cluster_size_autoscaling.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 autoscaling

import (
	"context"
	"fmt"
	"io"
	"math"
	"net/http"
	"os"
	"os/exec"
	"regexp"
	"strconv"
	"strings"
	"time"

	v1 "k8s.io/api/core/v1"
	policyv1 "k8s.io/api/policy/v1"
	schedulingv1 "k8s.io/api/scheduling/v1"
	apierrors "k8s.io/apimachinery/pkg/api/errors"
	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
	"k8s.io/apimachinery/pkg/fields"
	"k8s.io/apimachinery/pkg/labels"
	utilerrors "k8s.io/apimachinery/pkg/util/errors"
	"k8s.io/apimachinery/pkg/util/intstr"
	"k8s.io/apimachinery/pkg/util/sets"
	"k8s.io/apimachinery/pkg/util/uuid"
	"k8s.io/apimachinery/pkg/util/wait"
	clientset "k8s.io/client-go/kubernetes"
	"k8s.io/klog/v2"
	"k8s.io/kubernetes/test/e2e/feature"
	"k8s.io/kubernetes/test/e2e/framework"
	e2ekubectl "k8s.io/kubernetes/test/e2e/framework/kubectl"
	e2emanifest "k8s.io/kubernetes/test/e2e/framework/manifest"
	e2enetwork "k8s.io/kubernetes/test/e2e/framework/network"
	e2enode "k8s.io/kubernetes/test/e2e/framework/node"
	e2epv "k8s.io/kubernetes/test/e2e/framework/pv"
	e2erc "k8s.io/kubernetes/test/e2e/framework/rc"
	e2eskipper "k8s.io/kubernetes/test/e2e/framework/skipper"
	"k8s.io/kubernetes/test/e2e/scheduling"
	testutils "k8s.io/kubernetes/test/utils"
	imageutils "k8s.io/kubernetes/test/utils/image"
	admissionapi "k8s.io/pod-security-admission/api"

	"github.com/onsi/ginkgo/v2"
	"github.com/onsi/gomega"
)

const (
	defaultTimeout         = 3 * time.Minute
	resizeTimeout          = 5 * time.Minute
	manualResizeTimeout    = 6 * time.Minute
	scaleUpTimeout         = 5 * time.Minute
	scaleUpTriggerTimeout  = 2 * time.Minute
	scaleDownTimeout       = 20 * time.Minute
	podTimeout             = 2 * time.Minute
	nodesRecoverTimeout    = 5 * time.Minute
	rcCreationRetryTimeout = 4 * time.Minute
	rcCreationRetryDelay   = 20 * time.Second
	makeSchedulableTimeout = 10 * time.Minute
	makeSchedulableDelay   = 20 * time.Second
	freshStatusLimit       = 20 * time.Second

	gkeUpdateTimeout   = 15 * time.Minute
	gkeNodepoolNameKey = "cloud.google.com/gke-nodepool"

	disabledTaint             = "DisabledForAutoscalingTest"
	criticalAddonsOnlyTaint   = "CriticalAddonsOnly"
	newNodesForScaledownTests = 2
	unhealthyClusterThreshold = 4

	caNoScaleUpStatus      = "NoActivity"
	caOngoingScaleUpStatus = "InProgress"
	timestampFormat        = "2006-01-02 15:04:05 -0700 MST"

	expendablePriorityClassName = "expendable-priority"
	highPriorityClassName       = "high-priority"

	gpuLabel = "cloud.google.com/gke-accelerator"

	nonExistingBypassedSchedulerName = "non-existing-bypassed-scheduler"
)

var _ = SIGDescribe("Cluster size autoscaling", framework.WithSlow(), func() {
	f := framework.NewDefaultFramework("autoscaling")
	f.NamespacePodSecurityLevel = admissionapi.LevelPrivileged
	var c clientset.Interface
	var nodeCount int
	var memAllocatableMb int
	var originalSizes map[string]int

	ginkgo.BeforeEach(func(ctx context.Context) {
		c = f.ClientSet
		e2eskipper.SkipUnlessProviderIs("gce", "gke")

		originalSizes = make(map[string]int)
		sum := 0
		for _, mig := range strings.Split(framework.TestContext.CloudConfig.NodeInstanceGroup, ",") {
			size, err := framework.GroupSize(mig)
			framework.ExpectNoError(err)
			ginkgo.By(fmt.Sprintf("Initial size of %s: %d", mig, size))
			originalSizes[mig] = size
			sum += size
		}
		// Give instances time to spin up
		framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, c, sum, scaleUpTimeout))

		nodes, err := e2enode.GetReadySchedulableNodes(ctx, f.ClientSet)
		framework.ExpectNoError(err)
		nodeCount = len(nodes.Items)
		ginkgo.By(fmt.Sprintf("Initial number of schedulable nodes: %v", nodeCount))
		gomega.Expect(nodes.Items).ToNot(gomega.BeEmpty())
		mem := nodes.Items[0].Status.Allocatable[v1.ResourceMemory]
		memAllocatableMb = int((&mem).Value() / 1024 / 1024)

		gomega.Expect(nodeCount).To(gomega.Equal(sum))

		if framework.ProviderIs("gke") {
			val, err := isAutoscalerEnabled(5)
			framework.ExpectNoError(err)
			if !val {
				err = enableAutoscaler("default-pool", 3, 5)
				framework.ExpectNoError(err)
			}
		}
	})

	ginkgo.AfterEach(func(ctx context.Context) {
		e2eskipper.SkipUnlessProviderIs("gce", "gke")
		ginkgo.By(fmt.Sprintf("Restoring initial size of the cluster"))
		setMigSizes(originalSizes)
		expectedNodes := 0
		for _, size := range originalSizes {
			expectedNodes += size
		}
		framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, c, expectedNodes, scaleDownTimeout))
		nodes, err := c.CoreV1().Nodes().List(ctx, metav1.ListOptions{})
		framework.ExpectNoError(err)

		s := time.Now()
	makeSchedulableLoop:
		for start := time.Now(); time.Since(start) < makeSchedulableTimeout; time.Sleep(makeSchedulableDelay) {
			for _, n := range nodes.Items {
				err = makeNodeSchedulable(ctx, c, &n, true)
				switch err.(type) {
				case CriticalAddonsOnlyError:
					continue makeSchedulableLoop
				default:
					framework.ExpectNoError(err)
				}
			}
			break
		}
		klog.Infof("Made nodes schedulable again in %v", time.Since(s).String())
	})

	f.It("shouldn't increase cluster size if pending pod is too large", feature.ClusterSizeAutoscalingScaleUp, func(ctx context.Context) {
		ginkgo.By("Creating unschedulable pod")
		ReserveMemory(ctx, f, "memory-reservation", 1, int(1.1*float64(memAllocatableMb)), false, defaultTimeout)
		ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, "memory-reservation")

		ginkgo.By("Waiting for scale up hoping it won't happen")
		// Verify that the appropriate event was generated
		eventFound := false
	EventsLoop:
		for start := time.Now(); time.Since(start) < scaleUpTimeout; time.Sleep(20 * time.Second) {
			ginkgo.By("Waiting for NotTriggerScaleUp event")
			events, err := f.ClientSet.CoreV1().Events(f.Namespace.Name).List(ctx, metav1.ListOptions{})
			framework.ExpectNoError(err)

			for _, e := range events.Items {
				if e.InvolvedObject.Kind == "Pod" && e.Reason == "NotTriggerScaleUp" {
					ginkgo.By("NotTriggerScaleUp event found")
					eventFound = true
					break EventsLoop
				}
			}
		}
		if !eventFound {
			framework.Failf("Expected event with kind 'Pod' and reason 'NotTriggerScaleUp' not found.")
		}
		// Verify that cluster size is not changed
		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size <= nodeCount }, time.Second))
	})

	simpleScaleUpTest := func(ctx context.Context, unready int) {
		ReserveMemory(ctx, f, "memory-reservation", 100, nodeCount*memAllocatableMb, false, 1*time.Second)
		ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, "memory-reservation")

		// Verify that cluster size is increased
		framework.ExpectNoError(WaitForClusterSizeFuncWithUnready(ctx, f.ClientSet,
			func(size int) bool { return size >= nodeCount+1 }, scaleUpTimeout, unready))
		framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, c))
	}

	f.It("should increase cluster size if pending pods are small", feature.ClusterSizeAutoscalingScaleUp, func(ctx context.Context) {
		simpleScaleUpTest(ctx, 0)
	})

	gpuType := os.Getenv("TESTED_GPU_TYPE")

	f.It(fmt.Sprintf("Should scale up GPU pool from 0 [GpuType:%s]", gpuType), feature.ClusterSizeAutoscalingGpu, func(ctx context.Context) {
		e2eskipper.SkipUnlessProviderIs("gke")
		if gpuType == "" {
			framework.Failf("TEST_GPU_TYPE not defined")
			return
		}

		const gpuPoolName = "gpu-pool"
		addGpuNodePool(gpuPoolName, gpuType, 1, 0)
		defer deleteNodePool(gpuPoolName)

		installNvidiaDriversDaemonSet(ctx, f)

		ginkgo.By("Enable autoscaler")
		framework.ExpectNoError(enableAutoscaler(gpuPoolName, 0, 1))
		defer disableAutoscaler(gpuPoolName, 0, 1)
		gomega.Expect(getPoolNodes(ctx, f, gpuPoolName)).To(gomega.BeEmpty())

		ginkgo.By("Schedule a pod which requires GPU")
		framework.ExpectNoError(ScheduleAnySingleGpuPod(ctx, f, "gpu-pod-rc"))
		ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, "gpu-pod-rc")

		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size == nodeCount+1 }, scaleUpTimeout))
		gomega.Expect(getPoolNodes(ctx, f, gpuPoolName)).To(gomega.HaveLen(1))
	})

	f.It(fmt.Sprintf("Should scale up GPU pool from 1 [GpuType:%s]", gpuType), feature.ClusterSizeAutoscalingGpu, func(ctx context.Context) {
		e2eskipper.SkipUnlessProviderIs("gke")
		if gpuType == "" {
			framework.Failf("TEST_GPU_TYPE not defined")
			return
		}

		const gpuPoolName = "gpu-pool"
		addGpuNodePool(gpuPoolName, gpuType, 1, 1)
		defer deleteNodePool(gpuPoolName)

		installNvidiaDriversDaemonSet(ctx, f)

		ginkgo.By("Schedule a single pod which requires GPU")
		framework.ExpectNoError(ScheduleAnySingleGpuPod(ctx, f, "gpu-pod-rc"))
		ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, "gpu-pod-rc")

		ginkgo.By("Enable autoscaler")
		framework.ExpectNoError(enableAutoscaler(gpuPoolName, 0, 2))
		defer disableAutoscaler(gpuPoolName, 0, 2)
		gomega.Expect(getPoolNodes(ctx, f, gpuPoolName)).To(gomega.HaveLen(1))

		ginkgo.By("Scale GPU deployment")
		e2erc.ScaleRC(ctx, f.ClientSet, f.ScalesGetter, f.Namespace.Name, "gpu-pod-rc", 2, true)

