Implement graceful shutdown based on Pod priority

pkg/kubelet/apis/config/helpers_test.go

@@ -267,6 +267,8 @@ var (
 		"SeccompDefault",
 		"SerializeImagePulls",
 		"ShowHiddenMetricsForVersion",
+		"ShutdownGracePeriodByPodPriority[*].Priority",
+		"ShutdownGracePeriodByPodPriority[*].ShutdownGracePeriodSeconds",
 		"StreamingConnectionIdleTimeout.Duration",
 		"SyncFrequency.Duration",
 		"SystemCgroups",

pkg/kubelet/kubelet.go

@@ -867,14 +867,15 @@ func NewMainKubelet(kubeCfg *kubeletconfiginternal.KubeletConfiguration,
 
 	// setup node shutdown manager
 	shutdownManager, shutdownAdmitHandler := nodeshutdown.NewManager(&nodeshutdown.Config{
-		ProbeManager:                    klet.probeManager,
+		ProbeManager:                     klet.probeManager,
-		Recorder:                        kubeDeps.Recorder,
+		Recorder:                         kubeDeps.Recorder,
-		NodeRef:                         nodeRef,
+		NodeRef:                          nodeRef,
-		GetPodsFunc:                     klet.GetActivePods,
+		GetPodsFunc:                      klet.GetActivePods,
-		KillPodFunc:                     killPodNow(klet.podWorkers, kubeDeps.Recorder),
+		KillPodFunc:                      killPodNow(klet.podWorkers, kubeDeps.Recorder),
-		SyncNodeStatusFunc:              klet.syncNodeStatus,
+		SyncNodeStatusFunc:               klet.syncNodeStatus,
-		ShutdownGracePeriodRequested:    kubeCfg.ShutdownGracePeriod.Duration,
+		ShutdownGracePeriodRequested:     kubeCfg.ShutdownGracePeriod.Duration,
-		ShutdownGracePeriodCriticalPods: kubeCfg.ShutdownGracePeriodCriticalPods.Duration,
+		ShutdownGracePeriodCriticalPods:  kubeCfg.ShutdownGracePeriodCriticalPods.Duration,
+		ShutdownGracePeriodByPodPriority: kubeCfg.ShutdownGracePeriodByPodPriority,
 	})
 	klet.shutdownManager = shutdownManager
 	klet.admitHandlers.AddPodAdmitHandler(shutdownAdmitHandler)

pkg/kubelet/nodeshutdown/nodeshutdown_manager.go

@@ -21,6 +21,7 @@ import (
 
 	v1 "k8s.io/api/core/v1"
 	"k8s.io/client-go/tools/record"
+	kubeletconfig "k8s.io/kubernetes/pkg/kubelet/apis/config"
 	"k8s.io/kubernetes/pkg/kubelet/eviction"
 	"k8s.io/kubernetes/pkg/kubelet/lifecycle"
 	"k8s.io/kubernetes/pkg/kubelet/prober"
@@ -36,15 +37,16 @@ type Manager interface {
 
 // Config represents Manager configuration
 type Config struct {
-	ProbeManager                    prober.Manager
+	ProbeManager                     prober.Manager
-	Recorder                        record.EventRecorder
+	Recorder                         record.EventRecorder
-	NodeRef                         *v1.ObjectReference
+	NodeRef                          *v1.ObjectReference
-	GetPodsFunc                     eviction.ActivePodsFunc
+	GetPodsFunc                      eviction.ActivePodsFunc
-	KillPodFunc                     eviction.KillPodFunc
+	KillPodFunc                      eviction.KillPodFunc
-	SyncNodeStatusFunc              func()
+	SyncNodeStatusFunc               func()
-	ShutdownGracePeriodRequested    time.Duration
+	ShutdownGracePeriodRequested     time.Duration
-	ShutdownGracePeriodCriticalPods time.Duration
+	ShutdownGracePeriodCriticalPods  time.Duration
-	Clock                           clock.Clock
+	ShutdownGracePeriodByPodPriority []kubeletconfig.ShutdownGracePeriodByPodPriority
+	Clock                            clock.Clock
 }
 
 // managerStub is a fake node shutdown managerImpl .

pkg/kubelet/nodeshutdown/nodeshutdown_manager_linux.go

@@ -22,6 +22,7 @@ package nodeshutdown
 
 import (
 	"fmt"
+	"sort"
 	"sync"
 	"time"
 
