開發十年，就只剩下這套架構體系了！ >>>   

Author: [email protected] | Version: Kubernetes 1.13

摘要：DaemonSet是Kubernetes中使用者最常用的物件之一，我們用它來部署Nodes上守護應用，比如日誌元件、節點監控元件等。從使用者的使用角度來講，DaemonSet看似簡單，但實際上它涉及的點非常多，比如DaemonSet Pod滿足什麼條件才能在Node上執行、Node出現MemoryPressure或者其他異常Condition時是否能執行、排程的邏輯是怎樣的、滾動更新的邏輯是怎樣的等等，本文講從DaemonSet Controller的原始碼著手，分析其中關鍵邏輯。

DaemonSet Controller

DaemonSet Controller Struct

DaemonSet Controller的核心結構包括：

• burstReplcas int: 每次sync時，Create和Delete Pods的數量上限，程式碼中寫死為250。
• queue workqueue.RateLimitingInterface: 存放待同步DaemonSet Key(namespaces/name)的Delaying Queue。
• syncHandler func(dsKey string) error: 負責同步DaemonSet Queue中物件，包括Replicas管理、UpdateStrategy升級、更新DaemonSet Status等工作，是DaemonSet Controller中最核心的邏輯。
• expectations controller.ControllerExpectationsInterface: 維護每個DaemonSet物件每次Sync期望Create/Delete Pods數的TTLCache。
• suspendedDaemonPods map[string]sets.String: key為NodeName，value是DaemonSet集合，這些DaemonSet包含該Node上'wantToRun & !shouldSchedule'的Pod。
  • wantToRun: 為True，當DaemonSet Controller去Simulate排程時，Predicate（主要是GeneralPredicates和PodToleratesNodeTaints）時忽略如下PredicateFailureError（都是些資源類的Error）時成功，有其他PredicateFailureError為False。如果DaemonSet的Spec中指定了NodeName，則根據其是否與node.Name匹配成功來決定wantToRun的值。
    • ErrDiskConflict;
    • ErrVolumeZoneConflict;
    • ErrMaxVolumeCountExceeded;
    • ErrNodeUnderMemoryPressure;
    • ErrNodeUnderDiskPressure;
    • InsufficientResourceError;
  • shouldSchedule:
    • 如果DaemonSet的Spec中指定了NodeName，則根據其是否與node.Name匹配成功來決定shouldSchedule的值。
    • 如果Predicate時出現所有型別的PredicateFailureError之一，則shouldSchedule都為false。
    • 如果出現InsufficientResourceError，則shouldSchedule也為false。
• failedPodsBackoff *flowcontrol.Backoff: DaemonSet Controller Run時會啟動一個協程，每隔2*MaxDuration(2*15Min)會強制進行一次failedPods GC清理。每次syncDaemonSet處理該刪除的Pods時，會按照1s,2s,4s,8s,.....15min的Backoff機制做一定的delay處理，實現流控的效果。防止kubelet拒絕某些DaemonSet Pods後，馬上又被拒絕，如此就會出現很多無效的迴圈，因此加入了Backoff機制。

DaemonSet Controller的建立和啟動

NewDaemonSetsController負責建立Controller，其中很重要的工作就是註冊以下Informer的EventHandler：

• daemonSetInformer: AddFunc/DeleteFunc/UpdateFunc最終其實都主要是enqueue DaemonSet；
• historyInformer:
  • AddFunc: addHistory;
  • UpdateFunc: updateHistory;
  • DeleteFunc: deleteHistory;
• podInformer:
  • AddFunc: addPod;
  • UpdateFunc: updatePod;
  • DeleteFunc: deletePod;
• nodeInformer:
  • AddFunc: addNode;
  • UpdateFunc: updateNode;

DamonSet Controller Run啟動時，主要幹兩件事：

• 啟動2個workers協程，每個worker負責從queue中取DaemonSet Key進行sync。

• 啟動1個failedPodsBackoff GC協程，每隔1Min清理一次叢集中所有DaemonSet/Node對應的Failed Pods。

只有deletePod時，才會requeueSuspendedDaemonPods。-- 為什麼？

DaemonSet的同步

worker會從queue中取待同步的DamonSet Key，呼叫syncDaemonSet完成自動管理，syncDaemonSet是DaemonSet管理的核心入口。

pkg/controller/daemon/daemon_controller.go:1208

func (dsc *DaemonSetsController) syncDaemonSet(key string) error {
	...
	ds, err := dsc.dsLister.DaemonSets(namespace).Get(name)
	if errors.IsNotFound(err) {
		klog.V(3).Infof("daemon set has been deleted %v", key)
		dsc.expectations.DeleteExpectations(key)
		return nil
	}
	if err != nil {
		return fmt.Errorf("unable to retrieve ds %v from store: %v", key, err)
	}

	everything := metav1.LabelSelector{}
	if reflect.DeepEqual(ds.Spec.Selector, &everything) {
		dsc.eventRecorder.Eventf(ds, v1.EventTypeWarning, SelectingAllReason, "This daemon set is selecting all pods. A non-empty selector is required.")
		return nil
	}

