K8S源码分析Kubelet - 6、Pod驱逐

为了保证Node节点的稳定性，当资源(memory/storage)出现紧缺时，kubelet会主动选择驱逐一些Pods来释放资源。实现该功能的组件是Eviction Manager。

当驱逐一个Pod时，kubelet会将pod内的所有containers都kill掉，并把pod的状态设置为Failed。被kill掉的pod，可能被调度到其他node上。

可以人为定义Thresholds来告诉Kubelet在什么情况下驱逐pods。有两种类型的thresholds：

Soft Eviction Thresholds - 到达阈值时，并不会马上触发驱逐操作，而是会等待一个用户配置的grace period之后再触发。

Hard Eviction Thresholds - 立刻Kill Pods。

关于Eviction相关的内容可以阅读官方文档 https://kubernetes.io/docs/tasks/administer-cluster/out-of-resource/

实现

Eviction Manager相关的代码在包/pkg/kubelet/eviction中，核心逻辑是managerImpl.synchronize方法。EvictionManager会在一个单独的协程中周期性调用synchronize方法，实现驱逐。

synchronize方法主要包含以下几个步骤：

1.初始化配置

func (m *managerImpl) synchronize(diskInfoProvider DiskInfoProvider, podFunc ActivePodsFunc, capacityProvider CapacityProvider) []*v1.Pod {    // 1. 从配置中读取所有Thresholds    thresholds := m.config.Thresholds    if len(thresholds) == 0 {       return nil    }      ....    // 2. 初始化rank funcs/reclaim funcs等    if m.dedicatedImageFs == nil {       hasImageFs, ok := diskInfoProvider.HasDedicatedImageFs()       if ok != nil {          return nil       }       m.dedicatedImageFs = &hasImageFs       m.resourceToRankFunc = buildResourceToRankFunc(hasImageFs)       m.resourceToNodeReclaimFuncs = buildResourceToNodeReclaimFuncs(m.imageGC, m.containerGC, hasImageFs)    }      // 3. 通过初始化传入的func获取Pods    activePods := podFunc()    // 4. 通过summary provider获得当前资源使用情况    observations, statsFunc, err := makeSignalObservations(m.summaryProvider, capacityProvider, activePods)    ....      // 5. 通过memcg加速内存占用通知，只在notifiersInitialized 为false时进入    if m.config.KernelMemcgNotification && !m.notifiersInitialized {       ....       m.notifiersInitialized = true  // 初始化完成       err = startMemoryThresholdNotifier(m.config.Thresholds, observations, true, func(desc string) {          // 回调函数，memcg的通知会立即触发synchronize函数         glog.Infof("hard memory eviction threshold crossed at %s", desc)          m.synchronize(diskInfoProvider, podFunc, capacityProvider)       })       ....    }      .... }

2.计算Thresholds

func (m *managerImpl) synchronize(diskInfoProvider DiskInfoProvider, podFunc ActivePodsFunc, capacityProvider CapacityProvider) []*v1.Pod {     ....     // 1. 根据配置的thresholds参数以及当前资源使用情况，计算出超出的thresholds     thresholds = thresholdsMet(thresholds, observations, false)       // 2. 合并上次计算的thresholds结果     if len(m.thresholdsMet) > 0 {         thresholdsNotYetResolved := thresholdsMet(m.thresholdsMet, observations, true)         thresholds = mergeThresholds(thresholds, thresholdsNotYetResolved)     }       // 3. 过滤未真正激活的soft thresholds     now := m.clock.Now()     thresholdsFirstObservedAt := thresholdsFirstObservedAt(thresholds, m.thresholdsFirstObservedAt, now)     ....     thresholds = thresholdsMetGracePeriod(thresholdsFirstObservedAt, now)        // 4. 更新计算结果     m.Lock()     m.nodeConditions = nodeConditions     m.thresholdsFirstObservedAt = thresholdsFirstObservedAt     m.nodeConditionsLastObservedAt = nodeConditionsLastObservedAt     m.thresholdsMet = thresholds       // determine the set of thresholds whose stats have been updated since the last sync     thresholds = thresholdsUpdatedStats(thresholds, observations, m.lastObservations)     debugLogThresholdsWithObservation("thresholds - updated stats", thresholds, observations)       m.lastObservations = observations     m.Unlock()      ... }

3.计算本轮Eviction考察的Resource

在每一轮Eviction中，kubelet至多只会kill一个Pod。由于Eviction Manager会同时处理多种资源(memory/storage)的紧缺情况，因此在选择Pod之前，首先会选出本轮Eviction参考的资源类型，再将Pods对该种资源的使用量进行排序，选出kill掉的Pod。

func (m *managerImpl) synchronize(diskInfoProvider DiskInfoProvider, podFunc ActivePodsFunc, capacityProvider CapacityProvider) []*v1.Pod {     ....     // 1. 收集当前所有发生紧缺的resource类型     starvedResources := getStarvedResources(thresholds)     if len(starvedResources) == 0 {         glog.V(3).Infof("eviction manager: no resources are starved")         return nil     }       // 2. 排序并选择其中一种resource     sort.Sort(byEvictionPriority(starvedResources))     resourceToReclaim := starvedResources[0]       // determine if this is a soft or hard eviction associated with the resource     softEviction := isSoftEvictionThresholds(thresholds, resourceToReclaim)     .....       // 3. 根据选中Resource的使用量来排序Pods     rank, ok := m.resourceToRankFunc[resourceToReclaim]     ....     rank(activePods, statsFunc)       .... }

4. Kill Pod

在排好序的Pods中选择第一个Kill掉：

func (m *managerImpl) synchronize(diskInfoProvider DiskInfoProvider, podFunc ActivePodsFunc, capacityProvider CapacityProvider) []*v1.Pod {     ....     for i := range activePods {         pod := activePods[i]         ...         status := v1.PodStatus{             Phase:   v1.PodFailed,             Message: fmt.Sprintf(message, resourceToReclaim),             Reason:  reason,         }         ....         gracePeriodOverride := int64(0)         if softEviction {             gracePeriodOverride = m.config.MaxPodGracePeriodSeconds         }         // 真正KillPod         err := m.killPodFunc(pod, status, &gracePeriodOverride)         if err != nil {             glog.Warningf("eviction manager: error while evicting pod %s: %v", format.Pod(pod), err)         }         return []*v1.Pod{pod}     }     glog.Infof("eviction manager: unable to evict any pods from the node")     return nil }

转载于:https://my.oschina.net/zhuhui/blog/3081731