摘要： Knative Serving以Kubernetes和Istio为基础，支持无服务器应用程序和函数的部署并提供服务。我们从部署一个HelloWorld示例入手来分析Knative Serving的代码细节。

概念先知

官方给出的这几个资源的关系图还是比较清晰的：

1.Service: 自动管理工作负载整个生命周期。负责创建route,configuration以及每个service更新的revision。通过Service可以指定路由流量使用最新的revision，还是固定的revision。
2.Route：负责映射网络端点到一个或多个revision。可以通过多种方式管理流量。包括灰度流量和重命名路由。
3.Configuration:负责保持deployment的期望状态，提供了代码和配置之间清晰的分离，并遵循应用开发的12要素。修改一次Configuration产生一个revision。
4.Revision：Revision资源是对工作负载进行的每个修改的代码和配置的时间点快照。Revision是不可变对象，可以长期保留。

看一个简单的示例

我们开始运行官方hello-world示例，看看会发生什么事情：

apiVersion: serving.knative.dev/v1alpha1
kind: Service
metadata:
  name: helloworld-go
  namespace: default
spec:
  runLatest: // RunLatest defines a simple Service. It will automatically configure a route that keeps the latest ready revision from the supplied configuration running.
    configuration:
      revisionTemplate:
        spec:
          container:
            image: registry.cn-shanghai.aliyuncs.com/larus/helloworld-go
            env:
            - name: TARGET
              value: "Go Sample v1"

查看 knative-ingressgateway：

kubectl get svc knative-ingressgateway -n istio-system

查看服务访问：DOMAIN

kubectl get ksvc helloworld-go  --output=custom-columns=NAME:.metadata.name,DOMAIN:.status.domain

这里直接使用cluster ip即可访问

curl -H "Host: helloworld-go.default.example.com" http://10.96.199.35

目前看一下服务是部署ok的。那我们看一下k8s里面创建了哪些资源：

我们可以发现通过Serving,在k8s中创建了2个service和1个deployment：

那么究竟Serving中做了哪些处理，接下来我们分析一下Serving源代码

源代码分析

Main

先看一下各个组件的控制器启动代码，这个比较好找，在/cmd/controller/main.go中。
依次启动configuration、revision、route、labeler、service和clusteringress控制器。

...
controllers := []*controller.Impl{
        configuration.NewController(
            opt,
            configurationInformer,
            revisionInformer,
        ),
        revision.NewController(
            opt,
            revisionInformer,
            kpaInformer,
            imageInformer,
            deploymentInformer,
            coreServiceInformer,
            endpointsInformer,
            configMapInformer,
            buildInformerFactory,
        ),
        route.NewController(
            opt,
            routeInformer,
            configurationInformer,
            revisionInformer,
            coreServiceInformer,
            clusterIngressInformer,
        ),
        labeler.NewRouteToConfigurationController(
            opt,
            routeInformer,
            configurationInformer,
            revisionInformer,
        ),
        service.NewController(
            opt,
            serviceInformer,
            configurationInformer,
            routeInformer,
        ),
        clusteringress.NewController(
            opt,
            clusterIngressInformer,
            virtualServiceInformer,
        ),
    }
...

Service

首先我们要从Service来看，因为我们一开始的输入就是Service资源。在/pkg/reconciler/v1alpha1/service/service.go。
比较简单，就是根据Service创建Configuration和Route资源

func (c *Reconciler) reconcile(ctx context.Context, service *v1alpha1.Service) error {
    ...
    configName := resourcenames.Configuration(service)
    config, err := c.configurationLister.Configurations(service.Namespace).Get(configName)
    if errors.IsNotFound(err) {
        config, err = c.createConfiguration(service)
    ...
    routeName := resourcenames.Route(service)
    route, err := c.routeLister.Routes(service.Namespace).Get(routeName)
    if errors.IsNotFound(err) {
        route, err = c.createRoute(service)
    ...
}