@@ -29,13 +30,14 @@ import (
 	utilfeature "k8s.io/apiserver/pkg/util/feature"
 	"k8s.io/client-go/tools/record"
 	"k8s.io/klog/v2"
+	"k8s.io/kubernetes/pkg/apis/scheduling"
 	"k8s.io/kubernetes/pkg/features"
+	kubeletconfig "k8s.io/kubernetes/pkg/kubelet/apis/config"
 	kubeletevents "k8s.io/kubernetes/pkg/kubelet/events"
 	"k8s.io/kubernetes/pkg/kubelet/eviction"
 	"k8s.io/kubernetes/pkg/kubelet/lifecycle"
 	"k8s.io/kubernetes/pkg/kubelet/nodeshutdown/systemd"
 	"k8s.io/kubernetes/pkg/kubelet/prober"
-	kubelettypes "k8s.io/kubernetes/pkg/kubelet/types"
 	"k8s.io/utils/clock"
 )
 
@@ -66,8 +68,7 @@ type managerImpl struct {
 	nodeRef      *v1.ObjectReference
 	probeManager prober.Manager
 
-	shutdownGracePeriodRequested    time.Duration
+	shutdownGracePeriodByPodPriority []kubeletconfig.ShutdownGracePeriodByPodPriority
-	shutdownGracePeriodCriticalPods time.Duration
 
 	getPods        eviction.ActivePodsFunc
 	killPodFunc    eviction.KillPodFunc
@@ -84,28 +85,46 @@ type managerImpl struct {
 
 // NewManager returns a new node shutdown manager.
 func NewManager(conf *Config) (Manager, lifecycle.PodAdmitHandler) {
-	if !utilfeature.DefaultFeatureGate.Enabled(features.GracefulNodeShutdown) ||
+	if !utilfeature.DefaultFeatureGate.Enabled(features.GracefulNodeShutdown) {
-		(conf.ShutdownGracePeriodRequested == 0 && conf.ShutdownGracePeriodCriticalPods == 0) {
 		m := managerStub{}
 		return m, m
 	}
+
+	shutdownGracePeriodByPodPriority := conf.ShutdownGracePeriodByPodPriority
+	// Migration from the original configuration
+	if !utilfeature.DefaultFeatureGate.Enabled(features.GracefulNodeShutdownBasedOnPodPriority) ||
+		len(shutdownGracePeriodByPodPriority) == 0 {
+		shutdownGracePeriodByPodPriority = migrateConfig(conf.ShutdownGracePeriodRequested, conf.ShutdownGracePeriodCriticalPods)
+	}
+
+	// Disable if the configuration is empty
+	if len(shutdownGracePeriodByPodPriority) == 0 {
+		m := managerStub{}
+		return m, m
+	}
+
+	// Sort by priority from low to high
+	sort.Slice(shutdownGracePeriodByPodPriority, func(i, j int) bool {
+		return shutdownGracePeriodByPodPriority[i].Priority < shutdownGracePeriodByPodPriority[j].Priority
+	})
+
 	if conf.Clock == nil {
 		conf.Clock = clock.RealClock{}
 	}
 	manager := &managerImpl{
-		probeManager:                    conf.ProbeManager,
+		probeManager:                     conf.ProbeManager,
-		recorder:                        conf.Recorder,
+		recorder:                         conf.Recorder,
-		nodeRef:                         conf.NodeRef,
+		nodeRef:                          conf.NodeRef,
-		getPods:                         conf.GetPodsFunc,
+		getPods:                          conf.GetPodsFunc,
-		killPodFunc:                     conf.KillPodFunc,
+		killPodFunc:                      conf.KillPodFunc,
-		syncNodeStatus:                  conf.SyncNodeStatusFunc,
+		syncNodeStatus:                   conf.SyncNodeStatusFunc,
-		shutdownGracePeriodRequested:    conf.ShutdownGracePeriodRequested,
+		shutdownGracePeriodByPodPriority: shutdownGracePeriodByPodPriority,
-		shutdownGracePeriodCriticalPods: conf.ShutdownGracePeriodCriticalPods,
+		clock:                            conf.Clock,
-		clock:                           conf.Clock,
 	}
 	klog.InfoS("Creating node shutdown manager",
 		"shutdownGracePeriodRequested", conf.ShutdownGracePeriodRequested,
 		"shutdownGracePeriodCriticalPods", conf.ShutdownGracePeriodCriticalPods,
+		"shutdownGracePeriodByPodPriority", shutdownGracePeriodByPodPriority,
 	)
 	return manager, manager
 }
@@ -159,9 +178,9 @@ func (m *managerImpl) start() (chan struct{}, error) {
 		return nil, err
 	}
 