		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size == nodeCount+2 }, scaleUpTimeout))
		gomega.Expect(getPoolNodes(ctx, f, gpuPoolName)).To(gomega.HaveLen(2))
	})

	f.It(fmt.Sprintf("Should not scale GPU pool up if pod does not require GPUs [GpuType:%s]", gpuType), feature.ClusterSizeAutoscalingGpu, func(ctx context.Context) {
		e2eskipper.SkipUnlessProviderIs("gke")
		if gpuType == "" {
			framework.Failf("TEST_GPU_TYPE not defined")
			return
		}

		const gpuPoolName = "gpu-pool"
		addGpuNodePool(gpuPoolName, gpuType, 1, 0)
		defer deleteNodePool(gpuPoolName)

		installNvidiaDriversDaemonSet(ctx, f)

		ginkgo.By("Enable autoscaler")
		framework.ExpectNoError(enableAutoscaler(gpuPoolName, 0, 1))
		defer disableAutoscaler(gpuPoolName, 0, 1)
		gomega.Expect(getPoolNodes(ctx, f, gpuPoolName)).To(gomega.BeEmpty())

		ginkgo.By("Schedule bunch of pods beyond point of filling default pool but do not request any GPUs")
		ReserveMemory(ctx, f, "memory-reservation", 100, nodeCount*memAllocatableMb, false, 1*time.Second)
		ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, "memory-reservation")
		// Verify that cluster size is increased
		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size >= nodeCount+1 }, scaleUpTimeout))

		// Expect gpu pool to stay intact
		gomega.Expect(getPoolNodes(ctx, f, gpuPoolName)).To(gomega.BeEmpty())
	})

	f.It(fmt.Sprintf("Should scale down GPU pool from 1 [GpuType:%s]", gpuType), feature.ClusterSizeAutoscalingGpu, func(ctx context.Context) {
		e2eskipper.SkipUnlessProviderIs("gke")
		if gpuType == "" {
			framework.Failf("TEST_GPU_TYPE not defined")
			return
		}

		const gpuPoolName = "gpu-pool"
		addGpuNodePool(gpuPoolName, gpuType, 1, 1)
		defer deleteNodePool(gpuPoolName)

		installNvidiaDriversDaemonSet(ctx, f)

		ginkgo.By("Schedule a single pod which requires GPU")
		framework.ExpectNoError(ScheduleAnySingleGpuPod(ctx, f, "gpu-pod-rc"))
		ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, "gpu-pod-rc")

		ginkgo.By("Enable autoscaler")
		framework.ExpectNoError(enableAutoscaler(gpuPoolName, 0, 1))
		defer disableAutoscaler(gpuPoolName, 0, 1)
		gomega.Expect(getPoolNodes(ctx, f, gpuPoolName)).To(gomega.HaveLen(1))

		ginkgo.By("Remove the only POD requiring GPU")
		e2erc.DeleteRCAndWaitForGC(ctx, f.ClientSet, f.Namespace.Name, "gpu-pod-rc")

		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size == nodeCount }, scaleDownTimeout))
		gomega.Expect(getPoolNodes(ctx, f, gpuPoolName)).To(gomega.BeEmpty())
	})

	f.It("should increase cluster size if pending pods are small and one node is broken", feature.ClusterSizeAutoscalingScaleUp, func(ctx context.Context) {
		e2enetwork.TestUnderTemporaryNetworkFailure(ctx, c, "default", getAnyNode(ctx, c), func(ctx context.Context) { simpleScaleUpTest(ctx, 1) })
	})

	f.It("shouldn't trigger additional scale-ups during processing scale-up", feature.ClusterSizeAutoscalingScaleUp, func(ctx context.Context) {
		// Wait for the situation to stabilize - CA should be running and have up-to-date node readiness info.
		status, err := waitForScaleUpStatus(ctx, c, func(s *scaleUpStatus) bool {
			return s.ready == s.target && s.ready <= nodeCount
		}, scaleUpTriggerTimeout)
		framework.ExpectNoError(err)

		unmanagedNodes := nodeCount - status.ready

		ginkgo.By("Schedule more pods than can fit and wait for cluster to scale-up")
		ReserveMemory(ctx, f, "memory-reservation", 100, nodeCount*memAllocatableMb, false, 1*time.Second)
		ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, "memory-reservation")

		status, err = waitForScaleUpStatus(ctx, c, func(s *scaleUpStatus) bool {
			return s.status == caOngoingScaleUpStatus
		}, scaleUpTriggerTimeout)
		framework.ExpectNoError(err)
		target := status.target
		framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, c))

		ginkgo.By("Expect no more scale-up to be happening after all pods are scheduled")

		// wait for a while until scale-up finishes; we cannot read CA status immediately
		// after pods are scheduled as status config map is updated by CA once every loop iteration
		status, err = waitForScaleUpStatus(ctx, c, func(s *scaleUpStatus) bool {
			return s.status == caNoScaleUpStatus
		}, 2*freshStatusLimit)
		framework.ExpectNoError(err)

		if status.target != target {
			klog.Warningf("Final number of nodes (%v) does not match initial scale-up target (%v).", status.target, target)
		}
		gomega.Expect(status.timestamp.Add(freshStatusLimit)).To(gomega.BeTemporally(">=", time.Now()))
		gomega.Expect(status.status).To(gomega.Equal(caNoScaleUpStatus))
		gomega.Expect(status.ready).To(gomega.Equal(status.target))
		nodes, err := e2enode.GetReadySchedulableNodes(ctx, f.ClientSet)
		framework.ExpectNoError(err)
		gomega.Expect(nodes.Items).To(gomega.HaveLen(status.target + unmanagedNodes))
	})

	f.It("should increase cluster size if pending pods are small and there is another node pool that is not autoscaled", feature.ClusterSizeAutoscalingScaleUp, func(ctx context.Context) {
		e2eskipper.SkipUnlessProviderIs("gke")

		ginkgo.By("Creating new node-pool with e2-standard-4 machines")
		const extraPoolName = "extra-pool"
		addNodePool(extraPoolName, "e2-standard-4", 1)
		defer deleteNodePool(extraPoolName)
		extraNodes := getPoolInitialSize(extraPoolName)
		framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, c, nodeCount+extraNodes, resizeTimeout))
		// We wait for nodes to become schedulable to make sure the new nodes
		// will be returned by getPoolNodes below.
		framework.ExpectNoError(e2enode.WaitForAllNodesSchedulable(ctx, c, resizeTimeout))
		klog.Infof("Not enabling cluster autoscaler for the node pool (on purpose).")

		ginkgo.By("Getting memory available on new nodes, so we can account for it when creating RC")
		nodes := getPoolNodes(ctx, f, extraPoolName)
		gomega.Expect(nodes).To(gomega.HaveLen(extraNodes))
		extraMemMb := 0
		for _, node := range nodes {
			mem := node.Status.Allocatable[v1.ResourceMemory]
			extraMemMb += int((&mem).Value() / 1024 / 1024)
		}

		ginkgo.By("Reserving 0.1x more memory than the cluster holds to trigger scale up")
		totalMemoryReservation := int(1.1 * float64(nodeCount*memAllocatableMb+extraMemMb))
		ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, "memory-reservation")
		ReserveMemory(ctx, f, "memory-reservation", 100, totalMemoryReservation, false, defaultTimeout)

		// Verify, that cluster size is increased
		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size >= nodeCount+extraNodes+1 }, scaleUpTimeout))
		framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, c))
	})

	f.It("should disable node pool autoscaling", feature.ClusterSizeAutoscalingScaleUp, func(ctx context.Context) {
		e2eskipper.SkipUnlessProviderIs("gke")

		ginkgo.By("Creating new node-pool with e2-standard-4 machines")
		const extraPoolName = "extra-pool"
		addNodePool(extraPoolName, "e2-standard-4", 1)
		defer deleteNodePool(extraPoolName)
		extraNodes := getPoolInitialSize(extraPoolName)
		framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, c, nodeCount+extraNodes, resizeTimeout))
		framework.ExpectNoError(enableAutoscaler(extraPoolName, 1, 2))
		framework.ExpectNoError(disableAutoscaler(extraPoolName, 1, 2))
	})

	f.It("should increase cluster size if pods are pending due to host port conflict", feature.ClusterSizeAutoscalingScaleUp, func(ctx context.Context) {
		scheduling.CreateHostPortPods(ctx, f, "host-port", nodeCount+2, false)
		ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, "host-port")

		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size >= nodeCount+2 }, scaleUpTimeout))
		framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, c))
	})

	f.It("should increase cluster size if pods are pending due to pod anti-affinity", feature.ClusterSizeAutoscalingScaleUp, func(ctx context.Context) {
		pods := nodeCount
		newPods := 2
		labels := map[string]string{
			"anti-affinity": "yes",
		}
		ginkgo.By("starting a pod with anti-affinity on each node")
		framework.ExpectNoError(runAntiAffinityPods(ctx, f, f.Namespace.Name, pods, "some-pod", labels, labels))
		ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, "some-pod")
		framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, c))

		ginkgo.By("scheduling extra pods with anti-affinity to existing ones")
		framework.ExpectNoError(runAntiAffinityPods(ctx, f, f.Namespace.Name, newPods, "extra-pod", labels, labels))
		ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, "extra-pod")

		framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, c))
		framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, c, nodeCount+newPods, scaleUpTimeout))
	})

	f.It("should increase cluster size if pod requesting EmptyDir volume is pending", feature.ClusterSizeAutoscalingScaleUp, func(ctx context.Context) {
		ginkgo.By("creating pods")
		pods := nodeCount
		newPods := 1
		labels := map[string]string{
			"anti-affinity": "yes",
		}
		framework.ExpectNoError(runAntiAffinityPods(ctx, f, f.Namespace.Name, pods, "some-pod", labels, labels))
		ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, "some-pod")

		ginkgo.By("waiting for all pods before triggering scale up")
		framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, c))

		ginkgo.By("creating a pod requesting EmptyDir")
		framework.ExpectNoError(runVolumeAntiAffinityPods(ctx, f, f.Namespace.Name, newPods, "extra-pod", labels, labels, emptyDirVolumes))
		ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, "extra-pod")

		framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, c))
		framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, c, nodeCount+newPods, scaleUpTimeout))
	})

	f.It("should increase cluster size if pod requesting volume is pending", feature.ClusterSizeAutoscalingScaleUp, func(ctx context.Context) {
		e2eskipper.SkipUnlessProviderIs("gce", "gke")

		volumeLabels := labels.Set{
			e2epv.VolumeSelectorKey: f.Namespace.Name,
		}
		selector := metav1.SetAsLabelSelector(volumeLabels)

		ginkgo.By("creating volume & pvc")
		diskName, err := e2epv.CreatePDWithRetry(ctx)
		framework.ExpectNoError(err)
		pvConfig := e2epv.PersistentVolumeConfig{
			NamePrefix: "gce-",
			Labels:     volumeLabels,
			PVSource: v1.PersistentVolumeSource{
				GCEPersistentDisk: &v1.GCEPersistentDiskVolumeSource{
					PDName:   diskName,
					FSType:   "ext3",
					ReadOnly: false,
				},
			},
			Prebind: nil,
		}
		emptyStorageClass := ""
		pvcConfig := e2epv.PersistentVolumeClaimConfig{
			Selector:         selector,
			StorageClassName: &emptyStorageClass,
		}

		pv, pvc, err := e2epv.CreatePVPVC(ctx, c, f.Timeouts, pvConfig, pvcConfig, f.Namespace.Name, false)
		framework.ExpectNoError(err)
		framework.ExpectNoError(e2epv.WaitOnPVandPVC(ctx, c, f.Timeouts, f.Namespace.Name, pv, pvc))

