autoscaler/cluster-autoscaler/core/scale_up.go

/*
Copyright 2016 The Kubernetes Authors.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
*/

package core

import (
	"bytes"
	"fmt"
	"math"
	"strings"
	"time"

	"k8s.io/autoscaler/cluster-autoscaler/core/utils"
	"k8s.io/autoscaler/cluster-autoscaler/utils/taints"

	appsv1 "k8s.io/api/apps/v1"
	apiv1 "k8s.io/api/core/v1"
	"k8s.io/apimachinery/pkg/util/sets"
	"k8s.io/autoscaler/cluster-autoscaler/cloudprovider"
	"k8s.io/autoscaler/cluster-autoscaler/clusterstate"
	"k8s.io/autoscaler/cluster-autoscaler/context"
	"k8s.io/autoscaler/cluster-autoscaler/expander"
	"k8s.io/autoscaler/cluster-autoscaler/metrics"
	ca_processors "k8s.io/autoscaler/cluster-autoscaler/processors"
	"k8s.io/autoscaler/cluster-autoscaler/processors/nodegroups"
	"k8s.io/autoscaler/cluster-autoscaler/processors/nodegroupset"
	"k8s.io/autoscaler/cluster-autoscaler/processors/status"
	"k8s.io/autoscaler/cluster-autoscaler/utils/errors"
	"k8s.io/autoscaler/cluster-autoscaler/utils/gpu"
	"k8s.io/autoscaler/cluster-autoscaler/utils/klogx"
	schedulerframework "k8s.io/kubernetes/pkg/scheduler/framework"

	klog "k8s.io/klog/v2"
)

type scaleUpResourcesLimits map[string]int64
type scaleUpResourcesDelta map[string]int64

// used as a value in scaleUpResourcesLimits if actual limit could not be obtained due to errors talking to cloud provider
const scaleUpLimitUnknown = math.MaxInt64

func computeScaleUpResourcesLeftLimits(
	cp cloudprovider.CloudProvider,
	nodeGroups []cloudprovider.NodeGroup,
	nodeInfos map[string]*schedulerframework.NodeInfo,
	nodesFromNotAutoscaledGroups []*apiv1.Node,
	resourceLimiter *cloudprovider.ResourceLimiter) (scaleUpResourcesLimits, errors.AutoscalerError) {
	totalCores, totalMem, errCoresMem := calculateScaleUpCoresMemoryTotal(nodeGroups, nodeInfos, nodesFromNotAutoscaledGroups)

	var totalGpus map[string]int64
	var totalGpusErr error
	if cloudprovider.ContainsGpuResources(resourceLimiter.GetResources()) {
		totalGpus, totalGpusErr = calculateScaleUpGpusTotal(cp.GPULabel(), nodeGroups, nodeInfos, nodesFromNotAutoscaledGroups)
	}

	resultScaleUpLimits := make(scaleUpResourcesLimits)
	for _, resource := range resourceLimiter.GetResources() {
		max := resourceLimiter.GetMax(resource)

		// we put only actual limits into final map. No entry means no limit.
		if max > 0 {
			if (resource == cloudprovider.ResourceNameCores || resource == cloudprovider.ResourceNameMemory) && errCoresMem != nil {
				// core resource info missing - no reason to proceed with scale up
				return scaleUpResourcesLimits{}, errCoresMem
			}
			switch {
			case resource == cloudprovider.ResourceNameCores:
				if errCoresMem != nil {
					resultScaleUpLimits[resource] = scaleUpLimitUnknown
				} else {
					resultScaleUpLimits[resource] = computeBelowMax(totalCores, max)
				}

			case resource == cloudprovider.ResourceNameMemory:
				if errCoresMem != nil {
					resultScaleUpLimits[resource] = scaleUpLimitUnknown
				} else {
					resultScaleUpLimits[resource] = computeBelowMax(totalMem, max)
				}

			case cloudprovider.IsGpuResource(resource):
				if totalGpusErr != nil {
					resultScaleUpLimits[resource] = scaleUpLimitUnknown
				} else {
					resultScaleUpLimits[resource] = computeBelowMax(totalGpus[resource], max)
				}

			default:
				klog.Errorf("Scale up limits defined for unsupported resource '%s'", resource)
			}
		}
	}