-	// If the logind's InhibitDelayMaxUSec as configured in (logind.conf) is less than shutdownGracePeriodRequested, attempt to update the value to shutdownGracePeriodRequested.
+	// If the logind's InhibitDelayMaxUSec as configured in (logind.conf) is less than periodRequested, attempt to update the value to periodRequested.
-	if m.shutdownGracePeriodRequested > currentInhibitDelay {
+	if periodRequested := m.periodRequested(); periodRequested > currentInhibitDelay {
-		err := m.dbusCon.OverrideInhibitDelay(m.shutdownGracePeriodRequested)
+		err := m.dbusCon.OverrideInhibitDelay(periodRequested)
 		if err != nil {
 			return nil, fmt.Errorf("unable to override inhibit delay by shutdown manager: %v", err)
 		}
@@ -177,8 +196,8 @@ func (m *managerImpl) start() (chan struct{}, error) {
 			return nil, err
 		}
 
-		if m.shutdownGracePeriodRequested > updatedInhibitDelay {
+		if periodRequested > updatedInhibitDelay {
-			return nil, fmt.Errorf("node shutdown manager was unable to update logind InhibitDelayMaxSec to %v (ShutdownGracePeriod), current value of InhibitDelayMaxSec (%v) is less than requested ShutdownGracePeriod", m.shutdownGracePeriodRequested, updatedInhibitDelay)
+			return nil, fmt.Errorf("node shutdown manager was unable to update logind InhibitDelayMaxSec to %v (ShutdownGracePeriod), current value of InhibitDelayMaxSec (%v) is less than requested ShutdownGracePeriod", periodRequested, updatedInhibitDelay)
 		}
 	}
 
@@ -270,54 +289,54 @@ func (m *managerImpl) processShutdownEvent() error {
 	klog.V(1).InfoS("Shutdown manager processing shutdown event")
 	activePods := m.getPods()
 
-	nonCriticalPodGracePeriod := m.shutdownGracePeriodRequested - m.shutdownGracePeriodCriticalPods
+	groups := groupByPriority(m.shutdownGracePeriodByPodPriority, activePods)
+	for _, group := range groups {
+		// If there are no pods in a particular range,
+		// then do not wait for pods in that priority range.
+		if len(group.Pods) == 0 {
+			continue
+		}
 
-	var wg sync.WaitGroup
+		var wg sync.WaitGroup
-	wg.Add(len(activePods))
+		wg.Add(len(group.Pods))
-	for _, pod := range activePods {
+		for _, pod := range group.Pods {
-		go func(pod *v1.Pod) {
+			go func(pod *v1.Pod, group podShutdownGroup) {
-			defer wg.Done()
+				defer wg.Done()
 
-			var gracePeriodOverride int64
+				gracePeriodOverride := group.ShutdownGracePeriodSeconds
-			if kubelettypes.IsCriticalPod(pod) {
-				gracePeriodOverride = int64(m.shutdownGracePeriodCriticalPods.Seconds())
-				m.clock.Sleep(nonCriticalPodGracePeriod)
-			} else {
-				gracePeriodOverride = int64(nonCriticalPodGracePeriod.Seconds())
-			}
 
-			// Stop probes for the pod
+				// Stop probes for the pod
-			m.probeManager.RemovePod(pod)
+				m.probeManager.RemovePod(pod)
 
-			// If the pod's spec specifies a termination gracePeriod which is less than the gracePeriodOverride calculated, use the pod spec termination gracePeriod.
+				// If the pod's spec specifies a termination gracePeriod which is less than the gracePeriodOverride calculated, use the pod spec termination gracePeriod.
-			if pod.Spec.TerminationGracePeriodSeconds != nil && *pod.Spec.TerminationGracePeriodSeconds <= gracePeriodOverride {
+				if pod.Spec.TerminationGracePeriodSeconds != nil && *pod.Spec.TerminationGracePeriodSeconds <= gracePeriodOverride {
-				gracePeriodOverride = *pod.Spec.TerminationGracePeriodSeconds
+					gracePeriodOverride = *pod.Spec.TerminationGracePeriodSeconds
-			}
+				}
 
