示例#1
0
func newLifecycleTest(t *testing.T) lifecycleTest {
	assert := &EventAssertions{*assert.New(t)}

	// create a fake pod watch. We use that below to submit new pods to the scheduler
	podsListWatch := NewMockPodsListWatch(api.PodList{})

	// create fake apiserver
	apiServer := NewTestServer(t, api.NamespaceDefault, podsListWatch)

	// create ExecutorInfo with some data for static pods if set
	ei := mesosutil.NewExecutorInfo(
		mesosutil.NewExecutorID("executor-id"),
		mesosutil.NewCommandInfo("executor-cmd"),
	)
	ei.Data = []byte{0, 1, 2}

	// create framework
	client := client.NewOrDie(&client.Config{
		Host:         apiServer.server.URL,
		GroupVersion: testapi.Default.GroupVersion(),
	})
	c := *schedcfg.CreateDefaultConfig()
	fw := framework.New(framework.Config{
		Executor:        ei,
		Client:          client,
		SchedulerConfig: c,
		LookupNode:      apiServer.LookupNode,
	})

	// TODO(sttts): re-enable the following tests
	// assert.NotNil(framework.client, "client is nil")
	// assert.NotNil(framework.executor, "executor is nil")
	// assert.NotNil(framework.offers, "offer registry is nil")

	// create pod scheduler
	strategy := podschedulers.NewAllocationStrategy(
		podtask.NewDefaultPredicate(
			mresource.DefaultDefaultContainerCPULimit,
			mresource.DefaultDefaultContainerMemLimit,
		),
		podtask.NewDefaultProcurement(
			mresource.DefaultDefaultContainerCPULimit,
			mresource.DefaultDefaultContainerMemLimit,
		),
	)
	fcfs := podschedulers.NewFCFSPodScheduler(strategy, apiServer.LookupNode)

	// create scheduler process
	schedulerProc := ha.New(fw)

	// create scheduler
	eventObs := NewEventObserver()
	scheduler := components.New(&c, fw, fcfs, client, eventObs, schedulerProc.Terminal(), http.DefaultServeMux, &podsListWatch.ListWatch)
	assert.NotNil(scheduler)

	// create mock mesos scheduler driver
	driver := &framework.JoinableDriver{}

	return lifecycleTest{
		apiServer:     apiServer,
		driver:        driver,
		eventObs:      eventObs,
		podsListWatch: podsListWatch,
		framework:     fw,
		schedulerProc: schedulerProc,
		sched:         scheduler,
		t:             t,
	}
}
示例#2
0
// Test to create the scheduler plugin with the config returned by the scheduler,
// and play through the whole life cycle of the plugin while creating pods, deleting
// and failing them.
func TestPlugin_LifeCycle(t *testing.T) {
	t.Skip("This test is flaky, see #11901")
	assert := &EventAssertions{*assert.New(t)}

	// create a fake pod watch. We use that below to submit new pods to the scheduler
	podListWatch := NewMockPodsListWatch(api.PodList{})

	// create fake apiserver
	testApiServer := NewTestServer(t, api.NamespaceDefault, podListWatch)
	defer testApiServer.server.Close()

	// create executor with some data for static pods if set
	executor := util.NewExecutorInfo(
		util.NewExecutorID("executor-id"),
		util.NewCommandInfo("executor-cmd"),
	)
	executor.Data = []byte{0, 1, 2}

	// create scheduler
	nodeStore := cache.NewStore(cache.MetaNamespaceKeyFunc)
	as := NewAllocationStrategy(
		podtask.DefaultPredicate,
		podtask.NewDefaultProcurement(mresource.DefaultDefaultContainerCPULimit, mresource.DefaultDefaultContainerMemLimit))
	testScheduler := New(Config{
		Executor: executor,
		Client:   client.NewOrDie(&client.Config{Host: testApiServer.server.URL, Version: testapi.Default.Version()}),
		Scheduler: NewFCFSPodScheduler(as, func(node string) *api.Node {
			obj, _, _ := nodeStore.GetByKey(node)
			if obj == nil {
				return nil
			}
			return obj.(*api.Node)
		}),
		Schedcfg: *schedcfg.CreateDefaultConfig(),
	})

	assert.NotNil(testScheduler.client, "client is nil")
	assert.NotNil(testScheduler.executor, "executor is nil")
	assert.NotNil(testScheduler.offers, "offer registry is nil")

	// create scheduler process
	schedulerProcess := ha.New(testScheduler)

	// get plugin config from it
	c := testScheduler.NewPluginConfig(schedulerProcess.Terminal(), http.DefaultServeMux, &podListWatch.ListWatch)
	assert.NotNil(c)

	// make events observable
	eventObserver := NewEventObserver()
	c.Recorder = eventObserver

	// create plugin
	p := NewPlugin(c).(*schedulingPlugin)
	assert.NotNil(p)

	// run plugin
	p.Run(schedulerProcess.Terminal())
	defer schedulerProcess.End()

	// init scheduler
	err := testScheduler.Init(schedulerProcess.Master(), p, http.DefaultServeMux)
	assert.NoError(err)

