When a node becomes unreachable, do not evict DaemonSet-managed pods.

Part of the graduation requirement for DaemonSet spelled out in #15310.

pkg/controller/node/nodecontroller.go

@@ -111,6 +111,9 @@ type NodeController struct {
 	// Node framework and store
 	nodeController *framework.Controller
 	nodeStore      cache.StoreToNodeLister
+	// DaemonSet framework and store
+	daemonSetController *framework.Controller
+	daemonSetStore      cache.StoreToDaemonSetLister
 
 	forcefullyDeletePod func(*api.Pod)
 }
@@ -191,6 +194,19 @@ func NewNodeController(
 		controller.NoResyncPeriodFunc(),
 		framework.ResourceEventHandlerFuncs{},
 	)
+	nc.daemonSetStore.Store, nc.daemonSetController = framework.NewInformer(
+		&cache.ListWatch{
+			ListFunc: func(options api.ListOptions) (runtime.Object, error) {
+				return nc.kubeClient.Extensions().DaemonSets(api.NamespaceAll).List(options)
+			},
+			WatchFunc: func(options api.ListOptions) (watch.Interface, error) {
+				return nc.kubeClient.Extensions().DaemonSets(api.NamespaceAll).Watch(options)
+			},
+		},
+		&api.Node{},
+		controller.NoResyncPeriodFunc(),
+		framework.ResourceEventHandlerFuncs{},
+	)
 	return nc
 }
 
@@ -198,6 +214,8 @@ func NewNodeController(
 func (nc *NodeController) Run(period time.Duration) {
 	go nc.nodeController.Run(util.NeverStop)
 	go nc.podController.Run(util.NeverStop)
+	go nc.daemonSetController.Run(util.NeverStop)
+
 	// Incorporate the results of node status pushed from kubelet to master.
 	go util.Until(func() {
 		if err := nc.monitorNodeStatus(); err != nil {
@@ -752,6 +770,11 @@ func (nc *NodeController) deletePods(nodeName string) (bool, error) {
 		if pod.DeletionGracePeriodSeconds != nil {
 			continue
 		}
+		// if the pod is managed by a daemonset, ignore it
+		_, err := nc.daemonSetStore.GetPodDaemonSets(&pod)
+		if err == nil { // No error means at least one daemonset was found
+			continue
+		}
 
 		glog.V(2).Infof("Starting deletion of pod %v", pod.Name)
 		nc.recorder.Eventf(&pod, api.EventTypeNormal, "NodeControllerEviction", "Marking for deletion Pod %s from Node %s", pod.Name, nodeName)

pkg/controller/node/nodecontroller_test.go

@@ -26,6 +26,7 @@ import (
 	apierrors "k8s.io/kubernetes/pkg/api/errors"
 	"k8s.io/kubernetes/pkg/api/resource"
 	"k8s.io/kubernetes/pkg/api/unversioned"
+	"k8s.io/kubernetes/pkg/apis/extensions"
 	"k8s.io/kubernetes/pkg/client/cache"
 	client "k8s.io/kubernetes/pkg/client/unversioned"
 	"k8s.io/kubernetes/pkg/client/unversioned/testclient"
@@ -144,6 +145,7 @@ func TestMonitorNodeStatusEvictPods(t *testing.T) {
 