	return resultScaleUpLimits, nil
}

func calculateScaleUpCoresMemoryTotal(
	nodeGroups []cloudprovider.NodeGroup,
	nodeInfos map[string]*schedulerframework.NodeInfo,
	nodesFromNotAutoscaledGroups []*apiv1.Node) (int64, int64, errors.AutoscalerError) {
	var coresTotal int64
	var memoryTotal int64

	for _, nodeGroup := range nodeGroups {
		currentSize, err := nodeGroup.TargetSize()
		if err != nil {
			return 0, 0, errors.ToAutoscalerError(errors.CloudProviderError, err).AddPrefix("Failed to get node group size of %v:", nodeGroup.Id())
		}
		nodeInfo, found := nodeInfos[nodeGroup.Id()]
		if !found {
			return 0, 0, errors.NewAutoscalerError(errors.CloudProviderError, "No node info for: %s", nodeGroup.Id())
		}
		if currentSize > 0 {
			nodeCPU, nodeMemory := getNodeInfoCoresAndMemory(nodeInfo)
			coresTotal = coresTotal + int64(currentSize)*nodeCPU
			memoryTotal = memoryTotal + int64(currentSize)*nodeMemory
		}
	}

	for _, node := range nodesFromNotAutoscaledGroups {
		cores, memory := utils.GetNodeCoresAndMemory(node)
		coresTotal += cores
		memoryTotal += memory
	}

	return coresTotal, memoryTotal, nil
}

func calculateScaleUpGpusTotal(
	GPULabel string,
	nodeGroups []cloudprovider.NodeGroup,
	nodeInfos map[string]*schedulerframework.NodeInfo,
	nodesFromNotAutoscaledGroups []*apiv1.Node) (map[string]int64, errors.AutoscalerError) {

	result := make(map[string]int64)
	for _, nodeGroup := range nodeGroups {
		currentSize, err := nodeGroup.TargetSize()
		if err != nil {
			return nil, errors.ToAutoscalerError(errors.CloudProviderError, err).AddPrefix("Failed to get node group size of %v:", nodeGroup.Id())
		}
		nodeInfo, found := nodeInfos[nodeGroup.Id()]
		if !found {
			return nil, errors.NewAutoscalerError(errors.CloudProviderError, "No node info for: %s", nodeGroup.Id())
		}
		if currentSize > 0 {
			gpuType, gpuCount, err := gpu.GetNodeTargetGpus(GPULabel, nodeInfo.Node(), nodeGroup)
			if err != nil {
				return nil, errors.ToAutoscalerError(errors.CloudProviderError, err).AddPrefix("Failed to get target gpu for node group %v:", nodeGroup.Id())
			}
			if gpuType == "" {
				continue
			}
			result[gpuType] += gpuCount * int64(currentSize)
		}
	}

	for _, node := range nodesFromNotAutoscaledGroups {
		gpuType, gpuCount, err := gpu.GetNodeTargetGpus(GPULabel, node, nil)
		if err != nil {
			return nil, errors.ToAutoscalerError(errors.CloudProviderError, err).AddPrefix("Failed to get target gpu for node gpus count for node %v:", node.Name)
		}
		result[gpuType] += gpuCount
	}

	return result, nil
}

func computeBelowMax(total int64, max int64) int64 {
	if total < max {
		return max - total
	}
	return 0
}

func computeScaleUpResourcesDelta(cp cloudprovider.CloudProvider, nodeInfo *schedulerframework.NodeInfo, nodeGroup cloudprovider.NodeGroup, resourceLimiter *cloudprovider.ResourceLimiter) (scaleUpResourcesDelta, errors.AutoscalerError) {
	resultScaleUpDelta := make(scaleUpResourcesDelta)

	nodeCPU, nodeMemory := getNodeInfoCoresAndMemory(nodeInfo)
	resultScaleUpDelta[cloudprovider.ResourceNameCores] = nodeCPU
	resultScaleUpDelta[cloudprovider.ResourceNameMemory] = nodeMemory