	// Don't process a daemon set until all its creations and deletions have been processed.
	// For example if daemon set foo asked for 3 new daemon pods in the previous call to manage,
	// then we do not want to call manage on foo until the daemon pods have been created.
	...
	if ds.DeletionTimestamp != nil {
		return nil
	}

	// Construct histories of the DaemonSet, and get the hash of current history
	cur, old, err := dsc.constructHistory(ds)
	if err != nil {
		return fmt.Errorf("failed to construct revisions of DaemonSet: %v", err)
	}
	hash := cur.Labels[apps.DefaultDaemonSetUniqueLabelKey]

	if !dsc.expectations.SatisfiedExpectations(dsKey) {
		// Only update status. Don't raise observedGeneration since controller didn't process object of that generation.
		return dsc.updateDaemonSetStatus(ds, hash, false)
	}

	err = dsc.manage(ds, hash)
	if err != nil {
		return err
	}

	// Process rolling updates if we're ready.
	if dsc.expectations.SatisfiedExpectations(dsKey) {
		switch ds.Spec.UpdateStrategy.Type {
		case apps.OnDeleteDaemonSetStrategyType:
		case apps.RollingUpdateDaemonSetStrategyType:
			err = dsc.rollingUpdate(ds, hash)
		}
		if err != nil {
			return err
		}
	}

	err = dsc.cleanupHistory(ds, old)
	if err != nil {
		return fmt.Errorf("failed to clean up revisions of DaemonSet: %v", err)
	}

	return dsc.updateDaemonSetStatus(ds, hash, true)
}

核心的流程如下：

• 首先檢查該DaemonSet物件在本地Store中是否被刪除，如果是，則從expectations中刪除該DaemonSet對應的資料。
• 檢查該DaemonSet物件的LabelSelector是否為空，如果是，則syncDaemonSet返回結束，不進行同步，那麼DaemonSet對應的Pod也不會被建立了。
• 如果其DeletionTimestamp非空，意味著使用者觸發了刪除，則syncDaemonSet返回結束，不進行同步。DaemonSet對應的Pod交由GC Controller去完成刪除。
• 然後constructHistory獲取該DaemonSet的Current ControllerRevision和所有Old ControllerRevisions，並確保所有ControllerRevisions都打上Label: "controller-revision-hash: ControllerRevision.Name"，更新Current ControllerRevision的Revision = maxRevision(old) + 1。
• 檢查當前expectations是否已經滿足，當不滿足時，只更新DaemonSet Status，同步流程結束。
  • expectations中add和del都不大於0，表示Controller expectations已經實現，則當前expectations已經滿足。
  • expectations已經超時，超時時間是5min（不可配置），如果超時，則表示需要進行同步。
  • 如果expectations中還沒有該DaemonSet的資訊，則表示也滿足了，將觸發DaemonSet同步。
  • 此處updateDaemonSetStatus會更新該Daemonset.Status的如下欄位，注意不會更新ObservedGeneration（也沒發生變化）。
    • DesiredNumberScheduled：使用者期望排程的DaemonSet Pods數量，對應前面提到的wantToRun為true的pods數量。
    • CurrentNumberScheduled：使用者期望排程的，並且當前已經執行在Node上的Pods數量。
    • NumberMisscheduled：使用者不期望排程的（wantToRun為false），並且已經執行在對應Node上Pods數量，即已經錯誤排程的Pods數量。
    • NumberReady：CurrentNumberScheduled中，Pod Type Ready Condition為true的Pods數量。
    • UpdatedNumberScheduled：CurrentNumberScheduled中，Pod Label controller-revision-hash對應的hash值與Current ControllerRevision的該hash值相等的Pods數量，即Pod Template已經更新的Pods數量。
    • NumberAvailable：CurrentNumberScheduled中，Pod Type Ready Condition為true，並且Available（Ready時間超過minReadySeconds）的Pods數量。
    • NumberUnavailable：desiredNumberScheduled - numberAvailable。
• 呼叫manage進行DaemonSet Pod的管理：計算待刪除和建立的Pod列表，然後呼叫syncNodes分批次(1,2,4,8,..)的完成Pod的建立和刪除。如果syncNodes之前發現某些Node上對應DaemonSet Pod是Failed，那麼syncNodes後返回error。syncNode會將expectations中的add/del都歸零甚至負數，只有這樣，才會在syncDaemonSet中呼叫manage進行Pod管理。
• 如果manage返回error，則syncDaemonSet流程結束。否則會繼續下面的流程。
• 檢查當前expectations是否已經滿足，如果滿足，則根據UpdateStrategy觸發DaemonSet更新：
  • 如果UpdateStrategy是OnDelete，則等待使用者delete Pod，觸發對應的DaemonSet的enqueue，在syncNodes時更新最新的Pod Template建立新Pod。
  • 如果UpdateStrategy是RollingUpdate，則呼叫rollingUpdate進行滾動更新，後面會詳細分析。
• 如果DaemonSet更新成功，則根據需要（Old ControllerRevisions數量是否超過Spec.RevisionHistoryLimit，預設為10）清理超過RevisionHistoryLimit的最老的ControllerRevisions。
• updateDaemonSetStatus會更新該Daemonset.Status，跟前面不同的是，這裡還需要更新Status.ObservedGeneration。

