k8s由master节点和node节点组成，master节点和node节点中的k8s组件各不一样。
node节点：kubelet、kube-proxy
master节点：kube-controller-manager、cloud-controller-manager、
kube-apiserver、etcd（ceph）、kube-scheduler
各组件的作用：
- kube-controller-manager：负责集群内的Node、Pod、Endpoint、Namespace、ServiceAccount、ResourceQuota的管理，当某个Node意外宕机时，Controller Manager会及时发现并执行自动化修复流程，确保集群始终处于预期的工作状态。
- kube-apiserver：k8s通过api server来提供restful api，包括认证授权、数据校验以及集群状态变更等。提供其他模块之间的数据交互和通信的枢纽（其他模块通过 API Server 查询或修改数据，只有 API Server 才直接操作 etcd），在各节点上可以通过kubectl或者curl命令来与API通信。
- kube-scheduler：用于为运行中的容器转发API请求、用于判断哪个node节点适合运行pod。
- etcd（或ceph）：用于保存集群状态、容器配置和网络配置的存储系统。
- kubelet ：node节点代理工具，用于处理任何更新podsepc（a JSON or YAML file that describes a pod）的API请求(主要来自api-server的)，控制容器运行，管理资源，并起到在本地node上监控容器和资源的作用。kubelet 是基于 PodSpec 来工作的。每个 PodSpec 是一个描述 Pod 的 YAML 或 JSON 对象。
- kube-proxy：创建和管理网络规则，达到在k8s内部网络中暴露容器的作用。
control plane 控制平面：负责确保集群的当前状态与所需状态匹配的the various pods
/etc/kubernetes/manifests/
Fluentd：集群范围的日志功能，Kubernetes does not have cluster-wide logging yet. Instead, another CNCF project is used, called Fluentd. When implemented, it provides a unified logging layer for the cluster, which filters, buffers, and routes messages.
Prometheus：集群范围内的指标监控，Cluster-wide metrics is another area with limited functionality. The metrics-server SIG provides basic node and pod CPU and memory utilization. For more metrics, many use the Prometheus project.
英语原文介绍：
The kube-apiserver is central to the operation of the Kubernetes cluster. All calls, both internal and external traffic, are handled via this agent. All actions are accepted and validated by this agent, and it is the only connection to the etcd database. It validates and configures data for API objects, and services REST operations. As a result, it acts as a master process for the entire cluster, and acts as a frontend of the cluster's shared state.
Starting as an alpha feature in v1.16 is the ability to separate user-initiated traffic from server-initiated traffic. Until these features are developed, most network plugins commingle the traffic, which has performance, capacity, and security ramifications.
The kube-scheduler uses an algorithm to determine which node will host a Pod of containers. The scheduler will try to view available resources (such as volumes) to bind, and then try and retry to deploy the Pod based on availability and success. There are several ways you can affect the algorithm, or a custom scheduler could be used instead. You can also bind a Pod to a particular node, though the Pod may remain in a pending state due to other settings. One of the first settings referenced is if the Pod can be deployed within the current quota restrictions. If so, then the taints and tolerations, and labels of the Pods are used along with those of the nodes to determine the proper placement.
The details of the scheduler can be found on GitHub.
The state of the cluster, networking, and other persistent information is kept in an etcd database, or, more accurately, a b+tree key-value store. Rather than finding and changing an entry, values are always appended to the end. Previous copies of the data are then marked for future removal by a compaction process. It works with curl and other HTTP libraries, and provides reliable watch queries.
Simultaneous requests to update a value all travel via the kube-apiserver, which then passes along the request to etcd in a series. The first request would update the database. The second request would no longer have the same version number, in which case the kube-apiserver would reply with an error 409 to the requester. There is no logic past that response on the server side, meaning the client needs to expect this and act upon the denial to update.
There is a master database along with possible followers. They communicate with each other on an ongoing basis to determine which will be master, and determine another in the event of failure. While very fast and potentially durable, there have been some hiccups with new tools, such as kubeadm, and features like whole cluster upgrades.
While most Kubernetes objects are designed to be decoupled, a transient microservice which can be terminated without much concern etcd is the exception. As it is, the persistent state of the entire cluster must be protected and secured. Before upgrades or maintenance, you should plan on backing up etcd. The etcdctl command allows for snapshot save and snapshot restore.
The** kube-controller-manager** is a core control loop daemon which interacts with the kube-apiserver to determine the state of the cluster. If the state does not match, the manager will contact the necessary controller to match the desired state. There are several controllers in use, such as endpoints, namespace, and replication. The full list has expanded as Kubernetes has matured.
Remaining in beta in v1.19, the cloud-controller-manager (ccm) interacts with agents outside of the cloud. It handles tasks once handled by kube-controller-manager. This allows faster changes without altering the core Kubernetes control process. Each kubelet must use the --cloud-provider-external settings passed to the binary. You can also develop your own ccm, which can be deployed as a daemonset as an in-tree deployment or as a free-standing out-of-tree installation. The cloud-controller-manager is an optional agent which takes a few steps to enable. You can learn more about the cloud-controller-manager online.
This graphic shows a pod with a primary container, App, with an optional sidecar Logger. Also seen is the pause container, which is used by the cluster to reserve the IP address in the namespace prior to starting the other pods. This container is not seen from within Kubernetes, but can be seen using docker and crictl.
This graphic also shows a ClusterIP which is used to connect inside the cluster, not the IP of the cluster. As the graphic shows, this can be used to connect to a NodePort for outside the cluster, an IngressController or proxy, or another ”backend” pod or pods.