美文网首页
kube-controller-manager分析之NodeLi

kube-controller-manager分析之NodeLi

作者: zhangzhifei | 来源:发表于2022-12-22 16:12 被阅读0次

    功能

    监控node的健康状态,异常后打上相应的noExection 污点,由TaintManager驱逐异常node上的污点。
    Kube-controller-manager支持了很多参数配置NodeLifecycleController,由于驱逐pod属于高危动作,很可能会引起集群崩溃,业务服务不可用,如node-master间网络异常,但pod能提供正常服务。如apiserver使用的外部lb异常,或者apiserver本身异常,都可能发生不可预知的问题。但是这个能力不能完全不用,当node真正异常时,需要相应的容错能力。本文将NodeLifecycleController所有的参数在源码层面进行了相关分析和解释,希望对大家有帮助

    相关配置

    NodeLifecycleControllerConfiguration

    
    // NodeLifecycleControllerConfiguration contains elements describing NodeLifecycleController.
    type NodeLifecycleControllerConfiguration struct {
       // If set to true enables NoExecute Taints and will evict all not-tolerating
       // Pod running on Nodes tainted with this kind of Taints.
       // 对应--enable-taint-manager 默认为 true,如果为 true,则表示NodeController 将会启动 TaintManager,当已经调度到该 node 上的 pod 不能容忍 node 的 taint 时,由 TaintManager 负责驱逐此类 pod,若不开启该特性则已调度到该 node 上的 pod 会继续存在
       EnableTaintManager bool 
       // nodeEvictionRate is the number of nodes per second on which pods are deleted in case of node failure when a zone is healthy
       // 通过--node-eviction-rate设置, 默认 0.1,表示当集群下某个 zone 为 unhealthy 时,每秒应该剔除的 node 数量,默认即每 10s 剔除1个 node
       NodeEvictionRate float32
       // secondaryNodeEvictionRate is the number of nodes per second on which pods are deleted in case of node failure when a zone is unhealthy
       // 通过 --secondary-node-eviction-rate设置,默认为 0.01,表示如果某个 zone 下的 unhealthy 节点的百分比超过 --unhealthy-zone-threshold (默认为 0.55)时,驱逐速率将会减小,如果集群较小(小于等于 --large-cluster-size-threshold 个 节点 - 默认为 50),驱逐操作将会停止,否则驱逐速率将降为每秒 --secondary-node-eviction-rate 个(默认为 0.01);
       SecondaryNodeEvictionRate float32
       // nodeStartupGracePeriod is the amount of time which we allow starting a node to
       // be unresponsive before marking it unhealthy.
       // --node-startup-grace-period 默认 60s,在 node 启动完成前标记节点为unhealthy 的允许无响应时间;
       NodeStartupGracePeriod metav1.Duration
       // NodeMonitorGracePeriod is the amount of time which we allow a running node to be
       // unresponsive before marking it unhealthy. Must be N times more than kubelet's
       // nodeStatusUpdateFrequency, where N means number of retries allowed for kubelet
       // to post node status.
       // 通过--node-monitor-grace-period 设置,默认 40s,表示在标记某个 node为 unhealthy 前,允许40s内该node无响应
       NodeMonitorGracePeriod metav1.Duration
       // podEvictionTimeout is the grace period for deleting pods on failed nodes.
       // 通过--pod-eviction-timeout 设置,默认 5 分钟,表示在强制删除 node 上的 pod 时,容忍 pod 时间;没有启动TaintBasedEvictions才有效,所以不会用到
       PodEvictionTimeout metav1.Duration
       // secondaryNodeEvictionRate is implicitly overridden to 0 for clusters smaller than or equal to largeClusterSizeThreshold
       // 通过--large-cluster-size-threshold 设置,默认为 50,当该 zone 的节点超过该阈值时,则认为该 zone 是一个大集群;对于小于或等于largeClusterSizeThreshold的集群,secondarynodeevtionrate将降级为0,即不进行驱逐
       LargeClusterSizeThreshold int32
       // Zone is treated as unhealthy in nodeEvictionRate and secondaryNodeEvictionRate when at least
       // unhealthyZoneThreshold (no less than 3) of Nodes in the zone are NotReady
       // --unhealthy-zone-threshold, 不健康zone阈值,会影响什么时候开启二级驱赶速率,默认为0.55,即当该zone中节点宕机数目超过55%,而认为该zone不健康
       UnhealthyZoneThreshold float32
    }
    

    NodeMonitorPeriod

    // KubeCloudSharedConfiguration contains elements shared by both kube-controller manager
    // and cloud-controller manager, but not genericconfig.
    type KubeCloudSharedConfiguration struct {
    
