示例#1
0
文件: updater.go 项目: Mic92/docker
func (u *Updater) worker(ctx context.Context, queue <-chan orchestrator.Slot) {
	for slot := range queue {
		// Do we have a task with the new spec in desired state = RUNNING?
		// If so, all we have to do to complete the update is remove the
		// other tasks. Or if we have a task with the new spec that has
		// desired state < RUNNING, advance it to running and remove the
		// other tasks.
		var (
			runningTask *api.Task
			cleanTask   *api.Task
		)
		for _, t := range slot {
			if !u.isTaskDirty(t) {
				if t.DesiredState == api.TaskStateRunning {
					runningTask = t
					break
				}
				if t.DesiredState < api.TaskStateRunning {
					cleanTask = t
				}
			}
		}
		if runningTask != nil {
			if err := u.useExistingTask(ctx, slot, runningTask); err != nil {
				log.G(ctx).WithError(err).Error("update failed")
			}
		} else if cleanTask != nil {
			if err := u.useExistingTask(ctx, slot, cleanTask); err != nil {
				log.G(ctx).WithError(err).Error("update failed")
			}
		} else {
			updated := orchestrator.NewTask(u.cluster, u.newService, slot[0].Slot, "")
			if orchestrator.IsGlobalService(u.newService) {
				updated = orchestrator.NewTask(u.cluster, u.newService, slot[0].Slot, slot[0].NodeID)
			}
			updated.DesiredState = api.TaskStateReady

			if err := u.updateTask(ctx, slot, updated); err != nil {
				log.G(ctx).WithError(err).WithField("task.id", updated.ID).Error("update failed")
			}
		}

		if u.newService.Spec.Update != nil && (u.newService.Spec.Update.Delay.Seconds != 0 || u.newService.Spec.Update.Delay.Nanos != 0) {
			delay, err := ptypes.Duration(&u.newService.Spec.Update.Delay)
			if err != nil {
				log.G(ctx).WithError(err).Error("invalid update delay")
				continue
			}
			select {
			case <-time.After(delay):
			case <-u.stopChan:
				return
			}
		}
	}
}
示例#2
0
文件: restart.go 项目: Mic92/docker
func (r *Supervisor) shouldRestart(ctx context.Context, t *api.Task, service *api.Service) bool {
	// TODO(aluzzardi): This function should not depend on `service`.

	condition := orchestrator.RestartCondition(t)

	if condition != api.RestartOnAny &&
		(condition != api.RestartOnFailure || t.Status.State == api.TaskStateCompleted) {
		return false
	}

	if t.Spec.Restart == nil || t.Spec.Restart.MaxAttempts == 0 {
		return true
	}

	instanceTuple := instanceTuple{
		instance:  t.Slot,
		serviceID: t.ServiceID,
	}

	// Instance is not meaningful for "global" tasks, so they need to be
	// indexed by NodeID.
	if orchestrator.IsGlobalService(service) {
		instanceTuple.nodeID = t.NodeID
	}

	r.mu.Lock()
	defer r.mu.Unlock()

	restartInfo := r.history[instanceTuple]
	if restartInfo == nil {
		return true
	}

	if t.Spec.Restart.Window == nil || (t.Spec.Restart.Window.Seconds == 0 && t.Spec.Restart.Window.Nanos == 0) {
		return restartInfo.totalRestarts < t.Spec.Restart.MaxAttempts
	}

	if restartInfo.restartedInstances == nil {
		return true
	}

	window, err := ptypes.Duration(t.Spec.Restart.Window)
	if err != nil {
		log.G(ctx).WithError(err).Error("invalid restart lookback window")
		return restartInfo.totalRestarts < t.Spec.Restart.MaxAttempts
	}
	lookback := time.Now().Add(-window)

	var next *list.Element
	for e := restartInfo.restartedInstances.Front(); e != nil; e = next {
		next = e.Next()

		if e.Value.(restartedInstance).timestamp.After(lookback) {
			break
		}
		restartInfo.restartedInstances.Remove(e)
	}

	numRestarts := uint64(restartInfo.restartedInstances.Len())

	if numRestarts == 0 {
		restartInfo.restartedInstances = nil
	}

	return numRestarts < t.Spec.Restart.MaxAttempts
}
示例#3
0
// Run contains the global orchestrator event loop
func (g *Orchestrator) Run(ctx context.Context) error {
	defer close(g.doneChan)

	// Watch changes to services and tasks
	queue := g.store.WatchQueue()
	watcher, cancel := queue.Watch()
	defer cancel()

	// lookup the cluster
	var err error
	g.store.View(func(readTx store.ReadTx) {
		var clusters []*api.Cluster
		clusters, err = store.FindClusters(readTx, store.ByName("default"))

		if len(clusters) != 1 {
			return // just pick up the cluster when it is created.
		}
		g.cluster = clusters[0]
	})
	if err != nil {
		return err
	}

	// Get list of nodes
	var nodes []*api.Node
	g.store.View(func(readTx store.ReadTx) {
		nodes, err = store.FindNodes(readTx, store.All)
	})
	if err != nil {
		return err
	}
	for _, n := range nodes {
		g.updateNode(n)
	}