DaemonSet Pod的排程

在Kubernetes 1.12之前的版本中，預設由DaemonSet Controller完成Daemon Pods的排程工作，即由DaemonSet Controller給待排程Pod的spec.nodeName設定值，然後對應Node的kubelet watch到該事件，再在本節點建立DaemonSet Pod。在Kubernetes 1.12+，預設啟用了ScheduleDaemonSetPods FeatureGate, DaemonSet的排程就交由default scheduler完成。

DamonSet Pods Should Be On Node

在manage daemonset時，通過呼叫podsShouldBeOnNode來計算出希望在該Node上啟動的DaemonSet Pods(nodesNeedingDaemonPods)、希望在該Node上刪除的DaemonSet Pods(podsToDelete)，以及在該Node上已經Failed DamonSetPods數量，然後在syncNodes中根據這三個資訊，去建立、刪除對應的Pods。

func (dsc *DaemonSetsController) manage(ds *apps.DaemonSet, hash string) error {
	// Find out the pods which are created for the nodes by DaemonSet.
	nodeToDaemonPods, err := dsc.getNodesToDaemonPods(ds)
	...
	for _, node := range nodeList {
		nodesNeedingDaemonPodsOnNode, podsToDeleteOnNode, failedPodsObservedOnNode, err := dsc.podsShouldBeOnNode(
			node, nodeToDaemonPods, ds)

		if err != nil {
			continue
		}

		nodesNeedingDaemonPods = append(nodesNeedingDaemonPods, nodesNeedingDaemonPodsOnNode...)
		podsToDelete = append(podsToDelete, podsToDeleteOnNode...)
		failedPodsObserved += failedPodsObservedOnNode
	}

	// Label new pods using the hash label value of the current history when creating them
	if err = dsc.syncNodes(ds, podsToDelete, nodesNeedingDaemonPods, hash); err != nil {
		return err
	}

	...

	return nil
}

podsShouldBeOnNode是如何計算出nodesNeedingDaemonPods、podsToDelete、failedPodsObserved的呢？—— 通過呼叫nodeShouldRunDaemonPod(node *v1.Node, ds *apps.DaemonSet)計算出如下三個狀態值：

• wantToRun: 當DaemonSet Controller去Simulate排程時，Predicate（主要是GeneralPredicates和PodToleratesNodeTaints）時忽略如下PredicateFailureError（都是些資源類的Error）時為True，有其他PredicateFailureError為False。如果DaemonSet的Spec中指定了NodeName，則根據其是否與node.Name匹配成功來決定wantToRun的值。 - ErrDiskConflict; - ErrVolumeZoneConflict; - ErrMaxVolumeCountExceeded; - ErrNodeUnderMemoryPressure; - ErrNodeUnderDiskPressure; - InsufficientResourceError;
• shouldSchedule:
  - 如果DaemonSet的Spec中指定了NodeName，則根據其是否與node.Name匹配成功來決定shouldSchedule的值。 - 如果Predicate時出現所有型別的PredicateFailureError之一，則shouldSchedule都為false。 - 如果出現InsufficientResourceError，則shouldSchedule也為false。
• failedPodsBackoff *flowcontrol.Backoff: 按照1s,2s,4s,8s,...的backoff週期去處理（刪除重建）Failed DaemonSet Pods，實現流控的效果。DaemonSet Controller Run時會啟動一個協程，每隔2*MaxDuration(2*15Min)會強制進行一次failedPods GC清理。

• shouldContinueRunning，如下情況之一出現，則該值為false，其他情況為true。
  • ErrNodeSelectorNotMatch,
  • ErrPodNotMatchHostName,
  • ErrNodeLabelPresenceViolated,
  • ErrPodNotFitsHostPorts:
  • ErrTaintsTolerationsNotMatch，如果是No Execute型別的Taint/Toleration匹配，則為true，否則為false，也就是說會忽略NoExecute型別的Taint/Toleration匹配。
  • ErrPodAffinityNotMatch,
  • ErrServiceAffinityViolated，
  • unknown predicate failure reason

然後根據這三個狀態值，得到nodesNeedingDaemonPods []string、podsToDelete []string、failedPodsObserved int。