	// create mock mesos scheduler driver
	mockDriver := &joinableDriver{}
	mockDriver.On("Start").Return(mesos.Status_DRIVER_RUNNING, nil).Once()
	started := mockDriver.Upon()

	mAny := mock.AnythingOfType
	mockDriver.On("ReconcileTasks", mAny("[]*mesosproto.TaskStatus")).Return(mesos.Status_DRIVER_RUNNING, nil)
	mockDriver.On("SendFrameworkMessage", mAny("*mesosproto.ExecutorID"), mAny("*mesosproto.SlaveID"), mAny("string")).
		Return(mesos.Status_DRIVER_RUNNING, nil)

	type LaunchedTask struct {
		offerId  mesos.OfferID
		taskInfo *mesos.TaskInfo
	}
	launchedTasks := make(chan LaunchedTask, 1)
	launchTasksCalledFunc := func(args mock.Arguments) {
		offerIDs := args.Get(0).([]*mesos.OfferID)
		taskInfos := args.Get(1).([]*mesos.TaskInfo)
		assert.Equal(1, len(offerIDs))
		assert.Equal(1, len(taskInfos))
		launchedTasks <- LaunchedTask{
			offerId:  *offerIDs[0],
			taskInfo: taskInfos[0],
		}
	}
	mockDriver.On("LaunchTasks", mAny("[]*mesosproto.OfferID"), mAny("[]*mesosproto.TaskInfo"), mAny("*mesosproto.Filters")).
		Return(mesos.Status_DRIVER_RUNNING, nil).Run(launchTasksCalledFunc)
	mockDriver.On("DeclineOffer", mAny("*mesosproto.OfferID"), mAny("*mesosproto.Filters")).
		Return(mesos.Status_DRIVER_RUNNING, nil)

	// elect master with mock driver
	driverFactory := ha.DriverFactory(func() (bindings.SchedulerDriver, error) {
		return mockDriver, nil
	})
	schedulerProcess.Elect(driverFactory)
	elected := schedulerProcess.Elected()

	// driver will be started
	<-started

	// tell scheduler to be registered
	testScheduler.Registered(
		mockDriver,
		util.NewFrameworkID("kubernetes-id"),
		util.NewMasterInfo("master-id", (192<<24)+(168<<16)+(0<<8)+1, 5050),
	)

	// wait for being elected
	<-elected

	//TODO(jdef) refactor things above here into a test suite setup of some sort

	// fake new, unscheduled pod
	pod, i := NewTestPod()
	podListWatch.Add(pod, true) // notify watchers

	// wait for failedScheduling event because there is no offer
	assert.EventWithReason(eventObserver, "failedScheduling", "failedScheduling event not received")

	// add some matching offer
	offers := []*mesos.Offer{NewTestOffer(fmt.Sprintf("offer%d", i))}
	testScheduler.ResourceOffers(nil, offers)

	// and wait for scheduled pod
	assert.EventWithReason(eventObserver, "scheduled")
	select {
	case launchedTask := <-launchedTasks:
		// report back that the task has been staged, and then started by mesos
		testScheduler.StatusUpdate(mockDriver, newTaskStatusForTask(launchedTask.taskInfo, mesos.TaskState_TASK_STAGING))
		testScheduler.StatusUpdate(mockDriver, newTaskStatusForTask(launchedTask.taskInfo, mesos.TaskState_TASK_RUNNING))

		// check that ExecutorInfo.data has the static pod data
		assert.Len(launchedTask.taskInfo.Executor.Data, 3)

		// report back that the task has been lost
		mockDriver.AssertNumberOfCalls(t, "SendFrameworkMessage", 0)
		testScheduler.StatusUpdate(mockDriver, newTaskStatusForTask(launchedTask.taskInfo, mesos.TaskState_TASK_LOST))

		// and wait that framework message is sent to executor
		mockDriver.AssertNumberOfCalls(t, "SendFrameworkMessage", 1)

	case <-time.After(5 * time.Second):
		t.Fatalf("timed out waiting for launchTasks call")
	}

	// Launch a pod and wait until the scheduler driver is called
	schedulePodWithOffers := func(pod *api.Pod, offers []*mesos.Offer) (*api.Pod, *LaunchedTask, *mesos.Offer) {
		// wait for failedScheduling event because there is no offer
		assert.EventWithReason(eventObserver, "failedScheduling", "failedScheduling event not received")

		// supply a matching offer
		testScheduler.ResourceOffers(mockDriver, offers)

		// and wait to get scheduled
		assert.EventWithReason(eventObserver, "scheduled")

		// wait for driver.launchTasks call
		select {
		case launchedTask := <-launchedTasks:
			for _, offer := range offers {
				if offer.Id.GetValue() == launchedTask.offerId.GetValue() {
					return pod, &launchedTask, offer
				}
			}
			t.Fatalf("unknown offer used to start a pod")
			return nil, nil, nil
		case <-time.After(5 * time.Second):
			t.Fatal("timed out waiting for launchTasks")
			return nil, nil, nil
		}
	}
	// Launch a pod and wait until the scheduler driver is called
	launchPodWithOffers := func(pod *api.Pod, offers []*mesos.Offer) (*api.Pod, *LaunchedTask, *mesos.Offer) {
		podListWatch.Add(pod, true)
		return schedulePodWithOffers(pod, offers)
	}

	// Launch a pod, wait until the scheduler driver is called and report back that it is running
	startPodWithOffers := func(pod *api.Pod, offers []*mesos.Offer) (*api.Pod, *LaunchedTask, *mesos.Offer) {
		// notify about pod, offer resources and wait for scheduling
		pod, launchedTask, offer := launchPodWithOffers(pod, offers)
		if pod != nil {
			// report back status
			testScheduler.StatusUpdate(mockDriver, newTaskStatusForTask(launchedTask.taskInfo, mesos.TaskState_TASK_STAGING))
			testScheduler.StatusUpdate(mockDriver, newTaskStatusForTask(launchedTask.taskInfo, mesos.TaskState_TASK_RUNNING))
			return pod, launchedTask, offer
		}

		return nil, nil, nil
	}

	startTestPod := func() (*api.Pod, *LaunchedTask, *mesos.Offer) {
		pod, i := NewTestPod()
		offers := []*mesos.Offer{NewTestOffer(fmt.Sprintf("offer%d", i))}
		return startPodWithOffers(pod, offers)
	}

