Kubernetes PodGC Controller源码分析

Author: xidianwangtao@gmail.com

PodGC Controller配置

关于PodGC Controller的相关配置（kube-controller-manager配置），一共只有两个：

PodGC Controller入口

PodGC Controller是在kube-controller-manager Run的时候启动的。CMServer Run时会invoke StartControllers将预先注册的enabled Controllers遍历并逐个启动。

cmd/kube-controller-manager/app/controllermanager.go:180

func Run(s *options.CMServer) error {
   ...
	err := StartControllers(newControllerInitializers(), s, rootClientBuilder, clientBuilder, stop)
	...
}

在newControllerInitializers注册了所有一些常规Controllers及其对应的start方法，为什么说这些是常规的Controllers呢，因为还有一部分Controllers没在这里进行注册，比如非常重要的service Controller，node Controller等，我把这些称为非常规Controllers。

func newControllerInitializers() map[string]InitFunc {
	controllers := map[string]InitFunc{}
	controllers["endpoint"] = startEndpointController
	...
	controllers["podgc"] = startPodGCController
	...

	return controllers
}

因此CMServer最终是invoke startPodGCController来启动PodGC Controller的。

cmd/kube-controller-manager/app/core.go:66

func startPodGCController(ctx ControllerContext) (bool, error) {
	go podgc.NewPodGC(
		ctx.ClientBuilder.ClientOrDie("pod-garbage-collector"),
		ctx.InformerFactory.Core().V1().Pods(),
		int(ctx.Options.TerminatedPodGCThreshold),
	).Run(ctx.Stop)
	return true, nil
}

startPodGCController内容很简单，启动一个goruntine协程，创建PodGC并启动执行。

PodGC Controller的创建

我们先来看看PodGCController的定义。

pkg/controller/podgc/gc_controller.go:44

type PodGCController struct {
	kubeClient clientset.Interface

	podLister       corelisters.PodLister
	podListerSynced cache.InformerSynced

	deletePod              func(namespace, name string) error
	terminatedPodThreshold int
}

• kubeClient: 用来跟APIServer通信的client。
• PodLister: PodLister helps list Pods.
• podListerSynced: 用来判断PodLister是否Has Synced。
• deletePod: 调用apiserver删除对应pod的接口。
• terminatedPodThreshold: 对应--terminated-pod-gc-threshold的配置，默认为12500。

pkg/controller/podgc/gc_controller.go:54

func NewPodGC(kubeClient clientset.Interface, podInformer coreinformers.PodInformer, terminatedPodThreshold int) *PodGCController {
	if kubeClient != nil && kubeClient.Core().RESTClient().GetRateLimiter() != nil {
		metrics.RegisterMetricAndTrackRateLimiterUsage("gc_controller", kubeClient.Core().RESTClient().GetRateLimiter())
	}
	gcc := &PodGCController{
		kubeClient:             kubeClient,
		terminatedPodThreshold: terminatedPodThreshold,
		deletePod: func(namespace, name string) error {
			glog.Infof("PodGC is force deleting Pod: %v:%v", namespace, name)
			return kubeClient.Core().Pods(namespace).Delete(name, metav1.NewDeleteOptions(0))
		},
	}

	gcc.podLister = podInformer.Lister()
	gcc.podListerSynced = podInformer.Informer().HasSynced

	return gcc
}

创建PodGC Controller时其实只是把相关的PodGCController元素进行赋值。注意deletePod方法定义时的参数metav1.NewDeleteOptions(0)，表示立即删除pod，没有grace period。

PodGC Controller的运行

创建完PodGC Controller后，接下来就是执行Run方法启动执行了。

pkg/controller/podgc/gc_controller.go:73

func (gcc *PodGCController) Run(stop <-chan struct{}) {
	if !cache.WaitForCacheSync(stop, gcc.podListerSynced) {
		utilruntime.HandleError(fmt.Errorf("timed out waiting for caches to sync"))
		return
	}

	go wait.Until(gcc.gc, gcCheckPeriod, stop)
	<-stop
}

• 每100ms都会去检查对应的PodLister是否Has Synced，直到Has Synced。
• 启动goruntine协程，每执行完一次gcc.gc进行Pod回收后，等待20s，再次执行gcc.gc，直到收到stop信号。

pkg/controller/podgc/gc_controller.go:83

func (gcc *PodGCController) gc() {
	pods, err := gcc.podLister.List(labels.Everything())
	if err != nil {
		glog.Errorf("Error while listing all Pods: %v", err)
		return
	}
	if gcc.terminatedPodThreshold > 0 {
		gcc.gcTerminated(pods)
	}
	gcc.gcOrphaned(pods)
	gcc.gcUnscheduledTerminating(pods)
}

gcc.gc是最终的pod回收逻辑：

• 调从PodLister中去除所有的pods（不设置过滤）
• 如果terminatedPodThreshold大于0，则调用gcc.gcTerminated(pods)回收那些超出Threshold的Pods。
• 调用gcc.gcOrphaned(pods)回收Orphaned pods。
• 调用gcc.gcUnscheduledTerminating(pods)回收UnscheduledTerminating pods。