		defer func() {
			errs := e2epv.PVPVCCleanup(ctx, c, f.Namespace.Name, pv, pvc)
			if len(errs) > 0 {
				framework.Failf("failed to delete PVC and/or PV. Errors: %v", utilerrors.NewAggregate(errs))
			}
			pv, pvc = nil, nil
			if diskName != "" {
				framework.ExpectNoError(e2epv.DeletePDWithRetry(ctx, diskName))
			}
		}()

		ginkgo.By("creating pods")
		pods := nodeCount
		labels := map[string]string{
			"anti-affinity": "yes",
		}
		framework.ExpectNoError(runAntiAffinityPods(ctx, f, f.Namespace.Name, pods, "some-pod", labels, labels))
		ginkgo.DeferCleanup(func(ctx context.Context) {
			e2erc.DeleteRCAndWaitForGC(ctx, f.ClientSet, f.Namespace.Name, "some-pod")
			klog.Infof("RC and pods not using volume deleted")
		})

		ginkgo.By("waiting for all pods before triggering scale up")
		framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, c))

		ginkgo.By("creating a pod requesting PVC")
		pvcPodName := "pvc-pod"
		newPods := 1
		volumes := buildVolumes(pv, pvc)
		framework.ExpectNoError(runVolumeAntiAffinityPods(ctx, f, f.Namespace.Name, newPods, pvcPodName, labels, labels, volumes))
		ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, pvcPodName)
		ginkgo.DeferCleanup(waitForAllCaPodsReadyInNamespace, f, c)

		framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, c))
		framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, c, nodeCount+newPods, scaleUpTimeout))
	})

	f.It("should add node to the particular mig", feature.ClusterSizeAutoscalingScaleUp, func(ctx context.Context) {
		labelKey := "cluster-autoscaling-test.special-node"
		labelValue := "true"

		ginkgo.By("Finding the smallest MIG")
		minMig := ""
		minSize := nodeCount
		for mig, size := range originalSizes {
			if size <= minSize {
				minMig = mig
				minSize = size
			}
		}

		if minSize == 0 {
			newSizes := make(map[string]int)
			for mig, size := range originalSizes {
				newSizes[mig] = size
			}
			newSizes[minMig] = 1
			setMigSizes(newSizes)
		}

		removeLabels := func(nodesToClean sets.String) {
			ginkgo.By("Removing labels from nodes")
			for node := range nodesToClean {
				e2enode.RemoveLabelOffNode(c, node, labelKey)
			}
		}

		nodes, err := framework.GetGroupNodes(minMig)
		framework.ExpectNoError(err)
		nodesSet := sets.NewString(nodes...)
		defer removeLabels(nodesSet)
		ginkgo.By(fmt.Sprintf("Annotating nodes of the smallest MIG(%s): %v", minMig, nodes))

		for node := range nodesSet {
			e2enode.AddOrUpdateLabelOnNode(c, node, labelKey, labelValue)
		}

		err = scheduling.CreateNodeSelectorPods(ctx, f, "node-selector", minSize+1, map[string]string{labelKey: labelValue}, false)
		framework.ExpectNoError(err)
		ginkgo.By("Waiting for new node to appear and annotating it")
		framework.WaitForGroupSize(minMig, int32(minSize+1))
		// Verify that cluster size is increased
		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size >= nodeCount+1 }, scaleUpTimeout))

		newNodes, err := framework.GetGroupNodes(minMig)
		framework.ExpectNoError(err)
		newNodesSet := sets.NewString(newNodes...)
		newNodesSet.Delete(nodes...)
		if len(newNodesSet) > 1 {
			ginkgo.By(fmt.Sprintf("Spotted following new nodes in %s: %v", minMig, newNodesSet))
			klog.Infof("Usually only 1 new node is expected, investigating")
			klog.Infof("Kubectl:%s\n", e2ekubectl.RunKubectlOrDie(f.Namespace.Name, "get", "nodes", "-o", "json"))
			if output, err := exec.Command("gcloud", "compute", "instances", "list",
				"--project="+framework.TestContext.CloudConfig.ProjectID,
				"--zone="+framework.TestContext.CloudConfig.Zone).Output(); err == nil {
				klog.Infof("Gcloud compute instances list: %s", output)
			} else {
				klog.Errorf("Failed to get instances list: %v", err)
			}

			for newNode := range newNodesSet {
				if output, err := execCmd("gcloud", "compute", "instances", "describe",
					newNode,
					"--project="+framework.TestContext.CloudConfig.ProjectID,
					"--zone="+framework.TestContext.CloudConfig.Zone).Output(); err == nil {
					klog.Infof("Gcloud compute instances describe: %s", output)
				} else {
					klog.Errorf("Failed to get instances describe: %v", err)
				}
			}

			// TODO: possibly remove broken node from newNodesSet to prevent removeLabel from crashing.
			// However at this moment we DO WANT it to crash so that we don't check all test runs for the
			// rare behavior, but only the broken ones.
		}
		ginkgo.By(fmt.Sprintf("New nodes: %v\n", newNodesSet))
		registeredNodes := sets.NewString()
		for nodeName := range newNodesSet {
			node, err := f.ClientSet.CoreV1().Nodes().Get(ctx, nodeName, metav1.GetOptions{})
			if err == nil && node != nil {
				registeredNodes.Insert(nodeName)
			} else {
				klog.Errorf("Failed to get node %v: %v", nodeName, err)
			}
		}
		ginkgo.By(fmt.Sprintf("Setting labels for registered new nodes: %v", registeredNodes.List()))
		for node := range registeredNodes {
			e2enode.AddOrUpdateLabelOnNode(c, node, labelKey, labelValue)
		}

		defer removeLabels(registeredNodes)

		framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, c))
		framework.ExpectNoError(e2erc.DeleteRCAndWaitForGC(ctx, f.ClientSet, f.Namespace.Name, "node-selector"))
	})

	f.It("should scale up correct target pool", feature.ClusterSizeAutoscalingScaleUp, func(ctx context.Context) {
		e2eskipper.SkipUnlessProviderIs("gke")

		ginkgo.By("Creating new node-pool with e2-standard-4 machines")
		const extraPoolName = "extra-pool"
		addNodePool(extraPoolName, "e2-standard-4", 1)
		defer deleteNodePool(extraPoolName)
		extraNodes := getPoolInitialSize(extraPoolName)
		framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, c, nodeCount+extraNodes, resizeTimeout))
		framework.ExpectNoError(enableAutoscaler(extraPoolName, 1, 2))
		defer disableAutoscaler(extraPoolName, 1, 2)

		extraPods := extraNodes + 1
		totalMemoryReservation := int(float64(extraPods) * 1.5 * float64(memAllocatableMb))
		ginkgo.By(fmt.Sprintf("Creating rc with %v pods too big to fit default-pool but fitting extra-pool", extraPods))
		ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, "memory-reservation")
		ReserveMemory(ctx, f, "memory-reservation", extraPods, totalMemoryReservation, false, defaultTimeout)

		// Apparently GKE master is restarted couple minutes after the node pool is added
		// resetting all the timers in scale down code. Adding 5 extra minutes to workaround
		// this issue.
		// TODO: Remove the extra time when GKE restart is fixed.
		framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, c, nodeCount+extraNodes+1, scaleUpTimeout+5*time.Minute))
	})

	simpleScaleDownTest := func(ctx context.Context, unready int) {
		err := addKubeSystemPdbs(ctx, f)
		framework.ExpectNoError(err)

		ginkgo.By("Manually increase cluster size")
		increasedSize := 0
		newSizes := make(map[string]int)
		for key, val := range originalSizes {
			newSizes[key] = val + 2 + unready
			increasedSize += val + 2 + unready
		}
		setMigSizes(newSizes)
		framework.ExpectNoError(WaitForClusterSizeFuncWithUnready(ctx, f.ClientSet,
			func(size int) bool { return size >= increasedSize }, manualResizeTimeout, unready))

		ginkgo.By("Some node should be removed")
		framework.ExpectNoError(WaitForClusterSizeFuncWithUnready(ctx, f.ClientSet,
			func(size int) bool { return size < increasedSize }, scaleDownTimeout, unready))
	}

	f.It("should correctly scale down after a node is not needed", feature.ClusterSizeAutoscalingScaleDown,
		func(ctx context.Context) { simpleScaleDownTest(ctx, 0) })

	f.It("should correctly scale down after a node is not needed and one node is broken", feature.ClusterSizeAutoscalingScaleDown, func(ctx context.Context) {
		e2eskipper.SkipUnlessSSHKeyPresent()
		e2enetwork.TestUnderTemporaryNetworkFailure(ctx, c, "default", getAnyNode(ctx, c), func(ctx context.Context) { simpleScaleDownTest(ctx, 1) })
	})

	f.It("should correctly scale down after a node is not needed when there is non autoscaled pool", feature.ClusterSizeAutoscalingScaleDown, func(ctx context.Context) {
		e2eskipper.SkipUnlessProviderIs("gke")

		increasedSize := manuallyIncreaseClusterSize(ctx, f, originalSizes)

		const extraPoolName = "extra-pool"
		addNodePool(extraPoolName, "e2-standard-2", 3)
		defer deleteNodePool(extraPoolName)
		extraNodes := getPoolInitialSize(extraPoolName)

		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size >= increasedSize+extraNodes }, scaleUpTimeout))

		ginkgo.By("Some node should be removed")
		// Apparently GKE master is restarted couple minutes after the node pool is added
		// resetting all the timers in scale down code. Adding 10 extra minutes to workaround
		// this issue.
		// TODO: Remove the extra time when GKE restart is fixed.
		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size < increasedSize+extraNodes }, scaleDownTimeout+10*time.Minute))
	})

	f.It("should be able to scale down when rescheduling a pod is required and pdb allows for it", feature.ClusterSizeAutoscalingScaleDown, func(ctx context.Context) {
		runDrainTest(ctx, f, originalSizes, f.Namespace.Name, 1, 1, func(increasedSize int) {
			ginkgo.By("Some node should be removed")
			framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
				func(size int) bool { return size < increasedSize }, scaleDownTimeout))
		})
	})

	f.It("shouldn't be able to scale down when rescheduling a pod is required, but pdb doesn't allow drain", feature.ClusterSizeAutoscalingScaleDown, func(ctx context.Context) {
		runDrainTest(ctx, f, originalSizes, f.Namespace.Name, 1, 0, func(increasedSize int) {
			ginkgo.By("No nodes should be removed")
			time.Sleep(scaleDownTimeout)
			nodes, err := e2enode.GetReadySchedulableNodes(ctx, f.ClientSet)
			framework.ExpectNoError(err)
			gomega.Expect(nodes.Items).To(gomega.HaveLen(increasedSize))
		})
	})

	f.It("should be able to scale down by draining multiple pods one by one as dictated by pdb", feature.ClusterSizeAutoscalingScaleDown, func(ctx context.Context) {
		runDrainTest(ctx, f, originalSizes, f.Namespace.Name, 2, 1, func(increasedSize int) {
			ginkgo.By("Some node should be removed")
			framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
				func(size int) bool { return size < increasedSize }, scaleDownTimeout))
		})
	})