// podsShouldBeOnNode figures out the DaemonSet pods to be created and deleted on the given node:
func (dsc *DaemonSetsController) podsShouldBeOnNode(
	node *v1.Node,
	nodeToDaemonPods map[string][]*v1.Pod,
	ds *apps.DaemonSet,
) (nodesNeedingDaemonPods, podsToDelete []string, failedPodsObserved int, err error) {

	wantToRun, shouldSchedule, shouldContinueRunning, err := dsc.nodeShouldRunDaemonPod(node, ds)
	if err != nil {
		return
	}

	daemonPods, exists := nodeToDaemonPods[node.Name]
	dsKey, _ := cache.MetaNamespaceKeyFunc(ds)

	dsc.removeSuspendedDaemonPods(node.Name, dsKey)

	switch {
	case wantToRun && !shouldSchedule:
		// If daemon pod is supposed to run, but can not be scheduled, add to suspended list.
		dsc.addSuspendedDaemonPods(node.Name, dsKey)
	case shouldSchedule && !exists:
		// If daemon pod is supposed to be running on node, but isn't, create daemon pod.
		nodesNeedingDaemonPods = append(nodesNeedingDaemonPods, node.Name)
	case shouldContinueRunning:
		// If a daemon pod failed, delete it
		// If there's non-daemon pods left on this node, we will create it in the next sync loop
		var daemonPodsRunning []*v1.Pod
		for _, pod := range daemonPods {
			if pod.DeletionTimestamp != nil {
				continue
			}
			if pod.Status.Phase == v1.PodFailed {
				failedPodsObserved++

				// This is a critical place where DS is often fighting with kubelet that rejects pods.
				// We need to avoid hot looping and backoff.
				backoffKey := failedPodsBackoffKey(ds, node.Name)

				now := dsc.failedPodsBackoff.Clock.Now()
				inBackoff := dsc.failedPodsBackoff.IsInBackOffSinceUpdate(backoffKey, now)
				if inBackoff {
					delay := dsc.failedPodsBackoff.Get(backoffKey)
					klog.V(4).Infof("Deleting failed pod %s/%s on node %s has been limited by backoff - %v remaining",
						pod.Namespace, pod.Name, node.Name, delay)
					dsc.enqueueDaemonSetAfter(ds, delay)
					continue
				}

				dsc.failedPodsBackoff.Next(backoffKey, now)

				msg := fmt.Sprintf("Found failed daemon pod %s/%s on node %s, will try to kill it", pod.Namespace, pod.Name, node.Name)
				klog.V(2).Infof(msg)
				// Emit an event so that it's discoverable to users.
				dsc.eventRecorder.Eventf(ds, v1.EventTypeWarning, FailedDaemonPodReason, msg)
				podsToDelete = append(podsToDelete, pod.Name)
			} else {
				daemonPodsRunning = append(daemonPodsRunning, pod)
			}
		}
		// If daemon pod is supposed to be running on node, but more than 1 daemon pod is running, delete the excess daemon pods.
		// Sort the daemon pods by creation time, so the oldest is preserved.
		if len(daemonPodsRunning) > 1 {
			sort.Sort(podByCreationTimestampAndPhase(daemonPodsRunning))
			for i := 1; i < len(daemonPodsRunning); i++ {
				podsToDelete = append(podsToDelete, daemonPodsRunning[i].Name)
			}
		}
	case !shouldContinueRunning && exists:
		// If daemon pod isn't supposed to run on node, but it is, delete all daemon pods on node.
		for _, pod := range daemonPods {
			podsToDelete = append(podsToDelete, pod.Name)
		}
	}

	return nodesNeedingDaemonPods, podsToDelete, failedPodsObserved, nil
}

// nodeShouldRunDaemonPod checks a set of preconditions against a (node,daemonset) and returns a summary. 
func (dsc *DaemonSetsController) nodeShouldRunDaemonPod(node *v1.Node, ds *apps.DaemonSet) (wantToRun, shouldSchedule, shouldContinueRunning bool, err error) {
	newPod := NewPod(ds, node.Name)

	// Because these bools require an && of all their required conditions, we start
	// with all bools set to true and set a bool to false if a condition is not met.
	// A bool should probably not be set to true after this line.
	wantToRun, shouldSchedule, shouldContinueRunning = true, true, true
	// If the daemon set specifies a node name, check that it matches with node.Name.
	if !(ds.Spec.Template.Spec.NodeName == "" || ds.Spec.Template.Spec.NodeName == node.Name) {
		return false, false, false, nil
	}

	reasons, nodeInfo, err := dsc.simulate(newPod, node, ds)
	if err != nil {
		klog.Warningf("DaemonSet Predicates failed on node %s for ds '%s/%s' due to unexpected error: %v", node.Name, ds.ObjectMeta.Namespace, ds.ObjectMeta.Name, err)
		return false, false, false, err
	}