 	table := []struct {
 		fakeNodeHandler   *FakeNodeHandler
+		daemonSets        []extensions.DaemonSet
 		timeToPass        time.Duration
 		newNodeStatus     api.NodeStatus
 		expectedEvictPods bool
@@ -162,6 +164,7 @@ func TestMonitorNodeStatusEvictPods(t *testing.T) {
 				},
 				Fake: testclient.NewSimpleFake(&api.PodList{Items: []api.Pod{*newPod("pod0", "node0")}}),
 			},
+			daemonSets:        nil,
 			timeToPass:        0,
 			newNodeStatus:     api.NodeStatus{},
 			expectedEvictPods: false,
@@ -190,6 +193,7 @@ func TestMonitorNodeStatusEvictPods(t *testing.T) {
 				},
 				Fake: testclient.NewSimpleFake(&api.PodList{Items: []api.Pod{*newPod("pod0", "node0")}}),
 			},
+			daemonSets: nil,
 			timeToPass: evictionTimeout,
 			newNodeStatus: api.NodeStatus{
 				Conditions: []api.NodeCondition{
@@ -205,6 +209,72 @@ func TestMonitorNodeStatusEvictPods(t *testing.T) {
 			expectedEvictPods: false,
 			description:       "Node created long time ago, and kubelet posted NotReady for a short period of time.",
 		},
+		// Pod is ds-managed, and kubelet posted NotReady for a long period of time.
+		{
+			fakeNodeHandler: &FakeNodeHandler{
+				Existing: []*api.Node{
+					{
+						ObjectMeta: api.ObjectMeta{
+							Name:              "node0",
+							CreationTimestamp: unversioned.Date(2012, 1, 1, 0, 0, 0, 0, time.UTC),
+						},
+						Status: api.NodeStatus{
+							Conditions: []api.NodeCondition{
+								{
+									Type:               api.NodeReady,
+									Status:             api.ConditionFalse,
+									LastHeartbeatTime:  unversioned.Date(2015, 1, 1, 12, 0, 0, 0, time.UTC),
+									LastTransitionTime: unversioned.Date(2015, 1, 1, 12, 0, 0, 0, time.UTC),
+								},
+							},
+						},
+					},
+				},
+				Fake: testclient.NewSimpleFake(
+					&api.PodList{
+						Items: []api.Pod{
+							{
+								ObjectMeta: api.ObjectMeta{
+									Name:      "pod0",
+									Namespace: "default",
+									Labels:    map[string]string{"daemon": "yes"},
+								},
+								Spec: api.PodSpec{
+									NodeName: "node0",
+								},
+							},
+						},
+					},
+				),
+			},
+			daemonSets: []extensions.DaemonSet{
+				{
+					ObjectMeta: api.ObjectMeta{
+						Name:      "ds0",
+						Namespace: "default",
+					},
+					Spec: extensions.DaemonSetSpec{
+						Selector: &extensions.LabelSelector{
+							MatchLabels: map[string]string{"daemon": "yes"},
+						},
+					},
+				},
+			},
+			timeToPass: time.Hour,
+			newNodeStatus: api.NodeStatus{
+				Conditions: []api.NodeCondition{
+					{
+						Type:   api.NodeReady,
+						Status: api.ConditionFalse,
+						// Node status has just been updated, and is NotReady for 1hr.
+						LastHeartbeatTime:  unversioned.Date(2015, 1, 1, 12, 59, 0, 0, time.UTC),
+						LastTransitionTime: unversioned.Date(2015, 1, 1, 12, 0, 0, 0, time.UTC),
+					},
+				},
+			},
+			expectedEvictPods: false,
+			description:       "Pod is ds-managed, and kubelet posted NotReady for a long period of time.",
+		},
 		// Node created long time ago, and kubelet posted NotReady for a long period of time.
 		{
 			fakeNodeHandler: &FakeNodeHandler{
@@ -228,6 +298,7 @@ func TestMonitorNodeStatusEvictPods(t *testing.T) {
 				},
 				Fake: testclient.NewSimpleFake(&api.PodList{Items: []api.Pod{*newPod("pod0", "node0")}}),
 			},
+			daemonSets: nil,
 			timeToPass: time.Hour,
 			newNodeStatus: api.NodeStatus{
 				Conditions: []api.NodeCondition{
@@ -266,6 +337,7 @@ func TestMonitorNodeStatusEvictPods(t *testing.T) {
 				},
 				Fake: testclient.NewSimpleFake(&api.PodList{Items: []api.Pod{*newPod("pod0", "node0")}}),
 			},
+			daemonSets: nil,
 			timeToPass: evictionTimeout - testNodeMonitorGracePeriod,
 			newNodeStatus: api.NodeStatus{
 				Conditions: []api.NodeCondition{
@@ -304,6 +376,7 @@ func TestMonitorNodeStatusEvictPods(t *testing.T) {
 				},
 				Fake: testclient.NewSimpleFake(&api.PodList{Items: []api.Pod{*newPod("pod0", "node0")}}),
 			},
+			daemonSets: nil,
 			timeToPass: 60 * time.Minute,
 			newNodeStatus: api.NodeStatus{
 				Conditions: []api.NodeCondition{
@@ -326,6 +399,9 @@ func TestMonitorNodeStatusEvictPods(t *testing.T) {
 			evictionTimeout, util.NewFakeRateLimiter(), util.NewFakeRateLimiter(), testNodeMonitorGracePeriod,
 			testNodeStartupGracePeriod, testNodeMonitorPeriod, nil, false)
 		nodeController.now = func() unversioned.Time { return fakeNow }
+		for _, ds := range item.daemonSets {
+			nodeController.daemonSetStore.Add(&ds)
+		}
 		if err := nodeController.monitorNodeStatus(); err != nil {
 			t.Errorf("unexpected error: %v", err)
 		}