	if cloudprovider.ContainsGpuResources(resourceLimiter.GetResources()) {
		gpuType, gpuCount, err := gpu.GetNodeTargetGpus(cp.GPULabel(), nodeInfo.Node(), nodeGroup)
		if err != nil {
			return scaleUpResourcesDelta{}, errors.ToAutoscalerError(errors.CloudProviderError, err).AddPrefix("Failed to get target gpu for node group %v:", nodeGroup.Id())
		}
		resultScaleUpDelta[gpuType] = gpuCount
	}

	return resultScaleUpDelta, nil
}

type scaleUpLimitsCheckResult struct {
	exceeded          bool
	exceededResources []string
}

func scaleUpLimitsNotExceeded() scaleUpLimitsCheckResult {
	return scaleUpLimitsCheckResult{false, []string{}}
}

func (limits *scaleUpResourcesLimits) checkScaleUpDeltaWithinLimits(delta scaleUpResourcesDelta) scaleUpLimitsCheckResult {
	exceededResources := sets.NewString()
	for resource, resourceDelta := range delta {
		resourceLeft, found := (*limits)[resource]
		if found {
			if (resourceDelta > 0) && (resourceLeft == scaleUpLimitUnknown || resourceDelta > resourceLeft) {
				exceededResources.Insert(resource)
			}
		}
	}
	if len(exceededResources) > 0 {
		return scaleUpLimitsCheckResult{true, exceededResources.List()}
	}

	return scaleUpLimitsNotExceeded()
}

func getNodeInfoCoresAndMemory(nodeInfo *schedulerframework.NodeInfo) (int64, int64) {
	return utils.GetNodeCoresAndMemory(nodeInfo.Node())
}

type skippedReasons struct {
	message []string
}

func (sr *skippedReasons) Reasons() []string {
	return sr.message
}

var (
	backoffReason         = &skippedReasons{[]string{"in backoff after failed scale-up"}}
	maxLimitReachedReason = &skippedReasons{[]string{"max node group size reached"}}
	notReadyReason        = &skippedReasons{[]string{"not ready for scale-up"}}
)

func maxResourceLimitReached(resources []string) *skippedReasons {
	return &skippedReasons{[]string{fmt.Sprintf("max cluster %s limit reached", strings.Join(resources, ", "))}}
}

func computeExpansionOption(context *context.AutoscalingContext, podEquivalenceGroups []*podEquivalenceGroup, nodeGroup cloudprovider.NodeGroup, nodeInfo *schedulerframework.NodeInfo, upcomingNodes []*schedulerframework.NodeInfo) (expander.Option, error) {
	option := expander.Option{
		NodeGroup: nodeGroup,
		Pods:      make([]*apiv1.Pod, 0),
	}

	if err := context.ClusterSnapshot.Fork(); err != nil {
		klog.Errorf("Error while calling ClusterSnapshot.Fork; %v", err)
		return expander.Option{}, err
	}

	// add test node to snapshot
	var pods []*apiv1.Pod
	for _, podInfo := range nodeInfo.Pods {
		pods = append(pods, podInfo.Pod)
	}
	if err := context.ClusterSnapshot.AddNodeWithPods(nodeInfo.Node(), pods); err != nil {
		klog.Errorf("Error while adding test Node; %v", err)
		if err := context.ClusterSnapshot.Revert(); err != nil {
			klog.Fatalf("Error while calling ClusterSnapshot.Revert; %v", err)
		}
		// TODO: Or should I just skip the node group? specifically if Revert fails it is fatal error.
		//       Maybe we should not return error from Revert as we cannot handle it in any way on the caller side?
		return expander.Option{}, err
	}

	for _, eg := range podEquivalenceGroups {
		samplePod := eg.pods[0]
		if err := context.PredicateChecker.CheckPredicates(context.ClusterSnapshot, samplePod, nodeInfo.Node().Name); err == nil {
			// add pods to option
			option.Pods = append(option.Pods, eg.pods...)
			// mark pod group as (theoretically) schedulable
			eg.schedulable = true
		} else {
			klog.V(2).Infof("Pod %s can't be scheduled on %s, predicate checking error: %v", samplePod.Name, nodeGroup.Id(), err.VerboseMessage())
			if podCount := len(eg.pods); podCount > 1 {
				klog.V(2).Infof("%d other pods similar to %s can't be scheduled on %s", podCount-1, samplePod.Name, nodeGroup.Id())
			}
			eg.schedulingErrors[nodeGroup.Id()] = err
		}
	}