Route

/pkg/reconciler/v1alpha1/route/route.go
看一下Route中reconcile做了哪些处理：
1.判断是否有Ready的Revision可进行traffic
2.设置目标流量的Revision（runLatest:使用最新的版本;pinned：固定版本，不过已弃用;release:通过允许在两个修订版之间拆分流量，逐步扩大到新修订版，用于替换pinned。manual：手动模式，目前来看并未实现）
3.创建ClusterIngress:Route不直接依赖于VirtualService[https://istio.io/docs/reference/config/istio.networking.v1alpha3/#VirtualService] ,而是依赖一个中间资源ClusterIngress，它可以针对不同的网络平台进行不同的协调。目前实现是基于istio网络平台。
4.创建k8s service:这个Service主要为Istio路由提供域名访问。

func (c *Reconciler) reconcile(ctx context.Context, r *v1alpha1.Route) error {
    ....
    // 基于是否有Ready的Revision
    traffic, err := c.configureTraffic(ctx, r)
    if traffic == nil || err != nil {
        // Traffic targets aren't ready, no need to configure child resources.
        return err
    }

    logger.Info("Updating targeted revisions.")
    // In all cases we will add annotations to the referred targets.  This is so that when they become
    // routable we can know (through a listener) and attempt traffic configuration again.
    if err := c.reconcileTargetRevisions(ctx, traffic, r); err != nil {
        return err
    }

    // Update the information that makes us Addressable.
    r.Status.Domain = routeDomain(ctx, r)
    r.Status.DeprecatedDomainInternal = resourcenames.K8sServiceFullname(r)
    r.Status.Address = &duckv1alpha1.Addressable{
        Hostname: resourcenames.K8sServiceFullname(r),
    }

    // Add the finalizer before creating the ClusterIngress so that we can be sure it gets cleaned up.
    if err := c.ensureFinalizer(r); err != nil {
        return err
    }

    logger.Info("Creating ClusterIngress.")
    desired := resources.MakeClusterIngress(r, traffic, ingressClassForRoute(ctx, r))
    clusterIngress, err := c.reconcileClusterIngress(ctx, r, desired)
    if err != nil {
        return err
    }
    r.Status.PropagateClusterIngressStatus(clusterIngress.Status)

    logger.Info("Creating/Updating placeholder k8s services")
    if err := c.reconcilePlaceholderService(ctx, r, clusterIngress); err != nil {
        return err
    }

    r.Status.ObservedGeneration = r.Generation
    logger.Info("Route successfully synced")
    return nil
}

看一下helloworld-go生成的Route资源文件：

apiVersion: serving.knative.dev/v1alpha1
kind: Route
metadata:
  name: helloworld-go
  namespace: default
...
spec:
  generation: 1
  traffic:
  - configurationName: helloworld-go 
    percent: 100
status:
...
  domain: helloworld-go.default.example.com
  domainInternal: helloworld-go.default.svc.cluster.local
  traffic:
  - percent: 100 # 所有的流量通过这个revision
    revisionName: helloworld-go-00001 # 使用helloworld-go-00001 revision

这里可以看到通过helloworld-go配置, 找到了已经ready的helloworld-go-00001(Revision)。

Configuration

/pkg/reconciler/v1alpha1/configuration/configuration.go
1.获取当前Configuration对应的Revision, 若不存在则创建。
2.为Configuration设置最新的Revision
3.根据Revision是否readiness，设置Configuration的状态LatestReadyRevisionName

func (c *Reconciler) reconcile(ctx context.Context, config *v1alpha1.Configuration) error {
    ...
    // First, fetch the revision that should exist for the current generation.
    lcr, err := c.latestCreatedRevision(config)
    if errors.IsNotFound(err) {
        lcr, err = c.createRevision(ctx, config)
    ...    
    revName := lcr.Name
    // Second, set this to be the latest revision that we have created.
    config.Status.SetLatestCreatedRevisionName(revName)
    config.Status.ObservedGeneration = config.Generation

    // Last, determine whether we should set LatestReadyRevisionName to our
    // LatestCreatedRevision based on its readiness.
    rc := lcr.Status.GetCondition(v1alpha1.RevisionConditionReady)
    switch {
    case rc == nil || rc.Status == corev1.ConditionUnknown:
        logger.Infof("Revision %q of configuration %q is not ready", revName, config.Name)

    case rc.Status == corev1.ConditionTrue:
        logger.Infof("Revision %q of configuration %q is ready", revName, config.Name)

        created, ready := config.Status.LatestCreatedRevisionName, config.Status.LatestReadyRevisionName
        if ready == "" {
            // Surface an event for the first revision becoming ready.
            c.Recorder.Event(config, corev1.EventTypeNormal, "ConfigurationReady",
                "Configuration becomes ready")
        }
        // Update the LatestReadyRevisionName and surface an event for the transition.
        config.Status.SetLatestReadyRevisionName(lcr.Name)
        if created != ready {
            c.Recorder.Eventf(config, corev1.EventTypeNormal, "LatestReadyUpdate",
                "LatestReadyRevisionName updated to %q", lcr.Name)
        }
...
}