	// Lookup global services
	var existingServices []*api.Service
	g.store.View(func(readTx store.ReadTx) {
		existingServices, err = store.FindServices(readTx, store.All)
	})
	if err != nil {
		return err
	}

	var reconcileServiceIDs []string
	for _, s := range existingServices {
		if orchestrator.IsGlobalService(s) {
			g.updateService(s)
			reconcileServiceIDs = append(reconcileServiceIDs, s.ID)
		}
	}
	g.reconcileServices(ctx, reconcileServiceIDs)

	for {
		select {
		case event := <-watcher:
			// TODO(stevvooe): Use ctx to limit running time of operation.
			switch v := event.(type) {
			case state.EventUpdateCluster:
				g.cluster = v.Cluster
			case state.EventCreateService:
				if !orchestrator.IsGlobalService(v.Service) {
					continue
				}
				g.updateService(v.Service)
				g.reconcileServices(ctx, []string{v.Service.ID})
			case state.EventUpdateService:
				if !orchestrator.IsGlobalService(v.Service) {
					continue
				}
				g.updateService(v.Service)
				g.reconcileServices(ctx, []string{v.Service.ID})
			case state.EventDeleteService:
				if !orchestrator.IsGlobalService(v.Service) {
					continue
				}
				orchestrator.DeleteServiceTasks(ctx, g.store, v.Service)
				// delete the service from service map
				delete(g.globalServices, v.Service.ID)
				g.restarts.ClearServiceHistory(v.Service.ID)
			case state.EventCreateNode:
				g.updateNode(v.Node)
				g.reconcileOneNode(ctx, v.Node)
			case state.EventUpdateNode:
				g.updateNode(v.Node)
				switch v.Node.Status.State {
				// NodeStatus_DISCONNECTED is a transient state, no need to make any change
				case api.NodeStatus_DOWN:
					g.removeTasksFromNode(ctx, v.Node)
				case api.NodeStatus_READY:
					// node could come back to READY from DOWN or DISCONNECT
					g.reconcileOneNode(ctx, v.Node)
				}
			case state.EventDeleteNode:
				g.removeTasksFromNode(ctx, v.Node)
				delete(g.nodes, v.Node.ID)
			case state.EventUpdateTask:
				if _, exists := g.globalServices[v.Task.ServiceID]; !exists {
					continue
				}
				// global orchestrator needs to inspect when a task has terminated
				// it should ignore tasks whose DesiredState is past running, which
				// means the task has been processed
				if isTaskTerminated(v.Task) {
					g.restartTask(ctx, v.Task.ID, v.Task.ServiceID)
				}
			case state.EventDeleteTask:
				// CLI allows deleting task
				if _, exists := g.globalServices[v.Task.ServiceID]; !exists {
					continue
				}
				g.reconcileServicesOneNode(ctx, []string{v.Task.ServiceID}, v.Task.NodeID)
			}
		case <-g.stopChan:
			return nil
		}
	}
}
示例#4
0
文件: restart.go 项目: Mic92/docker
// Restart initiates a new task to replace t if appropriate under the service's
// restart policy.
func (r *Supervisor) Restart(ctx context.Context, tx store.Tx, cluster *api.Cluster, service *api.Service, t api.Task) error {
	// TODO(aluzzardi): This function should not depend on `service`.

	// Is the old task still in the process of restarting? If so, wait for
	// its restart delay to elapse, to avoid tight restart loops (for
	// example, when the image doesn't exist).
	r.mu.Lock()
	oldDelay, ok := r.delays[t.ID]
	if ok {
		if !oldDelay.waiter {
			oldDelay.waiter = true
			go r.waitRestart(ctx, oldDelay, cluster, t.ID)
		}
		r.mu.Unlock()
		return nil
	}
	r.mu.Unlock()

	// Sanity check: was the task shut down already by a separate call to
	// Restart? If so, we must avoid restarting it, because this will create
	// an extra task. This should never happen unless there is a bug.
	if t.DesiredState > api.TaskStateRunning {
		return errors.New("Restart called on task that was already shut down")
	}

	t.DesiredState = api.TaskStateShutdown
	err := store.UpdateTask(tx, &t)
	if err != nil {
		log.G(ctx).WithError(err).Errorf("failed to set task desired state to dead")
		return err
	}

	if !r.shouldRestart(ctx, &t, service) {
		return nil
	}

	var restartTask *api.Task

	if orchestrator.IsReplicatedService(service) {
		restartTask = orchestrator.NewTask(cluster, service, t.Slot, "")
	} else if orchestrator.IsGlobalService(service) {
		restartTask = orchestrator.NewTask(cluster, service, 0, t.NodeID)
	} else {
		log.G(ctx).Error("service not supported by restart supervisor")
		return nil
	}

	n := store.GetNode(tx, t.NodeID)

	restartTask.DesiredState = api.TaskStateReady

	var restartDelay time.Duration
	// Restart delay is not applied to drained nodes
	if n == nil || n.Spec.Availability != api.NodeAvailabilityDrain {
		if t.Spec.Restart != nil && t.Spec.Restart.Delay != nil {
			var err error
			restartDelay, err = ptypes.Duration(t.Spec.Restart.Delay)
			if err != nil {
				log.G(ctx).WithError(err).Error("invalid restart delay; using default")
				restartDelay = orchestrator.DefaultRestartDelay
			}
		} else {
			restartDelay = orchestrator.DefaultRestartDelay
		}
	}

	waitStop := true

	// Normally we wait for the old task to stop running, but we skip this
	// if the old task is already dead or the node it's assigned to is down.
	if (n != nil && n.Status.State == api.NodeStatus_DOWN) || t.Status.State > api.TaskStateRunning {
		waitStop = false
	}

	if err := store.CreateTask(tx, restartTask); err != nil {
		log.G(ctx).WithError(err).WithField("task.id", restartTask.ID).Error("task create failed")
		return err
	}

	r.recordRestartHistory(restartTask)

	r.DelayStart(ctx, tx, &t, restartTask.ID, restartDelay, waitStop)
	return nil
}