Advanced StatefulSet

This controller enhances the rolling update workflow of default StatefulSet controller from aspects, such as adding maxUnavailable and introducing in-place update strategy.

Note that Advanced StatefulSet extends the same CRD schema of default StatefulSet with newly added fields. The CRD kind name is still StatefulSet. This is done on purpose so that user can easily migrate workload to the Advanced StatefulSet from the default StatefulSet. For example, one may simply replace the value of apiVersion in the StatefulSet yaml file from apps/v1 to apps.kruise.io/v1alpha1 or apps.kruise.io/v1beta1 after installing Kruise manager.

  1. - apiVersion: apps/v1
  2. + apiVersion: apps.kruise.io/v1beta1
  3. kind: StatefulSet
  4. metadata:
  5. name: sample
  6. spec:
  7. #...

Note that since Kruise v0.7.0, Advanced StatefulSet has been promoted to v1beta1, which is compatible with v1alpha1. And for Kruise version lower than v0.7.0, you can only use v1alpha1.

MaxUnavailable Rolling Update Strategy

Advanced StatefulSet adds a maxUnavailable capability in the RollingUpdateStatefulSetStrategy to allow parallel Pod updates with the guarantee that the number of unavailable pods during the update cannot exceed this value. It is only allowed to use when the podManagementPolicy is Parallel.

This feature achieves similar update efficiency like Deployment for cases where the order of update is not critical to the workload. Without this feature, the native StatefulSet controller can only update Pods one by one even if the podManagementPolicy is Parallel.

  1. apiVersion: apps.kruise.io/v1alpha1
  2. kind: StatefulSet
  3. spec:
  4. # ...
  5. podManagementPolicy: Parallel
  6. updateStrategy:
  7. type: RollingUpdate
  8. rollingUpdate:
  9. maxUnavailable: 20%

For example, assuming an Advanced StatefulSet has five Pods named P0 to P4, and the appplication can tolerate losing three replicas temporally. If we want to update the StatefulSet Pod spec from v1 to v2, we can perform the following steps using the MaxUnavailable feature for fast update.

  1. Set MaxUnavailable to 3 to allow three unavailable Pods maximally.
  2. Optionally, Set Partition to 4 in case canary update is needed. Partition means all Pods with an ordinal that is greater than or equal to the partition will be updated. In this case P4 will be updated even though MaxUnavailable is 3.
  3. After P4 finish update, change Partition to 0. The controller will update P1,P2 and P3 concurrently. Note that with default StatefulSet, the Pods will be updated sequentially in the order of P3, P2, P1.
  4. Once one of P1, P2 and P3 finishes update, P0 will be updated immediately.

In-Place Pod Update Strategy

Advanced StatefulSet adds a podUpdatePolicy field in spec.updateStrategy.rollingUpdate which controls recreate or in-place update for Pods.

  • ReCreate controller will delete old Pods and create new ones. This is the same behavior as default StatefulSet.
  • InPlaceIfPossible controller will try to in-place update Pod instead of recreating them if possible. Currently, only spec.template.metadata.* and spec.template.spec.containers[x].image field can be in-place updated.
  • InPlaceOnly controller will in-place update Pod instead of recreating them. With InPlaceOnly policy, user cannot modify any fields in spec.template other than spec.template.spec.containers[x].image.

When a Pod being in-place update, controller will firstly update Pod status to make it become not-ready using readinessGate, and then update images in Pod spec to trigger Kubelet recreate the container on Node. However, sometimes Kubelet recreate containers so fast that other controllers such as endpoints-controller in kcm have not noticed the Pod has turned to not-ready. This may lead to requests damaged.

So we bring graceful period into in-place update. Advanced StatefulSet has supported gracePeriodSeconds, which is a period duration between controller update pod status and update pod images. So that endpoints-controller could have enough time to remove this Pod from endpoints.

  1. apiVersion: apps.kruise.io/v1alpha1
  2. kind: StatefulSet
  3. spec:
  4. # ...
  5. podManagementPolicy: Parallel
  6. updateStrategy:
  7. type: RollingUpdate
  8. rollingUpdate:
  9. podUpdatePolicy: InPlaceIfPossible
  10. inPlaceUpdateStrategy:
  11. gracePeriodSeconds: 10

More importantly, a readiness-gate named InPlaceUpdateReady must be added into template.spec.readinessGates when using InPlaceIfPossible or InPlaceOnly. The condition InPlaceUpdateReady in podStatus will be updated to False before in-place update and updated to True after the update is finished. This ensures that pod remain at NotReady state while the in-place update is happening.

