本文介绍了Kubelet在Predicate Admit准入检查时对CriticalPod的资源抢占的原理，以及Priority Admission Controller对CriticalPod的PriorityClassName特殊处理。
深入分析Kubernetes Critical Pod系列：
深入分析Kubernetes Critical Pod（一）
深入分析Kubernetes Critical Pod（二）
深入分析Kubernetes Critical Pod（三）
深入分析Kubernetes Critical Pod（四）
Kubelet Predicate Admit时对Critical的资源抢占处理
kubelet 在Predicate Admit流程中，会对Pods进行各种Predicate准入检查，包括GeneralPredicates检查本节点是否有足够的cpu,mem,gpu资源。如果GeneralPredicates准入检测失败，对于nonCriticalPod则直接Admit失败，但如果是CriticalPod则会触发kubelet preemption进行资源抢占，按照一定规则杀死一些Pods释放资源，抢占成功，则Admit成功。
流程的源头应该从kubelet初始化的流程开始。
pkg/kubelet/kubelet.go:315
// NewMainKubelet instantiates a new Kubelet object along with all the required internal modules.
// No initialization of Kubelet and its modules should happen here.
func NewMainKubelet(...) (*Kubelet, error) {
...
criticalPodAdmissionHandler := preemption.NewCriticalPodAdmissionHandler(klet.GetActivePods, killPodNow(klet.podWorkers, kubeDeps.Recorder), kubeDeps.Recorder)
klet.admitHandlers.AddPodAdmitHandler(lifecycle.NewPredicateAdmitHandler(klet.getNodeAnyWay, criticalPodAdmissionHandler, klet.containerManager.UpdatePluginResources))
// apply functional Option's
for _, opt := range kubeDeps.Options {
opt(klet)
}
...
return klet, nil
}在NewMainKubelet对kubelet进行初始化时，通过AddPodAdmitHandler注册了criticalPodAdmissionHandler，CriticalPod的Admit的特殊之处就体现在criticalPodAdmissionHandler。
然后，我们进入kubelet的predicateAdmitHandler流程中，看看GeneralPredicates失败后的处理逻辑。
pkg/kubelet/lifecycle/predicate.go:58
func (w *predicateAdmitHandler) Admit(attrs *PodAdmitAttributes) PodAdmitResult {
...
fit, reasons, err := predicates.GeneralPredicates(podWithoutMissingExtendedResources, nil, nodeInfo)
if err != nil {
message := fmt.Sprintf("GeneralPredicates failed due to %v, which is unexpected.", err)
glog.Warningf("Failed to admit pod %v - %s", format.Pod(pod), message)
return PodAdmitResult{
Admit: fit,
Reason: "UnexpectedAdmissionError",
Message: message,
}
}
if !fit {
fit, reasons, err = w.admissionFailureHandler.HandleAdmissionFailure(pod, reasons)
if err != nil {
message := fmt.Sprintf("Unexpected error while attempting to recover from admission failure: %v", err)
glog.Warningf("Failed to admit pod %v - %s", format.Pod(pod), message)
return PodAdmitResult{
Admit: fit,
Reason: "UnexpectedAdmissionError",
Message: message,
}
}
}
...
return PodAdmitResult{
Admit: true,
}
}在kubelet predicateAdmitHandler中对Pod进行GeneralPredicates检查cpu,mem,gpu资源时，如果发现资源不足导致Admit失败，则接着调用HandleAdmissionFailure进行额外处理。前提提到，kubelet初始化时注册了criticalPodAdmissionHandler为HandleAdmissionFailure。
CriticalPodAdmissionHandler struct定义如下：
pkg/kubelet/preemption/preemption.go:41
type CriticalPodAdmissionHandler struct {
getPodsFunc eviction.ActivePodsFunc
killPodFunc eviction.KillPodFunc
recorder record.EventRecorder
}CriticalPodAdmissionHandler的HandleAdmissionFailure方法就是处理CriticalPod特殊的逻辑所在。
pkg/kubelet/preemption/preemption.go:66
// HandleAdmissionFailure gracefully handles admission rejection, and, in some cases,
// to allow admission of the pod despite its previous failure.
func (c *CriticalPodAdmissionHandler) HandleAdmissionFailure(pod *v1.Pod, failureReasons []algorithm.PredicateFailureReason) (bool, []algorithm.PredicateFailureReason, error) {
if !kubetypes.IsCriticalPod(pod) || !utilfeature.DefaultFeatureGate.Enabled(features.ExperimentalCriticalPodAnnotation) {
return false, failureReasons, nil
}
// InsufficientResourceError is not a reason to reject a critical pod.
