kubernetes的rolling update机制解析

浏览数：32 / 时间：2015年06月09日

commit: d577db99873cbf04b8e17b78f17ec8f3a27eca30
Date: Fri Apr 10 23:45:36 2015 -0700

0.命令行和依赖的基础知识

Synopsis

Perform a rolling update of the given ReplicationController.

Replaces the specified controller with new controller, updating one pod at a time to use the
new PodTemplate. The new-controller.json must specify the same namespace as the
existing controller and overwrite at least one (common) label in its replicaSelector.


kubectl rolling-update OLD_CONTROLLER_NAME -f NEW_CONTROLLER_SPEC

Examples

// Update pods of frontend-v1 using new controller data in frontend-v2.json.
$ kubectl rolling-update frontend-v1 -f frontend-v2.json

// Update pods of frontend-v1 using JSON data passed into stdin.
$ cat frontend-v2.json | kubectl rolling-update frontend-v1 -f -

ReplicationController，简称rc，是kubernet体系中某一种类型pod的集合，rc有一个关键参数叫做replicas，也是就是pod的数量。

那么rc有什么用呢？这是为了解决在集群上一堆pod中有些如果挂了，那么就在别的宿主机上把容器启动起来，并让业务流量导入到正确启动的pod上。也就是说，rc保证了集群服务的可用性，当你有很多个服务启动在一个集群中，你需要用程序去监控这些服务的运行状况，并动态保证服务可用。

rc和pod的对应关系是怎么样的？rc通过selector来选择一些pod作为他的控制范围。只要pod的标签（label）符合seletor，则属于这个rc，下面是pod和rc的示例。

xx-controller.json

   "spec":{
      "replicas":1,
      "selector":{
         "name":"redis",
         "role":"master"
      },

xx-pod.json

  "labels": {
    "name": "redis"
  },

kubernetes被我们简称为k8s，如果对其中的基础概念有兴趣可以看这篇

1.kubctl入口

/cmd/kubectl/kubctl.go

func main() {
    runtime.GOMAXPROCS(runtime.NumCPU())
    cmd := cmd.NewKubectlCommand(cmdutil.NewFactory(nil), os.Stdin, os.Stdout, os.Stderr)
    if err := cmd.Execute(); err != nil {
        os.Exit(1)
    }
}

2.实际调用

源代码在pkg包内，/pkg/kubectl/cmd/cmd.go，每个子命令都实现统一的接口，rollingupdate这行是：

    cmds.AddCommand(NewCmdRollingUpdate(f, out))

这个函数的实现在：/pkg/kubectl/cmd/rollingupdate.go

func NewCmdRollingUpdate(f *cmdutil.Factory, out io.Writer) *cobra.Command {
    cmd := &cobra.Command{
        Use: "rolling-update OLD_CONTROLLER_NAME -f NEW_CONTROLLER_SPEC",
        // rollingupdate is deprecated.
        Aliases: []string{"rollingupdate"},
        Short:   "Perform a rolling update of the given ReplicationController.",
        Long:    rollingUpdate_long,
        Example: rollingUpdate_example,
        Run: func(cmd *cobra.Command, args []string) {
            err := RunRollingUpdate(f, out, cmd, args)
            cmdutil.CheckErr(err)
        },
    }
}

可以看到实际调用时的执行函数是RunRollingUpdate，算是进入正题了

func RunRollingUpdate(f *cmdutil.Factory, out io.Writer, cmd *cobra.Command, args []string) error {
...
    mapper, typer := f.Object()
    // TODO: use resource.Builder instead
    obj, err := resource.NewBuilder(mapper, typer, f.ClientMapperForCommand()).
        NamespaceParam(cmdNamespace).RequireNamespace().
        FilenameParam(filename).
        Do().
        Object()
    if err != nil {
        return err
    }
    newRc, ok := obj.(*api.ReplicationController)
    if !ok {
        return cmdutil.UsageError(cmd, "%s does not specify a valid ReplicationController", filename)
    }

这是建立一个新的rc的代码，其中resource是kubneter所有资源(pod,service,rc)的基类。可以看到新的rc从json参数文件中获取所有信息，然后转义为ReplicationController这个类。

    if oldName == newName {
        return cmdutil.UsageError(cmd, "%s cannot have the same name as the existing ReplicationController %s",
            filename, oldName)
    }

    var hasLabel bool
    for key, oldValue := range oldRc.Spec.Selector {
        if newValue, ok := newRc.Spec.Selector[key]; ok && newValue != oldValue {
            hasLabel = true
            break
        }
    }
    if !hasLabel {
        return cmdutil.UsageError(cmd, "%s must specify a matching key with non-equal value in Selector for %s",
            filename, oldName)
    }

这里可以看到，对于新的rc和旧的rc，有2项限制，一个是新旧名字需要不同，另一个是rc的selector中需要至少有一项的值不一样。

    updater := kubectl.NewRollingUpdater(newRc.Namespace, client)

    // fetch rc
    oldRc, err := client.ReplicationControllers(newRc.Namespace).Get(oldName)
    if err != nil {
        return err
    }
...
    err = updater.Update(out, oldRc, newRc, period, interval, timeout)
    if err != nil {
        return err
    }

在做rolling update的时候，有两个条件限制，一个是新的rc的名字需要和旧的不一样，第二是至少有个一个标签的值不一样。其中namespace是k8s用来做多租户资源隔离的，可以先忽略不计。