-			klog.V(1).InfoS("Shutdown manager killing pod with gracePeriod", "pod", klog.KObj(pod), "gracePeriod", gracePeriodOverride)
+				klog.V(1).InfoS("Shutdown manager killing pod with gracePeriod", "pod", klog.KObj(pod), "gracePeriod", gracePeriodOverride)
-			if err := m.killPodFunc(pod, false, &gracePeriodOverride, func(status *v1.PodStatus) {
-				status.Message = nodeShutdownMessage
-				status.Reason = nodeShutdownReason
-			}); err != nil {
-				klog.V(1).InfoS("Shutdown manager failed killing pod", "pod", klog.KObj(pod), "err", err)
-			} else {
-				klog.V(1).InfoS("Shutdown manager finished killing pod", "pod", klog.KObj(pod))
-			}
-		}(pod)
-	}
 
-	c := make(chan struct{})
+				if err := m.killPodFunc(pod, false, &gracePeriodOverride, func(status *v1.PodStatus) {
-	go func() {
+					status.Message = nodeShutdownMessage
-		defer close(c)
+					status.Reason = nodeShutdownReason
-		wg.Wait()
+				}); err != nil {
-	}()
+					klog.V(1).InfoS("Shutdown manager failed killing pod", "pod", klog.KObj(pod), "err", err)
+				} else {
+					klog.V(1).InfoS("Shutdown manager finished killing pod", "pod", klog.KObj(pod))
+				}
+			}(pod, group)
+		}
 
-	// We want to ensure that inhibitLock is released, so only wait up to the shutdownGracePeriodRequested timeout.
+		c := make(chan struct{})
-	select {
+		go func() {
-	case <-c:
+			defer close(c)
-		break
+			wg.Wait()
-	case <-time.After(m.shutdownGracePeriodRequested):
+		}()
-		klog.V(1).InfoS("Shutdown manager pod killing time out", "gracePeriod", m.shutdownGracePeriodRequested)
+
+		select {
+		case <-c:
+		case <-time.After(time.Duration(group.ShutdownGracePeriodSeconds) * time.Second):
+			klog.V(1).InfoS("Shutdown manager pod killing time out", "gracePeriod", group.ShutdownGracePeriodSeconds, "priority", group.Priority)
+		}
 	}
 
 	m.dbusCon.ReleaseInhibitLock(m.inhibitLock)
@@ -325,3 +344,78 @@ func (m *managerImpl) processShutdownEvent() error {
 
 	return nil
 }
+
+func (m *managerImpl) periodRequested() time.Duration {
+	var sum int64
+	for _, period := range m.shutdownGracePeriodByPodPriority {
+		sum += period.ShutdownGracePeriodSeconds
+	}
+	return time.Duration(sum) * time.Second
+}
+
+func migrateConfig(shutdownGracePeriodRequested, shutdownGracePeriodCriticalPods time.Duration) []kubeletconfig.ShutdownGracePeriodByPodPriority {
+	if shutdownGracePeriodRequested == 0 {
+		return nil
+	}
+	defaultPriority := shutdownGracePeriodRequested - shutdownGracePeriodCriticalPods
+	if defaultPriority < 0 {
+		return nil
+	}
+	criticalPriority := shutdownGracePeriodRequested - defaultPriority
+	if criticalPriority < 0 {
+		return nil
+	}
+	return []kubeletconfig.ShutdownGracePeriodByPodPriority{
+		{
+			Priority:                   scheduling.DefaultPriorityWhenNoDefaultClassExists,
+			ShutdownGracePeriodSeconds: int64(defaultPriority / time.Second),
+		},
+		{
+			Priority:                   scheduling.SystemCriticalPriority,
+			ShutdownGracePeriodSeconds: int64(criticalPriority / time.Second),
+		},
+	}
+}
+
+func groupByPriority(shutdownGracePeriodByPodPriority []kubeletconfig.ShutdownGracePeriodByPodPriority, pods []*v1.Pod) []podShutdownGroup {
+	groups := make([]podShutdownGroup, 0, len(shutdownGracePeriodByPodPriority))
+	for _, period := range shutdownGracePeriodByPodPriority {
+		groups = append(groups, podShutdownGroup{
+			ShutdownGracePeriodByPodPriority: period,
+		})
+	}
+
+	for _, pod := range pods {
+		var priority int32
+		if pod.Spec.Priority != nil {
+			priority = *pod.Spec.Priority
+		}
+
+		// Find the group index according to the priority.
+		index := sort.Search(len(groups), func(i int) bool {
+			return groups[i].Priority >= priority
+		})
+
+		// 1. Those higher than the highest priority default to the highest priority
+		// 2. Those lower than the lowest priority default to the lowest priority
+		// 3. Those boundary priority default to the lower priority
+		// if priority of pod is:
+		//   groups[index-1].Priority <= pod priority < groups[index].Priority
+		// in which case we want to pick lower one (i.e index-1)
+		if index == len(groups) {
+			index = len(groups) - 1
+		} else if index < 0 {
+			index = 0
+		} else if index > 0 && groups[index].Priority > priority {
+			index--
+		}
+
+		groups[index].Pods = append(groups[index].Pods, pod)
+	}
+	return groups
+}
+
+type podShutdownGroup struct {
+	kubeletconfig.ShutdownGracePeriodByPodPriority
+	Pods []*v1.Pod
+}