	f.It("should be able to scale down by draining system pods with pdb", feature.ClusterSizeAutoscalingScaleDown, func(ctx context.Context) {
		runDrainTest(ctx, f, originalSizes, "kube-system", 2, 1, func(increasedSize int) {
			ginkgo.By("Some node should be removed")
			framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
				func(size int) bool { return size < increasedSize }, scaleDownTimeout))
		})
	})

	f.It("Should be able to scale a node group up from 0", feature.ClusterSizeAutoscalingScaleUp, func(ctx context.Context) {
		// Provider-specific setup
		if framework.ProviderIs("gke") {
			// GKE-specific setup
			ginkgo.By("Add a new node pool with 0 nodes and min size 0")
			const extraPoolName = "extra-pool"
			addNodePool(extraPoolName, "e2-standard-4", 0)
			defer deleteNodePool(extraPoolName)
			framework.ExpectNoError(enableAutoscaler(extraPoolName, 0, 1))
			defer disableAutoscaler(extraPoolName, 0, 1)
		} else {
			// on GCE, run only if there are already at least 2 node groups
			e2eskipper.SkipUnlessAtLeast(len(originalSizes), 2, "At least 2 node groups are needed for scale-to-0 tests")

			ginkgo.By("Manually scale smallest node group to 0")
			minMig := ""
			minSize := nodeCount
			for mig, size := range originalSizes {
				if size <= minSize {
					minMig = mig
					minSize = size
				}
			}
			framework.ExpectNoError(framework.ResizeGroup(minMig, int32(0)))
			framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, c, nodeCount-minSize, resizeTimeout))
		}

		ginkgo.By("Make remaining nodes unschedulable")
		nodes, err := f.ClientSet.CoreV1().Nodes().List(ctx, metav1.ListOptions{FieldSelector: fields.Set{
			"spec.unschedulable": "false",
		}.AsSelector().String()})
		framework.ExpectNoError(err)

		for _, node := range nodes.Items {
			err = makeNodeUnschedulable(ctx, f.ClientSet, &node)

			n := node
			ginkgo.DeferCleanup(makeNodeSchedulable, f.ClientSet, &n, false)

			framework.ExpectNoError(err)
		}

		ginkgo.By("Run a scale-up test")
		ReserveMemory(ctx, f, "memory-reservation", 1, 100, false, 1*time.Second)
		ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, "memory-reservation")

		// Verify that cluster size is increased
		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size >= len(nodes.Items)+1 }, scaleUpTimeout))
		framework.ExpectNoError(waitForAllCaPodsReadyInNamespace(ctx, f, c))
	})

	// Scale to 0 test is split into two functions (for GKE & GCE.)
	// The reason for it is that scenario is exactly the same,
	// but setup & verification use different APIs.
	//
	// Scenario:
	// (GKE only) add an extra node pool with size 1 & enable autoscaling for it
	// (GCE only) find the smallest MIG & resize it to 1
	// manually drain the single node from this node pool/MIG
	// wait for cluster size to decrease
	// verify the targeted node pool/MIG is of size 0
	gkeScaleToZero := func(ctx context.Context) {
		// GKE-specific setup
		ginkgo.By("Add a new node pool with size 1 and min size 0")
		const extraPoolName = "extra-pool"
		addNodePool(extraPoolName, "e2-standard-4", 1)
		defer deleteNodePool(extraPoolName)
		extraNodes := getPoolInitialSize(extraPoolName)
		framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, c, nodeCount+extraNodes, resizeTimeout))
		framework.ExpectNoError(enableAutoscaler(extraPoolName, 0, 1))
		defer disableAutoscaler(extraPoolName, 0, 1)

		ngNodes := getPoolNodes(ctx, f, extraPoolName)
		gomega.Expect(ngNodes).To(gomega.HaveLen(extraNodes))
		for _, node := range ngNodes {
			ginkgo.By(fmt.Sprintf("Target node for scale-down: %s", node.Name))
		}

		for _, node := range ngNodes {
			drainNode(ctx, f, node)
		}
		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size <= nodeCount }, scaleDownTimeout))

		// GKE-specific check
		newSize := getPoolSize(ctx, f, extraPoolName)
		gomega.Expect(newSize).To(gomega.BeEmpty())
	}

	gceScaleToZero := func(ctx context.Context) {
		// non-GKE only
		ginkgo.By("Find smallest node group and manually scale it to a single node")
		minMig := ""
		minSize := nodeCount
		for mig, size := range originalSizes {
			if size <= minSize {
				minMig = mig
				minSize = size
			}
		}
		framework.ExpectNoError(framework.ResizeGroup(minMig, int32(1)))
		framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, c, nodeCount-minSize+1, resizeTimeout))
		ngNodes, err := framework.GetGroupNodes(minMig)
		framework.ExpectNoError(err)
		if len(ngNodes) != 1 {
			framework.Failf("Expected one node, got instead: %v", ngNodes)
		}
		node, err := f.ClientSet.CoreV1().Nodes().Get(ctx, ngNodes[0], metav1.GetOptions{})
		ginkgo.By(fmt.Sprintf("Target node for scale-down: %s", node.Name))
		framework.ExpectNoError(err)

		// this part is identical
		drainNode(ctx, f, node)
		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size < nodeCount-minSize+1 }, scaleDownTimeout))

		// non-GKE only
		newSize, err := framework.GroupSize(minMig)
		framework.ExpectNoError(err)
		gomega.Expect(newSize).To(gomega.BeEmpty())
	}

	f.It("Should be able to scale a node group down to 0", feature.ClusterSizeAutoscalingScaleDown, func(ctx context.Context) {
		if framework.ProviderIs("gke") { // In GKE, we can just add a node pool
			gkeScaleToZero(ctx)
		} else if len(originalSizes) >= 2 {
			gceScaleToZero(ctx)
		} else {
			e2eskipper.Skipf("At least 2 node groups are needed for scale-to-0 tests")
		}
	})

	f.It("Shouldn't perform scale up operation and should list unhealthy status if most of the cluster is broken", feature.ClusterSizeAutoscalingScaleUp, func(ctx context.Context) {
		e2eskipper.SkipUnlessSSHKeyPresent()

		clusterSize := nodeCount
		for clusterSize < unhealthyClusterThreshold+1 {
			clusterSize = manuallyIncreaseClusterSize(ctx, f, originalSizes)
		}

		// If new nodes are disconnected too soon, they'll be considered not started
		// instead of unready, and cluster won't be considered unhealthy.
		//
		// More precisely, Cluster Autoscaler will never consider a
		// node to be unhealthy unless it was created more than 15m
		// ago. Within that 15m window, it'll assume node is just
		// starting and not unhealthy.
		//
		// However, waiting for 15m would allow scale down to kick in
		// and remove recently added nodes, so here we just wait 2m for
		// nodes to come up (1m should be enough, another 1m added as
		// an extra buffer. Then, we break connectivity to a subset of
		// nodes and only after that we wait for 15m, since scale down
		// shouldn't happen when the cluster is unhealthy.
		time.Sleep(2 * time.Minute)

		ginkgo.By("Block network connectivity to some nodes to simulate unhealthy cluster")
		nodesToBreakCount := int(math.Ceil(math.Max(float64(unhealthyClusterThreshold), 0.5*float64(clusterSize))))
		nodes, err := f.ClientSet.CoreV1().Nodes().List(ctx, metav1.ListOptions{FieldSelector: fields.Set{
			"spec.unschedulable": "false",
		}.AsSelector().String()})
		framework.ExpectNoError(err)
		if nodesToBreakCount > len(nodes.Items) {
			framework.Failf("Expected at most %d nodes to break, got %d", len(nodes.Items), nodesToBreakCount)
		}
		nodesToBreak := nodes.Items[:nodesToBreakCount]

		// TestUnderTemporaryNetworkFailure only removes connectivity to a single node,
		// and accepts func() callback. This is expanding the loop to recursive call
		// to avoid duplicating TestUnderTemporaryNetworkFailure
		var testFunction func(ctx context.Context)
		testFunction = func(ctx context.Context) {
			if len(nodesToBreak) > 0 {
				ntb := &nodesToBreak[0]
				nodesToBreak = nodesToBreak[1:]
				e2enetwork.TestUnderTemporaryNetworkFailure(ctx, c, "default", ntb, testFunction)
			} else {
				ReserveMemory(ctx, f, "memory-reservation", 100, nodeCount*memAllocatableMb, false, defaultTimeout)
				ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, f.Namespace.Name, "memory-reservation")
				// Wait for 15m to ensure Cluster Autoscaler won't consider broken nodes as still starting.
				time.Sleep(15 * time.Minute)
				currentNodes, err := e2enode.GetReadySchedulableNodes(ctx, f.ClientSet)
				framework.ExpectNoError(err)
				framework.Logf("Currently available nodes: %v, nodes available at the start of test: %v, disabled nodes: %v", len(currentNodes.Items), len(nodes.Items), nodesToBreakCount)
				gomega.Expect(currentNodes.Items).To(gomega.HaveLen(len(nodes.Items) - nodesToBreakCount))
				status, err := getClusterwideStatus(ctx, c)
				framework.Logf("Clusterwide status: %v", status)
				framework.ExpectNoError(err)
				gomega.Expect(status).To(gomega.Equal("Unhealthy"))
			}
		}
		testFunction(ctx)
		// Give nodes time to recover from network failure
		framework.ExpectNoError(e2enode.WaitForReadyNodes(ctx, c, len(nodes.Items), nodesRecoverTimeout))
	})

	f.It("shouldn't scale up when expendable pod is created", feature.ClusterSizeAutoscalingScaleUp, func(ctx context.Context) {
		createPriorityClasses(ctx, f)
		// Create nodesCountAfterResize+1 pods allocating 0.7 allocatable on present nodes. One more node will have to be created.
		ginkgo.DeferCleanup(ReserveMemoryWithPriority, f, "memory-reservation", nodeCount+1, int(float64(nodeCount+1)*float64(0.7)*float64(memAllocatableMb)), false, time.Second, expendablePriorityClassName)
		ginkgo.By(fmt.Sprintf("Waiting for scale up hoping it won't happen, sleep for %s", scaleUpTimeout.String()))
		time.Sleep(scaleUpTimeout)
		// Verify that cluster size is not changed
		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size == nodeCount }, time.Second))
	})

	f.It("should scale up when non expendable pod is created", feature.ClusterSizeAutoscalingScaleUp, func(ctx context.Context) {
		createPriorityClasses(ctx, f)
		// Create nodesCountAfterResize+1 pods allocating 0.7 allocatable on present nodes. One more node will have to be created.
		cleanupFunc := ReserveMemoryWithPriority(ctx, f, "memory-reservation", nodeCount+1, int(float64(nodeCount+1)*float64(0.7)*float64(memAllocatableMb)), true, scaleUpTimeout, highPriorityClassName)
		defer cleanupFunc()
		// Verify that cluster size is not changed
		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size > nodeCount }, time.Second))
	})