// Instead of rejecting, we free up resources to admit it, if no other reasons for rejection exist.
nonResourceReasons := []algorithm.PredicateFailureReason{}
resourceReasons := []*admissionRequirement{}
for _, reason := range failureReasons {
if r, ok := reason.(*predicates.InsufficientResourceError); ok {
resourceReasons = append(resourceReasons, &admissionRequirement{
resourceName: r.ResourceName,
quantity: r.GetInsufficientAmount(),
})
} else {
nonResourceReasons = append(nonResourceReasons, reason)
}
}
if len(nonResourceReasons) > 0 {
// Return only reasons that are not resource related, since critical pods cannot fail admission for resource reasons.
return false, nonResourceReasons, nil
}
err := c.evictPodsToFreeRequests(admissionRequirementList(resourceReasons))
// if no error is returned, preemption succeeded and the pod is safe to admit.
return err == nil, nil, err
}- 如果Pod不是CriticalPod，或者ExperimentalCriticalPodAnnotation Feature Gate是关闭的，则直接返回false，表示Admit失败。
- 判断Admit的failureReasons是否包含
predicate.InsufficientResourceError，如果包含，则调用evictPodsToFreeRequests触发kubelet preemption。注意这里的抢占不同于scheduler preemtion，不要混淆了。
evictPodsToFreeRequests就是kubelet preemption进行资源抢占的逻辑实现，其核心就是调用getPodsToPreempt挑选合适的待杀死的Pods(podsToPreempt)。
pkg/kubelet/preemption/preemption.go:121
// getPodsToPreempt returns a list of pods that could be preempted to free requests >= requirements
func getPodsToPreempt(pods []*v1.Pod, requirements admissionRequirementList) ([]*v1.Pod, error) {
bestEffortPods, burstablePods, guaranteedPods := sortPodsByQOS(pods)
// make sure that pods exist to reclaim the requirements
unableToMeetRequirements := requirements.subtract(append(append(bestEffortPods, burstablePods...), guaranteedPods...)...)
if len(unableToMeetRequirements) > 0 {
return nil, fmt.Errorf("no set of running pods found to reclaim resources: %v", unableToMeetRequirements.toString())
}
// find the guaranteed pods we would need to evict if we already evicted ALL burstable and besteffort pods.
guarateedToEvict, err := getPodsToPreemptByDistance(guaranteedPods, requirements.subtract(append(bestEffortPods, burstablePods...)...))
if err != nil {
return nil, err
}
// Find the burstable pods we would need to evict if we already evicted ALL besteffort pods, and the required guaranteed pods.
burstableToEvict, err := getPodsToPreemptByDistance(burstablePods, requirements.subtract(append(bestEffortPods, guarateedToEvict...)...))
if err != nil {
return nil, err
}
// Find the besteffort pods we would need to evict if we already evicted the required guaranteed and burstable pods.
bestEffortToEvict, err := getPodsToPreemptByDistance(bestEffortPods, requirements.subtract(append(burstableToEvict, guarateedToEvict...)...))
