      When the problem has been resolved, you can retry removing broker 2 by removing the pause label.

      status:
      ...
        brokersState:
          "2":
           ...
            gracefulActionState:
              cruiseControlOperationReference:
                name: kafka-removebroker-lg7qm
              cruiseControlState: GracefulDownscaleRunning
           ...
      

      If everything goes well, the broker is removed.

Last modified October 26, 2023: Standalone site (#125) (6eafe48)