pkg/kubelet/nodeshutdown/nodeshutdown_manager_linux_test.go

@@ -35,6 +35,7 @@ import (
 	featuregatetesting "k8s.io/component-base/featuregate/testing"
 	"k8s.io/kubernetes/pkg/apis/scheduling"
 	pkgfeatures "k8s.io/kubernetes/pkg/features"
+	kubeletconfig "k8s.io/kubernetes/pkg/kubelet/apis/config"
 	"k8s.io/kubernetes/pkg/kubelet/nodeshutdown/systemd"
 	probetest "k8s.io/kubernetes/pkg/kubelet/prober/testing"
 	testingclock "k8s.io/utils/clock/testing"
@@ -81,12 +82,7 @@ func (f *fakeDbus) OverrideInhibitDelay(inhibitDelayMax time.Duration) error {
 	return nil
 }
 
-func makePod(name string, criticalPod bool, terminationGracePeriod *int64) *v1.Pod {
+func makePod(name string, priority int32, terminationGracePeriod *int64) *v1.Pod {
-	var priority int32
-	if criticalPod {
-		priority = scheduling.SystemCriticalPriority
-	}
-
 	return &v1.Pod{
 		ObjectMeta: metav1.ObjectMeta{
 			Name: name,
@@ -104,15 +100,15 @@ func TestManager(t *testing.T) {
 	defer func() {
 		systemDbus = systemDbusTmp
 	}()
-	normalPodNoGracePeriod := makePod("normal-pod-nil-grace-period", false /* criticalPod */, nil /* terminationGracePeriod */)
+	normalPodNoGracePeriod := makePod("normal-pod-nil-grace-period", scheduling.DefaultPriorityWhenNoDefaultClassExists, nil /* terminationGracePeriod */)
-	criticalPodNoGracePeriod := makePod("critical-pod-nil-grace-period", true /* criticalPod */, nil /* terminationGracePeriod */)
+	criticalPodNoGracePeriod := makePod("critical-pod-nil-grace-period", scheduling.SystemCriticalPriority, nil /* terminationGracePeriod */)
 
 	shortGracePeriod := int64(2)
-	normalPodGracePeriod := makePod("normal-pod-grace-period", false /* criticalPod */, &shortGracePeriod /* terminationGracePeriod */)
+	normalPodGracePeriod := makePod("normal-pod-grace-period", scheduling.DefaultPriorityWhenNoDefaultClassExists, &shortGracePeriod /* terminationGracePeriod */)
-	criticalPodGracePeriod := makePod("critical-pod-grace-period", true /* criticalPod */, &shortGracePeriod /* terminationGracePeriod */)
+	criticalPodGracePeriod := makePod("critical-pod-grace-period", scheduling.SystemCriticalPriority, &shortGracePeriod /* terminationGracePeriod */)
 