看一下helloworld-go生成的Configuration资源文件：

apiVersion: serving.knative.dev/v1alpha1
kind: Configuration
metadata:
  name: helloworld-go
  namespace: default
  ...
spec:
  generation: 1
  revisionTemplate:
    metadata:
      creationTimestamp: null
    spec:
      container:
        env:
        - name: TARGET
          value: Go Sample v1
        image: registry.cn-shanghai.aliyuncs.com/larus/helloworld-go
        name: ""
        resources: {}
      timeoutSeconds: 300
status:
  ...
  latestCreatedRevisionName: helloworld-go-00001
  latestReadyRevisionName: helloworld-go-00001
  observedGeneration: 1

我们可以发现LatestReadyRevisionName设置了helloworld-go-00001(Revision)。

Revision

/pkg/reconciler/v1alpha1/revision/revision.go
1.获取build进度
2.设置镜像摘要
3.创建deployment
4.创建k8s service：根据Revision构建服务访问Service
5.创建fluentd configmap
6.创建KPA
感觉这段代码写的很优雅，函数执行过程写的很清晰，值得借鉴。另外我们也可以发现，目前knative只支持deployment的工作负载

func (c *Reconciler) reconcile(ctx context.Context, rev *v1alpha1.Revision) error {
    ...
    if err := c.reconcileBuild(ctx, rev); err != nil {
        return err
    }

    bc := rev.Status.GetCondition(v1alpha1.RevisionConditionBuildSucceeded)
    if bc == nil || bc.Status == corev1.ConditionTrue {
        // There is no build, or the build completed successfully.

        phases := []struct {
            name string
            f    func(context.Context, *v1alpha1.Revision) error
        }{{
            name: "image digest",
            f:    c.reconcileDigest,
        }, {
            name: "user deployment",
            f:    c.reconcileDeployment,
        }, {
            name: "user k8s service",
            f:    c.reconcileService,
        }, {
            // Ensures our namespace has the configuration for the fluentd sidecar.
            name: "fluentd configmap",
            f:    c.reconcileFluentdConfigMap,
        }, {
            name: "KPA",
            f:    c.reconcileKPA,
        }}

        for _, phase := range phases {
            if err := phase.f(ctx, rev); err != nil {
                logger.Errorf("Failed to reconcile %s: %v", phase.name, zap.Error(err))
                return err
            }
        }
    }
...
}

最后我们看一下生成的Revision资源：

apiVersion: serving.knative.dev/v1alpha1
kind: Service
metadata:
  name: helloworld-go
  namespace: default
  ...
spec:
  generation: 1
  runLatest:
    configuration:
      revisionTemplate:
        spec:
          container:
            env:
            - name: TARGET
              value: Go Sample v1
            image: registry.cn-shanghai.aliyuncs.com/larus/helloworld-go
          timeoutSeconds: 300
status:
  address:
    hostname: helloworld-go.default.svc.cluster.local
 ...
  domain: helloworld-go.default.example.com
  domainInternal: helloworld-go.default.svc.cluster.local
  latestCreatedRevisionName: helloworld-go-00001
  latestReadyRevisionName: helloworld-go-00001
  observedGeneration: 1
  traffic:
  - percent: 100
    revisionName: helloworld-go-00001

这里我们可以看到访问域名helloworld-go.default.svc.cluster.local，以及当前revision的流量配比(100%)
这样我们分析完之后，现在打开Serving这个黑盒：

最后

这里只是基于简单的例子，分析了主要的业务流程处理代码。对于activator(如何唤醒业务容器)，autoscaler（Pod如何自动缩为0）等代码实现有兴趣的同学可以一起交流。

参考

https://github.com/knative/docs/tree/master/docs/serving

本文作者：元毅

原文链接

本文为云栖社区原创内容，未经允许不得转载。