	// TODO(k82cn): When 'ScheduleDaemonSetPods' upgrade to beta or GA, remove unnecessary check on failure reason,
	//              e.g. InsufficientResourceError; and simplify "wantToRun, shouldSchedule, shouldContinueRunning"
	//              into one result, e.g. selectedNode.
	var insufficientResourceErr error
	for _, r := range reasons {
		klog.V(4).Infof("DaemonSet Predicates failed on node %s for ds '%s/%s' for reason: %v", node.Name, ds.ObjectMeta.Namespace, ds.ObjectMeta.Name, r.GetReason())
		switch reason := r.(type) {
		case *predicates.InsufficientResourceError:
			insufficientResourceErr = reason
		case *predicates.PredicateFailureError:
			var emitEvent bool
			// we try to partition predicates into two partitions here: intentional on the part of the operator and not.
			switch reason {
			// intentional
			case
				predicates.ErrNodeSelectorNotMatch,
				predicates.ErrPodNotMatchHostName,
				predicates.ErrNodeLabelPresenceViolated,
				// this one is probably intentional since it's a workaround for not having
				// pod hard anti affinity.
				predicates.ErrPodNotFitsHostPorts:
				return false, false, false, nil
			case predicates.ErrTaintsTolerationsNotMatch:
				// DaemonSet is expected to respect taints and tolerations
				fitsNoExecute, _, err := predicates.PodToleratesNodeNoExecuteTaints(newPod, nil, nodeInfo)
				if err != nil {
					return false, false, false, err
				}
				if !fitsNoExecute {
					return false, false, false, nil
				}
				wantToRun, shouldSchedule = false, false
			// unintentional
			case
				predicates.ErrDiskConflict,
				predicates.ErrVolumeZoneConflict,
				predicates.ErrMaxVolumeCountExceeded,
				predicates.ErrNodeUnderMemoryPressure,
				predicates.ErrNodeUnderDiskPressure:
				// wantToRun and shouldContinueRunning are likely true here. They are
				// absolutely true at the time of writing the comment. See first comment
				// of this method.
				shouldSchedule = false
				emitEvent = true
			// unexpected
			case
				predicates.ErrPodAffinityNotMatch,
				predicates.ErrServiceAffinityViolated:
				klog.Warningf("unexpected predicate failure reason: %s", reason.GetReason())
				return false, false, false, fmt.Errorf("unexpected reason: DaemonSet Predicates should not return reason %s", reason.GetReason())
			default:
				klog.V(4).Infof("unknown predicate failure reason: %s", reason.GetReason())
				wantToRun, shouldSchedule, shouldContinueRunning = false, false, false
				emitEvent = true
			}
			if emitEvent {
				dsc.eventRecorder.Eventf(ds, v1.EventTypeWarning, FailedPlacementReason, "failed to place pod on %q: %s", node.ObjectMeta.Name, reason.GetReason())
			}
		}
	}
	// only emit this event if insufficient resource is the only thing
	// preventing the daemon pod from scheduling
	if shouldSchedule && insufficientResourceErr != nil {
		dsc.eventRecorder.Eventf(ds, v1.EventTypeWarning, FailedPlacementReason, "failed to place pod on %q: %s", node.ObjectMeta.Name, insufficientResourceErr.Error())
		shouldSchedule = false
	}
	return
}

• 如果shouldSchedule && !exists，則會把該Pod加入到nodesNeedingDaemonPods中。

• 如果shouldContinueRunning && pod.DeletionTimestamp == nil && pod.Status.Phase == v1.PodFailed則檢查是否在流控週期（15min, hardcode）中，如果已經超過流控週期，會把該Pod加入到podsToDelete中，否則將再次入佇列。

• 如果shouldContinueRunning && pod.DeletionTimestamp == nil && pod.Status.Phase != v1.PodFailed則會把該Pod加入到daemonPodsRunning中記錄著該DamonSet在該Node上正在執行的非Failed的Pods，如果daemonPodsRunning不止一個，則需要按照建立時間排序，將不是最早建立的其他所有DaemonSet Pods都加入到podsToDelete中。

在nodeShouldRunDaemonPod中呼叫simulate模擬排程返回Pod和Node的匹配結果，根據algorithm.PredicateFailureReason結果知道wantToRun，shouldSchedule,shouldContinueRunning的值。下面我們看看simulate中的排程邏輯。

// Predicates checks if a DaemonSet's pod can be scheduled on a node using GeneralPredicates
// and PodToleratesNodeTaints predicate
func Predicates(pod *v1.Pod, nodeInfo *schedulercache.NodeInfo) (bool, []algorithm.PredicateFailureReason, error) {
	var predicateFails []algorithm.PredicateFailureReason

	// If ScheduleDaemonSetPods is enabled, only check nodeSelector, nodeAffinity and toleration/taint match.
	if utilfeature.DefaultFeatureGate.Enabled(features.ScheduleDaemonSetPods) {
		fit, reasons, err := checkNodeFitness(pod, nil, nodeInfo)
		if err != nil {
			return false, predicateFails, err
		}
		if !fit {
			predicateFails = append(predicateFails, reasons...)
		}

		return len(predicateFails) == 0, predicateFails, nil
	}

	critical := kubelettypes.IsCriticalPod(pod)

	fit, reasons, err := predicates.PodToleratesNodeTaints(pod, nil, nodeInfo)
	if err != nil {
		return false, predicateFails, err
	}
	if !fit {
		predicateFails = append(predicateFails, reasons...)
	}
	if critical {
		// If the pod is marked as critical and support for critical pod annotations is enabled,
		// check predicates for critical pods only.
		fit, reasons, err = predicates.EssentialPredicates(pod, nil, nodeInfo)
	} else {
		fit, reasons, err = predicates.GeneralPredicates(pod, nil, nodeInfo)
	}
	if err != nil {
		return false, predicateFails, err
	}
	if !fit {
		predicateFails = append(predicateFails, reasons...)
	}