	// start another pod
	pod, launchedTask, _ := startTestPod()

	// mock drvier.KillTask, should be invoked when a pod is deleted
	mockDriver.On("KillTask", mAny("*mesosproto.TaskID")).Return(mesos.Status_DRIVER_RUNNING, nil).Run(func(args mock.Arguments) {
		killedTaskId := *(args.Get(0).(*mesos.TaskID))
		assert.Equal(*launchedTask.taskInfo.TaskId, killedTaskId, "expected same TaskID as during launch")
	})
	killTaskCalled := mockDriver.Upon()

	// stop it again via the apiserver mock
	podListWatch.Delete(pod, true) // notify watchers

	// and wait for the driver killTask call with the correct TaskId
	select {
	case <-killTaskCalled:
		// report back that the task is finished
		testScheduler.StatusUpdate(mockDriver, newTaskStatusForTask(launchedTask.taskInfo, mesos.TaskState_TASK_FINISHED))

	case <-time.After(5 * time.Second):
		t.Fatal("timed out waiting for KillTask")
	}

	// start a pod with on a given NodeName and check that it is scheduled to the right host
	pod, i = NewTestPod()
	pod.Spec.NodeName = "hostname1"
	offers = []*mesos.Offer{}
	for j := 0; j < 3; j++ {
		offer := NewTestOffer(fmt.Sprintf("offer%d_%d", i, j))
		hostname := fmt.Sprintf("hostname%d", j)
		offer.Hostname = &hostname
		offers = append(offers, offer)
	}

	_, _, usedOffer := startPodWithOffers(pod, offers)

	assert.Equal(offers[1].Id.GetValue(), usedOffer.Id.GetValue())
	assert.Equal(pod.Spec.NodeName, *usedOffer.Hostname)

	testScheduler.OfferRescinded(mockDriver, offers[0].Id)
	testScheduler.OfferRescinded(mockDriver, offers[2].Id)

	// start pods:
	// - which are failing while binding,
	// - leading to reconciliation
	// - with different states on the apiserver

	failPodFromExecutor := func(task *mesos.TaskInfo) {
		beforePodLookups := testApiServer.Stats(pod.Name)
		status := newTaskStatusForTask(task, mesos.TaskState_TASK_FAILED)
		message := messages.CreateBindingFailure
		status.Message = &message
		testScheduler.StatusUpdate(mockDriver, status)

		// wait until pod is looked up at the apiserver
		assertext.EventuallyTrue(t, time.Second, func() bool {
			return testApiServer.Stats(pod.Name) == beforePodLookups+1
		}, "expect that reconcileTask will access apiserver for pod %v", pod.Name)
	}

	launchTestPod := func() (*api.Pod, *LaunchedTask, *mesos.Offer) {
		pod, i := NewTestPod()
		offers := []*mesos.Offer{NewTestOffer(fmt.Sprintf("offer%d", i))}
		return launchPodWithOffers(pod, offers)
	}

	// 1. with pod deleted from the apiserver
	//    expected: pod is removed from internal task registry
	pod, launchedTask, _ = launchTestPod()
	podListWatch.Delete(pod, false) // not notifying the watchers
	failPodFromExecutor(launchedTask.taskInfo)

	podKey, _ := podtask.MakePodKey(api.NewDefaultContext(), pod.Name)
	assertext.EventuallyTrue(t, time.Second, func() bool {
		t, _ := p.api.tasks().ForPod(podKey)
		return t == nil
	})

	// 2. with pod still on the apiserver, not bound
	//    expected: pod is rescheduled
	pod, launchedTask, _ = launchTestPod()
	failPodFromExecutor(launchedTask.taskInfo)

	retryOffers := []*mesos.Offer{NewTestOffer("retry-offer")}
	schedulePodWithOffers(pod, retryOffers)

	// 3. with pod still on the apiserver, bound, notified via ListWatch
	// expected: nothing, pod updates not supported, compare ReconcileTask function
	pod, launchedTask, usedOffer = startTestPod()
	pod.Annotations = map[string]string{
		meta.BindingHostKey: *usedOffer.Hostname,
	}
	pod.Spec.NodeName = *usedOffer.Hostname
	podListWatch.Modify(pod, true) // notifying the watchers
	time.Sleep(time.Second / 2)
	failPodFromExecutor(launchedTask.taskInfo)
}
示例#3
0
func (s *SchedulerServer) bootstrap(hks hyperkube.Interface, sc *schedcfg.Config) (*ha.SchedulerProcess, ha.DriverFactory, etcd.Client, *mesos.ExecutorID) {
	s.frameworkName = strings.TrimSpace(s.frameworkName)
	if s.frameworkName == "" {
		log.Fatalf("framework-name must be a non-empty string")
	}
	s.frameworkWebURI = strings.TrimSpace(s.frameworkWebURI)

	metrics.Register()
	runtime.Register()
	s.mux.Handle("/metrics", prometheus.Handler())
	healthz.InstallHandler(s.mux)

	if len(s.etcdServerList) == 0 {
		log.Fatalf("specify --etcd-servers must be specified")
	}

	if len(s.apiServerList) < 1 {
		log.Fatal("No api servers specified.")
	}

	clientConfig, err := s.createAPIServerClientConfig()
	if err != nil {
		log.Fatalf("Unable to make apiserver client config: %v", err)
	}
	s.client, err = clientset.NewForConfig(clientConfig)
	if err != nil {
		log.Fatalf("Unable to make apiserver clientset: %v", err)
	}

	if s.reconcileCooldown < defaultReconcileCooldown {
		s.reconcileCooldown = defaultReconcileCooldown
		log.Warningf("user-specified reconcile cooldown too small, defaulting to %v", s.reconcileCooldown)
	}

	eiPrototype, err := s.prepareExecutorInfo(hks)
	if err != nil {
		log.Fatalf("misconfigured executor: %v", err)
	}