        // nodeMonitorPeriod is the period for syncing NodeStatus in NodeController.    
        // 通过--node-monitor-period 设置,默认为 5s,表示在 NodeController 中同步NodeStatus 的周期,多长时间Controller检查一次。这个值应该小于--node-monitor-grace-period 
        NodeMonitorPeriod metav1.Duration
    }
    

    代码

    代码来自1.21

    startNodeLifecycleController->
        NewNodeLifecycleController->
        lifecycleController.Run->
            nc.taintManager.Run->
            nc.doNodeProcessingPassWorker->
            nc.doPodProcessingWorker->
            nc.doNoExecuteTaintingPass(EnableTaintManager)/nc.doEvictionPass->
            nc.monitorNodeHealth->
    

    taintManager

    主要功能:负责删除pod

    1. 监听集群中所有的Node和pod,当node上存在taints并且该node上的pod不能容忍所有taint,或者pod配置了tolerationSeconds,并且倒计时完成,则delete 该pod
    2. 相关参数,apiserver准入控制添加的,时间可配置
      tolerations:
      - effect: NoExecute
        key: node.kubernetes.io/not-ready
        operator: Exists
        tolerationSeconds: 300
      - effect: NoExecute
        key: node.kubernetes.io/unreachable
        operator: Exists
        tolerationSeconds: 300
    
    • --default-not-ready-toleration-seconds // apiserver admit阶段默认给非daemontsetpod添加,默认300s。
    • --default-unreachable-toleration-seconds // apiserver admit默认给非daemontsetpod添加,默认300s


      image.png

    doNodeProcessingPassWorker

    对该node执行doNoScheduleTaintingPass–根据node status里的condition设置taint(noschedule)

    
    // map {NodeConditionType: {ConditionStatus: TaintKey}}
    // represents which NodeConditionType under which ConditionStatus should be
    // tainted with which TaintKey
    // for certain NodeConditionType, there are multiple {ConditionStatus,TaintKey} pairs
    nodeConditionToTaintKeyStatusMap = map[v1.NodeConditionType]map[v1.ConditionStatus]string{
       v1.NodeReady: {
          v1.ConditionFalse:   v1.TaintNodeNotReady,
          v1.ConditionUnknown: v1.TaintNodeUnreachable,
       },
       v1.NodeMemoryPressure: {
          v1.ConditionTrue: v1.TaintNodeMemoryPressure,
       },
       v1.NodeDiskPressure: {
          v1.ConditionTrue: v1.TaintNodeDiskPressure,
       },
       v1.NodeNetworkUnavailable: {
          v1.ConditionTrue: v1.TaintNodeNetworkUnavailable,
       },
       v1.NodePIDPressure: {
          v1.ConditionTrue: v1.TaintNodePIDPressure,
       },
    }
    
    image.png

    doPodProcessingWorker

    主要的功能是当node ReadyCondition不是true时,将pod的ReadyCondition更新为false.这里也比较重要,node notreay后pod都会被设置成not ready,service对象的endpoint的pod就会被摘除,依赖service的服务需要关注这块。

    // pod的spec.NodeName不是nil并且pod的spec.NodeName发生变化则会加入到podUpdateQueue。
    func (nc *Controller) podUpdated(oldPod, newPod *v1.Pod) {
       if newPod == nil {
          return
       }
       if len(newPod.Spec.NodeName) != 0 && (oldPod == nil || newPod.Spec.NodeName != oldPod.Spec.NodeName) {
          podItem := podUpdateItem{newPod.Namespace, newPod.Name}
          nc.podUpdateQueue.Add(podItem)
       }
    }
    
    image.png

    doNoExecuteTaintingPass

    处理基于taint的evictions方式,驱逐的时候是不会做限速的,所以这里要现实添加taint的速度(monitorNodeHealth侧实现限速)。真正的驱逐pod还是在taintManager中。RateLimitedTimedQueue令牌桶限速队列

    • 从zoneNoExecuteTainter中获得一个zone的node队列,从队列中获取一个node
    • 如果node ready condition为false,移除“node.kubernetes.io/unreachable”的taint,添加“node.kubernetes.io/not-ready” 的taint,Effect为NoExecute。
    • 如果node ready condition为unknown,移除“node.kubernetes.io/not-ready” 的taint,添加“node.kubernetes.io/unreachable” 的taint,Effect为NoExecute。
    • nc.zoneNoExecuteTainter[k].Try() 过程中限速会起作用,限速是由monitorNodeHealth->handleDisruption设置的