	if err := context.ClusterSnapshot.Revert(); err != nil {
		klog.Fatalf("Error while calling ClusterSnapshot.Revert; %v", err)
		return expander.Option{}, err
	}

	if len(option.Pods) > 0 {
		estimator := context.EstimatorBuilder(context.PredicateChecker, context.ClusterSnapshot)
		option.NodeCount = estimator.Estimate(option.Pods, nodeInfo)
	}

	return option, nil
}

// 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 occurred. Assumes that all nodes in the cluster are
// ready and in sync with instance groups.
func ScaleUp(context *context.AutoscalingContext, processors *ca_processors.AutoscalingProcessors, clusterStateRegistry *clusterstate.ClusterStateRegistry, unschedulablePods []*apiv1.Pod,
	nodes []*apiv1.Node, daemonSets []*appsv1.DaemonSet, nodeInfos map[string]*schedulerframework.NodeInfo, ignoredTaints taints.TaintKeySet) (*status.ScaleUpStatus, errors.AutoscalerError) {
	// 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 {
		klog.V(1).Info("No unschedulable pods")
		return &status.ScaleUpStatus{Result: status.ScaleUpNotNeeded}, nil
	}

	now := time.Now()

	loggingQuota := klogx.PodsLoggingQuota()

	for _, pod := range unschedulablePods {
		klogx.V(1).UpTo(loggingQuota).Infof("Pod %s/%s is unschedulable", pod.Namespace, pod.Name)
	}
	klogx.V(1).Over(loggingQuota).Infof("%v other pods are also unschedulable", -loggingQuota.Left())

	nodesFromNotAutoscaledGroups, err := utils.FilterOutNodesFromNotAutoscaledGroups(nodes, context.CloudProvider)
	if err != nil {
		return &status.ScaleUpStatus{Result: status.ScaleUpError}, err.AddPrefix("failed to filter out nodes which are from not autoscaled groups: ")
	}

	nodeGroups := context.CloudProvider.NodeGroups()
	gpuLabel := context.CloudProvider.GPULabel()
	availableGPUTypes := context.CloudProvider.GetAvailableGPUTypes()

	resourceLimiter, errCP := context.CloudProvider.GetResourceLimiter()
	if errCP != nil {
		return &status.ScaleUpStatus{Result: status.ScaleUpError}, errors.ToAutoscalerError(
			errors.CloudProviderError,
			errCP)
	}

	scaleUpResourcesLeft, errLimits := computeScaleUpResourcesLeftLimits(context.CloudProvider, nodeGroups, nodeInfos, nodesFromNotAutoscaledGroups, resourceLimiter)
	if errLimits != nil {
		return &status.ScaleUpStatus{Result: status.ScaleUpError}, errLimits.AddPrefix("Could not compute total resources: ")
	}

	upcomingNodes := make([]*schedulerframework.NodeInfo, 0)
	for nodeGroup, numberOfNodes := range clusterStateRegistry.GetUpcomingNodes() {
		nodeTemplate, found := nodeInfos[nodeGroup]
		if !found {
			return &status.ScaleUpStatus{Result: status.ScaleUpError}, errors.NewAutoscalerError(
				errors.InternalError,
				"failed to find template node for node group %s",
				nodeGroup)
		}
		for i := 0; i < numberOfNodes; i++ {
			upcomingNodes = append(upcomingNodes, nodeTemplate)
		}
	}
	klog.V(4).Infof("Upcoming %d nodes", len(upcomingNodes))

	expansionOptions := make(map[string]expander.Option, 0)

	if processors != nil && processors.NodeGroupListProcessor != nil {
		var errProc error
		nodeGroups, nodeInfos, errProc = processors.NodeGroupListProcessor.Process(context, nodeGroups, nodeInfos, unschedulablePods)
		if errProc != nil {
			return &status.ScaleUpStatus{Result: status.ScaleUpError}, errors.ToAutoscalerError(errors.InternalError, errProc)
		}
	}

	podEquivalenceGroups := buildPodEquivalenceGroups(unschedulablePods)