	// TODO(jdef): remove the dependency on etcd as soon as
	// (1) the generic config store is available for the FrameworkId storage
	// (2) the generic master election is provided by the apiserver
	// Compare docs/proposals/high-availability.md
	etcdClient, err := newEtcd(s.etcdServerList)
	if err != nil {
		log.Fatalf("misconfigured etcd: %v", err)
	}
	keysAPI := etcd.NewKeysAPI(etcdClient)

	// mirror all nodes into the nodeStore
	var eiRegistry executorinfo.Registry
	nodesClientConfig := *clientConfig
	nodesClient, err := clientset.NewForConfig(&nodesClientConfig)
	if err != nil {
		log.Fatalf("Cannot create client to watch nodes: %v", err)
	}
	nodeLW := cache.NewListWatchFromClient(nodesClient.CoreClient, "nodes", api.NamespaceAll, fields.Everything())
	nodeStore, nodeCtl := controllerfw.NewInformer(nodeLW, &api.Node{}, s.nodeRelistPeriod, &controllerfw.ResourceEventHandlerFuncs{
		DeleteFunc: func(obj interface{}) {
			if eiRegistry != nil {
				// TODO(jdef) use controllerfw.DeletionHandlingMetaNamespaceKeyFunc at some point?
				nodeName := ""
				if tombstone, ok := obj.(cache.DeletedFinalStateUnknown); ok {
					nodeName = tombstone.Key
				} else if node, ok := obj.(*api.Node); ok {
					nodeName = node.Name
				}
				if nodeName != "" {
					log.V(2).Infof("deleting node %q from registry", nodeName)
					eiRegistry.Invalidate(nodeName)
				}
			}
		},
	})

	lookupNode := func(hostName string) *api.Node {
		n, _, _ := nodeStore.GetByKey(hostName) // ignore error and return nil then
		if n == nil {
			return nil
		}
		return n.(*api.Node)
	}

	execInfoCache, err := executorinfo.NewCache(defaultExecutorInfoCacheSize)
	if err != nil {
		log.Fatalf("cannot create executorinfo cache: %v", err)
	}

	eiRegistry, err = executorinfo.NewRegistry(lookupNode, eiPrototype, execInfoCache)
	if err != nil {
		log.Fatalf("cannot create executorinfo registry: %v", err)
	}

	pr := podtask.NewDefaultProcurement(eiPrototype, eiRegistry)
	fcfs := podschedulers.NewFCFSPodScheduler(pr, lookupNode)
	frameworkIDStorage, err := s.frameworkIDStorage(keysAPI)
	if err != nil {
		log.Fatalf("cannot init framework ID storage: %v", err)
	}
	framework := framework.New(framework.Config{
		SchedulerConfig:   *sc,
		Client:            s.client,
		FailoverTimeout:   s.failoverTimeout,
		ReconcileInterval: s.reconcileInterval,
		ReconcileCooldown: s.reconcileCooldown,
		LookupNode:        lookupNode,
		StoreFrameworkId:  frameworkIDStorage.Set,
		ExecutorId:        eiPrototype.GetExecutorId(),
	})
	masterUri := s.mesosMaster
	info, cred, err := s.buildFrameworkInfo()
	if err != nil {
		log.Fatalf("Misconfigured mesos framework: %v", err)
	}

	schedulerProcess := ha.New(framework)

	// try publishing on the same IP as the slave
	var publishedAddress net.IP
	if libprocessIP := os.Getenv("LIBPROCESS_IP"); libprocessIP != "" {
		publishedAddress = net.ParseIP(libprocessIP)
	}
	if publishedAddress != nil {
		log.V(1).Infof("driver will publish address %v", publishedAddress)
	}

	dconfig := &bindings.DriverConfig{
		Scheduler:        schedulerProcess,
		Framework:        info,
		Master:           masterUri,
		Credential:       cred,
		BindingAddress:   s.address,
		BindingPort:      uint16(s.driverPort),
		PublishedAddress: publishedAddress,
		HostnameOverride: s.hostnameOverride,
		WithAuthContext: func(ctx context.Context) context.Context {
			ctx = auth.WithLoginProvider(ctx, s.mesosAuthProvider)
			ctx = sasl.WithBindingAddress(ctx, s.address)
			return ctx
		},
	}

	// create event recorder sending events to the "" namespace of the apiserver
	eventsClientConfig := *clientConfig
	eventsClient, err := clientset.NewForConfig(&eventsClientConfig)
	if err != nil {
		log.Fatalf("Invalid API configuration: %v", err)
	}
	broadcaster := record.NewBroadcaster()
	recorder := broadcaster.NewRecorder(api.EventSource{Component: api.DefaultSchedulerName})
	broadcaster.StartLogging(log.Infof)
	broadcaster.StartRecordingToSink(&unversionedcore.EventSinkImpl{Interface: eventsClient.Events("")})

	lw := cache.NewListWatchFromClient(s.client.CoreClient, "pods", api.NamespaceAll, fields.Everything())

	hostPortStrategy := hostport.StrategyFixed
	if s.useHostPortEndpoints {
		hostPortStrategy = hostport.StrategyWildcard
	}