      image.png
    
    func (nc *Controller) doNoExecuteTaintingPass() {
       nc.evictorLock.Lock()
       defer nc.evictorLock.Unlock()
       for k := range nc.zoneNoExecuteTainter {
          // Function should return 'false' and a time after which it should be retried, or 'true' if it shouldn't (it succeeded).
          nc.zoneNoExecuteTainter[k].Try(func(value scheduler.TimedValue) (bool, time.Duration) {
             node, err := nc.nodeLister.Get(value.Value)
             if apierrors.IsNotFound(err) {
                klog.Warningf("Node %v no longer present in nodeLister!", value.Value)
                return true, 0
             } else if err != nil {
                klog.Warningf("Failed to get Node %v from the nodeLister: %v", value.Value, err)
                // retry in 50 millisecond
                return false, 50 * time.Millisecond
             }
             _, condition := nodeutil.GetNodeCondition(&node.Status, v1.NodeReady)
             // Because we want to mimic NodeStatus.Condition["Ready"] we make "unreachable" and "not ready" taints mutually exclusive.
             taintToAdd := v1.Taint{}
             oppositeTaint := v1.Taint{}
             switch condition.Status {
             case v1.ConditionFalse:
                taintToAdd = *NotReadyTaintTemplate
                oppositeTaint = *UnreachableTaintTemplate
             case v1.ConditionUnknown:
                taintToAdd = *UnreachableTaintTemplate
                oppositeTaint = *NotReadyTaintTemplate
             default:
                // It seems that the Node is ready again, so there's no need to taint it.
                klog.V(4).Infof("Node %v was in a taint queue, but it's ready now. Ignoring taint request.", value.Value)
                return true, 0
             }
    
             result := nodeutil.SwapNodeControllerTaint(nc.kubeClient, []*v1.Taint{&taintToAdd}, []*v1.Taint{&oppositeTaint}, node)
             if result {
                //count the evictionsNumber
                zone := utilnode.GetZoneKey(node)
                evictionsNumber.WithLabelValues(zone).Inc()
             }
    
             return result, 0
          })
       }
    }
    

    monitorNodeHealth

    每隔nodeMonitorPeriod周期,执行一次monitorNodeHealth,维护node状态和zone的状态,当 node 处于异常状态时更新 node 的 taint 。根据集群不同状态设置zone的速率。
    NodeLifecycleController 会为每一个 node 划分一个 zoneStates,不同的zoneStates 分别对应着不同的驱逐速率

    • Initial:新添的node state
    • fullyDisrupted:zone 下所有 node 都处于 notReady 状态;
    • partiallyDisrupted:notReady node 占比 >= unhealthyZoneThreshold 的值(默认为0.55,通过--unhealthy-zone-threshold设置)且 notReady node 数超过2个;
    • normal:以上两种情况之外的;
      关键参数
    • --node-monitor-period ,nodeMonitorPeriod这么长时间,monitorNodeHealth执行一次,也就是检查node健康请况的周期


      image.png
    • 注意:只要node notready,就会将pod更新为notready*
    
    switch {
    case currentReadyCondition.Status != v1.ConditionTrue && observedReadyCondition.Status == v1.ConditionTrue:
       // Report node event only once when status changed.
       nodeutil.RecordNodeStatusChange(nc.recorder, node, "NodeNotReady")
       fallthrough
    case needsRetry && observedReadyCondition.Status != v1.ConditionTrue:
       if err = nodeutil.MarkPodsNotReady(nc.kubeClient, nc.recorder, pods, node.Name); err != nil {
          utilruntime.HandleError(fmt.Errorf("unable to mark all pods NotReady on node %v: %v; queuing for retry", node.Name, err))
          nc.nodesToRetry.Store(node.Name, struct{}{})
          continue
       }
    }
    

    tryUpdateNodeHealth

    • tryUpdateNodeHealth 会根据当前获取的 node status 更新 nodeHealthMap,nodeHealthMap 保存 node 最近一次的状态。然后根据nodeHealthMap和node最新的状态判断 node是否已经处于 unknown 状态(probeTimestamp是否超过nodeMonitorGracePeriod/nodeStartupGracePeriod),并更新到apiserver。

    • 如果node没有上报status,但是lease更新,依然认为node是健康的。
      主要参数:

    • --node-monitor-grace-period nodeMonitorGracePeriod,改时间内node没有更新状态/lease就会被设置为unknown

    • --node-startup-grace-period nodeStartupGracePeriod,新节点允许的优雅时间往往要长一些
      变量解释:

    • observedReadyCondition是上个nodeMonitorPeriod周期时node的condition

    • currentReadyCondition是node当前的condition
      主要逻辑:

    1. 从nodeHealthMap获取上一次存储的nodeHealth(nodeHealthData)
    type nodeHealthData struct {
       probeTimestamp           metav1.Time
       readyTransitionTimestamp metav1.Time
       status                   *v1.NodeStatus
       lease                    *coordv1.Lease
    }
    
    1. 从node(apiserver中获取的)获取currentReadyCondition(ReadyCondition),如果是nil说明kubelet或者nodecontroller(也就是NodeLifecycleController)还没有上报过状态,这里能看出controller-manager也会更新node的status。
    • node.Status的里没有ReadyCondition,就fake一个,并赋值给observedReadyCondition,因为之前肯定没有值
    • nodeHealth也用当前的node.Status创建。这里是一个弄的在中首次更新
    1. 根据savedCondition、currentReadyCondition、observedLease更新nodeHealth的probeTimestamp
    2. 判断nodeHealth的probeTimestamp是不是已经超过了gracePeriod(nodeMonitorGracePeriod/nodeStartupGracePeriod),如果超时了就将node 的condition设置为"Unknown"然后更新到apiserver
    3. 将gracePeriod, observedReadyCondition, currentReadyCondition返回
      handleDisruption
      根据各zone中unhealthy node的情况(依据zoneToNodeConditions),给 zone 设置不同的驱逐速率。
      关键概念:
    • allAreFullyDisrupted代表现在所有zone状态stateFullDisruption全挂
    • allWasFullyDisrupted为true代表过去所有zone状态stateFullDisruption全挂
      主要逻辑:
    1. 当allAreFullyDisrupted为false,allWasFullyDisrupted为true,之前zone未全挂,现在所有zone全挂:
    • 执行markNodeAsReachable删除所有node的taint
    • 并将zoneNoExecuteTainter的QPS设置为0,也就是不再打taint。
    1. 当allAreFullyDisrupted为true,allWasFullyDisrupted为false,过去所有zone全挂,现在所有zone未全挂:
    • 更新所有node的nodeHealthMap里的probeTImestamp、readyTransitiontimestamp的时间戳
    • 遍历zoneStates,设置QPS,每个zone的每秒给几个弄的加taint
      • 当zone的状态为stateNormal,则zoneNoExecuteTainter速率设置为evictionLimiterQPS(--node-eviction-rate)
      • 当zone状态为statePartialDisruption,根据zone里的node数量,当node数量大于largeClusterThreshold(--large-cluster-size-threshold),设置zoneNoExecuteTainter速率为SecondEvictionLimiterQPS(--secondary-node-eviction-rate);小于等于largeClusterThreshold,设置zoneNoExecuteTainter速率为0。
      • 当zone状态为stateFullDisruption,则zoneNoExecuteTainter速率设置为evictionLimiterQPS。也就是说如果一共有个两个zone,az1不是stateFullDisruption,az2是stateFullDisruption,那么az2不会全部停止驱逐,驱逐速率为evictionLimiterQPS,然后向SecondEvictionLimiterQPS过度。
        computeZoneStateFunc
        计算zone的state,计算每个zone中notReady的node并将 zone 分为三种:
    • fullyDisrupted:所有的node都是notReady
    • partiallyDisrupted:notReady超过unhealthyZoneThreshold,notReady大于两个
    • normal:以上两种情况之外的情况
    // ComputeZoneState returns a slice of NodeReadyConditions for all Nodes in a given zone.
    // The zone is considered:
    // - fullyDisrupted if there're no Ready Nodes,
    // - partiallyDisrupted if at least than nc.unhealthyZoneThreshold percent of Nodes are not Ready,
    // - normal otherwise
    func (nc *Controller) ComputeZoneState(nodeReadyConditions []*v1.NodeCondition) (int, ZoneState) {
            readyNodes := 0
            notReadyNodes := 0
            for i := range nodeReadyConditions {
                    if nodeReadyConditions[i] != nil && nodeReadyConditions[i].Status == v1.ConditionTrue {
                            readyNodes++
                    } else {
                            notReadyNodes++
                    }
            }
            switch {
            case readyNodes == 0 && notReadyNodes > 0:
                    return notReadyNodes, stateFullDisruption
            case notReadyNodes > 2 && float32(notReadyNodes)/float32(notReadyNodes+readyNodes) >= nc.unhealthyZoneThreshold:
                    return notReadyNodes, statePartialDisruption
            default:
                    return notReadyNodes, stateNormal
            }
    }
    

    isNodeExcludedFromDisruptionChecks
    只是用来排除一些node,使其不参与限速相关的计算。

    相关文章

      网友评论

          本文标题:kube-controller-manager分析之NodeLi

          本文链接:https://www.haomeiwen.com/subject/meivqdtx.html