	f.It("shouldn't scale up when expendable pod is preempted", feature.ClusterSizeAutoscalingScaleUp, func(ctx context.Context) {
		createPriorityClasses(ctx, f)
		// Create nodesCountAfterResize pods allocating 0.7 allocatable on present nodes - one pod per node.
		cleanupFunc1 := ReserveMemoryWithPriority(ctx, f, "memory-reservation1", nodeCount, int(float64(nodeCount)*float64(0.7)*float64(memAllocatableMb)), true, defaultTimeout, expendablePriorityClassName)
		defer cleanupFunc1()
		// Create nodesCountAfterResize pods allocating 0.7 allocatable on present nodes - one pod per node. Pods created here should preempt pods created above.
		cleanupFunc2 := ReserveMemoryWithPriority(ctx, f, "memory-reservation2", nodeCount, int(float64(nodeCount)*float64(0.7)*float64(memAllocatableMb)), true, defaultTimeout, highPriorityClassName)
		defer cleanupFunc2()
		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size == nodeCount }, time.Second))
	})

	f.It("should scale down when expendable pod is running", feature.ClusterSizeAutoscalingScaleDown, func(ctx context.Context) {
		createPriorityClasses(ctx, f)
		increasedSize := manuallyIncreaseClusterSize(ctx, f, originalSizes)
		// Create increasedSize pods allocating 0.7 allocatable on present nodes - one pod per node.
		cleanupFunc := ReserveMemoryWithPriority(ctx, f, "memory-reservation", increasedSize, int(float64(increasedSize)*float64(0.7)*float64(memAllocatableMb)), true, scaleUpTimeout, expendablePriorityClassName)
		defer cleanupFunc()
		ginkgo.By("Waiting for scale down")
		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size == nodeCount }, scaleDownTimeout))
	})

	f.It("shouldn't scale down when non expendable pod is running", feature.ClusterSizeAutoscalingScaleDown, func(ctx context.Context) {
		createPriorityClasses(ctx, f)
		increasedSize := manuallyIncreaseClusterSize(ctx, f, originalSizes)
		// Create increasedSize pods allocating 0.7 allocatable on present nodes - one pod per node.
		cleanupFunc := ReserveMemoryWithPriority(ctx, f, "memory-reservation", increasedSize, int(float64(increasedSize)*float64(0.7)*float64(memAllocatableMb)), true, scaleUpTimeout, highPriorityClassName)
		defer cleanupFunc()
		ginkgo.By(fmt.Sprintf("Waiting for scale down hoping it won't happen, sleep for %s", scaleDownTimeout.String()))
		time.Sleep(scaleDownTimeout)
		framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet,
			func(size int) bool { return size == increasedSize }, time.Second))
	})

	f.It("should scale up when unprocessed pod is created and is going to be unschedulable", feature.ClusterScaleUpBypassScheduler, func(ctx context.Context) {
		// 70% of allocatable memory of a single node * replica count, forcing a scale up in case of normal pods
		replicaCount := 2 * nodeCount
		reservedMemory := int(float64(replicaCount) * float64(0.7) * float64(memAllocatableMb))
		cleanupFunc := ReserveMemoryWithSchedulerName(ctx, f, "memory-reservation", replicaCount, reservedMemory, false, 1, nonExistingBypassedSchedulerName)
		defer framework.ExpectNoError(cleanupFunc())
		// Verify that cluster size is increased
		ginkgo.By("Waiting for cluster scale-up")
		sizeFunc := func(size int) bool {
			// Softly checks scale-up since other types of machines can be added which would affect #nodes
			return size > nodeCount
		}
		framework.ExpectNoError(WaitForClusterSizeFuncWithUnready(ctx, f.ClientSet, sizeFunc, scaleUpTimeout, 0))
	})
	f.It("shouldn't scale up when unprocessed pod is created and is going to be schedulable", feature.ClusterScaleUpBypassScheduler, func(ctx context.Context) {
		// 50% of allocatable memory of a single node, so that no scale up would trigger in normal cases
		replicaCount := 1
		reservedMemory := int(float64(0.5) * float64(memAllocatableMb))
		cleanupFunc := ReserveMemoryWithSchedulerName(ctx, f, "memory-reservation", replicaCount, reservedMemory, false, 1, nonExistingBypassedSchedulerName)
		defer framework.ExpectNoError(cleanupFunc())
		// Verify that cluster size is the same
		ginkgo.By(fmt.Sprintf("Waiting for scale up hoping it won't happen, polling cluster size for %s", scaleUpTimeout.String()))
		sizeFunc := func(size int) bool {
			return size == nodeCount
		}
		gomega.Consistently(ctx, func() error {
			return WaitForClusterSizeFunc(ctx, f.ClientSet, sizeFunc, time.Second)
		}).WithTimeout(scaleUpTimeout).WithPolling(framework.Poll).ShouldNot(gomega.HaveOccurred())
	})
	f.It("shouldn't scale up when unprocessed pod is created and scheduler is not specified to be bypassed", feature.ClusterScaleUpBypassScheduler, func(ctx context.Context) {
		// 70% of allocatable memory of a single node * replica count, forcing a scale up in case of normal pods
		replicaCount := 2 * nodeCount
		reservedMemory := int(float64(replicaCount) * float64(0.7) * float64(memAllocatableMb))
		schedulerName := "non-existent-scheduler-" + f.UniqueName
		cleanupFunc := ReserveMemoryWithSchedulerName(ctx, f, "memory-reservation", replicaCount, reservedMemory, false, 1, schedulerName)
		defer framework.ExpectNoError(cleanupFunc())
		// Verify that cluster size is the same
		ginkgo.By(fmt.Sprintf("Waiting for scale up hoping it won't happen, polling cluster size for %s", scaleUpTimeout.String()))
		sizeFunc := func(size int) bool {
			return size == nodeCount
		}
		gomega.Consistently(ctx, func() error {
			return WaitForClusterSizeFunc(ctx, f.ClientSet, sizeFunc, time.Second)
		}).WithTimeout(scaleUpTimeout).WithPolling(framework.Poll).ShouldNot(gomega.HaveOccurred())
	})
})

func installNvidiaDriversDaemonSet(ctx context.Context, f *framework.Framework) {
	ginkgo.By("Add daemonset which installs nvidia drivers")

	dsYamlURL := "https://raw.githubusercontent.com/GoogleCloudPlatform/container-engine-accelerators/master/daemonset.yaml"
	framework.Logf("Using %v", dsYamlURL)
	// Creates the DaemonSet that installs Nvidia Drivers.
	ds, err := e2emanifest.DaemonSetFromURL(ctx, dsYamlURL)
	framework.ExpectNoError(err)
	ds.Namespace = f.Namespace.Name

	_, err = f.ClientSet.AppsV1().DaemonSets(f.Namespace.Name).Create(ctx, ds, metav1.CreateOptions{})
	framework.ExpectNoError(err, "failed to create nvidia-driver-installer daemonset")
}

func execCmd(args ...string) *exec.Cmd {
	klog.Infof("Executing: %s", strings.Join(args, " "))
	return exec.Command(args[0], args[1:]...)
}

func runDrainTest(ctx context.Context, f *framework.Framework, migSizes map[string]int, namespace string, podsPerNode, pdbSize int, verifyFunction func(int)) {
	increasedSize := manuallyIncreaseClusterSize(ctx, f, migSizes)

	nodes, err := f.ClientSet.CoreV1().Nodes().List(ctx, metav1.ListOptions{FieldSelector: fields.Set{
		"spec.unschedulable": "false",
	}.AsSelector().String()})
	framework.ExpectNoError(err)
	numPods := len(nodes.Items) * podsPerNode
	testID := string(uuid.NewUUID()) // So that we can label and find pods
	labelMap := map[string]string{"test_id": testID}
	framework.ExpectNoError(runReplicatedPodOnEachNode(ctx, f, nodes.Items, namespace, podsPerNode, "reschedulable-pods", labelMap, 0))

	ginkgo.DeferCleanup(e2erc.DeleteRCAndWaitForGC, f.ClientSet, namespace, "reschedulable-pods")

	ginkgo.By("Create a PodDisruptionBudget")
	minAvailable := intstr.FromInt32(int32(numPods - pdbSize))
	pdb := &policyv1.PodDisruptionBudget{
		ObjectMeta: metav1.ObjectMeta{
			Name:      "test_pdb",
			Namespace: namespace,
		},
		Spec: policyv1.PodDisruptionBudgetSpec{
			Selector:     &metav1.LabelSelector{MatchLabels: labelMap},
			MinAvailable: &minAvailable,
		},
	}
	_, err = f.ClientSet.PolicyV1().PodDisruptionBudgets(namespace).Create(ctx, pdb, metav1.CreateOptions{})

	ginkgo.DeferCleanup(framework.IgnoreNotFound(f.ClientSet.PolicyV1().PodDisruptionBudgets(namespace).Delete), pdb.Name, metav1.DeleteOptions{})

	framework.ExpectNoError(err)
	verifyFunction(increasedSize)
}

func getGkeAPIEndpoint() string {
	gkeAPIEndpoint := os.Getenv("CLOUDSDK_API_ENDPOINT_OVERRIDES_CONTAINER")
	if gkeAPIEndpoint == "" {
		gkeAPIEndpoint = "https://test-container.sandbox.googleapis.com"
	}
	if strings.HasSuffix(gkeAPIEndpoint, "/") {
		gkeAPIEndpoint = gkeAPIEndpoint[:len(gkeAPIEndpoint)-1]
	}
	return gkeAPIEndpoint
}

func getGKEURL(apiVersion string, suffix string) string {
	out, err := execCmd("gcloud", "auth", "print-access-token").Output()
	framework.ExpectNoError(err)
	token := strings.Replace(string(out), "\n", "", -1)

	return fmt.Sprintf("%s/%s/%s?access_token=%s",
		getGkeAPIEndpoint(),
		apiVersion,
		suffix,
		token)
}

func getGKEClusterURL(apiVersion string) string {
	if isRegionalCluster() {
		// TODO(bskiba): Use locations API for all clusters once it's graduated to v1.
		return getGKEURL(apiVersion, fmt.Sprintf("projects/%s/locations/%s/clusters/%s",
			framework.TestContext.CloudConfig.ProjectID,
			framework.TestContext.CloudConfig.Region,
			framework.TestContext.CloudConfig.Cluster))
	}
	return getGKEURL(apiVersion, fmt.Sprintf("projects/%s/zones/%s/clusters/%s",
		framework.TestContext.CloudConfig.ProjectID,
		framework.TestContext.CloudConfig.Zone,
		framework.TestContext.CloudConfig.Cluster))
}

func getCluster(apiVersion string) (string, error) {
	resp, err := http.Get(getGKEClusterURL(apiVersion))
	if err != nil {
		return "", err
	}
	defer resp.Body.Close()
	body, err := io.ReadAll(resp.Body)
	if err != nil {
		return "", err
	}
	if resp.StatusCode != http.StatusOK {
		return "", fmt.Errorf("error: %s %s", resp.Status, body)
	}

	return string(body), nil
}

func isAutoscalerEnabled(expectedMaxNodeCountInTargetPool int) (bool, error) {
	apiVersion := "v1"
	if isRegionalCluster() {
		apiVersion = "v1beta1"
	}
	strBody, err := getCluster(apiVersion)
	if err != nil {
		return false, err
	}
	if strings.Contains(strBody, "\"maxNodeCount\": "+strconv.Itoa(expectedMaxNodeCountInTargetPool)) {
		return true, nil
	}
	return false, nil
}

func getClusterLocation() string {
	if isRegionalCluster() {
		return "--region=" + framework.TestContext.CloudConfig.Region
	}
	return "--zone=" + framework.TestContext.CloudConfig.Zone
}