	// create scheduler core with all components arranged around it
	sched := components.New(
		sc,
		framework,
		fcfs,
		s.client,
		recorder,
		schedulerProcess.Terminal(),
		s.mux,
		lw,
		podtask.Config{
			DefaultPodRoles:              s.defaultPodRoles,
			FrameworkRoles:               s.frameworkRoles,
			GenerateTaskDiscoveryEnabled: s.generateTaskDiscovery,
			HostPortStrategy:             hostPortStrategy,
			Prototype:                    eiPrototype,
		},
		s.defaultContainerCPULimit,
		s.defaultContainerMemLimit,
	)

	runtime.On(framework.Registration(), func() { sched.Run(schedulerProcess.Terminal()) })
	runtime.On(framework.Registration(), s.newServiceWriter(publishedAddress, schedulerProcess.Terminal()))
	runtime.On(framework.Registration(), func() { nodeCtl.Run(schedulerProcess.Terminal()) })

	driverFactory := ha.DriverFactory(func() (drv bindings.SchedulerDriver, err error) {
		log.V(1).Infoln("performing deferred initialization")
		if err = framework.Init(sched, schedulerProcess.Master(), s.mux); err != nil {
			return nil, fmt.Errorf("failed to initialize pod scheduler: %v", err)
		}

		log.V(1).Infoln("deferred init complete")
		if s.failoverTimeout > 0 {
			// defer obtaining framework ID to prevent multiple schedulers
			// from overwriting each other's framework IDs
			var frameworkID string
			frameworkID, err = frameworkIDStorage.Get(context.TODO())
			if err != nil {
				return nil, fmt.Errorf("failed to fetch framework ID from storage: %v", err)
			}
			if frameworkID != "" {
				log.Infof("configuring FrameworkInfo with ID found in storage: %q", frameworkID)
				dconfig.Framework.Id = &mesos.FrameworkID{Value: &frameworkID}
			} else {
				log.V(1).Infof("did not find framework ID in storage")
			}
		} else {
			// TODO(jdef) this is a hack, really for development, to simplify clean up of old framework IDs
			frameworkIDStorage.Remove(context.TODO())
		}

		log.V(1).Infoln("constructing mesos scheduler driver")
		drv, err = bindings.NewMesosSchedulerDriver(*dconfig)
		if err != nil {
			return nil, fmt.Errorf("failed to construct scheduler driver: %v", err)
		}

		log.V(1).Infoln("constructed mesos scheduler driver:", drv)
		s.setDriver(drv)
		return drv, nil
	})

	return schedulerProcess, driverFactory, etcdClient, eiPrototype.GetExecutorId()
}
示例#4
0
func (s *SchedulerServer) bootstrap(hks hyperkube.Interface, sc *schedcfg.Config) (*ha.SchedulerProcess, ha.DriverFactory, tools.EtcdClient, *uid.UID) {

	s.FrameworkName = strings.TrimSpace(s.FrameworkName)
	if s.FrameworkName == "" {
		log.Fatalf("framework-name must be a non-empty string")
	}
	s.FrameworkWebURI = strings.TrimSpace(s.FrameworkWebURI)

	metrics.Register()
	runtime.Register()
	s.mux.Handle("/metrics", prometheus.Handler())
	healthz.InstallHandler(s.mux)

	if (s.EtcdConfigFile != "" && len(s.EtcdServerList) != 0) || (s.EtcdConfigFile == "" && len(s.EtcdServerList) == 0) {
		log.Fatalf("specify either --etcd-servers or --etcd-config")
	}

	if len(s.APIServerList) < 1 {
		log.Fatal("No api servers specified.")
	}

	client, err := s.createAPIServerClient()
	if err != nil {
		log.Fatalf("Unable to make apiserver client: %v", err)
	}
	s.client = client

	if s.ReconcileCooldown < defaultReconcileCooldown {
		s.ReconcileCooldown = defaultReconcileCooldown
		log.Warningf("user-specified reconcile cooldown too small, defaulting to %v", s.ReconcileCooldown)
	}

	executor, eid, err := s.prepareExecutorInfo(hks)
	if err != nil {
		log.Fatalf("misconfigured executor: %v", err)
	}

	// TODO(jdef): remove the dependency on etcd as soon as
	// (1) the generic config store is available for the FrameworkId storage
	// (2) the generic master election is provided by the apiserver
	// Compare docs/proposals/high-availability.md
	etcdClient, err := newEtcd(s.EtcdConfigFile, s.EtcdServerList)
	if err != nil {
		log.Fatalf("misconfigured etcd: %v", err)
	}

	as := scheduler.NewAllocationStrategy(
		podtask.NewDefaultPredicate(
			s.DefaultContainerCPULimit,
			s.DefaultContainerMemLimit,
		),
		podtask.NewDefaultProcurement(
			s.DefaultContainerCPULimit,
			s.DefaultContainerMemLimit,
		),
	)

	// downgrade allocation strategy if user disables "account-for-pod-resources"
	if !s.AccountForPodResources {
		as = scheduler.NewAllocationStrategy(
			podtask.DefaultMinimalPredicate,
			podtask.DefaultMinimalProcurement)
	}

	// mirror all nodes into the nodeStore
	nodesClient, err := s.createAPIServerClient()
	if err != nil {
		log.Fatalf("Cannot create client to watch nodes: %v", err)
	}
	nodeStore := cache.NewStore(cache.MetaNamespaceKeyFunc)
	nodeLW := cache.NewListWatchFromClient(nodesClient, "nodes", api.NamespaceAll, fields.Everything())
	cache.NewReflector(nodeLW, &api.Node{}, nodeStore, s.nodeRelistPeriod).Run()

	lookupNode := func(hostName string) *api.Node {
		n, _, _ := nodeStore.GetByKey(hostName) // ignore error and return nil then
		if n == nil {
			return nil
		}
		return n.(*api.Node)
	}

	fcfs := scheduler.NewFCFSPodScheduler(as, lookupNode)
	mesosPodScheduler := scheduler.New(scheduler.Config{
		Schedcfg:          *sc,
		Executor:          executor,
		Scheduler:         fcfs,
		Client:            client,
		EtcdClient:        etcdClient,
		FailoverTimeout:   s.FailoverTimeout,
		ReconcileInterval: s.ReconcileInterval,
		ReconcileCooldown: s.ReconcileCooldown,
		LookupNode:        lookupNode,
	})

