Author: xidianwangtao@gmail.com，Based on Kubernetes 1.9

摘要：Kubernetes StatefulSet在1.9版本中stable了，相信以后会有越老越多的企业会使用它来部署有状态应用，比如Mysql、Zookeeper、ElasticSearch、Redis等等。本文是对StatefulSet的源码分析，包括其Inner Structure、Sync的核心逻辑、Update的主要流程说明、完整的Code Logic Diagram及一些思考。

Inner Structure

下面是简单的StatefulSet Controller工作的内部结构图。

NewStatefulSetController

同其他Controller一样，StatefulSet Controller也是由ControllerManager初始化时负责启动。

// NewStatefulSetController creates a new statefulset controller.
func NewStatefulSetController(podInformer coreinformers.PodInformer,setInformer appsinformers.StatefulSetInformer,pvcInformer coreinformers.PersistentVolumeClaimInformer,revInformer appsinformers.ControllerRevisionInformer,kubeClient clientset.Interface,
) *StatefulSetController {...ssc := &StatefulSetController{kubeClient: kubeClient,control: NewDefaultStatefulSetControl(NewRealStatefulPodControl(kubeClient,setInformer.Lister(),podInformer.Lister(),pvcInformer.Lister(),recorder),NewRealStatefulSetStatusUpdater(kubeClient, setInformer.Lister()),history.NewHistory(kubeClient, revInformer.Lister()),),pvcListerSynced: pvcInformer.Informer().HasSynced,queue:           workqueue.NewNamedRateLimitingQueue(workqueue.DefaultControllerRateLimiter(), "statefulset"),podControl:      controller.RealPodControl{KubeClient: kubeClient, Recorder: recorder},revListerSynced: revInformer.Informer().HasSynced,}podInformer.Informer().AddEventHandler(cache.ResourceEventHandlerFuncs{// lookup the statefulset and enqueueAddFunc: ssc.addPod,// lookup current and old statefulset if labels changedUpdateFunc: ssc.updatePod,// lookup statefulset accounting for deletion tombstonesDeleteFunc: ssc.deletePod,})ssc.podLister = podInformer.Lister()ssc.podListerSynced = podInformer.Informer().HasSyncedsetInformer.Informer().AddEventHandlerWithResyncPeriod(cache.ResourceEventHandlerFuncs{AddFunc: ssc.enqueueStatefulSet,UpdateFunc: func(old, cur interface{}) {oldPS := old.(*apps.StatefulSet)curPS := cur.(*apps.StatefulSet)if oldPS.Status.Replicas != curPS.Status.Replicas {glog.V(4).Infof("Observed updated replica count for StatefulSet: %v, %d->%d", curPS.Name, oldPS.Status.Replicas, curPS.Status.Replicas)}ssc.enqueueStatefulSet(cur)},DeleteFunc: ssc.enqueueStatefulSet,},statefulSetResyncPeriod,)ssc.setLister = setInformer.Lister()ssc.setListerSynced = setInformer.Informer().HasSynced// TODO: Watch volumesreturn ssc
}

很熟悉的代码风格，也是创建对应的eventBroadcaster，然后给对应的objectInformer注册对应的eventHandler：

• StatefulSetController主要ListWatch Pod和StatefulSet对象；
• Pod Informer注册了add/update/delete EventHandler，这三个EventHandler都会将Pod对应的StatefulSet加入到StatefulSet Queue中。
• StatefulSet Informer同样注册了add/update/event EventHandler，也都会将StatefulSet加入到StatefulSet Queue中。
• 目前StatefulSetController还未感知PVC Informer的EventHandler，这里继续按照PVC Controller全部处理。在StatefulSet Controller创建和删除Pod时，会调用apiserver创建和删除对应的PVC。
• RevisionController类似，在StatefulSet Controller Reconcile时会创建或者删除对应的Revision。

StatefulSetController sync

接下来，会进入StatefulSetController的worker（只有一个worker，也就是只一个go routine），worker会从StatefulSet Queue中pop out一个StatefulSet对象，然后执行sync进行Reconcile操作。