	return len(predicateFails) == 0, predicateFails, nil
}

• 如果是啟用了ScheduleDaemonSetPods FeatureGate，則Predicate邏輯如下。這裡並沒有真正的完成排程，只是做了三個predicate檢查，最終的排程還是會交給default scheduler。default scheduler又是如何控制DaemonSet Pod和Node繫結關係的呢，先買個關子。
  • PodFitsHost: 檢查Pod.spec.nodeName非空時是否與Node Name匹配；
  • PodMatchNodeSelector: 檢查Pod的NodeSelector和NodeAffinity是否與Node匹配；
  • PodToleratesNodeTaints: 檢查Pod的NoExecute和NoSchedule型別的Toleration是否與Node Taint匹配。
• 如果是沒啟用ScheduleDaemonSetPods FeatureGate，則Predicate邏輯如下。這裡並沒有真正的完成排程，只是做了幾個predicate檢查，最終的排程還是會交給DaemonSet Controller。
  • PodToleratesNodeTaints：檢查Pod的NoExecute和NoSchedule型別的Toleration是否與Node Taint匹配。
  • 如果是Critical DaemonSet Pod，則再進行EssentialPredicates，包括：
    • PodFitsHost：檢查Pod.spec.nodeName非空時是否與Node Name匹配；
    • PodFitsHostPorts：檢查DaemonSet Pods請求的協議&Host埠是否已經被佔用；
    • PodMatchNodeSelector: 檢查Pod的NodeSelector和NodeAffinity是否與Node匹配；
  • 如果不是Critical DaemonSet Pod，則再進行GeneralPredicates，
    • PodFitsResources:檢查Node剩餘可分配資源是否能滿足Pod請求；
    • PodFitsHost: 檢查Pod.spec.nodeName非空時是否與Node Name匹配；
    • PodFitsHostPorts: 檢查DaemonSet Pods請求的協議&Host埠是否已經被佔用；
    • PodMatchNodeSelector: 檢查Pod的NodeSelector和NodeAffinity是否與Node匹配；

Sync Nodes

前面通過podsShouldBeOnNode得到了nodesNeedingDaemonPods []string, podsToDelete []string, failedPodsObserved int，接下來就該去建立和刪除對應的Pods了。

// syncNodes deletes given pods and creates new daemon set pods on the given nodes
// returns slice with erros if any
func (dsc *DaemonSetsController) syncNodes(ds *apps.DaemonSet, podsToDelete, nodesNeedingDaemonPods []string, hash string) error {
	// We need to set expectations before creating/deleting pods to avoid race conditions.
	dsKey, err := controller.KeyFunc(ds)
	if err != nil {
		return fmt.Errorf("couldn't get key for object %#v: %v", ds, err)
	}

	createDiff := len(nodesNeedingDaemonPods)
	deleteDiff := len(podsToDelete)

	if createDiff > dsc.burstReplicas {
		createDiff = dsc.burstReplicas
	}
	if deleteDiff > dsc.burstReplicas {
		deleteDiff = dsc.burstReplicas
	}

	dsc.expectations.SetExpectations(dsKey, createDiff, deleteDiff)

	// error channel to communicate back failures.  make the buffer big enough to avoid any blocking
	errCh := make(chan error, createDiff+deleteDiff)

	klog.V(4).Infof("Nodes needing daemon pods for daemon set %s: %+v, creating %d", ds.Name, nodesNeedingDaemonPods, createDiff)
	createWait := sync.WaitGroup{}
	// If the returned error is not nil we have a parse error.
	// The controller handles this via the hash.
	generation, err := util.GetTemplateGeneration(ds)
	if err != nil {
		generation = nil
	}
	template := util.CreatePodTemplate(ds.Namespace, ds.Spec.Template, generation, hash)
	// Batch the pod creates. Batch sizes start at SlowStartInitialBatchSize
	// and double with each successful iteration in a kind of "slow start".
	// This handles attempts to start large numbers of pods that would
	// likely all fail with the same error. For example a project with a
	// low quota that attempts to create a large number of pods will be
	// prevented from spamming the API service with the pod create requests
	// after one of its pods fails.  Conveniently, this also prevents the
	// event spam that those failures would generate.
	batchSize := integer.IntMin(createDiff, controller.SlowStartInitialBatchSize)
	for pos := 0; createDiff > pos; batchSize, pos = integer.IntMin(2*batchSize, createDiff-(pos+batchSize)), pos+batchSize {
		errorCount := len(errCh)
		createWait.Add(batchSize)
		for i := pos; i < pos+batchSize; i++ {
			go func(ix int) {
				defer createWait.Done()
				var err error

				podTemplate := &template

				if utilfeature.DefaultFeatureGate.Enabled(features.ScheduleDaemonSetPods) {
					podTemplate = template.DeepCopy()
					// The pod's NodeAffinity will be updated to make sure the Pod is bound
					// to the target node by default scheduler. It is safe to do so because there
					// should be no conflicting node affinity with the target node.
					podTemplate.Spec.Affinity = util.ReplaceDaemonSetPodNodeNameNodeAffinity(
						podTemplate.Spec.Affinity, nodesNeedingDaemonPods[ix])