	masterUri := s.MesosMaster
	info, cred, err := s.buildFrameworkInfo()
	if err != nil {
		log.Fatalf("Misconfigured mesos framework: %v", err)
	}

	schedulerProcess := ha.New(mesosPodScheduler)
	dconfig := &bindings.DriverConfig{
		Scheduler:        schedulerProcess,
		Framework:        info,
		Master:           masterUri,
		Credential:       cred,
		BindingAddress:   s.Address,
		BindingPort:      uint16(s.DriverPort),
		HostnameOverride: s.HostnameOverride,
		WithAuthContext: func(ctx context.Context) context.Context {
			ctx = auth.WithLoginProvider(ctx, s.MesosAuthProvider)
			ctx = sasl.WithBindingAddress(ctx, s.Address)
			return ctx
		},
	}

	kpl := scheduler.NewPlugin(mesosPodScheduler.NewDefaultPluginConfig(schedulerProcess.Terminal(), s.mux))
	runtime.On(mesosPodScheduler.Registration(), func() { kpl.Run(schedulerProcess.Terminal()) })
	runtime.On(mesosPodScheduler.Registration(), s.newServiceWriter(schedulerProcess.Terminal()))

	driverFactory := ha.DriverFactory(func() (drv bindings.SchedulerDriver, err error) {
		log.V(1).Infoln("performing deferred initialization")
		if err = mesosPodScheduler.Init(schedulerProcess.Master(), kpl, s.mux); err != nil {
			return nil, fmt.Errorf("failed to initialize pod scheduler: %v", err)
		}
		log.V(1).Infoln("deferred init complete")
		// defer obtaining framework ID to prevent multiple schedulers
		// from overwriting each other's framework IDs
		dconfig.Framework.Id, err = s.fetchFrameworkID(etcdClient)
		if err != nil {
			return nil, fmt.Errorf("failed to fetch framework ID from etcd: %v", err)
		}
		log.V(1).Infoln("constructing mesos scheduler driver")
		drv, err = bindings.NewMesosSchedulerDriver(*dconfig)
		if err != nil {
			return nil, fmt.Errorf("failed to construct scheduler driver: %v", err)
		}
		log.V(1).Infoln("constructed mesos scheduler driver:", drv)
		s.setDriver(drv)
		return drv, nil
	})

	return schedulerProcess, driverFactory, etcdClient, eid
}
示例#5
0
func (s *SchedulerServer) bootstrap(hks hyperkube.Interface, sc *schedcfg.Config) (*ha.SchedulerProcess, ha.DriverFactory, tools.EtcdClient, *uid.UID) {

	s.FrameworkName = strings.TrimSpace(s.FrameworkName)
	if s.FrameworkName == "" {
		log.Fatalf("framework-name must be a non-empty string")
	}
	s.FrameworkWebURI = strings.TrimSpace(s.FrameworkWebURI)

	metrics.Register()
	runtime.Register()
	s.mux.Handle("/metrics", prometheus.Handler())

	if (s.EtcdConfigFile != "" && len(s.EtcdServerList) != 0) || (s.EtcdConfigFile == "" && len(s.EtcdServerList) == 0) {
		log.Fatalf("specify either --etcd-servers or --etcd-config")
	}

	if len(s.APIServerList) < 1 {
		log.Fatal("No api servers specified.")
	}

	client, err := s.createAPIServerClient()
	if err != nil {
		log.Fatalf("Unable to make apiserver client: %v", err)
	}
	s.client = client

	if s.ReconcileCooldown < defaultReconcileCooldown {
		s.ReconcileCooldown = defaultReconcileCooldown
		log.Warningf("user-specified reconcile cooldown too small, defaulting to %v", s.ReconcileCooldown)
	}

	executor, eid, err := s.prepareExecutorInfo(hks)
	if err != nil {
		log.Fatalf("misconfigured executor: %v", err)
	}

	// TODO(jdef): remove the dependency on etcd as soon as
	// (1) the generic config store is available for the FrameworkId storage
	// (2) the generic master election is provided by the apiserver
	// Compare docs/proposals/high-availability.md
	etcdClient, err := newEtcd(s.EtcdConfigFile, s.EtcdServerList)
	if err != nil {
		log.Fatalf("misconfigured etcd: %v", err)
	}

	mesosPodScheduler := scheduler.New(scheduler.Config{
		Schedcfg:                 *sc,
		Executor:                 executor,
		ScheduleFunc:             scheduler.FCFSScheduleFunc,
		Client:                   client,
		EtcdClient:               etcdClient,
		FailoverTimeout:          s.FailoverTimeout,
		ReconcileInterval:        s.ReconcileInterval,
		ReconcileCooldown:        s.ReconcileCooldown,
		DefaultContainerCPULimit: s.DefaultContainerCPULimit,
		DefaultContainerMemLimit: s.DefaultContainerMemLimit,
	})

	masterUri := s.MesosMaster
	info, cred, err := s.buildFrameworkInfo()
	if err != nil {
		log.Fatalf("Misconfigured mesos framework: %v", err)
	}

	schedulerProcess := ha.New(mesosPodScheduler)
	dconfig := &bindings.DriverConfig{
		Scheduler:        schedulerProcess,
		Framework:        info,
		Master:           masterUri,
		Credential:       cred,
		BindingAddress:   net.IP(s.Address),
		BindingPort:      uint16(s.DriverPort),
		HostnameOverride: s.HostnameOverride,
		WithAuthContext: func(ctx context.Context) context.Context {
			ctx = auth.WithLoginProvider(ctx, s.MesosAuthProvider)
			ctx = sasl.WithBindingAddress(ctx, net.IP(s.Address))
			return ctx
		},
	}