// sync syncs the given statefulset.
func (ssc *StatefulSetController) sync(key string) error {startTime := time.Now()defer func() {glog.V(4).Infof("Finished syncing statefulset %q (%v)", key, time.Now().Sub(startTime))}()namespace, name, err := cache.SplitMetaNamespaceKey(key)if err != nil {return err}set, err := ssc.setLister.StatefulSets(namespace).Get(name)if errors.IsNotFound(err) {glog.Infof("StatefulSet has been deleted %v", key)return nil}if err != nil {utilruntime.HandleError(fmt.Errorf("unable to retrieve StatefulSet %v from store: %v", key, err))return err}selector, err := metav1.LabelSelectorAsSelector(set.Spec.Selector)if err != nil {utilruntime.HandleError(fmt.Errorf("error converting StatefulSet %v selector: %v", key, err))// This is a non-transient error, so don't retry.return nil}if err := ssc.adoptOrphanRevisions(set); err != nil {return err}pods, err := ssc.getPodsForStatefulSet(set, selector)if err != nil {return err}return ssc.syncStatefulSet(set, pods)
}

• sync中根据setLabel匹配出所有revisions、然后检查这些revisions中是否有OwnerReference为空的，如果有，那说明存在Orphaned的Revisions。

  注意：只要检查到有一个History Revision就会触发给所有的Resivions打上Patch：
  {"metadata":{"ownerReferences":[{"apiVersion":"%s","kind":"%s","name":"%s","uid":"%s","controller":true,"blockOwnerDeletion":true}],"uid":"%s"}}

• 调用getPodsForStatefulSet获取这个StatefulSet应该管理的Pods。

  • 获取该StatefulSet对应Namesapce下所有的Pods；
  • 执行ClaimPods操作：检查set和pod的Label是否匹配上，如果Label不匹配，那么需要release这个Pod，然后检查pod的name和StatefulSet name的格式是否能匹配上。对于都匹配上的，并且ControllerRef UID也相同的，则不需要处理。
  • 如果Selector和ControllerRef都匹配不上，则执行ReleasePod操作，给Pod打Patch: {“metadata":{"ownerReferences":[{"$patch":"delete","uid":"%s"}],"uid":"%s"}}
  • 对于Label和name格式能匹配上的，但是controllerRef为空的Pods,就执行AdoptPod，给Pod打上Patch： {“metadata":{"ownerReferences":[{"apiVersion":"%s","kind":"%s","name":"%s","uid":"%s","controller":true,"blockOwnerDeletion":true}],"uid":"%s"}}

UpdateStatefulSet

syncStatefulSet的实现只是调用UpdateStatefulSet。

func (ssc *defaultStatefulSetControl) UpdateStatefulSet(set *apps.StatefulSet, pods []*v1.Pod) error {// list all revisions and sort themrevisions, err := ssc.ListRevisions(set)if err != nil {return err}history.SortControllerRevisions(revisions)// get the current, and update revisionscurrentRevision, updateRevision, collisionCount, err := ssc.getStatefulSetRevisions(set, revisions)if err != nil {return err}// perform the main update function and get the statusstatus, err := ssc.updateStatefulSet(set, currentRevision, updateRevision, collisionCount, pods)if err != nil {return err}// update the set's statuserr = ssc.updateStatefulSetStatus(set, status)if err != nil {return err}glog.V(4).Infof("StatefulSet %s/%s pod status replicas=%d ready=%d current=%d updated=%d",set.Namespace,set.Name,status.Replicas,status.ReadyReplicas,status.CurrentReplicas,status.UpdatedReplicas)glog.V(4).Infof("StatefulSet %s/%s revisions current=%s update=%s",set.Namespace,set.Name,status.CurrentRevision,status.UpdateRevision)// maintain the set's revision history limitreturn ssc.truncateHistory(set, pods, revisions, currentRevision, updateRevision)
}

UpdateStatefulSet主要流程为：

• ListRevisions获取该StatefulSet的所有Revisions，并按照Revision从小到大进行排序。
• getStatefulSetRevisions获取currentRevison和UpdateRevision。
  • 只有当RollingUpdate策略时Partition不为0时，才会有部分Pods是updateRevision。
  • 其他情况，所有Pods都得维持currentRevision。
• updateStatefulSet是StatefulSet Controller的核心逻辑，负责创建、更新、删除Pods，使得声明式target得以维护：
  • 使得target state始终有Spec.Replicas个Running And Ready的Pods。
  • 如果更新策略是RollingUpdate，并且Partition为0，则保证所有Pods都对应Status.CurrentRevision。
  • 如果更新策略是RollingUpdate，并且Partition不为0，则ordinal小于Partition的Pods保持Status.CurrentRevision，而ordinal大于等于Partition的Pods更新到Status.UpdateRevision。
  • 如果更新策略是OnDelete，则只有删除Pods时才会触发对应Pods的更新，也就是说与Revisions不关联。
• truncateHistory维护History Revision个数不超过.Spec.RevisionHistoryLimit。