	skippedNodeGroups := map[string]status.Reasons{}
	for _, nodeGroup := range nodeGroups {
		// Autoprovisioned node groups without nodes are created later so skip check for them.
		if nodeGroup.Exist() && !clusterStateRegistry.IsNodeGroupSafeToScaleUp(nodeGroup, now) {
			// Hack that depends on internals of IsNodeGroupSafeToScaleUp.
			if !clusterStateRegistry.IsNodeGroupHealthy(nodeGroup.Id()) {
				klog.Warningf("Node group %s is not ready for scaleup - unhealthy", nodeGroup.Id())
				skippedNodeGroups[nodeGroup.Id()] = notReadyReason
			} else {
				klog.Warningf("Node group %s is not ready for scaleup - backoff", nodeGroup.Id())
				skippedNodeGroups[nodeGroup.Id()] = backoffReason
			}
			continue
		}

		currentTargetSize, err := nodeGroup.TargetSize()
		if err != nil {
			klog.Errorf("Failed to get node group size: %v", err)
			skippedNodeGroups[nodeGroup.Id()] = notReadyReason
			continue
		}
		if currentTargetSize >= nodeGroup.MaxSize() {
			klog.V(4).Infof("Skipping node group %s - max size reached", nodeGroup.Id())
			skippedNodeGroups[nodeGroup.Id()] = maxLimitReachedReason
			continue
		}

		nodeInfo, found := nodeInfos[nodeGroup.Id()]
		if !found {
			klog.Errorf("No node info for: %s", nodeGroup.Id())
			skippedNodeGroups[nodeGroup.Id()] = notReadyReason
			continue
		}

		scaleUpResourcesDelta, err := computeScaleUpResourcesDelta(context.CloudProvider, nodeInfo, nodeGroup, resourceLimiter)
		if err != nil {
			klog.Errorf("Skipping node group %s; error getting node group resources: %v", nodeGroup.Id(), err)
			skippedNodeGroups[nodeGroup.Id()] = notReadyReason
			continue
		}
		checkResult := scaleUpResourcesLeft.checkScaleUpDeltaWithinLimits(scaleUpResourcesDelta)
		if checkResult.exceeded {
			klog.V(4).Infof("Skipping node group %s; maximal limit exceeded for %v", nodeGroup.Id(), checkResult.exceededResources)
			skippedNodeGroups[nodeGroup.Id()] = maxResourceLimitReached(checkResult.exceededResources)
			continue
		}

		option, err := computeExpansionOption(context, podEquivalenceGroups, nodeGroup, nodeInfo, upcomingNodes)
		if err != nil {
			return &status.ScaleUpStatus{Result: status.ScaleUpError}, errors.ToAutoscalerError(errors.InternalError, err)
		}

		if len(option.Pods) > 0 {
			if option.NodeCount > 0 {
				expansionOptions[nodeGroup.Id()] = option
			} else {
				klog.V(4).Infof("No pod can fit to %s", nodeGroup.Id())
			}
		} else {
			klog.V(4).Infof("No pod can fit to %s", nodeGroup.Id())
		}
	}
	if len(expansionOptions) == 0 {
		klog.V(1).Info("No expansion options")
		return &status.ScaleUpStatus{
			Result:                  status.ScaleUpNoOptionsAvailable,
			PodsRemainUnschedulable: getRemainingPods(podEquivalenceGroups, skippedNodeGroups),
			ConsideredNodeGroups:    nodeGroups,
		}, nil
	}

	// Pick some expansion option.
	options := make([]expander.Option, 0, len(expansionOptions))
	for _, o := range expansionOptions {
		options = append(options, o)
	}
	bestOption := context.ExpanderStrategy.BestOption(options, nodeInfos)
	if bestOption != nil && bestOption.NodeCount > 0 {
		klog.V(1).Infof("Best option to resize: %s", bestOption.NodeGroup.Id())
		if len(bestOption.Debug) > 0 {
			klog.V(1).Info(bestOption.Debug)
		}
		klog.V(1).Infof("Estimated %d nodes needed in %s", bestOption.NodeCount, bestOption.NodeGroup.Id())

