Exemple #1
0
// CheckMigsAndNodes checks if all migs have all required nodes.
func CheckMigsAndNodes(nodes []*kube_api.Node, gceManager *gce.GceManager) error {
	migCount := make(map[string]int)
	migs := make(map[string]*config.MigConfig)
	for _, node := range nodes {
		instanceConfig, err := config.InstanceConfigFromProviderId(node.Spec.ProviderID)
		if err != nil {
			return err
		}

		migConfig, err := gceManager.GetMigForInstance(instanceConfig)
		if err != nil {
			return err
		}
		url := migConfig.Url()
		count, _ := migCount[url]
		migCount[url] = count + 1
		migs[url] = migConfig
	}
	for url, mig := range migs {
		size, err := gceManager.GetMigSize(mig)
		if err != nil {
			return err
		}
		count := migCount[url]
		if size != int64(count) {
			return fmt.Errorf("wrong number of nodes for mig: %s expected: %d actual: %d", url, size, count)
		}
	}
	return nil
}
Exemple #2
0
// ScaleUp tries to scale the cluster up. Return true if it found a way to increase the size,
// false if it didn't and error if an error occured.
func ScaleUp(unschedulablePods []*kube_api.Pod, nodes []*kube_api.Node, migConfigs []*config.MigConfig,
	gceManager *gce.GceManager, kubeClient *kube_client.Client,
	predicateChecker *simulator.PredicateChecker, recorder kube_record.EventRecorder) (bool, error) {

	// From now on we only care about unschedulable pods that were marked after the newest
	// node became available for the scheduler.
	if len(unschedulablePods) == 0 {
		glog.V(1).Info("No unschedulable pods")
		return false, nil
	}

	for _, pod := range unschedulablePods {
		glog.V(1).Infof("Pod %s/%s is unschedulable", pod.Namespace, pod.Name)
	}

	expansionOptions := make([]ExpansionOption, 0)
	nodeInfos, err := GetNodeInfosForMigs(nodes, gceManager, kubeClient)
	if err != nil {
		return false, fmt.Errorf("failed to build node infors for migs: %v", err)
	}

	podsRemainUnshedulable := make(map[*kube_api.Pod]struct{})
	for _, migConfig := range migConfigs {

		currentSize, err := gceManager.GetMigSize(migConfig)
		if err != nil {
			glog.Errorf("Failed to get MIG size: %v", err)
			continue
		}
		if currentSize >= int64(migConfig.MaxSize) {
			// skip this mig.
			glog.V(4).Infof("Skipping MIG %s - max size reached", migConfig.Url())
			continue
		}

		option := ExpansionOption{
			migConfig: migConfig,
			estimator: estimator.NewBasicNodeEstimator(),
		}
		migHelpsSomePods := false

		nodeInfo, found := nodeInfos[migConfig.Url()]
		if !found {
			glog.Errorf("No node info for: %s", migConfig.Url())
			continue
		}

		for _, pod := range unschedulablePods {
			err = predicateChecker.CheckPredicates(pod, nodeInfo)
			if err == nil {
				migHelpsSomePods = true
				option.estimator.Add(pod)
			} else {
				glog.V(2).Infof("Scale-up predicate failed: %v", err)
				podsRemainUnshedulable[pod] = struct{}{}
			}
		}
		if migHelpsSomePods {
			expansionOptions = append(expansionOptions, option)
		}
	}

	// Pick some expansion option.
	bestOption := BestExpansionOption(expansionOptions)
	if bestOption != nil && bestOption.estimator.GetCount() > 0 {
		glog.V(1).Infof("Best option to resize: %s", bestOption.migConfig.Url())
		nodeInfo, found := nodeInfos[bestOption.migConfig.Url()]
		if !found {
			return false, fmt.Errorf("no sample node for: %s", bestOption.migConfig.Url())