updateStatefulSet

updateStatefulSet是整个StatefulSetController的核心。

func (ssc *defaultStatefulSetControl) updateStatefulSet(set *apps.StatefulSet,currentRevision *apps.ControllerRevision,updateRevision *apps.ControllerRevision,collisionCount int32,pods []*v1.Pod) (*apps.StatefulSetStatus, error) {// get the current and update revisions of the set.currentSet, err := ApplyRevision(set, currentRevision)if err != nil {return nil, err}updateSet, err := ApplyRevision(set, updateRevision)if err != nil {return nil, err}// set the generation, and revisions in the returned statusstatus := apps.StatefulSetStatus{}status.ObservedGeneration = new(int64)*status.ObservedGeneration = set.Generationstatus.CurrentRevision = currentRevision.Namestatus.UpdateRevision = updateRevision.Namestatus.CollisionCount = new(int32)*status.CollisionCount = collisionCountreplicaCount := int(*set.Spec.Replicas)// slice that will contain all Pods such that 0 <= getOrdinal(pod) < set.Spec.Replicasreplicas := make([]*v1.Pod, replicaCount)// slice that will contain all Pods such that set.Spec.Replicas <= getOrdinal(pod)condemned := make([]*v1.Pod, 0, len(pods))unhealthy := 0firstUnhealthyOrdinal := math.MaxInt32var firstUnhealthyPod *v1.Pod// First we partition pods into two lists valid replicas and condemned Podsfor i := range pods {status.Replicas++// count the number of running and ready replicasif isRunningAndReady(pods[i]) {status.ReadyReplicas++}// count the number of current and update replicasif isCreated(pods[i]) && !isTerminating(pods[i]) {if getPodRevision(pods[i]) == currentRevision.Name {status.CurrentReplicas++} else if getPodRevision(pods[i]) == updateRevision.Name {status.UpdatedReplicas++}}if ord := getOrdinal(pods[i]); 0 <= ord && ord < replicaCount {// if the ordinal of the pod is within the range of the current number of replicas,// insert it at the indirection of its ordinalreplicas[ord] = pods[i]} else if ord >= replicaCount {// if the ordinal is greater than the number of replicas add it to the condemned listcondemned = append(condemned, pods[i])}// If the ordinal could not be parsed (ord < 0), ignore the Pod.}// for any empty indices in the sequence [0,set.Spec.Replicas) create a new Pod at the correct revisionfor ord := 0; ord < replicaCount; ord++ {if replicas[ord] == nil {replicas[ord] = newVersionedStatefulSetPod(currentSet,updateSet,currentRevision.Name,updateRevision.Name, ord)}}// sort the condemned Pods by their ordinalssort.Sort(ascendingOrdinal(condemned))// find the first unhealthy Podfor i := range replicas {if !isHealthy(replicas[i]) {unhealthy++if ord := getOrdinal(replicas[i]); ord < firstUnhealthyOrdinal {firstUnhealthyOrdinal = ordfirstUnhealthyPod = replicas[i]}}}for i := range condemned {if !isHealthy(condemned[i]) {unhealthy++if ord := getOrdinal(condemned[i]); ord < firstUnhealthyOrdinal {firstUnhealthyOrdinal = ordfirstUnhealthyPod = condemned[i]}}}if unhealthy > 0 {glog.V(4).Infof("StatefulSet %s/%s has %d unhealthy Pods starting with %s",set.Namespace,set.Name,unhealthy,firstUnhealthyPod.Name)}// If the StatefulSet is being deleted, don't do anything other than updating// status.if set.DeletionTimestamp != nil {return &status, nil}monotonic := !allowsBurst(set)// Examine each replica with respect to its ordinalfor i := range replicas {// delete and recreate failed podsif isFailed(replicas[i]) {glog.V(4).Infof("StatefulSet %s/%s is recreating failed Pod %s",set.Namespace,set.Name,replicas[i].Name)if err := ssc.podControl.DeleteStatefulPod(set, replicas[i]); err != nil {return &status, err}if getPodRevision(replicas[i]) == currentRevision.Name {status.CurrentReplicas--} else if getPodRevision(replicas[i]) == updateRevision.Name {status.UpdatedReplicas--}status.Replicas--replicas[i] = newVersionedStatefulSetPod(currentSet,updateSet,currentRevision.Name,updateRevision.Name,i)}// If we find a Pod that has not been created we create the Podif !isCreated(replicas[i]) {if err := ssc.podControl.CreateStatefulPod(set, replicas[i]); err != nil {return &status, err}status.Replicas++if