 	longGracePeriod := int64(1000)
-	normalPodLongGracePeriod := makePod("normal-pod-long-grace-period", false /* criticalPod */, &longGracePeriod /* terminationGracePeriod */)
+	normalPodLongGracePeriod := makePod("normal-pod-long-grace-period", scheduling.DefaultPriorityWhenNoDefaultClassExists, &longGracePeriod /* terminationGracePeriod */)
 
 	var tests = []struct {
 		desc                             string
@@ -256,7 +252,9 @@ func TestManager(t *testing.T) {
 			lock.Unlock()
 
 			if tc.expectedError != nil {
-				if !strings.Contains(err.Error(), tc.expectedError.Error()) {
+				if err == nil {
+					t.Errorf("unexpected error message. Got: <nil> want %s", tc.expectedError.Error())
+				} else if !strings.Contains(err.Error(), tc.expectedError.Error()) {
 					t.Errorf("unexpected error message. Got: %s want %s", err.Error(), tc.expectedError.Error())
 				}
 			} else {
@@ -266,7 +264,11 @@ func TestManager(t *testing.T) {
 				assert.Equal(t, manager.Admit(nil).Admit, true)
 
 				// Send fake shutdown event
-				fakeShutdownChan <- true
+				select {
+				case fakeShutdownChan <- true:
+				case <-time.After(1 * time.Second):
+					t.Fatal()
+				}
 