		}
		node := nodeInfo.Node()
		estimate, report := bestOption.estimator.Estimate(node)
		glog.V(1).Info(bestOption.estimator.GetDebug())
		glog.V(1).Info(report)
		glog.V(1).Infof("Estimated %d nodes needed in %s", estimate, bestOption.migConfig.Url())

		currentSize, err := gceManager.GetMigSize(bestOption.migConfig)
		if err != nil {
			return false, fmt.Errorf("failed to get MIG size: %v", err)
		}
		newSize := currentSize + int64(estimate)
		if newSize >= int64(bestOption.migConfig.MaxSize) {
			glog.V(1).Infof("Capping size to MAX (%d)", bestOption.migConfig.MaxSize)
			newSize = int64(bestOption.migConfig.MaxSize)
		}
		glog.V(1).Infof("Setting %s size to %d", bestOption.migConfig.Url(), newSize)

		if err := gceManager.SetMigSize(bestOption.migConfig, newSize); err != nil {
			return false, fmt.Errorf("failed to set MIG size: %v", err)
		}

		for pod := range bestOption.estimator.FittingPods {
			recorder.Eventf(pod, kube_api.EventTypeNormal, "TriggeredScaleUp",
				"pod triggered scale-up, mig: %s, sizes (current/new): %d/%d", bestOption.migConfig.Name, currentSize, newSize)
		}

		return true, nil
	}
	for pod := range podsRemainUnshedulable {
		recorder.Event(pod, kube_api.EventTypeNormal, "NotTriggerScaleUp",
			"pod didn't trigger scale-up (it wouldn't fit if a new node is added)")
	}

	return false, nil
}
Exemple #3
0
// ScaleDown tries to scale down the cluster. It returns ScaleDownResult indicating if any node was
// removed and error if such occured.
func ScaleDown(
	nodes []*kube_api.Node,
	unneededNodes map[string]time.Time,
	unneededTime time.Duration,
	pods []*kube_api.Pod,
	gceManager *gce.GceManager,
	client *kube_client.Client,
	predicateChecker *simulator.PredicateChecker) (ScaleDownResult, error) {

	now := time.Now()
	candidates := make([]*kube_api.Node, 0)
	for _, node := range nodes {
		if val, found := unneededNodes[node.Name]; found {

			glog.V(2).Infof("%s was unneeded for %s", node.Name, now.Sub(val).String())

			// Check how long the node was underutilized.
			if !val.Add(unneededTime).Before(now) {
				continue
			}

			// Check mig size.
			instance, err := config.InstanceConfigFromProviderId(node.Spec.ProviderID)
			if err != nil {
				glog.Errorf("Error while parsing providerid of %s: %v", node.Name, err)
				continue
			}
			migConfig, err := gceManager.GetMigForInstance(instance)
			if err != nil {
				glog.Errorf("Error while checking mig config for instance %v: %v", instance, err)
				continue
			}
			size, err := gceManager.GetMigSize(migConfig)
			if err != nil {
				glog.Errorf("Error while checking mig size for instance %v: %v", instance, err)
				continue
			}

			if size <= int64(migConfig.MinSize) {
				glog.V(1).Infof("Skipping %s - mig min size reached", node.Name)
				continue
			}

			candidates = append(candidates, node)
		}
	}
	if len(candidates) == 0 {
		glog.Infof("No candidates for scale down")
		return ScaleDownNoUnneeded, nil
	}

	nodesToRemove, err := simulator.FindNodesToRemove(candidates, nodes, pods, client, predicateChecker, 1, false)
	if err != nil {
		return ScaleDownError, fmt.Errorf("Find node to remove failed: %v", err)
	}
	if len(nodesToRemove) == 0 {
		glog.V(1).Infof("No node to remove")
		return ScaleDownNoNodeDeleted, nil
	}
	nodeToRemove := nodesToRemove[0]
	glog.Infof("Removing %s", nodeToRemove.Name)
	instanceConfig, err := config.InstanceConfigFromProviderId(nodeToRemove.Spec.ProviderID)
	if err != nil {
		return ScaleDownError, fmt.Errorf("Failed to get instance config for %s: %v", nodeToRemove.Name, err)
	}

	err = gceManager.DeleteInstances([]*config.InstanceConfig{instanceConfig})
	if err != nil {
		return ScaleDownError, fmt.Errorf("Failed to delete %v: %v", instanceConfig, err)
	}

	return ScaleDownNodeDeleted, nil
}