getPodRevision(replicas[i]) == currentRevision.Name {status.CurrentReplicas++} else if getPodRevision(replicas[i]) == updateRevision.Name {status.UpdatedReplicas++}// if the set does not allow bursting, return immediatelyif monotonic {return &status, nil}// pod created, no more work possible for this roundcontinue}// If we find a Pod that is currently terminating, we must wait until graceful deletion// completes before we continue to make progress.if isTerminating(replicas[i]) && monotonic {glog.V(4).Infof("StatefulSet %s/%s is waiting for Pod %s to Terminate",set.Namespace,set.Name,replicas[i].Name)return &status, nil}// If we have a Pod that has been created but is not running and ready we can not make progress.// We must ensure that all for each Pod, when we create it, all of its predecessors, with respect to its// ordinal, are Running and Ready.if !isRunningAndReady(replicas[i]) && monotonic {glog.V(4).Infof("StatefulSet %s/%s is waiting for Pod %s to be Running and Ready",set.Namespace,set.Name,replicas[i].Name)return &status, nil}// Enforce the StatefulSet invariantsif identityMatches(set, replicas[i]) && storageMatches(set, replicas[i]) {continue}// Make a deep copy so we don't mutate the shared cachereplica := replicas[i].DeepCopy()if err := ssc.podControl.UpdateStatefulPod(updateSet, replica); err != nil {return &status, err}}// At this point, all of the current Replicas are Running and Ready, we can consider termination.// We will wait for all predecessors to be Running and Ready prior to attempting a deletion.// We will terminate Pods in a monotonically decreasing order over [len(pods),set.Spec.Replicas).// Note that we do not resurrect Pods in this interval. Also not that scaling will take precedence over// updates.for target := len(condemned) - 1; target >= 0; target-- {// wait for terminating pods to expireif isTerminating(condemned[target]) {glog.V(4).Infof("StatefulSet %s/%s is waiting for Pod %s to Terminate prior to scale down",set.Namespace,set.Name,condemned[target].Name)// block if we are in monotonic modeif monotonic {return &status, nil}continue}// if we are in monotonic mode and the condemned target is not the first unhealthy Pod blockif !isRunningAndReady(condemned[target]) && monotonic && condemned[target] != firstUnhealthyPod {glog.V(4).Infof("StatefulSet %s/%s is waiting for Pod %s to be Running and Ready prior to scale down",set.Namespace,set.Name,firstUnhealthyPod.Name)return &status, nil}glog.V(4).Infof("StatefulSet %s/%s terminating Pod %s for scale dowm",set.Namespace,set.Name,condemned[target].Name)if err := ssc.podControl.DeleteStatefulPod(set, condemned[target]); err != nil {return &status, err}if getPodRevision(condemned[target]) == currentRevision.Name {status.CurrentReplicas--} else if getPodRevision(condemned[target]) == updateRevision.Name {status.UpdatedReplicas--}if monotonic {return &status, nil}}// for the OnDelete strategy we short circuit. Pods will be updated when they are manually deleted.if set.Spec.UpdateStrategy.Type == apps.OnDeleteStatefulSetStrategyType {return &status, nil}// we compute the minimum ordinal of the target sequence for a destructive update based on the strategy.updateMin := 0if set.Spec.UpdateStrategy.RollingUpdate != nil {updateMin = int(*set.Spec.UpdateStrategy.RollingUpdate.Partition)}// we terminate the Pod with the largest ordinal that does not match the update revision.for target := len(replicas) - 1; target >= updateMin; target-- {// delete the Pod if it is not already terminating and does not match the update revision.if getPodRevision(replicas[target]) != updateRevision.Name && !isTerminating(replicas[target]) {glog.V(4).Infof("StatefulSet %s/%s terminating Pod %s for update",set.Namespace,set.Name,replicas[target].Name)err := ssc.podControl.DeleteStatefulPod(set, replicas[target])status.CurrentReplicas--return &status, err}// wait for unhealthy Pods on updateif !isHealthy(replicas[target]) {glog.V(4).Infof("StatefulSet %s/%s is waiting for Pod %s to update",set.Namespace,set.Name,replicas[target].Name)return &status, nil}}return &status, nil
}