func getGcloudCommandFromTrack(commandTrack string, args []string) []string {
	command := []string{"gcloud"}
	if commandTrack == "beta" || commandTrack == "alpha" {
		command = append(command, commandTrack)
	}
	command = append(command, args...)
	command = append(command, getClusterLocation())
	command = append(command, "--project="+framework.TestContext.CloudConfig.ProjectID)
	return command
}

func getGcloudCommand(args []string) []string {
	track := ""
	if isRegionalCluster() {
		track = "beta"
	}
	return getGcloudCommandFromTrack(track, args)
}

func isRegionalCluster() bool {
	// TODO(bskiba): Use an appropriate indicator that the cluster is regional.
	return framework.TestContext.CloudConfig.MultiZone
}

func enableAutoscaler(nodePool string, minCount, maxCount int) error {
	klog.Infof("Using gcloud to enable autoscaling for pool %s", nodePool)

	args := []string{"container", "clusters", "update", framework.TestContext.CloudConfig.Cluster,
		"--enable-autoscaling",
		"--min-nodes=" + strconv.Itoa(minCount),
		"--max-nodes=" + strconv.Itoa(maxCount),
		"--node-pool=" + nodePool}
	output, err := execCmd(getGcloudCommand(args)...).CombinedOutput()

	if err != nil {
		klog.Errorf("Failed config update result: %s", output)
		return fmt.Errorf("Failed to enable autoscaling: %w", err)
	}
	klog.Infof("Config update result: %s", output)

	var finalErr error
	for startTime := time.Now(); startTime.Add(gkeUpdateTimeout).After(time.Now()); time.Sleep(30 * time.Second) {
		val, err := isAutoscalerEnabled(maxCount)
		if err == nil && val {
			return nil
		}
		finalErr = err
	}
	return fmt.Errorf("autoscaler not enabled, last error: %v", finalErr)
}

func disableAutoscaler(nodePool string, minCount, maxCount int) error {
	klog.Infof("Using gcloud to disable autoscaling for pool %s", nodePool)
	args := []string{"container", "clusters", "update", framework.TestContext.CloudConfig.Cluster,
		"--no-enable-autoscaling",
		"--node-pool=" + nodePool}
	output, err := execCmd(getGcloudCommand(args)...).CombinedOutput()

	if err != nil {
		klog.Errorf("Failed config update result: %s", output)
		return fmt.Errorf("Failed to disable autoscaling: %w", err)
	}
	klog.Infof("Config update result: %s", output)

	var finalErr error
	for startTime := time.Now(); startTime.Add(gkeUpdateTimeout).After(time.Now()); time.Sleep(30 * time.Second) {
		val, err := isAutoscalerEnabled(maxCount)
		if err == nil && !val {
			return nil
		}
		finalErr = err
	}
	return fmt.Errorf("autoscaler still enabled, last error: %v", finalErr)
}

func addNodePool(name string, machineType string, numNodes int) {
	args := []string{"container", "node-pools", "create", name, "--quiet",
		"--machine-type=" + machineType,
		"--num-nodes=" + strconv.Itoa(numNodes),
		"--cluster=" + framework.TestContext.CloudConfig.Cluster}
	output, err := execCmd(getGcloudCommand(args)...).CombinedOutput()
	klog.Infof("Creating node-pool %s: %s", name, output)
	framework.ExpectNoError(err, string(output))
}

func addGpuNodePool(name string, gpuType string, gpuCount int, numNodes int) {
	args := []string{"beta", "container", "node-pools", "create", name, "--quiet",
		"--accelerator", "type=" + gpuType + ",count=" + strconv.Itoa(gpuCount),
		"--num-nodes=" + strconv.Itoa(numNodes),
		"--cluster=" + framework.TestContext.CloudConfig.Cluster}
	output, err := execCmd(getGcloudCommand(args)...).CombinedOutput()
	klog.Infof("Creating node-pool %s: %s", name, output)
	framework.ExpectNoError(err, string(output))
}

func deleteNodePool(name string) {
	klog.Infof("Deleting node pool %s", name)
	args := []string{"container", "node-pools", "delete", name, "--quiet",
		"--cluster=" + framework.TestContext.CloudConfig.Cluster}
	err := wait.ExponentialBackoff(
		wait.Backoff{Duration: 1 * time.Minute, Factor: float64(3), Steps: 3},
		func() (bool, error) {
			output, err := execCmd(getGcloudCommand(args)...).CombinedOutput()
			if err != nil {
				klog.Warningf("Error deleting nodegroup - error:%v, output: %s", err, output)
				return false, nil
			}
			klog.Infof("Node-pool deletion output: %s", output)
			return true, nil
		})
	framework.ExpectNoError(err)
}

func getPoolNodes(ctx context.Context, f *framework.Framework, poolName string) []*v1.Node {
	nodes := make([]*v1.Node, 0, 1)
	nodeList, err := e2enode.GetReadyNodesIncludingTainted(ctx, f.ClientSet)
	if err != nil {
		framework.Logf("Unexpected error occurred: %v", err)
	}
	framework.ExpectNoErrorWithOffset(0, err)
	for _, node := range nodeList.Items {
		if node.Labels[gkeNodepoolNameKey] == poolName {
			node := node
			nodes = append(nodes, &node)
		}
	}
	return nodes
}

// getPoolInitialSize returns the initial size of the node pool taking into
// account that it may span multiple zones. In that case, node pool consists of
// multiple migs all containing initialNodeCount nodes.
func getPoolInitialSize(poolName string) int {
	// get initial node count
	args := []string{"container", "node-pools", "describe", poolName, "--quiet",
		"--cluster=" + framework.TestContext.CloudConfig.Cluster,
		"--format=value(initialNodeCount)"}
	output, err := execCmd(getGcloudCommand(args)...).CombinedOutput()
	klog.Infof("Node-pool initial size: %s", output)
	framework.ExpectNoError(err, string(output))
	fields := strings.Fields(string(output))
	gomega.Expect(fields).To(gomega.HaveLen(1))
	size, err := strconv.ParseInt(fields[0], 10, 64)
	framework.ExpectNoError(err)

	// get number of node pools
	args = []string{"container", "node-pools", "describe", poolName, "--quiet",
		"--cluster=" + framework.TestContext.CloudConfig.Cluster,
		"--format=value(instanceGroupUrls)"}
	output, err = execCmd(getGcloudCommand(args)...).CombinedOutput()
	framework.ExpectNoError(err, string(output))
	nodeGroupCount := len(strings.Split(string(output), ";"))
	return int(size) * nodeGroupCount
}

func getPoolSize(ctx context.Context, f *framework.Framework, poolName string) int {
	size := 0
	nodeList, err := e2enode.GetReadySchedulableNodes(ctx, f.ClientSet)
	framework.ExpectNoError(err)
	for _, node := range nodeList.Items {
		if node.Labels[gkeNodepoolNameKey] == poolName {
			size++
		}
	}
	return size
}

func reserveMemory(ctx context.Context, f *framework.Framework, id string, replicas, megabytes int, expectRunning bool, timeout time.Duration, selector map[string]string, tolerations []v1.Toleration, priorityClassName, schedulerName string) func() error {
	ginkgo.By(fmt.Sprintf("Running RC which reserves %v MB of memory", megabytes))
	request := int64(1024 * 1024 * megabytes / replicas)
	config := &testutils.RCConfig{
		Client:            f.ClientSet,
		Name:              id,
		Namespace:         f.Namespace.Name,
		Timeout:           timeout,
		Image:             imageutils.GetPauseImageName(),
		Replicas:          replicas,
		MemRequest:        request,
		NodeSelector:      selector,
		Tolerations:       tolerations,
		PriorityClassName: priorityClassName,
		SchedulerName:     schedulerName,
	}
	for start := time.Now(); time.Since(start) < rcCreationRetryTimeout; time.Sleep(rcCreationRetryDelay) {
		err := e2erc.RunRC(ctx, *config)
		if err != nil && strings.Contains(err.Error(), "Error creating replication controller") {
			klog.Warningf("Failed to create memory reservation: %v", err)
			continue
		}
		if expectRunning {
			framework.ExpectNoError(err)
		}
		return func() error {
			return e2erc.DeleteRCAndWaitForGC(ctx, f.ClientSet, f.Namespace.Name, id)
		}
	}
	framework.Failf("Failed to reserve memory within timeout")
	return nil
}

// ReserveMemoryWithPriority creates a replication controller with pods with priority that, in summation,
// request the specified amount of memory.
func ReserveMemoryWithPriority(ctx context.Context, f *framework.Framework, id string, replicas, megabytes int, expectRunning bool, timeout time.Duration, priorityClassName string) func() error {
	return reserveMemory(ctx, f, id, replicas, megabytes, expectRunning, timeout, nil, nil, priorityClassName, "")
}

// ReserveMemoryWithSelectorAndTolerations creates a replication controller with pods with node selector that, in summation,
// request the specified amount of memory.
func ReserveMemoryWithSelectorAndTolerations(ctx context.Context, f *framework.Framework, id string, replicas, megabytes int, expectRunning bool, timeout time.Duration, selector map[string]string, tolerations []v1.Toleration) func() error {
	return reserveMemory(ctx, f, id, replicas, megabytes, expectRunning, timeout, selector, tolerations, "", "")
}

// ReserveMemoryWithSchedulerName creates a replication controller with pods with scheduler name that, in summation,
// request the specified amount of memory.
func ReserveMemoryWithSchedulerName(ctx context.Context, f *framework.Framework, id string, replicas, megabytes int, expectRunning bool, timeout time.Duration, schedulerName string) func() error {
	return reserveMemory(ctx, f, id, replicas, megabytes, expectRunning, timeout, nil, nil, "", schedulerName)
}

// ReserveMemory creates a replication controller with pods that, in summation,
// request the specified amount of memory.
func ReserveMemory(ctx context.Context, f *framework.Framework, id string, replicas, megabytes int, expectRunning bool, timeout time.Duration) func() error {
	return reserveMemory(ctx, f, id, replicas, megabytes, expectRunning, timeout, nil, nil, "", "")
}

// WaitForClusterSizeFunc waits until the cluster size matches the given function.
func WaitForClusterSizeFunc(ctx context.Context, c clientset.Interface, sizeFunc func(int) bool, timeout time.Duration) error {
	return WaitForClusterSizeFuncWithUnready(ctx, c, sizeFunc, timeout, 0)
}

// WaitForClusterSizeFuncWithUnready waits until the cluster size matches the given function and assumes some unready nodes.
func WaitForClusterSizeFuncWithUnready(ctx context.Context, c clientset.Interface, sizeFunc func(int) bool, timeout time.Duration, expectedUnready int) error {
	for start := time.Now(); time.Since(start) < timeout && ctx.Err() == nil; time.Sleep(20 * time.Second) {
		nodes, err := c.CoreV1().Nodes().List(ctx, metav1.ListOptions{FieldSelector: fields.Set{
			"spec.unschedulable": "false",
		}.AsSelector().String()})
		if err != nil {
			klog.Warningf("Failed to list nodes: %v", err)
			continue
		}
		numNodes := len(nodes.Items)