 				// Wait for all the pods to be killed
 				killedPodsToGracePeriods := map[string]int64{}
@@ -413,3 +415,196 @@ func TestRestart(t *testing.T) {
 		shutdownChanMut.Unlock()
 	}
 }
+
+func Test_migrateConfig(t *testing.T) {
+	type shutdownConfig struct {
+		shutdownGracePeriodRequested    time.Duration
+		shutdownGracePeriodCriticalPods time.Duration
+	}
+	tests := []struct {
+		name string
+		args shutdownConfig
+		want []kubeletconfig.ShutdownGracePeriodByPodPriority
+	}{
+		{
+			name: "both shutdownGracePeriodRequested and shutdownGracePeriodCriticalPods",
+			args: shutdownConfig{
+				shutdownGracePeriodRequested:    300 * time.Second,
+				shutdownGracePeriodCriticalPods: 120 * time.Second,
+			},
+			want: []kubeletconfig.ShutdownGracePeriodByPodPriority{
+				{
+					Priority:                   scheduling.DefaultPriorityWhenNoDefaultClassExists,
+					ShutdownGracePeriodSeconds: 180,
+				},
+				{
+					Priority:                   scheduling.SystemCriticalPriority,
+					ShutdownGracePeriodSeconds: 120,
+				},
+			},
+		},
+		{
+			name: "only shutdownGracePeriodRequested",
+			args: shutdownConfig{
+				shutdownGracePeriodRequested:    100 * time.Second,
+				shutdownGracePeriodCriticalPods: 0 * time.Second,
+			},
+			want: []kubeletconfig.ShutdownGracePeriodByPodPriority{
+				{
+					Priority:                   scheduling.DefaultPriorityWhenNoDefaultClassExists,
+					ShutdownGracePeriodSeconds: 100,
+				},
+				{
+					Priority:                   scheduling.SystemCriticalPriority,
+					ShutdownGracePeriodSeconds: 0,
+				},
+			},
+		},
+		{
+			name: "empty configuration",
+			args: shutdownConfig{
+				shutdownGracePeriodRequested:    0 * time.Second,
+				shutdownGracePeriodCriticalPods: 0 * time.Second,
+			},
+			want: nil,
+		},
+		{
+			name: "wrong configuration",
+			args: shutdownConfig{
+				shutdownGracePeriodRequested:    1 * time.Second,
+				shutdownGracePeriodCriticalPods: 100 * time.Second,
+			},
+			want: nil,
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := migrateConfig(tt.args.shutdownGracePeriodRequested, tt.args.shutdownGracePeriodCriticalPods); !assert.Equal(t, tt.want, got) {
+				t.Errorf("migrateConfig() = %v, want %v", got, tt.want)
+			}
+		})
+	}
+}
+
+func Test_groupByPriority(t *testing.T) {
+	type args struct {
+		shutdownGracePeriodByPodPriority []kubeletconfig.ShutdownGracePeriodByPodPriority
+		pods                             []*v1.Pod
+	}
+	tests := []struct {
+		name string
+		args args
+		want []podShutdownGroup
+	}{
+		{
+			name: "migrate config",
+			args: args{
+				shutdownGracePeriodByPodPriority: migrateConfig(300*time.Second /* shutdownGracePeriodRequested */, 120*time.Second /* shutdownGracePeriodCriticalPods */),
+				pods: []*v1.Pod{
+					makePod("normal-pod", scheduling.DefaultPriorityWhenNoDefaultClassExists, nil),
+					makePod("highest-user-definable-pod", scheduling.HighestUserDefinablePriority, nil),
+					makePod("critical-pod", scheduling.SystemCriticalPriority, nil),
+				},
+			},
+			want: []podShutdownGroup{
+				{
+					ShutdownGracePeriodByPodPriority: kubeletconfig.ShutdownGracePeriodByPodPriority{
+						Priority:                   scheduling.DefaultPriorityWhenNoDefaultClassExists,
+						ShutdownGracePeriodSeconds: 180,
+					},
+					Pods: []*v1.Pod{
+						makePod("normal-pod", scheduling.DefaultPriorityWhenNoDefaultClassExists, nil),
+						makePod("highest-user-definable-pod", scheduling.HighestUserDefinablePriority, nil),
+					},
+				},
+				{
+					ShutdownGracePeriodByPodPriority: kubeletconfig.ShutdownGracePeriodByPodPriority{
+						Priority:                   scheduling.SystemCriticalPriority,
+						ShutdownGracePeriodSeconds: 120,
+					},
+					Pods: []*v1.Pod{
+						makePod("critical-pod", scheduling.SystemCriticalPriority, nil),
+					},
+				},
+			},
+		},
+		{
+			name: "pod priority",
+			args: args{
+				shutdownGracePeriodByPodPriority: []kubeletconfig.ShutdownGracePeriodByPodPriority{
+					{
+						Priority:                   1,
+						ShutdownGracePeriodSeconds: 10,
+					},
+					{
+						Priority:                   2,
+						ShutdownGracePeriodSeconds: 20,
+					},
+					{
+						Priority:                   3,
+						ShutdownGracePeriodSeconds: 30,
+					},
+					{
+						Priority:                   4,
+						ShutdownGracePeriodSeconds: 40,
+					},
+				},
+				pods: []*v1.Pod{
+					makePod("pod-0", 0, nil),
+					makePod("pod-1", 1, nil),
+					makePod("pod-2", 2, nil),
+					makePod("pod-3", 3, nil),
+					makePod("pod-4", 4, nil),
+					makePod("pod-5", 5, nil),
+				},
+			},
+			want: []podShutdownGroup{
+				{
+					ShutdownGracePeriodByPodPriority: kubeletconfig.ShutdownGracePeriodByPodPriority{
+						Priority:                   1,
+						ShutdownGracePeriodSeconds: 10,
+					},
+					Pods: []*v1.Pod{
+						makePod("pod-0", 0, nil),
+						makePod("pod-1", 1, nil),
+					},
+				},
+				{
+					ShutdownGracePeriodByPodPriority: kubeletconfig.ShutdownGracePeriodByPodPriority{
+						Priority:                   2,
+						ShutdownGracePeriodSeconds: 20,
+					},
+					Pods: []*v1.Pod{
+						makePod("pod-2", 2, nil),
+					},
+				},
+				{
+					ShutdownGracePeriodByPodPriority: kubeletconfig.ShutdownGracePeriodByPodPriority{
+						Priority:                   3,
+						ShutdownGracePeriodSeconds: 30,
+					},
+					Pods: []*v1.Pod{
+						makePod("pod-3", 3, nil),
+					},
+				},
+				{
+					ShutdownGracePeriodByPodPriority: kubeletconfig.ShutdownGracePeriodByPodPriority{
+						Priority:                   4,
+						ShutdownGracePeriodSeconds: 40,
+					},
+					Pods: []*v1.Pod{
+						makePod("pod-4", 4, nil),
+						makePod("pod-5", 5, nil),
+					},
+				},
+			},
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if got := groupByPriority(tt.args.shutdownGracePeriodByPodPriority, tt.args.pods); !assert.Equal(t, tt.want, got) {
+				t.Errorf("groupByPriority() = %v, want %v", got, tt.want)
+			}
+		})
+	}
+}