主要流程：

• 获取currentRevision和updateRevision对应的StatefulSet Object，并设置generation，currentRevision, updateRevision等信息到StatefulSet status。

• 将前面getPodsForStatefulSet获取到的pods分成两个slice：

  • valid replicas slice: : 0 <= getOrdinal(pod) < set.Spec.Replicas
  • condemned pods slice: set.Spec.Replicas <= getOrdinal(pod)

• 如果valid replicas中存在某些ordinal没有对应的Pod，则创建对应Revision的Pods Object，后面会检测到该Pod没有真实创建就会去创建对应的Pod实例：

  • 如果更新策略是RollingUpdate且Partition为0或者ordinal < Partition，则使用currentRevision创建该Pod Object。
  • 如果更新策略时RollingUpdate且Partition不为0且ordinal >= Partition，则使用updateRevision创建该Pod Object。

• 从valid repilcas和condemned pods两个slices中找出第一个unhealthy的Pod。（ordinal最小的unhealth pod）

  healthy pods means：pods is running and ready, and not terminating.

• 对于正在删除(DeletionTimestamp非空)的StatefulSet，不做任何操作，直接返回当前status。

• 遍历valid replicas中pods，保证valid replicas中index在[0，spec.replicas）的pod都是Running And Ready的：

  • 如果检测到某个pod Failed (pod.Status.Phase = Failed), 则删除这个Pod，并重新new这个pod object（注意revisions匹配）
  • 如果这个pod还没有recreate,则Create it。
  • 如果ParallelPodManagement = "OrderedReady”，则直接返回当前status。否则ParallelPodManagement = "Parallel”,则循环检测下一个。
  • 如果pod正在删除并且ParallelPodManagement = "OrderedReady”，则返回status结束。
  • 如果pod不是RunningAndReady状态，并且ParallelPodManagement = "OrderedReady”，则返回status结束。
  • 检测该pod与statefulset的identity和storage是否匹配，如果有一个不匹配，则调用apiserver Update Stateful Pod进行updateIdentity和updateStorage（并创建对应的PVC），返回status，结束。

  Pod is Running and Ready means:
  pod.Status.Phase = Runnin,
  pod.Status.Condition = Ready

• 遍历condemned replicas中pods，index由大到小的顺序，确保这些pods最终都被删除：

  • 如果这个Pod正在删除(DeletionTimestamp)，并且Pod Management是OrderedReady，则进行Block住，返回status，流程结束。
  • 如果是OrderedReady策略，Pod不是处于Running and Ready状态，且该pod不是first unhealthy pod，则返回status，流程结束。
  • 其他情况，则删除该statefulset pod。
  • 根据该pod的controller-revision-hash Label获取Revision，如果等于currentRevision，则更新status.CurrentReplicas；如果等于updateRevision，则更新status.UpdatedReplicas；
  • 如果是OrderedReady策略，则返回status，流程结束。

• OnDelete更新策略：删除Pod才会触发更新这个ordinal的更新 如果UpdateStrategy Type是OnDelete, 意味着只有当对应的Pods被手动删除后，才会触发Recreate，因此直接返回status，流程结束。

• RollingUpdate更新策略：（Partition不设置就相当于0，意味着全部pods进行滚动更新） 如果UpdateStrategy Type是RollingUpdate, 根据RollingUpdate中Partition配置得到updateMin作为update replicas index区间最小值，遍历valid replicas，index从最大值到updateMin递减的顺序：

  • 如果pod revision不是updateRevision，并且不是正在删除的，则删除这个pod，并更新status.CurrentReplicas，然后返回status，流程结束。
  • 如果pod不是healthy的，那么将等待它变成healthy，因此这里就直接返回status，流程结束。

Identity Match

updateStatefulSet Reconcile中，会检查identity match的情况，具体包含哪些？

StatefulSetPodNameLabel        = "statefulset.kubernetes.io/pod-name"// identityMatches returns true if pod has a valid identity and network identity for a member of set.
func identityMatches(set *apps.StatefulSet, pod *v1.Pod) bool {parent, ordinal := getParentNameAndOrdinal(pod)return ordinal >= 0 &&set.Name == parent &&pod.Name == getPodName(set, ordinal) &&pod.Namespace == set.Namespace &&pod.Labels[apps.StatefulSetPodNameLabel] == pod.Name
}