		newNodes := bestOption.NodeCount

		if context.MaxNodesTotal > 0 && len(nodes)+newNodes+len(upcomingNodes) > context.MaxNodesTotal {
			klog.V(1).Infof("Capping size to max cluster total size (%d)", context.MaxNodesTotal)
			newNodes = context.MaxNodesTotal - len(nodes) - len(upcomingNodes)
			context.LogRecorder.Eventf(apiv1.EventTypeWarning, "MaxNodesTotalReached", "Max total nodes in cluster reached: %v", context.MaxNodesTotal)
			if newNodes < 1 {
				return &status.ScaleUpStatus{Result: status.ScaleUpError}, errors.NewAutoscalerError(
					errors.TransientError,
					"max node total count already reached")
			}
		}

		createNodeGroupResults := make([]nodegroups.CreateNodeGroupResult, 0)
		if !bestOption.NodeGroup.Exist() {
			oldId := bestOption.NodeGroup.Id()
			createNodeGroupResult, err := processors.NodeGroupManager.CreateNodeGroup(context, bestOption.NodeGroup)
			if err != nil {
				return &status.ScaleUpStatus{Result: status.ScaleUpError}, err
			}
			createNodeGroupResults = append(createNodeGroupResults, createNodeGroupResult)
			bestOption.NodeGroup = createNodeGroupResult.MainCreatedNodeGroup

			// If possible replace candidate node-info with node info based on crated node group. The latter
			// one should be more in line with nodes which will be created by node group.
			mainCreatedNodeInfo, err := utils.GetNodeInfoFromTemplate(createNodeGroupResult.MainCreatedNodeGroup, daemonSets, context.PredicateChecker, ignoredTaints)
			if err == nil {
				nodeInfos[createNodeGroupResult.MainCreatedNodeGroup.Id()] = mainCreatedNodeInfo
			} else {
				klog.Warningf("Cannot build node info for newly created main node group %v; balancing similar node groups may not work; err=%v", createNodeGroupResult.MainCreatedNodeGroup.Id(), err)
				// Use node info based on expansion candidate but upadte Id which likely changed when node group was created.
				nodeInfos[bestOption.NodeGroup.Id()] = nodeInfos[oldId]
			}

			if oldId != createNodeGroupResult.MainCreatedNodeGroup.Id() {
				delete(nodeInfos, oldId)
			}

			for _, nodeGroup := range createNodeGroupResult.ExtraCreatedNodeGroups {
				nodeInfo, err := utils.GetNodeInfoFromTemplate(nodeGroup, daemonSets, context.PredicateChecker, ignoredTaints)

				if err != nil {
					klog.Warningf("Cannot build node info for newly created extra node group %v; balancing similar node groups will not work; err=%v", nodeGroup.Id(), err)
					continue
				}
				nodeInfos[nodeGroup.Id()] = nodeInfo

				option, err2 := computeExpansionOption(context, podEquivalenceGroups, nodeGroup, nodeInfo, upcomingNodes)
				if err2 != nil {
					return &status.ScaleUpStatus{Result: status.ScaleUpError}, errors.ToAutoscalerError(errors.InternalError, err)
				}

				if len(option.Pods) > 0 && option.NodeCount > 0 {
					expansionOptions[nodeGroup.Id()] = option
				}
			}

			// Update ClusterStateRegistry so similar nodegroups rebalancing works.
			// TODO(lukaszos) when pursuing scalability update this call with one which takes list of changed node groups so we do not
			//                do extra API calls. (the call at the bottom of ScaleUp() could be also changed then)
			clusterStateRegistry.Recalculate()
		}

		nodeInfo, found := nodeInfos[bestOption.NodeGroup.Id()]
		if !found {
			// This should never happen, as we already should have retrieved
			// nodeInfo for any considered nodegroup.
			klog.Errorf("No node info for: %s", bestOption.NodeGroup.Id())
			return &status.ScaleUpStatus{Result: status.ScaleUpError, CreateNodeGroupResults: createNodeGroupResults}, errors.NewAutoscalerError(
				errors.CloudProviderError,
				"No node info for best expansion option!")
		}