	kpl := scheduler.NewPlugin(mesosPodScheduler.NewDefaultPluginConfig(schedulerProcess.Terminal(), s.mux))
	runtime.On(mesosPodScheduler.Registration(), func() { kpl.Run(schedulerProcess.Terminal()) })
	runtime.On(mesosPodScheduler.Registration(), s.newServiceWriter(schedulerProcess.Terminal()))

	driverFactory := ha.DriverFactory(func() (drv bindings.SchedulerDriver, err error) {
		log.V(1).Infoln("performing deferred initialization")
		if err = mesosPodScheduler.Init(schedulerProcess.Master(), kpl, s.mux); err != nil {
			return nil, fmt.Errorf("failed to initialize pod scheduler: %v", err)
		}
		log.V(1).Infoln("deferred init complete")
		// defer obtaining framework ID to prevent multiple schedulers
		// from overwriting each other's framework IDs
		dconfig.Framework.Id, err = s.fetchFrameworkID(etcdClient)
		if err != nil {
			return nil, fmt.Errorf("failed to fetch framework ID from etcd: %v", err)
		}
		log.V(1).Infoln("constructing mesos scheduler driver")
		drv, err = bindings.NewMesosSchedulerDriver(*dconfig)
		if err != nil {
			return nil, fmt.Errorf("failed to construct scheduler driver: %v", err)
		}
		log.V(1).Infoln("constructed mesos scheduler driver:", drv)
		s.setDriver(drv)
		return drv, nil
	})

	return schedulerProcess, driverFactory, etcdClient, eid
}
示例#6
0
func (s *SchedulerServer) bootstrap(hks hyperkube.Interface, sc *schedcfg.Config) (*ha.SchedulerProcess, ha.DriverFactory, *etcd.Client, *mesos.ExecutorID) {
	s.frameworkName = strings.TrimSpace(s.frameworkName)
	if s.frameworkName == "" {
		log.Fatalf("framework-name must be a non-empty string")
	}
	s.frameworkWebURI = strings.TrimSpace(s.frameworkWebURI)

	metrics.Register()
	runtime.Register()
	s.mux.Handle("/metrics", prometheus.Handler())
	healthz.InstallHandler(s.mux)

	if (s.etcdConfigFile != "" && len(s.etcdServerList) != 0) || (s.etcdConfigFile == "" && len(s.etcdServerList) == 0) {
		log.Fatalf("specify either --etcd-servers or --etcd-config")
	}

	if len(s.apiServerList) < 1 {
		log.Fatal("No api servers specified.")
	}

	client, err := s.createAPIServerClient()
	if err != nil {
		log.Fatalf("Unable to make apiserver client: %v", err)
	}
	s.client = client

	if s.reconcileCooldown < defaultReconcileCooldown {
		s.reconcileCooldown = defaultReconcileCooldown
		log.Warningf("user-specified reconcile cooldown too small, defaulting to %v", s.reconcileCooldown)
	}

	eiPrototype, err := s.prepareExecutorInfo(hks)
	if err != nil {
		log.Fatalf("misconfigured executor: %v", err)
	}

	// TODO(jdef): remove the dependency on etcd as soon as
	// (1) the generic config store is available for the FrameworkId storage
	// (2) the generic master election is provided by the apiserver
	// Compare docs/proposals/high-availability.md
	etcdClient, err := newEtcd(s.etcdConfigFile, s.etcdServerList)
	if err != nil {
		log.Fatalf("misconfigured etcd: %v", err)
	}

	// mirror all nodes into the nodeStore
	var eiRegistry executorinfo.Registry
	nodesClient, err := s.createAPIServerClient()
	if err != nil {
		log.Fatalf("Cannot create client to watch nodes: %v", err)
	}
	nodeLW := cache.NewListWatchFromClient(nodesClient, "nodes", api.NamespaceAll, fields.Everything())
	nodeStore, nodeCtl := controllerfw.NewInformer(nodeLW, &api.Node{}, s.nodeRelistPeriod, &controllerfw.ResourceEventHandlerFuncs{
		DeleteFunc: func(obj interface{}) {
			node := obj.(*api.Node)
			if eiRegistry != nil {
				log.V(2).Infof("deleting node %q from registry", node.Name)
				eiRegistry.Invalidate(node.Name)
			}
		},
	})

	lookupNode := func(hostName string) *api.Node {
		n, _, _ := nodeStore.GetByKey(hostName) // ignore error and return nil then
		if n == nil {
			return nil
		}
		return n.(*api.Node)
	}

	execInfoCache, err := executorinfo.NewCache(defaultExecutorInfoCacheSize)
	if err != nil {
		log.Fatalf("cannot create executorinfo cache: %v", err)
	}

	eiRegistry, err = executorinfo.NewRegistry(lookupNode, eiPrototype, execInfoCache)
	if err != nil {
		log.Fatalf("cannot create executorinfo registry: %v", err)
	}

	pr := podtask.NewDefaultProcurement(eiPrototype, eiRegistry)
	fcfs := podschedulers.NewFCFSPodScheduler(pr, lookupNode)

	framework := framework.New(framework.Config{
		SchedulerConfig:   *sc,
		Client:            client,
		FailoverTimeout:   s.failoverTimeout,
		ReconcileInterval: s.reconcileInterval,
		ReconcileCooldown: s.reconcileCooldown,
		LookupNode:        lookupNode,
		StoreFrameworkId: func(id string) {
			// TODO(jdef): port FrameworkId store to generic Kubernetes config store as soon as available
			_, err := etcdClient.Set(meta.FrameworkIDKey, id, uint64(s.failoverTimeout))
			if err != nil {
				log.Errorf("failed to renew frameworkId TTL: %v", err)
			}
		},
		ExecutorId: eiPrototype.GetExecutorId(),
	})

