K8S集群基礎架構的有效管理實踐

1、K8S集群基礎架構的有效管理實踐How we Manage our Widely Varied Kubernetes Infrastructures in AlibabaAgendaBackgroundAlibaba Kubernetes ArchitectureInfrastructure ManagementCI/CD PipelinesQuick DemoBackgroundWho are we?Scale of Alibaba Kubernetes Clusters (handreds of internal clusters, 5k-10k nodes each)Variety of

2、 Cluster Infrastructures (200+ addons)Significance of keeping the stability in large-scale clusters.Tenant ClusterMeta ClusterArchitecture of Alibaba Kubernetes InfrastructureKubeletPouch-ContainerPodPodPodCNIalinetultron-pluginAlibaba ECSKubeletcontainerdPodPodPodCNIterwaycsi-pluginData PlaneBare M

3、entalKubeletcontainerdkatakatakataCNIalinetCSIultron-pluginMulti-tenantNetwork ControllerStorage ControllerKruiseDefender OperatorKubeNode Operatorkube-apiserverkube-controller-managerControl Planekube-scheduleretcdalphaCustomized SchedulerAlert OpeatorMonitoring OpeatorMetrics OperatorAdd-onsRepair

4、 OperatorCustomized OperatorInfrastructure Management - MasterapiVersion: /v1alpha1 kind: Clustermetadata:labels:cluster.id: c3f1b726caecf4d0ca076f73ee781e312 name: kubernetes-clusternamespace: c3f1b726caecf4d0ca076f73ee781e312 spec:kubernetes:kcm:commit: 0bfce06name: kubernetes.kdm.kcm replicas: 3v

5、ersion: v1.16.3-alibaba.2kore:name: kubernetes.kdm.korepanelreplicas: 3version: v1.16.3-alibaba.2 rols:name: kubernetes.kdm.roles version: v1.16.3-alibaba.2scheduler:commit: 0bfce06name: kubernetes.kdm.schedulerreplicas: 3version: v1.16.3-alibaba.2Kubernetes VersionOpsCICluster APIKube-ApiserverKube

6、-Controller-ManagerKube-SchedulerCluster Spec1. push k8s version2. update cluster specSimplified logic of managing master versionswatch3. Upgrade master versionUse Cluster API manage master versionOperator manager infrastructureInfrastructure Management - AddonSimplified logic of managing addon vers

7、ionsOpsCIOperator-Manager1. push operator version2. Call Operator manager to trick canary grayOld PodCanary PodNew Pod4. upgrade versionfirst create a canary pod andthen update operator rules and watching the canary podstatuscall UpdateOperatorRule to empty the rules and delete the canary podupgrade

8、 to new version3. operate canary podInfrastructure Management - DataplaneSimplified logic of managing data plane versionsOpsCI1. push rpm version2. create machine component setMachine OperatorMachineComponentSetwatchkube-node-agentRPMcall kubenode agent to upgade rpm versionupgrade rpmuse partition

9、to controller the batch of grayKubeNode: upgrade a dataplane component“Philosophy”Components varied from different clustersHow to manage componentsAlways provide the stable component versionHow to make stable releasesContinuous and non-disruptive cluster deliveryHow to build safe delivery pipelinesC

10、omponent ManagementImage-OrientedOnly patch container imageSimple but not fit to all casesYAML-OrientedHelm templateSeparate image and meta- configDesign for CIHelm + Version ControlComponent ManagementapiVersion: apps.kruise.io/v1alpha1kind: DaemonSetmetadata:name: asi-proxy-ds-1namespace: kube-sys

11、tem Spec:template:spec:containers:- image: .image.nginx.repository:.image.nginx.tagresource: toYaml .resource | indent 8 tolerations: toYaml .tolerations | indent 8 .nginx:repository: nginxtag: latestresource: requests: cpu: 1 memory: 2Gi limit:cpu: 2memory: 4Gitolerations:- operator: ExistsYAMLMeta

12、-Config: Varies from cluster to clusterInfrastructure Components = YAML = Template + Image + Meta-ConfigImage: expected to be the sameimage:Template: constants that never changesComponent ManagementDo things like that kubectl apply doesCompare with current spec/cluster specPATCH diff to apiserver7HP