		// apply upper limits for CPU and memory
		newNodes, err = applyScaleUpResourcesLimits(context.CloudProvider, newNodes, scaleUpResourcesLeft, nodeInfo, bestOption.NodeGroup, resourceLimiter)
		if err != nil {
			return &status.ScaleUpStatus{Result: status.ScaleUpError, CreateNodeGroupResults: createNodeGroupResults}, err
		}

		targetNodeGroups := []cloudprovider.NodeGroup{bestOption.NodeGroup}
		if context.BalanceSimilarNodeGroups {
			similarNodeGroups, typedErr := processors.NodeGroupSetProcessor.FindSimilarNodeGroups(context, bestOption.NodeGroup, nodeInfos)
			if typedErr != nil {
				return &status.ScaleUpStatus{Result: status.ScaleUpError, CreateNodeGroupResults: createNodeGroupResults}, typedErr.AddPrefix("Failed to find matching node groups: ")
			}
			similarNodeGroups = filterNodeGroupsByPods(similarNodeGroups, bestOption.Pods, expansionOptions)
			for _, ng := range similarNodeGroups {
				if clusterStateRegistry.IsNodeGroupSafeToScaleUp(ng, now) {
					targetNodeGroups = append(targetNodeGroups, ng)
				} else {
					// This should never happen, as we will filter out the node group earlier on
					// because of missing entry in podsPassingPredicates, but double checking doesn't
					// really cost us anything
					klog.V(2).Infof("Ignoring node group %s when balancing: group is not ready for scaleup", ng.Id())
				}
			}
			if len(targetNodeGroups) > 1 {
				var buffer bytes.Buffer
				for i, ng := range targetNodeGroups {
					if i > 0 {
						buffer.WriteString(", ")
					}
					buffer.WriteString(ng.Id())
				}
				klog.V(1).Infof("Splitting scale-up between %v similar node groups: {%v}", len(targetNodeGroups), buffer.String())
			}
		}
		scaleUpInfos, typedErr := processors.NodeGroupSetProcessor.BalanceScaleUpBetweenGroups(
			context, targetNodeGroups, newNodes)
		if typedErr != nil {
			return &status.ScaleUpStatus{Result: status.ScaleUpError, CreateNodeGroupResults: createNodeGroupResults}, typedErr
		}
		klog.V(1).Infof("Final scale-up plan: %v", scaleUpInfos)
		for _, info := range scaleUpInfos {
			typedErr := executeScaleUp(context, clusterStateRegistry, info, gpu.GetGpuTypeForMetrics(gpuLabel, availableGPUTypes, nodeInfo.Node(), nil), now)
			if typedErr != nil {
				return &status.ScaleUpStatus{Result: status.ScaleUpError, CreateNodeGroupResults: createNodeGroupResults}, typedErr
			}
		}

		clusterStateRegistry.Recalculate()
		return &status.ScaleUpStatus{
			Result:                  status.ScaleUpSuccessful,
			ScaleUpInfos:            scaleUpInfos,
			PodsRemainUnschedulable: getRemainingPods(podEquivalenceGroups, skippedNodeGroups),
			ConsideredNodeGroups:    nodeGroups,
			CreateNodeGroupResults:  createNodeGroupResults,
			PodsTriggeredScaleUp:    bestOption.Pods,
			PodsAwaitEvaluation:     getPodsAwaitingEvaluation(podEquivalenceGroups, bestOption.NodeGroup.Id()),
		}, nil
	}

	return &status.ScaleUpStatus{
		Result:                  status.ScaleUpNoOptionsAvailable,
		PodsRemainUnschedulable: getRemainingPods(podEquivalenceGroups, skippedNodeGroups),
		ConsideredNodeGroups:    nodeGroups,
	}, nil
}

func getRemainingPods(egs []*podEquivalenceGroup, skipped map[string]status.Reasons) []status.NoScaleUpInfo {
	remaining := []status.NoScaleUpInfo{}
	for _, eg := range egs {
		if eg.schedulable {
			continue
		}
		for _, pod := range eg.pods {
			noScaleUpInfo := status.NoScaleUpInfo{
				Pod:                pod,
				RejectedNodeGroups: eg.schedulingErrors,
				SkippedNodeGroups:  skipped,
			}
			remaining = append(remaining, noScaleUpInfo)
		}
	}
	return remaining
}