if err != nil {
return nil, err
}
return append(append(bestEffortToEvict, burstableToEvict...), guarateedToEvict...), nil
}kubelet preemtion时候挑选待杀死Pods的逻辑如下：
- 如果该Pod的某个Resource request quantity超过了现在的所有的bestEffortPods, burstablePods, guaranteedPods的该Resource request quantity，则podsToPreempt为nil，意味着无合适Pods以释放。
-
如果释放所有bestEffortPods, burstablePods的资源都不足够，则再挑选guaranteedPods（guarateedToEvict）。挑选的规则是：
- 规则一：越少的Pods被释放越好；
- 规则二：释放的资源越少越好；
- 规则一的优先级比规则二高；
- 如果释放所有bestEffortPods及guarateedToEvict的资源都不足够，则再挑选burstablePods(burstableToEvict)。挑选的规则同上。
- 如果释放所有burstableToEvict及guarateedToEvict的资源都不足够，则再挑选bestEffortPods(bestEffortToEvict)。挑选的规则同上。
也就是说：Pod Resource QoS优先级越低的越先被抢占，同一个QoS Level内挑选Pods按照如下规则：
- 规则一：越少的Pods被释放越好；
- 规则二：释放的资源越少越好；
- 规则一的优先级比规则二高；
Priority Admission Controller对CriticalPod的特殊处理
我们先看看几类特殊的、系统预留的CriticalPod：
- ClusterCriticalPod: PriorityClass Name是
system-cluster-critical的Pod。 - NodeCriticalPod:PriorityClass Name是
system-node-critical的Pod。
如果AdmissionController中启动了Priority Admission Controller，那么在创建Pod时对Priority的检查也存在CriticalPod的特殊处理。
Priority Admission Controller主要作用是根据Pod中指定的PriorityClassName替换成对应的Spec.Pritory数值。
plugin/pkg/admission/priority/admission.go:138
// admitPod makes sure a new pod does not set spec.Priority field. It also makes sure that the PriorityClassName exists if it is provided and resolves the pod priority from the PriorityClassName.
func (p *priorityPlugin) admitPod(a admission.Attributes) error {
operation := a.GetOperation()
pod, ok := a.GetObject().(*api.Pod)
if !ok {
return errors.NewBadRequest("resource was marked with kind Pod but was unable to be converted")
}
// Make sure that the client has not set `priority` at the time of pod creation.
if operation == admission.Create && pod.Spec.Priority != nil {
return admission.NewForbidden(a, fmt.Errorf("the integer value of priority must not be provided in pod spec. Priority admission controller populates the value from the given PriorityClass name"))
}
if utilfeature.DefaultFeatureGate.Enabled(features.PodPriority) {
var priority int32
// TODO: @ravig - This is for backwards compatibility to ensure that critical pods with annotations just work fine.
// Remove when no longer needed.
if len(pod.Spec.PriorityClassName) == 0 &&
utilfeature.DefaultFeatureGate.Enabled(features.ExperimentalCriticalPodAnnotation) &&
kubelettypes.IsCritical(a.GetNamespace(), pod.Annotations) {
pod.Spec.PriorityClassName = scheduling.SystemClusterCritical
}
if len(pod.Spec.PriorityClassName) == 0 {
var err error
priority, err = p.getDefaultPriority()
if err != nil {
return fmt.Errorf("failed to get default priority class: %v", err)
}
} else {
// Try resolving the priority class name.
pc, err := p.lister.Get(pod.Spec.PriorityClassName)
if err != nil {
if errors.IsNotFound(err) {
return admission.NewForbidden(a, fmt.Errorf("no PriorityClass with name %v was found", pod.Spec.PriorityClassName))
}
return fmt.Errorf("failed to get PriorityClass with name %s: %v", pod.Spec.PriorityClassName, err)
}
priority = pc.Value
}
pod.Spec.Priority = &priority
}
return nil
}同时满足以下所有条件时，给Pod的Spec.PriorityClassName赋值为system-cluster-critical,即认为是ClusterCriticalPod。
- 如果Enable了ExperimentalCriticalPodAnnotation和PodPriority Feature Gate；
- 该Pod没有指定PriorityClassName；
- 该Pod属于kube-system namespace；
- 该Pod打了
scheduler.alpha.kubernetes.io/critical-pod=""Annotation；
总结
本文介绍了Kubelet在Predicate Admit准入检查时对CriticalPod的资源抢占的原理，以及Priority Admission Controller对CriticalPod的PriorityClassName特殊处理。下一篇是最后一处关于Kubernetes对CriticalPod进行特殊待遇的地方——DaemonSet Controller。