					err = dsc.podControl.CreatePodsWithControllerRef(ds.Namespace, podTemplate,
						ds, metav1.NewControllerRef(ds, controllerKind))
				} else {
					err = dsc.podControl.CreatePodsOnNode(nodesNeedingDaemonPods[ix], ds.Namespace, podTemplate,
						ds, metav1.NewControllerRef(ds, controllerKind))
				}

				if err != nil && errors.IsTimeout(err) {
					// Pod is created but its initialization has timed out.
					// If the initialization is successful eventually, the
					// controller will observe the creation via the informer.
					// If the initialization fails, or if the pod keeps
					// uninitialized for a long time, the informer will not
					// receive any update, and the controller will create a new
					// pod when the expectation expires.
					return
				}
				if err != nil {
					klog.V(2).Infof("Failed creation, decrementing expectations for set %q/%q", ds.Namespace, ds.Name)
					dsc.expectations.CreationObserved(dsKey)
					errCh <- err
					utilruntime.HandleError(err)
				}
			}(i)
		}
		createWait.Wait()
		// any skipped pods that we never attempted to start shouldn't be expected.
		skippedPods := createDiff - batchSize
		if errorCount < len(errCh) && skippedPods > 0 {
			klog.V(2).Infof("Slow-start failure. Skipping creation of %d pods, decrementing expectations for set %q/%q", skippedPods, ds.Namespace, ds.Name)
			for i := 0; i < skippedPods; i++ {
				dsc.expectations.CreationObserved(dsKey)
			}
			// The skipped pods will be retried later. The next controller resync will
			// retry the slow start process.
			break
		}
	}

	klog.V(4).Infof("Pods to delete for daemon set %s: %+v, deleting %d", ds.Name, podsToDelete, deleteDiff)
	deleteWait := sync.WaitGroup{}
	deleteWait.Add(deleteDiff)
	for i := 0; i < deleteDiff; i++ {
		go func(ix int) {
			defer deleteWait.Done()
			if err := dsc.podControl.DeletePod(ds.Namespace, podsToDelete[ix], ds); err != nil {
				klog.V(2).Infof("Failed deletion, decrementing expectations for set %q/%q", ds.Namespace, ds.Name)
				dsc.expectations.DeletionObserved(dsKey)
				errCh <- err
				utilruntime.HandleError(err)
			}
		}(i)
	}
	deleteWait.Wait()

	// collect errors if any for proper reporting/retry logic in the controller
	errors := []error{}
	close(errCh)
	for err := range errCh {
		errors = append(errors, err)
	}
	return utilerrors.NewAggregate(errors)
}

• 每次刪除和建立的最大Pods個數分別為250個。
• 根據DaemonSet Object構建Pod Template，並且增加/更新以下Tolerations：
  • node.kubernetes.io/not-ready | exist | NoExecute
  • node.kubernetes.io/unreachable | exist | NoExecute
  • node.kubernetes.io/disk-pressure | exist | NoSchedule
  • node.kubernetes.io/memory-pressure | exist | NoSchedule
  • node.kubernetes.io/unschedulable | exist | NoSchedule
  • node.kubernetes.io/network-unavailable | exist | NoSchedule
  • 如果是Critical Pod，還會增加以下Tolerations：
    • node.kubernetes.io/out-of-disk | exist | NoExecute
    • node.kubernetes.io/out-of-disk | exist | NoSchedule
• 給Pod加上Label: controller-revision-hash=$DaemonSetControlelrHash
• 分批的建立DaemonSet Pods（按照1,2,4,8,...的batch size去Create DaemonSet Pods，防止大批量的一次性建立所有DaemonSet Pods時因同樣的錯誤導致失敗。對於建立失敗的Pods，注意更新expectations中的Adds值，每失敗一個就會將expectations.adds值減1。
  • 如果啟用了ScheduleDaemonSetPods FeatureGate，則往Pod Tempalete中新增/更新metadata.name=$NodeName的NodeAffinity。通過這種方式，來實現通過default scheduler來排程DaemonSet Pods的目的。
• 一次性的刪除podsToDelete的Pods。

DaemonSet的滾動更新

DaemonSet的滾動更新，跟Deployment的滾動更新略有不同，DaemonSet RollingUpdate只有MaxUnavailable這一個配置項，沒有MinAvailable。

// rollingUpdate deletes old daemon set pods making sure that no more than
// ds.Spec.UpdateStrategy.RollingUpdate.MaxUnavailable pods are unavailable
func (dsc *DaemonSetsController) rollingrollingrollingUpdate(ds *apps.DaemonSet, hash string) error {
	nodeToDaemonPods, err := dsc.getNodesToDaemonPods(ds)
	if err != nil {
		return fmt.Errorf("couldn't get node to daemon pod mapping for daemon set %q: %v", ds.Name, err)
	}