An example for StatefulSet using in-place update:

  1. apiVersion: apps.kruise.io/v1alpha1
  2. kind: StatefulSet
  3. metadata:
  4. name: sample
  5. spec:
  6. replicas: 3
  7. serviceName: fake-service
  8. selector:
  9. matchLabels:
  10. app: sample
  11. template:
  12. metadata:
  13. labels:
  14. app: sample
  15. spec:
  16. readinessGates:
  17. # A new condition that ensures the pod remains at NotReady state while the in-place update is happening
  18. - conditionType: InPlaceUpdateReady
  19. containers:
  20. - name: main
  21. image: nginx:alpine
  22. podManagementPolicy: Parallel # allow parallel updates, works together with maxUnavailable
  23. updateStrategy:
  24. type: RollingUpdate
  25. rollingUpdate:
  26. # Do in-place update if possible, currently only image update is supported for in-place update
  27. podUpdatePolicy: InPlaceIfPossible
  28. # Allow parallel updates with max number of unavailable instances equals to 2
  29. maxUnavailable: 2

Update sequence

Advanced StatefulSet adds a unorderedUpdate field in spec.updateStrategy.rollingUpdate, which contains strategies for non-ordered update. If unorderedUpdate is not nil, pods will be updated with non-ordered sequence. Noted that UnorderedUpdate can only be allowed to work with Parallel podManagementPolicy.

Currently unorderedUpdate only contains one field: priorityStrategy.

priority strategy

This strategy defines rules for calculating the priority of updating pods. All update candidates will be applied with the priority terms. priority can be calculated either by weight or by order.

  • weight: Priority is determined by the sum of weights for terms that match selector. For example,
  1. apiVersion: apps.kruise.io/v1alpha1
  2. kind: StatefulSet
  3. spec:
  4. # ...
  5. updateStrategy:
  6. rollingUpdate:
  7. unorderedUpdate:
  8. priorityStrategy:
  9. weightPriority:
  10. - weight: 50
  11. matchSelector:
  12. matchLabels:
  13. test-key: foo
  14. - weight: 30
  15. matchSelector:
  16. matchLabels:
  17. test-key: bar
  • order: Priority will be determined by the value of the orderKey. The update candidates are sorted based on the “int” part of the value string. For example, 5 in string “5” and 10 in string “sts-10”.
  1. apiVersion: apps.kruise.io/v1alpha1
  2. kind: StatefulSet
  3. spec:
  4. # ...
  5. updateStrategy:
  6. rollingUpdate:
  7. unorderedUpdate:
  8. priorityStrategy:
  9. orderPriority:
  10. - orderedKey: some-label-key

Paused update

paused indicates that Pods updating is paused, controller will not update Pods but just maintain the number of replicas.

  1. apiVersion: apps.kruise.io/v1alpha1
  2. kind: StatefulSet
  3. spec:
  4. # ...
  5. updateStrategy:
  6. rollingUpdate:
  7. paused: true

Ordinals reserve(skip)

Since Advanced StatefulSet v1beta1 (Kruise >= v0.7.0), it supports ordinals reserve.

By adding the ordinals to reserve into reserveOrdinals fields, Advanced StatefulSet will skip to create Pods with those ordinals. If these Pods have already existed, controller will delete them. Note that spec.replicas is the expectation number of running Pods and spec.reserveOrdinals is the ordinals that should be skipped.

  1. apiVersion: apps.kruise.io/v1beta1
  2. kind: StatefulSet
  3. spec:
  4. # ...
  5. replicas: 4
  6. reserveOrdinals:
  7. - 1

For an Advanced StatefulSet with replicas=4, reserveOrdinals=[1], the ordinals of running Pods will be [0,2,3,4].

  • If you want to migrate Pod-3 and reserve this ordinal, just append 3 into reserveOrdinals list. Then controller will delete Pod-3 and create Pod-5 (existing Pods will be [0,2,4,5]).
  • If you just want to delete Pod-3, you should append 3 into reserveOrdinals list and set replicas to 3. Then controller will delete Pod-3 (existing Pods will be [0,2,4]).