3. 数据结构和实现

这段代码出现了NewRollingUpdater，是在上一层的/pkg/kubectl/rollingupdate.go这个文件中，更加接近主体了

// RollingUpdater provides methods for updating replicated pods in a predictable,
// fault-tolerant way.
type RollingUpdater struct {
    // Client interface for creating and updating controllers
    c client.Interface
    // Namespace for resources
    ns string
}

可以看到这里的RollingUpdater里面是一个k8s的client的结构来向api server发送命令

func (r *RollingUpdater) Update(out io.Writer, oldRc, newRc *api.ReplicationController, updatePeriod, interval, timeout time.Duration) error {
    oldName := oldRc.ObjectMeta.Name
    newName := newRc.ObjectMeta.Name
    retry := &RetryParams{interval, timeout}
    waitForReplicas := &RetryParams{interval, timeout}
    if newRc.Spec.Replicas <= 0 {
        return fmt.Errorf("Invalid controller spec for %s; required: > 0 replicas, actual: %s\n", newName, newRc.Spec)
    }
    desired := newRc.Spec.Replicas
    sourceId := fmt.Sprintf("%s:%s", oldName, oldRc.ObjectMeta.UID)

    // look for existing newRc, incase this update was previously started but interrupted
    rc, existing, err := r.getExistingNewRc(sourceId, newName)
    if existing {
        fmt.Fprintf(out, "Continuing update with existing controller %s.\n", newName)
        if err != nil {
            return err
        }
        replicas := rc.ObjectMeta.Annotations[desiredReplicasAnnotation]
        desired, err = strconv.Atoi(replicas)
        if err != nil {
            return fmt.Errorf("Unable to parse annotation for %s: %s=%s",
                newName, desiredReplicasAnnotation, replicas)
        }
        newRc = rc
    } else {
        fmt.Fprintf(out, "Creating %s\n", newName)
        if newRc.ObjectMeta.Annotations == nil {
            newRc.ObjectMeta.Annotations = map[string]string{}
        }
        newRc.ObjectMeta.Annotations[desiredReplicasAnnotation] = fmt.Sprintf("%d", desired)
        newRc.ObjectMeta.Annotations[sourceIdAnnotation] = sourceId
        newRc.Spec.Replicas = 0
        newRc, err = r.c.ReplicationControllers(r.ns).Create(newRc)
        if err != nil {
            return err
        }
    }

    // +1, -1 on oldRc, newRc until newRc has desired number of replicas or oldRc has 0 replicas
    for newRc.Spec.Replicas < desired && oldRc.Spec.Replicas != 0 {
        newRc.Spec.Replicas += 1
        oldRc.Spec.Replicas -= 1
        fmt.Printf("At beginning of loop: %s replicas: %d, %s replicas: %d\n",
            oldName, oldRc.Spec.Replicas,
            newName, newRc.Spec.Replicas)
        fmt.Fprintf(out, "Updating %s replicas: %d, %s replicas: %d\n",
            oldName, oldRc.Spec.Replicas,
            newName, newRc.Spec.Replicas)

        newRc, err = r.resizeAndWait(newRc, retry, waitForReplicas)
        if err != nil {
            return err
        }
        time.Sleep(updatePeriod)
        oldRc, err = r.resizeAndWait(oldRc, retry, waitForReplicas)
        if err != nil {
            return err
        }
        fmt.Printf("At end of loop: %s replicas: %d, %s replicas: %d\n",
            oldName, oldRc.Spec.Replicas,
            newName, newRc.Spec.Replicas)
    }
    // delete remaining replicas on oldRc
    if oldRc.Spec.Replicas != 0 {
        fmt.Fprintf(out, "Stopping %s replicas: %d -> %d\n",
            oldName, oldRc.Spec.Replicas, 0)
        oldRc.Spec.Replicas = 0
        oldRc, err = r.resizeAndWait(oldRc, retry, waitForReplicas)
        // oldRc, err = r.resizeAndWait(oldRc, interval, timeout)
        if err != nil {
            return err
        }
    }
    // add remaining replicas on newRc
    if newRc.Spec.Replicas != desired {
        fmt.Fprintf(out, "Resizing %s replicas: %d -> %d\n",
            newName, newRc.Spec.Replicas, desired)
        newRc.Spec.Replicas = desired
        newRc, err = r.resizeAndWait(newRc, retry, waitForReplicas)
        if err != nil {
            return err
        }
    }
    // Clean up annotations
    if newRc, err = r.c.ReplicationControllers(r.ns).Get(newName); err != nil {
        return err
    }
    delete(newRc.ObjectMeta.Annotations, sourceIdAnnotation)
    delete(newRc.ObjectMeta.Annotations, desiredReplicasAnnotation)
    newRc, err = r.updateAndWait(newRc, interval, timeout)
    if err != nil {
        return err
    }
    // delete old rc
    fmt.Fprintf(out, "Update succeeded. Deleting %s\n", oldName)
    return r.c.ReplicationControllers(r.ns).Delete(oldName)
}

这段代码很长，但做的事情很简单：

如果新的rc没有被创建，就先创一下，如果已经创建了（在上次的rolling_update中创建了但超时了）
用几个循环，把新的rc的replicas增加上去，旧的rc的replicas降低下来，主要调用的函数是resizeAndWait和updateAndWait

4. 底层调用

接上一节的resizeAndWait，代码在/pkg/kubectl/resize.go，这里的具体代码就不贴了
其余的所有调用都发生/pkg/client这个目录下，这是一个http/json的client，主要功能就是向api-server发送请求
整体来说，上面的wait的实现都是比较土的，就是发一个update请求过去，后面轮询的调用get来检测状态是否符合最终需要的状态。