		// Filter out not-ready nodes.
		e2enode.Filter(nodes, func(node v1.Node) bool {
			return e2enode.IsConditionSetAsExpected(&node, v1.NodeReady, true)
		})
		numReady := len(nodes.Items)

		if numNodes == numReady+expectedUnready && sizeFunc(numNodes) {
			klog.Infof("Cluster has reached the desired size")
			return nil
		}
		klog.Infof("Waiting for cluster with func, current size %d, not ready nodes %d", numNodes, numNodes-numReady)
	}
	return fmt.Errorf("timeout waiting %v for appropriate cluster size", timeout)
}

func waitForCaPodsReadyInNamespace(ctx context.Context, f *framework.Framework, c clientset.Interface, tolerateUnreadyCount int) error {
	var notready []string
	for start := time.Now(); time.Now().Before(start.Add(scaleUpTimeout)) && ctx.Err() == nil; time.Sleep(20 * time.Second) {
		pods, err := c.CoreV1().Pods(f.Namespace.Name).List(ctx, metav1.ListOptions{})
		if err != nil {
			return fmt.Errorf("failed to get pods: %w", err)
		}
		notready = make([]string, 0)
		for _, pod := range pods.Items {
			ready := false
			for _, c := range pod.Status.Conditions {
				if c.Type == v1.PodReady && c.Status == v1.ConditionTrue {
					ready = true
				}
			}
			// Failed pods in this context generally mean that they have been
			// double scheduled onto a node, but then failed a constraint check.
			if pod.Status.Phase == v1.PodFailed {
				klog.Warningf("Pod has failed: %v", pod)
			}
			if !ready && pod.Status.Phase != v1.PodFailed {
				notready = append(notready, pod.Name)
			}
		}
		if len(notready) <= tolerateUnreadyCount {
			klog.Infof("sufficient number of pods ready. Tolerating %d unready", tolerateUnreadyCount)
			return nil
		}
		klog.Infof("Too many pods are not ready yet: %v", notready)
	}
	klog.Info("Timeout on waiting for pods being ready")
	klog.Info(e2ekubectl.RunKubectlOrDie(f.Namespace.Name, "get", "pods", "-o", "json", "--all-namespaces"))
	klog.Info(e2ekubectl.RunKubectlOrDie(f.Namespace.Name, "get", "nodes", "-o", "json"))

	// Some pods are still not running.
	return fmt.Errorf("Too many pods are still not running: %v", notready)
}

func waitForAllCaPodsReadyInNamespace(ctx context.Context, f *framework.Framework, c clientset.Interface) error {
	return waitForCaPodsReadyInNamespace(ctx, f, c, 0)
}

func getAnyNode(ctx context.Context, c clientset.Interface) *v1.Node {
	nodes, err := c.CoreV1().Nodes().List(ctx, metav1.ListOptions{FieldSelector: fields.Set{
		"spec.unschedulable": "false",
	}.AsSelector().String()})
	if err != nil {
		klog.Errorf("Failed to get node list: %v", err)
		return nil
	}
	if len(nodes.Items) == 0 {
		klog.Errorf("No nodes")
		return nil
	}
	return &nodes.Items[0]
}

func setMigSizes(sizes map[string]int) bool {
	madeChanges := false
	for mig, desiredSize := range sizes {
		currentSize, err := framework.GroupSize(mig)
		framework.ExpectNoError(err)
		if desiredSize != currentSize {
			ginkgo.By(fmt.Sprintf("Setting size of %s to %d", mig, desiredSize))
			err = framework.ResizeGroup(mig, int32(desiredSize))
			framework.ExpectNoError(err)
			madeChanges = true
		}
	}
	return madeChanges
}

func drainNode(ctx context.Context, f *framework.Framework, node *v1.Node) {
	ginkgo.By("Make the single node unschedulable")
	framework.ExpectNoError(makeNodeUnschedulable(ctx, f.ClientSet, node))

	ginkgo.By("Manually drain the single node")
	podOpts := metav1.ListOptions{FieldSelector: fields.OneTermEqualSelector("spec.nodeName", node.Name).String()}
	pods, err := f.ClientSet.CoreV1().Pods(metav1.NamespaceAll).List(ctx, podOpts)
	framework.ExpectNoError(err)
	for _, pod := range pods.Items {
		err = f.ClientSet.CoreV1().Pods(pod.Namespace).Delete(ctx, pod.Name, *metav1.NewDeleteOptions(0))
		framework.ExpectNoError(err)
	}
}

func makeNodeUnschedulable(ctx context.Context, c clientset.Interface, node *v1.Node) error {
	ginkgo.By(fmt.Sprintf("Taint node %s", node.Name))
	for j := 0; j < 3; j++ {
		freshNode, err := c.CoreV1().Nodes().Get(ctx, node.Name, metav1.GetOptions{})
		if err != nil {
			return err
		}
		for _, taint := range freshNode.Spec.Taints {
			if taint.Key == disabledTaint {
				return nil
			}
		}
		freshNode.Spec.Taints = append(freshNode.Spec.Taints, v1.Taint{
			Key:    disabledTaint,
			Value:  "DisabledForTest",
			Effect: v1.TaintEffectNoSchedule,
		})
		_, err = c.CoreV1().Nodes().Update(ctx, freshNode, metav1.UpdateOptions{})
		if err == nil {
			return nil
		}
		if !apierrors.IsConflict(err) {
			return err
		}
		klog.Warningf("Got 409 conflict when trying to taint node, retries left: %v", 3-j)
	}
	return fmt.Errorf("Failed to taint node in allowed number of retries")
}

// CriticalAddonsOnlyError implements the `error` interface, and signifies the
// presence of the `CriticalAddonsOnly` taint on the node.
type CriticalAddonsOnlyError struct{}

func (CriticalAddonsOnlyError) Error() string {
	return fmt.Sprintf("CriticalAddonsOnly taint found on node")
}

func makeNodeSchedulable(ctx context.Context, c clientset.Interface, node *v1.Node, failOnCriticalAddonsOnly bool) error {
	ginkgo.By(fmt.Sprintf("Remove taint from node %s", node.Name))
	for j := 0; j < 3; j++ {
		freshNode, err := c.CoreV1().Nodes().Get(ctx, node.Name, metav1.GetOptions{})
		if err != nil {
			return err
		}
		var newTaints []v1.Taint
		for _, taint := range freshNode.Spec.Taints {
			if failOnCriticalAddonsOnly && taint.Key == criticalAddonsOnlyTaint {
				return CriticalAddonsOnlyError{}
			}
			if taint.Key != disabledTaint {
				newTaints = append(newTaints, taint)
			}
		}

		if len(newTaints) == len(freshNode.Spec.Taints) {
			return nil
		}
		freshNode.Spec.Taints = newTaints
		_, err = c.CoreV1().Nodes().Update(ctx, freshNode, metav1.UpdateOptions{})
		if err == nil {
			return nil
		}
		if !apierrors.IsConflict(err) {
			return err
		}
		klog.Warningf("Got 409 conflict when trying to taint node, retries left: %v", 3-j)
	}
	return fmt.Errorf("Failed to remove taint from node in allowed number of retries")
}

// ScheduleAnySingleGpuPod schedules a pod which requires single GPU of any type
func ScheduleAnySingleGpuPod(ctx context.Context, f *framework.Framework, id string) error {
	return ScheduleGpuPod(ctx, f, id, "", 1)
}

// ScheduleGpuPod schedules a pod which requires a given number of gpus of given type
func ScheduleGpuPod(ctx context.Context, f *framework.Framework, id string, gpuType string, gpuLimit int64) error {
	config := &testutils.RCConfig{
		Client:    f.ClientSet,
		Name:      id,
		Namespace: f.Namespace.Name,
		Timeout:   3 * scaleUpTimeout, // spinning up GPU node is slow
		Image:     imageutils.GetPauseImageName(),
		Replicas:  1,
		GpuLimit:  gpuLimit,
		Labels:    map[string]string{"requires-gpu": "yes"},
	}

	if gpuType != "" {
		config.NodeSelector = map[string]string{gpuLabel: gpuType}
	}

	err := e2erc.RunRC(ctx, *config)
	if err != nil {
		return err
	}
	return nil
}

// Create an RC running a given number of pods with anti-affinity
func runAntiAffinityPods(ctx context.Context, f *framework.Framework, namespace string, pods int, id string, podLabels, antiAffinityLabels map[string]string) error {
	config := &testutils.RCConfig{
		Affinity:  buildAntiAffinity(antiAffinityLabels),
		Client:    f.ClientSet,
		Name:      id,
		Namespace: namespace,
		Timeout:   scaleUpTimeout,
		Image:     imageutils.GetPauseImageName(),
		Replicas:  pods,
		Labels:    podLabels,
	}
	err := e2erc.RunRC(ctx, *config)
	if err != nil {
		return err
	}
	_, err = f.ClientSet.CoreV1().ReplicationControllers(namespace).Get(ctx, id, metav1.GetOptions{})
	if err != nil {
		return err
	}
	return nil
}

func runVolumeAntiAffinityPods(ctx context.Context, f *framework.Framework, namespace string, pods int, id string, podLabels, antiAffinityLabels map[string]string, volumes []v1.Volume) error {
	config := &testutils.RCConfig{
		Affinity:  buildAntiAffinity(antiAffinityLabels),
		Volumes:   volumes,
		Client:    f.ClientSet,
		Name:      id,
		Namespace: namespace,
		Timeout:   scaleUpTimeout,
		Image:     imageutils.GetPauseImageName(),
		Replicas:  pods,
		Labels:    podLabels,
	}
	err := e2erc.RunRC(ctx, *config)
	if err != nil {
		return err
	}
	_, err = f.ClientSet.CoreV1().ReplicationControllers(namespace).Get(ctx, id, metav1.GetOptions{})
	if err != nil {
		return err
	}
	return nil
}

var emptyDirVolumes = []v1.Volume{
	{
		Name: "empty-volume",
		VolumeSource: v1.VolumeSource{
			EmptyDir: &v1.EmptyDirVolumeSource{},
		},
	},
}

func buildVolumes(pv *v1.PersistentVolume, pvc *v1.PersistentVolumeClaim) []v1.Volume {
	return []v1.Volume{
		{
			Name: pv.Name,
			VolumeSource: v1.VolumeSource{
				PersistentVolumeClaim: &v1.PersistentVolumeClaimVolumeSource{
					ClaimName: pvc.Name,
					ReadOnly:  false,
				},
			},
		},
	}
}

func buildAntiAffinity(labels map[string]string) *v1.Affinity {
	return &v1.Affinity{
		PodAntiAffinity: &v1.PodAntiAffinity{
			RequiredDuringSchedulingIgnoredDuringExecution: []v1.PodAffinityTerm{
				{
					LabelSelector: &metav1.LabelSelector{
						MatchLabels: labels,
					},
					TopologyKey: "kubernetes.io/hostname",
				},
			},
		},
	}
}