注意：

1. gcTerminated和gcOrphaned，gcUnscheduledTerminating这三个gc都是串行执行的。
2. gcTerminated删除超出阈值的pods的删除动作是并行的，通过sync.WaitGroup等待所有对应的pods删除完成后，gcTerminated才会结束返回，才能开始后面的gcOrphaned.
3. gcOrphaned，gcUnscheduledTerminatin，gcUnscheduledTerminatin内部都是串行gc pods的。

回收那些Terminated的pods

func (gcc *PodGCController) gcTerminated(pods []*v1.Pod) {
	terminatedPods := []*v1.Pod{}
	for _, pod := range pods {
		if isPodTerminated(pod) {
			terminatedPods = append(terminatedPods, pod)
		}
	}

	terminatedPodCount := len(terminatedPods)
	sort.Sort(byCreationTimestamp(terminatedPods))

	deleteCount := terminatedPodCount - gcc.terminatedPodThreshold

	if deleteCount > terminatedPodCount {
		deleteCount = terminatedPodCount
	}
	if deleteCount > 0 {
		glog.Infof("garbage collecting %v pods", deleteCount)
	}

	var wait sync.WaitGroup
	for i := 0; i < deleteCount; i++ {
		wait.Add(1)
		go func(namespace string, name string) {
			defer wait.Done()
			if err := gcc.deletePod(namespace, name); err != nil {
				// ignore not founds
				defer utilruntime.HandleError(err)
			}
		}(terminatedPods[i].Namespace, terminatedPods[i].Name)
	}
	wait.Wait()
}

• 遍历所有pods，过滤出所有Terminated Pods（Pod.Status.Phase不为Pending, Running, Unknow的Pods）.
• 计算terminated pods数与terminatedPodThreshold的(超出)差值deleteCount。
• 启动deleteCount数量的goruntine协程，并行调用gcc.deletePod（invoke apiserver's api）方法立刻删除对应的pod。

回收那些Binded的Nodes已经不存在的pods

// gcOrphaned deletes pods that are bound to nodes that don't exist.
func (gcc *PodGCController) gcOrphaned(pods []*v1.Pod) {
	glog.V(4).Infof("GC'ing orphaned")
	// We want to get list of Nodes from the etcd, to make sure that it's as fresh as possible.
	nodes, err := gcc.kubeClient.Core().Nodes().List(metav1.ListOptions{})
	if err != nil {
		return
	}
	nodeNames := sets.NewString()
	for i := range nodes.Items {
		nodeNames.Insert(nodes.Items[i].Name)
	}

	for _, pod := range pods {
		if pod.Spec.NodeName == "" {
			continue
		}
		if nodeNames.Has(pod.Spec.NodeName) {
			continue
		}
		glog.V(2).Infof("Found orphaned Pod %v assigned to the Node %v. Deleting.", pod.Name, pod.Spec.NodeName)
		if err := gcc.deletePod(pod.Namespace, pod.Name); err != nil {
			utilruntime.HandleError(err)
		} else {
			glog.V(0).Infof("Forced deletion of orphaned Pod %s succeeded", pod.Name)
		}
	}
}

gcOrphaned用来删除那些bind的node已经不存在的pods。

• 调用apiserver接口，获取所有的Nodes。
• 遍历所有pods，如果pod bind的NodeName不为空且不包含在刚刚获取的所有Nodes中，则串行逐个调用gcc.deletePod删除对应的pod。

回收Unscheduled并且Terminating的pods

pkg/controller/podgc/gc_controller.go:167

// gcUnscheduledTerminating deletes pods that are terminating and haven't been scheduled to a particular node.
func (gcc *PodGCController) gcUnscheduledTerminating(pods []*v1.Pod) {
	glog.V(4).Infof("GC'ing unscheduled pods which are terminating.")

	for _, pod := range pods {
		if pod.DeletionTimestamp == nil || len(pod.Spec.NodeName) > 0 {
			continue
		}

		glog.V(2).Infof("Found unscheduled terminating Pod %v not assigned to any Node. Deleting.", pod.Name)
		if err := gcc.deletePod(pod.Namespace, pod.Name); err != nil {
			utilruntime.HandleError(err)
		} else {
			glog.V(0).Infof("Forced deletion of unscheduled terminating Pod %s succeeded", pod.Name)
		}
	}
}

gcUnscheduledTerminating删除那些terminating并且还没调度到某个node的pods。

• 遍历所有pods，过滤那些terminating(pod.DeletionTimestamp != nil)并且未调度成功的(pod.Spec.NodeName为空)的pods。
• 串行逐个调用gcc.deletePod删除对应的pod。

总结

PodGC Controller作为Kubernetes默认启动的Controllers之一，在Master后台每隔20s进行一次Pod GC。

• 通过--controllers可以控制PodGC Controller的开关。
• 通过--terminated-pod-gc-threshold设置gcTerminated的阈值。
• PodGC Controller串行的执行以下三个gc子过程：
  • 回收超过阈值的Terminated Pods（Pod.Status.Phase不为Pending, Running, Unknow的Pods）。
  • 回收那些binded的node已经不存在（不在etcd中）的pods。
  • 回收那些terminating并且还没调度到某个node的pods。