	_, oldPods := dsc.getAllDaemonSetPods(ds, nodeToDaemonPods, hash)
	maxUnavailable, numUnavailable, err := dsc.getUnavailableNumbers(ds, nodeToDaemonPods)
	if err != nil {
		return fmt.Errorf("Couldn't get unavailable numbers: %v", err)
	}
	oldAvailablePods, oldUnavailablePods := util.SplitByAvailablePods(ds.Spec.MinReadySeconds, oldPods)

	// for oldPods delete all not running pods
	var oldPodsToDelete []string
	klog.V(4).Infof("Marking all unavailable old pods for deletion")
	for _, pod := range oldUnavailablePods {
		// Skip terminating pods. We won't delete them again
		if pod.DeletionTimestamp != nil {
			continue
		}
		klog.V(4).Infof("Marking pod %s/%s for deletion", ds.Name, pod.Name)
		oldPodsToDelete = append(oldPodsToDelete, pod.Name)
	}

	klog.V(4).Infof("Marking old pods for deletion")
	for _, pod := range oldAvailablePods {
		if numUnavailable >= maxUnavailable {
			klog.V(4).Infof("Number of unavailable DaemonSet pods: %d, is equal to or exceeds allowed maximum: %d", numUnavailable, maxUnavailable)
			break
		}
		klog.V(4).Infof("Marking pod %s/%s for deletion", ds.Name, pod.Name)
		oldPodsToDelete = append(oldPodsToDelete, pod.Name)
		numUnavailable++
	}
	return dsc.syncNodes(ds, oldPodsToDelete, []string{}, hash)
}

• 根據最新的Hash值選出所有的OldPods；
• 計算那些!available及那些期望排程但還沒執行的Pods之和，作為numUnavailable。
• 將OldPods分為oldAvailablePods和oldUnavailablePods，將DeletionTimestamp為空的oldUnavailablePods加入到待刪除Pods列表（oldPodsToDelete）。
• 遍歷oldAvailablePods，逐個加入到oldPodsToDelete中，直到numUnavailable達到maxUnavailable為止，從oldAvailablePods加入到oldPodsToDelete的Pods最大個數為(maxUnavailable - 1)。
• 因此，oldPodsToDelete包括所有的DeletionTimestamp為空的oldUnavailablePods及最多(maxUnavailable - 1)個oldAvailablePods。
• 最後呼叫syncNodes開始刪除oldPodsToDelete中的DaemonSet Pods。

Node更新

Node Add事件很簡單，遍歷所有DaemonSets物件，呼叫nodeShouldRunDaemonPod計算出每個DaemonSet是否應該在該Node上啟動。如果要啟動，則把DaemonSet加入到Queue，由syncDaemonSet進行處理。

對於Node Update事件，需要判斷Update的欄位等，然後根據情況決定是否要加入到Queue進行syncDaemonSet。

func (dsc *DaemonSetsController) updateNode(old, cur interface{}) {
	oldNode := old.(*v1.Node)
	curNode := cur.(*v1.Node)
	if shouldIgnoreNodeUpdate(*oldNode, *curNode) {
		return
	}

	dsList, err := dsc.dsLister.List(labels.Everything())
	if err != nil {
		klog.V(4).Infof("Error listing daemon sets: %v", err)
		return
	}
	// TODO: it'd be nice to pass a hint with these enqueues, so that each ds would only examine the added node (unless it has other work to do, too).
	for _, ds := range dsList {
		_, oldShouldSchedule, oldShouldContinueRunning, err := dsc.nodeShouldRunDaemonPod(oldNode, ds)
		if err != nil {
			continue
		}
		_, currentShouldSchedule, currentShouldContinueRunning, err := dsc.nodeShouldRunDaemonPod(curNode, ds)
		if err != nil {
			continue
		}
		if (oldShouldSchedule != currentShouldSchedule) || (oldShouldContinueRunning != currentShouldContinueRunning) {
			dsc.enqueueDaemonSet(ds)
		}
	}
}

• 如果Node Condition沒有發生變更，則不能忽略該Node變更事件。
• 除了Node Condition和ResourceVersion之外，如果新舊Node物件不一致，也不能忽略該變更事件。
• 對於不能忽略的變更，則分別對於oldNode，currentNode呼叫nodeShouldRunDaemonPod計算ShouldSchedule、ShouldContinueRunning是否一致，只要ShouldSchedule或者ShouldContinueRunning發生變更，則將該DaemonSet Object入佇列進入syncDaemonSet進行處理。

DaemonSet Controller主體邏輯

總結

本文主要對DaemonSet的結構、建立、同步、排程、滾動更新幾個方面進行了原始碼分析，在生產環境中使用DaemonSet進行大規模部署使用之前，加深這些瞭解是有幫助的。下一篇部落格，我將會從一些實際問題出發，從使用者角度分析DaemonSet的若干行為。比如，Node Taint變更後DaemonSet的行為、DaemonSet刪除時異常導致Hang住的原因及解決辦法、Node NotReady時DamonSet Pods會