// Create an RC running a given number of pods on each node without adding any constraint forcing
// such pod distribution. This is meant to create a bunch of underutilized (but not unused) nodes
// with pods that can be rescheduled on different nodes.
// This is achieved using the following method:
// 1. disable scheduling on each node
// 2. create an empty RC
// 3. for each node:
// 3a. enable scheduling on that node
// 3b. increase number of replicas in RC by podsPerNode
func runReplicatedPodOnEachNode(ctx context.Context, f *framework.Framework, nodes []v1.Node, namespace string, podsPerNode int, id string, labels map[string]string, memRequest int64) error {
	ginkgo.By("Run a pod on each node")
	for _, node := range nodes {
		err := makeNodeUnschedulable(ctx, f.ClientSet, &node)

		n := node
		ginkgo.DeferCleanup(makeNodeSchedulable, f.ClientSet, &n, false)

		if err != nil {
			return err
		}
	}
	config := &testutils.RCConfig{
		Client:     f.ClientSet,
		Name:       id,
		Namespace:  namespace,
		Timeout:    defaultTimeout,
		Image:      imageutils.GetPauseImageName(),
		Replicas:   0,
		Labels:     labels,
		MemRequest: memRequest,
	}
	err := e2erc.RunRC(ctx, *config)
	if err != nil {
		return err
	}
	rc, err := f.ClientSet.CoreV1().ReplicationControllers(namespace).Get(ctx, id, metav1.GetOptions{})
	if err != nil {
		return err
	}
	for i, node := range nodes {
		err = makeNodeSchedulable(ctx, f.ClientSet, &node, false)
		if err != nil {
			return err
		}

		// Update replicas count, to create new pods that will be allocated on node
		// (we retry 409 errors in case rc reference got out of sync)
		for j := 0; j < 3; j++ {
			*rc.Spec.Replicas = int32((i + 1) * podsPerNode)
			rc, err = f.ClientSet.CoreV1().ReplicationControllers(namespace).Update(ctx, rc, metav1.UpdateOptions{})
			if err == nil {
				break
			}
			if !apierrors.IsConflict(err) {
				return err
			}
			klog.Warningf("Got 409 conflict when trying to scale RC, retries left: %v", 3-j)
			rc, err = f.ClientSet.CoreV1().ReplicationControllers(namespace).Get(ctx, id, metav1.GetOptions{})
			if err != nil {
				return err
			}
		}

		err = wait.PollImmediate(5*time.Second, podTimeout, func() (bool, error) {
			rc, err = f.ClientSet.CoreV1().ReplicationControllers(namespace).Get(ctx, id, metav1.GetOptions{})
			if err != nil || rc.Status.ReadyReplicas < int32((i+1)*podsPerNode) {
				return false, nil
			}
			return true, nil
		})
		if err != nil {
			return fmt.Errorf("failed to coerce RC into spawning a pod on node %s within timeout", node.Name)
		}
		err = makeNodeUnschedulable(ctx, f.ClientSet, &node)
		if err != nil {
			return err
		}
	}
	return nil
}

// Increase cluster size by newNodesForScaledownTests to create some unused nodes
// that can be later removed by cluster autoscaler.
func manuallyIncreaseClusterSize(ctx context.Context, f *framework.Framework, originalSizes map[string]int) int {
	ginkgo.By("Manually increase cluster size")
	increasedSize := 0
	newSizes := make(map[string]int)
	for key, val := range originalSizes {
		newSizes[key] = val + newNodesForScaledownTests
		increasedSize += val + newNodesForScaledownTests
	}
	setMigSizes(newSizes)

	checkClusterSize := func(size int) bool {
		if size >= increasedSize {
			return true
		}
		resized := setMigSizes(newSizes)
		if resized {
			klog.Warning("Unexpected node group size while waiting for cluster resize. Setting size to target again.")
		}
		return false
	}

	framework.ExpectNoError(WaitForClusterSizeFunc(ctx, f.ClientSet, checkClusterSize, manualResizeTimeout))
	return increasedSize
}

// Try to get clusterwide health from CA status configmap.
// Status configmap is not parsing-friendly, so evil regexpery follows.
func getClusterwideStatus(ctx context.Context, c clientset.Interface) (string, error) {
	configMap, err := c.CoreV1().ConfigMaps("kube-system").Get(ctx, "cluster-autoscaler-status", metav1.GetOptions{})
	if err != nil {
		return "", err
	}
	status, ok := configMap.Data["status"]
	if !ok {
		return "", fmt.Errorf("Status information not found in configmap")
	}
	matcher, err := regexp.Compile("Cluster-wide:\\s*\n\\s*Health:\\s*([A-Za-z]+)")
	if err != nil {
		return "", err
	}
	result := matcher.FindStringSubmatch(status)
	if len(result) < 2 {
		return "", fmt.Errorf("Failed to parse CA status configmap, raw status: %v", status)
	}
	return result[1], nil
}

type scaleUpStatus struct {
	status    string
	ready     int
	target    int
	timestamp time.Time
}

// Try to get timestamp from status.
// Status configmap is not parsing-friendly, so evil regexpery follows.
func getStatusTimestamp(status string) (time.Time, error) {
	timestampMatcher, err := regexp.Compile("Cluster-autoscaler status at \\s*([0-9\\-]+ [0-9]+:[0-9]+:[0-9]+\\.[0-9]+ \\+[0-9]+ [A-Za-z]+)")
	if err != nil {
		return time.Time{}, err
	}

	timestampMatch := timestampMatcher.FindStringSubmatch(status)
	if len(timestampMatch) < 2 {
		return time.Time{}, fmt.Errorf("Failed to parse CA status timestamp, raw status: %v", status)
	}

	timestamp, err := time.Parse(timestampFormat, timestampMatch[1])
	if err != nil {
		return time.Time{}, err
	}
	return timestamp, nil
}

// Try to get scaleup statuses of all node groups.
// Status configmap is not parsing-friendly, so evil regexpery follows.
func getScaleUpStatus(ctx context.Context, c clientset.Interface) (*scaleUpStatus, error) {
	configMap, err := c.CoreV1().ConfigMaps("kube-system").Get(ctx, "cluster-autoscaler-status", metav1.GetOptions{})
	if err != nil {
		return nil, err
	}
	status, ok := configMap.Data["status"]
	if !ok {
		return nil, fmt.Errorf("Status information not found in configmap")
	}

	timestamp, err := getStatusTimestamp(status)
	if err != nil {
		return nil, err
	}

	matcher, err := regexp.Compile("s*ScaleUp:\\s*([A-Za-z]+)\\s*\\(ready=([0-9]+)\\s*cloudProviderTarget=([0-9]+)\\s*\\)")
	if err != nil {
		return nil, err
	}
	matches := matcher.FindAllStringSubmatch(status, -1)
	if len(matches) < 1 {
		return nil, fmt.Errorf("Failed to parse CA status configmap, raw status: %v", status)
	}

	result := scaleUpStatus{
		status:    caNoScaleUpStatus,
		ready:     0,
		target:    0,
		timestamp: timestamp,
	}
	for _, match := range matches {
		if match[1] == caOngoingScaleUpStatus {
			result.status = caOngoingScaleUpStatus
		}
		newReady, err := strconv.Atoi(match[2])
		if err != nil {
			return nil, err
		}
		result.ready += newReady
		newTarget, err := strconv.Atoi(match[3])
		if err != nil {
			return nil, err
		}
		result.target += newTarget
	}
	klog.Infof("Cluster-Autoscaler scale-up status: %v (%v, %v)", result.status, result.ready, result.target)
	return &result, nil
}

func waitForScaleUpStatus(ctx context.Context, c clientset.Interface, cond func(s *scaleUpStatus) bool, timeout time.Duration) (*scaleUpStatus, error) {
	var finalErr error
	var status *scaleUpStatus
	err := wait.PollUntilContextTimeout(ctx, 5*time.Second, timeout, true, func(ctx context.Context) (bool, error) {
		status, finalErr = getScaleUpStatus(ctx, c)
		if finalErr != nil {
			return false, nil
		}
		if status.timestamp.Add(freshStatusLimit).Before(time.Now()) {
			// stale status
			finalErr = fmt.Errorf("Status too old")
			return false, nil
		}
		return cond(status), nil
	})
	if err != nil {
		err = fmt.Errorf("Failed to find expected scale up status: %v, last status: %v, final err: %v", err, status, finalErr)
	}
	return status, err
}

// This is a temporary fix to allow CA to migrate some kube-system pods
// TODO: Remove this when the PDB is added for some of those components
func addKubeSystemPdbs(ctx context.Context, f *framework.Framework) error {
	ginkgo.By("Create PodDisruptionBudgets for kube-system components, so they can be migrated if required")

	var newPdbs []string
	cleanup := func(ctx context.Context) {
		var finalErr error
		for _, newPdbName := range newPdbs {
			ginkgo.By(fmt.Sprintf("Delete PodDisruptionBudget %v", newPdbName))
			err := f.ClientSet.PolicyV1().PodDisruptionBudgets("kube-system").Delete(ctx, newPdbName, metav1.DeleteOptions{})
			if err != nil {
				// log error, but attempt to remove other pdbs
				klog.Errorf("Failed to delete PodDisruptionBudget %v, err: %v", newPdbName, err)
				finalErr = err
			}
		}
		if finalErr != nil {
			framework.Failf("Error during PodDisruptionBudget cleanup: %v", finalErr)
		}
	}
	ginkgo.DeferCleanup(cleanup)

	type pdbInfo struct {
		label        string
		minAvailable int
	}
	pdbsToAdd := []pdbInfo{
		{label: "kube-dns", minAvailable: 1},
		{label: "kube-dns-autoscaler", minAvailable: 0},
		{label: "metrics-server", minAvailable: 0},
		{label: "kubernetes-dashboard", minAvailable: 0},
		{label: "glbc", minAvailable: 0},
	}
	for _, pdbData := range pdbsToAdd {
		ginkgo.By(fmt.Sprintf("Create PodDisruptionBudget for %v", pdbData.label))
		labelMap := map[string]string{"k8s-app": pdbData.label}
		pdbName := fmt.Sprintf("test-pdb-for-%v", pdbData.label)
		minAvailable := intstr.FromInt32(int32(pdbData.minAvailable))
		pdb := &policyv1.PodDisruptionBudget{
			ObjectMeta: metav1.ObjectMeta{
				Name:      pdbName,
				Namespace: "kube-system",
			},
			Spec: policyv1.PodDisruptionBudgetSpec{
				Selector:     &metav1.LabelSelector{MatchLabels: labelMap},
				MinAvailable: &minAvailable,
			},
		}
		_, err := f.ClientSet.PolicyV1().PodDisruptionBudgets("kube-system").Create(ctx, pdb, metav1.CreateOptions{})
		newPdbs = append(newPdbs, pdbName)

		if err != nil {
			return err
		}
	}
	return nil
}

func createPriorityClasses(ctx context.Context, f *framework.Framework) {
	priorityClasses := map[string]int32{
		expendablePriorityClassName: -15,
		highPriorityClassName:       1000,
	}
	for className, priority := range priorityClasses {
		_, err := f.ClientSet.SchedulingV1().PriorityClasses().Create(ctx, &schedulingv1.PriorityClass{ObjectMeta: metav1.ObjectMeta{Name: className}, Value: priority}, metav1.CreateOptions{})
		if err != nil {
			klog.Errorf("Error creating priority class: %v", err)
		}
		if err != nil && !apierrors.IsAlreadyExists(err) {
			framework.Failf("unexpected error while creating priority class: %v", err)
		}
	}

	ginkgo.DeferCleanup(func(ctx context.Context) {
		for className := range priorityClasses {
			err := f.ClientSet.SchedulingV1().PriorityClasses().Delete(ctx, className, metav1.DeleteOptions{})
			if err != nil {
				klog.Errorf("Error deleting priority class: %v", err)
			}
		}
	})
}