	masterUri := s.mesosMaster
	info, cred, err := s.buildFrameworkInfo()
	if err != nil {
		log.Fatalf("Misconfigured mesos framework: %v", err)
	}

	schedulerProcess := ha.New(framework)
	dconfig := &bindings.DriverConfig{
		Scheduler:        schedulerProcess,
		Framework:        info,
		Master:           masterUri,
		Credential:       cred,
		BindingAddress:   s.address,
		BindingPort:      uint16(s.driverPort),
		HostnameOverride: s.hostnameOverride,
		WithAuthContext: func(ctx context.Context) context.Context {
			ctx = auth.WithLoginProvider(ctx, s.mesosAuthProvider)
			ctx = sasl.WithBindingAddress(ctx, s.address)
			return ctx
		},
	}

	// create event recorder sending events to the "" namespace of the apiserver
	broadcaster := record.NewBroadcaster()
	recorder := broadcaster.NewRecorder(api.EventSource{Component: "scheduler"})
	broadcaster.StartRecordingToSink(client.Events(""))

	// create scheduler core with all components arranged around it
	lw := cache.NewListWatchFromClient(client, "pods", api.NamespaceAll, fields.Everything())
	sched := components.New(
		sc,
		framework,
		fcfs,
		client,
		recorder,
		schedulerProcess.Terminal(),
		s.mux,
		lw,
		eiPrototype,
		s.mesosRoles,
		s.defaultContainerCPULimit,
		s.defaultContainerMemLimit,
	)

	runtime.On(framework.Registration(), func() { sched.Run(schedulerProcess.Terminal()) })
	runtime.On(framework.Registration(), s.newServiceWriter(schedulerProcess.Terminal()))
	runtime.On(framework.Registration(), func() { nodeCtl.Run(schedulerProcess.Terminal()) })

	driverFactory := ha.DriverFactory(func() (drv bindings.SchedulerDriver, err error) {
		log.V(1).Infoln("performing deferred initialization")
		if err = framework.Init(sched, schedulerProcess.Master(), s.mux); err != nil {
			return nil, fmt.Errorf("failed to initialize pod scheduler: %v", err)
		}
		log.V(1).Infoln("deferred init complete")
		// defer obtaining framework ID to prevent multiple schedulers
		// from overwriting each other's framework IDs
		dconfig.Framework.Id, err = s.fetchFrameworkID(etcdClient)
		if err != nil {
			return nil, fmt.Errorf("failed to fetch framework ID from etcd: %v", err)
		}
		log.V(1).Infoln("constructing mesos scheduler driver")
		drv, err = bindings.NewMesosSchedulerDriver(*dconfig)
		if err != nil {
			return nil, fmt.Errorf("failed to construct scheduler driver: %v", err)
		}
		log.V(1).Infoln("constructed mesos scheduler driver:", drv)
		s.setDriver(drv)
		return drv, nil
	})

	return schedulerProcess, driverFactory, etcdClient, eiPrototype.GetExecutorId()
}
示例#7
0
func newLifecycleTest(t *testing.T) lifecycleTest {
	assert := &EventAssertions{*assert.New(t)}

	// create a fake pod watch. We use that below to submit new pods to the scheduler
	podsListWatch := NewMockPodsListWatch(api.PodList{})

	// create fake apiserver
	apiServer := NewTestServer(t, api.NamespaceDefault, podsListWatch)

	// create executor with some data for static pods if set
	executor := mesosutil.NewExecutorInfo(
		mesosutil.NewExecutorID("executor-id"),
		mesosutil.NewCommandInfo("executor-cmd"),
	)
	executor.Data = []byte{0, 1, 2}

	// create scheduler
	strategy := NewAllocationStrategy(
		podtask.NewDefaultPredicate(
			mresource.DefaultDefaultContainerCPULimit,
			mresource.DefaultDefaultContainerMemLimit,
		),
		podtask.NewDefaultProcurement(
			mresource.DefaultDefaultContainerCPULimit,
			mresource.DefaultDefaultContainerMemLimit,
		),
	)

	scheduler := New(Config{
		Executor: executor,
		Client: client.NewOrDie(&client.Config{
			Host:    apiServer.server.URL,
			Version: testapi.Default.Version(),
		}),
		Scheduler:  NewFCFSPodScheduler(strategy, apiServer.LookupNode),
		Schedcfg:   *schedcfg.CreateDefaultConfig(),
		LookupNode: apiServer.LookupNode,
	})

	assert.NotNil(scheduler.client, "client is nil")
	assert.NotNil(scheduler.executor, "executor is nil")
	assert.NotNil(scheduler.offers, "offer registry is nil")

	// create scheduler process
	schedulerProc := ha.New(scheduler)

	// get plugin config from it
	config := scheduler.NewPluginConfig(
		schedulerProc.Terminal(),
		http.DefaultServeMux,
		&podsListWatch.ListWatch,
	)
	assert.NotNil(config)

	// make events observable
	eventObs := NewEventObserver()
	config.Recorder = eventObs

	// create plugin
	plugin := NewPlugin(config).(*schedulingPlugin)
	assert.NotNil(plugin)

	// create mock mesos scheduler driver
	driver := &joinableDriver{}

	return lifecycleTest{
		apiServer:     apiServer,
		driver:        driver,
		eventObs:      eventObs,
		plugin:        plugin,
		podsListWatch: podsListWatch,
		scheduler:     scheduler,
		schedulerProc: schedulerProc,
		t:             t,
	}
}