Exemplo n.º 1
0
func TestProc_manyEndings(t *testing.T) {
	p := New()
	const COUNT = 20
	var wg sync.WaitGroup
	wg.Add(COUNT)
	for i := 0; i < COUNT; i++ {
		runtime.On(p.End(), wg.Done)
	}
	fatalAfter(t, runtime.After(wg.Wait), 5*time.Second, "timed out waiting for loose End()s")
	fatalAfter(t, p.Done(), 5*time.Second, "timed out waiting for process death")
}
Exemplo n.º 2
0
func (k *KubernetesScheduler) NewPluginConfig(terminate <-chan struct{}, mux *http.ServeMux,
	podsWatcher *cache.ListWatch) *PluginConfig {

	// Watch and queue pods that need scheduling.
	updates := make(chan queue.Entry, k.schedcfg.UpdatesBacklog)
	podUpdates := &podStoreAdapter{queue.NewHistorical(updates)}
	reflector := cache.NewReflector(podsWatcher, &api.Pod{}, podUpdates, 0)

	// lock that guards critial sections that involve transferring pods from
	// the store (cache) to the scheduling queue; its purpose is to maintain
	// an ordering (vs interleaving) of operations that's easier to reason about.
	kapi := &k8smScheduler{internal: k}
	q := newQueuer(podUpdates)
	podDeleter := &deleter{
		api: kapi,
		qr:  q,
	}
	eh := &errorHandler{
		api:     kapi,
		backoff: backoff.New(k.schedcfg.InitialPodBackoff.Duration, k.schedcfg.MaxPodBackoff.Duration),
		qr:      q,
	}
	startLatch := make(chan struct{})
	eventBroadcaster := record.NewBroadcaster()
	runtime.On(startLatch, func() {
		eventBroadcaster.StartRecordingToSink(k.client.Events(""))
		reflector.Run() // TODO(jdef) should listen for termination
		podDeleter.Run(updates, terminate)
		q.Run(terminate)

		q.installDebugHandlers(mux)
		podtask.InstallDebugHandlers(k.taskRegistry, mux)
	})
	return &PluginConfig{
		Config: &plugin.Config{
			MinionLister: nil,
			Algorithm: &kubeScheduler{
				api:                      kapi,
				podUpdates:               podUpdates,
				defaultContainerCPULimit: k.defaultContainerCPULimit,
				defaultContainerMemLimit: k.defaultContainerMemLimit,
			},
			Binder:   &binder{api: kapi},
			NextPod:  q.yield,
			Error:    eh.handleSchedulingError,
			Recorder: eventBroadcaster.NewRecorder(api.EventSource{Component: "scheduler"}),
		},
		api:      kapi,
		client:   k.client,
		qr:       q,
		deleter:  podDeleter,
		starting: startLatch,
	}
}
Exemplo n.º 3
0
func New(sched bindings.Scheduler) *SchedulerProcess {
	p := &SchedulerProcess{
		Process:   proc.New(),
		Scheduler: sched,
		stage:     initStage,
		elected:   make(chan struct{}),
		failover:  make(chan struct{}),
		standby:   make(chan struct{}),
		fin:       make(chan struct{}),
	}
	runtime.On(p.Running(), p.begin)
	return p
}
Exemplo n.º 4
0
// watch the scheduler process for failover signals and properly handle such. may never return.
func (s *SchedulerServer) awaitFailover(schedulerProcess schedulerProcessInterface, handler func() error) error {

	// we only want to return the first error (if any), everyone else can block forever
	errCh := make(chan error, 1)
	doFailover := func() error {
		// we really don't expect handler to return, if it does something went seriously wrong
		err := handler()
		if err != nil {
			defer schedulerProcess.End()
			err = fmt.Errorf("failover failed, scheduler will terminate: %v", err)
		}
		return err
	}

	// guard for failover signal processing, first signal processor wins
	failoverLatch := &runtime.Latch{}
	runtime.On(schedulerProcess.Terminal(), func() {
		if !failoverLatch.Acquire() {
			log.V(1).Infof("scheduler process ending, already failing over")
			select {}
		}
		var err error
		defer func() { errCh <- err }()
		select {
		case <-schedulerProcess.Failover():
			err = doFailover()
		default:
			if s.HA {
				err = fmt.Errorf("ha scheduler exiting instead of failing over")
			} else {
				log.Infof("exiting scheduler")
			}
		}
	})
	runtime.OnOSSignal(makeFailoverSigChan(), func(_ os.Signal) {
		if !failoverLatch.Acquire() {
			log.V(1).Infof("scheduler process signalled, already failing over")
			select {}
		}
		errCh <- doFailover()
	})
	return <-errCh
}
Exemplo n.º 5
0
func (s *SchedulerServer) bootstrap(hks hyperkube.Interface, sc *schedcfg.Config) (*ha.SchedulerProcess, ha.DriverFactory, tools.EtcdGetSet, *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,
	})

	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
}