• pod name和statefulset name内容匹配。
• namespace匹配。
• Pod的Label：statefulset.kubernetes.io/pod-name与Pod name真实匹配。

Storage Match

updateStatefulSet Reconcile中，会检查Storage match的情况，具体怎么匹配的呢？

// storageMatches returns true if pod's Volumes cover the set of PersistentVolumeClaims
func storageMatches(set *apps.StatefulSet, pod *v1.Pod) bool {ordinal := getOrdinal(pod)if ordinal < 0 {return false}volumes := make(map[string]v1.Volume, len(pod.Spec.Volumes))for _, volume := range pod.Spec.Volumes {volumes[volume.Name] = volume}for _, claim := range set.Spec.VolumeClaimTemplates {volume, found := volumes[claim.Name]if !found ||volume.VolumeSource.PersistentVolumeClaim == nil ||volume.VolumeSource.PersistentVolumeClaim.ClaimName !=getPersistentVolumeClaimName(set, &claim, ordinal) {return false}}return true
}

Code Logic Diagram

基于上述分析，下面是一个相对完整的StatefulSetController的代码逻辑图。 (不支持大于2MB的图片，所以不太清晰，不过基本在前面内容都提到了。)

思考

滚动更新过程中出现异常

在上一篇博文浅析Kubernetes StatefulSet中遗留了一个问题：StatefulSet滚动更新时，如果某个Pod更新失败，会怎么办呢？

通过上面源码分析中滚动更新部分的分析，我们知道：

• 如果UpdateStrategy Type是RollingUpdate, 根据RollingUpdate中Partition（Partition不设置就相当于0，意味着全部pods进行滚动更新）配置得到updateMin作为update replicas index区间最小值，遍历valid replicas，index从最大值到updateMin递减的顺序：
  • 如果pod revision不是updateRevision，并且不是正在删除的，则删除这个pod，并更新status.CurrentReplicas，然后返回status，流程结束。
  • 如果pod不是healthy的，那么将等待它变成healthy，因此这里就直接返回status，流程结束。

知道这一点后，就能回答这个问题了，答案很简单：

• 如果更新策略是RollingUpdate，则逐个滚动更新过程中，如果在更新某个ordinal replica时这个Pod一直无法达到Running and Ready状态，那么整个滚动更新流程将Block在这里。还没有更新的replicas将不会触发更新，已经更新成功的replicas就保持更新后的版本，并不存在什么自动回滚的机制。在下一次sync时，检测到这个Pod isFailed（pod.Status.Phase = Failed）,会delete and recreate这个failed pod。

podManagementPolicy设置为Parallel时，体现在哪？

问题：podManagementPolicy: "Parallel"体现在什么时候呢？Scale的时候？RollingUpdate的时候？

• 在前面代码分析中updateStatefulSet中那段-"遍历valid replicas中pods，保证valid replicas中index在[0，spec.replicas）的pod都是Running And Ready的"：如果发现某个ordinal replica应该创建但是还没被创建，则会触发create。如果podManagementPolicy设置为Parallel，则会继续delete then create其他应该创建的replicas，而不会等待前面创建的replicas成为Running and Ready。
• 在前面代码分析中updateStatefulSet中那段-”遍历condemned replicas中pods，index由大到小的顺序，确保这些pods最终都被删除"：podManagementPolicy设置为Parallel，如果发现某个ordinal replica正在删除，则继续删除其他应该删除的replicas，而不会等待之前删除的replica重建并成为Running and Ready状态。

因此Parallel体现在以下场景：

• 初始化部署StatefulSet时，并行create pods。
• 级联删除StatefulSet时，并行delete pods。
• Scale up时，并行create pods。
• Scale down时，并行delete pods。

而在滚动更新时，是不会受podManagementPolicy的配置影响的，都是按照逐个地、ordinal从大到小的的顺序，保证前者Running and Ready的原则，进行RollingUpdate。

如果更新策略是OnDelete呢？那情况就不同于RollingUpdate了，因为update的流程就体现在前面提到的两个阶段了，因此Parallel是会启作用的。