13、SODWH,PDJH0HWDDB SpecTarget SpecCluster SpecResource DiResource DiNew Cluster Specpodpodpodpodpod% 5HFRUGHG,PDJH 9HUVLRQ1HZ0HWD 9HUVLRQreplicas: 3cpu: 1 mem: 2Gi spec:replicas: 3 resource:request: UHSOLFD Cluster SpecFilter out danger fieldthree-way diTrigger operators reconcileVersion Release & Tes

14、tingBranch updateRun e2e testsRelease and deliveryControlplaneaddonaddonaddonaddonDeploy to e2e ClusterDev branchNew features & fixesRelease v1.0.0ClusterClusterClusterVersion Release & TestingKubernetes Conformance e2ee2e-cluster-1apiserver kcm scheduler cni-serviceextension-webhookextension-contro

15、ller Pouche2e-cluster-2apiserver kcm scheduler kube-proxy coredns containerde2e-cluster-3apiserver kcm scheduler terwaycloud-controller-manager containerd:KLWHER 7HVWLQJOperatorsguest-cluster-1Operatorsguest-cluster-2Operatorsguest-cluster-3Canary Test Sets 1%ODFNER 7HVWLQJCanary Test Sets 2Canary T

16、est Sets 3e2e testing is not enoughCanary tests runs continuouslyCreate/delete pod/sts/deployUpgrade sts/deployScale up/down sts/deployCreate JobCreate CustomResouceIntra-cluster upgradeRolling updates for Kubernetes WorkloadsDeployment (Kruise)StatefulSet (Kruise)DaemonSet (Kruise)Dataplane compone

17、nts (KubeNode)Rollout PolicyPause/ResumeMax unavailableDeploymentStatefulSetDaemonSetDataplane ComponentsRollout PolicyRollingUpdate Canary DeployRollingUpdate Canary DeployRollingUpdateRollingUpdatePause/Resum eYesYesYesYesMax unavailableNot yetYesYesYesPartitionNoNoYesYes/openkruise/kruiseRollout

18、for operatorsEnhance the ability of Operator (StatefulSet / Deployment)Implement operator as the way kubebuilder doesSidecar container which contains clientset, informer and pluginsServing operator with gRPC requests/openkruise/kruiseRollout for operatorsCanary deploy for OperatorsFlow control on a

19、monilithic managerFlow slice controlled by rule (Custom Resource)Rolling update/openkruise/kruiseRollout for DaemonSetOriginal DaemonSetLack of the ability of rolling updatealways updates all pods once image changesOnDelete ?Replicas:5Updated Replicas: 0Replicas:5Updated Replicas: 5Rollout for Daemo

20、nSetKruise: Enhance the ability of DaemonSetsPartition: the number of pods remained to be old versionMaxUnavailable: the maximum number of pods can be unavailable during rolling updateReplicas:5Updated Replicas: 0Partiton:5Replicas:5Updated Replicas: 5Partition:4Replicas:5Updated Replicas: 5Partitio

21、n:2/openkruise/kruiseRollout for DataplaneKubelet / Pouch / containerd Similar to Kruise Daemonset on patition controlNodeSet: a group of nodes which has the same characters, minimum rollout unitRolling update in each NodeSetUpgrade NodeSet sequentiallyNS - 1NS - 2NS - 3Cluster - 1NS - 5NS - 6Cluste

22、r - 2Bake timeNS - 4Inter-cluster upgradesInter-cluster rollout pipelinesOrchestrate clusters with scale / importance of upper biz appsBuild a gray release pipelineTekton-liked implementationTesting ClusterCanary ClusterSmallflow ClusterProduction ClusterContains small percent of service invocationI

23、nter-cluster upgradesInter-cluster rollout pipelinesSilent period between each clustersPre-checking and post-checkingtime window checker / rule-based blockerMetric monitoring / health checksCluster - 1Cluster - 2Cluster - 3Cluster - 4Pipelines inter-clusterPut all things together, here comes our pipeline journey ! Source CodeUnit TestBuild Imagee2e Cluster Ca