func getPodsAwaitingEvaluation(egs []*podEquivalenceGroup, bestOption string) []*apiv1.Pod {
	awaitsEvaluation := []*apiv1.Pod{}
	for _, eg := range egs {
		if eg.schedulable {
			if _, found := eg.schedulingErrors[bestOption]; found {
				// Schedulable, but not yet.
				awaitsEvaluation = append(awaitsEvaluation, eg.pods...)
			}
		}
	}
	return awaitsEvaluation
}

func filterNodeGroupsByPods(
	groups []cloudprovider.NodeGroup,
	podsRequiredToFit []*apiv1.Pod,
	expansionOptions map[string]expander.Option) []cloudprovider.NodeGroup {

	result := make([]cloudprovider.NodeGroup, 0)

groupsloop:
	for _, group := range groups {
		option, found := expansionOptions[group.Id()]
		if !found {
			klog.V(1).Infof("No info about pods passing predicates found for group %v, skipping it from scale-up consideration", group.Id())
			continue
		}
		fittingPods := option.Pods
		podSet := make(map[*apiv1.Pod]bool, len(fittingPods))
		for _, pod := range fittingPods {
			podSet[pod] = true
		}
		for _, pod := range podsRequiredToFit {
			if _, found := podSet[pod]; !found {
				klog.V(1).Infof("Group %v, can't fit pod %v/%v, removing from scale-up consideration", group.Id(), pod.Namespace, pod.Name)
				continue groupsloop
			}
		}
		result = append(result, group)
	}
	return result
}

func executeScaleUp(context *context.AutoscalingContext, clusterStateRegistry *clusterstate.ClusterStateRegistry, info nodegroupset.ScaleUpInfo, gpuType string, now time.Time) errors.AutoscalerError {
	klog.V(0).Infof("Scale-up: setting group %s size to %d", info.Group.Id(), info.NewSize)
	context.LogRecorder.Eventf(apiv1.EventTypeNormal, "ScaledUpGroup",
		"Scale-up: setting group %s size to %d", info.Group.Id(), info.NewSize)
	increase := info.NewSize - info.CurrentSize
	if err := info.Group.IncreaseSize(increase); err != nil {
		context.LogRecorder.Eventf(apiv1.EventTypeWarning, "FailedToScaleUpGroup", "Scale-up failed for group %s: %v", info.Group.Id(), err)
		clusterStateRegistry.RegisterFailedScaleUp(info.Group, metrics.CloudProviderError, now)
		return errors.NewAutoscalerError(errors.CloudProviderError,
			"failed to increase node group size: %v", err)
	}
	clusterStateRegistry.RegisterOrUpdateScaleUp(
		info.Group,
		increase,
		time.Now())
	metrics.RegisterScaleUp(increase, gpuType)
	context.LogRecorder.Eventf(apiv1.EventTypeNormal, "ScaledUpGroup",
		"Scale-up: group %s size set to %d", info.Group.Id(), info.NewSize)
	return nil
}

func applyScaleUpResourcesLimits(
	cp cloudprovider.CloudProvider,
	newNodes int,
	scaleUpResourcesLeft scaleUpResourcesLimits,
	nodeInfo *schedulerframework.NodeInfo,
	nodeGroup cloudprovider.NodeGroup,
	resourceLimiter *cloudprovider.ResourceLimiter) (int, errors.AutoscalerError) {

	delta, err := computeScaleUpResourcesDelta(cp, nodeInfo, nodeGroup, resourceLimiter)
	if err != nil {
		return 0, err
	}

	for resource, resourceDelta := range delta {
		limit, limitFound := scaleUpResourcesLeft[resource]
		if !limitFound {
			continue
		}
		if limit == scaleUpLimitUnknown {
			// should never happen - checked before
			return 0, errors.NewAutoscalerError(
				errors.InternalError,
				fmt.Sprintf("limit unknown for resource %s", resource))
		}
		if int64(newNodes)*resourceDelta <= limit {
			// no capping required
			continue
		}

		newNodes = int(limit / resourceDelta)
		klog.V(1).Infof("Capping scale-up size due to limit for resource %s", resource)
		if newNodes < 1 {
			// should never happen - checked before
			return 0, errors.NewAutoscalerError(
				errors.InternalError,
				fmt.Sprintf("cannot create any node; max limit for resource %s reached", resource))
		}
	}
	return newNodes, nil
}