Merge pull request #7668 from bprashanth/scheduling_minions

Scheduler ignores nodes that are in a bad state

pkg/client/cache/listers.go

@@ -21,6 +21,7 @@ import (
 
 	"github.com/GoogleCloudPlatform/kubernetes/pkg/api"
 	"github.com/GoogleCloudPlatform/kubernetes/pkg/labels"
+	"github.com/golang/glog"
 )
 
 //  TODO: generate these classes and methods for all resources of interest using
@@ -106,6 +107,54 @@ func (s *StoreToNodeLister) List() (machines api.NodeList, err error) {
 	return machines, nil
 }
 
+// NodeCondition returns a storeToNodeConditionLister
+func (s *StoreToNodeLister) NodeCondition(conditionType api.NodeConditionType, conditionStatus api.ConditionStatus) storeToNodeConditionLister {
+	// TODO: Move this filtering server side. Currently our selectors don't facilitate searching through a list so we
+	// have the reflector filter out the Unschedulable field and sift through node conditions in the lister.
+	return storeToNodeConditionLister{s.Store, conditionType, conditionStatus}
+}
+
+// storeToNodeConditionLister filters and returns nodes matching the given type and status from the store.
+type storeToNodeConditionLister struct {
+	store           Store
+	conditionType   api.NodeConditionType
+	conditionStatus api.ConditionStatus
+}
+
+// List returns a list of nodes that match the condition type/status in the storeToNodeConditionLister.
+func (s storeToNodeConditionLister) List() (nodes api.NodeList, err error) {
+	for _, m := range s.store.List() {
+		node := *m.(*api.Node)
+
+		// We currently only use a conditionType of "Ready". If the kubelet doesn't
+		// periodically report the status of a node, the nodecontroller sets its
+		// ConditionStatus to "Unknown". If the kubelet thinks a node is unhealthy
+		// it can (in theory) set its ConditionStatus to "False".
+		var nodeCondition *api.NodeCondition
+
+		// Get the last condition of the required type
+		for _, cond := range node.Status.Conditions {
+			if cond.Type == s.conditionType {
+				nodeCondition = &cond
+			} else {
+				glog.V(4).Infof("Ignoring condition type %v for node %v", cond.Type, node.Name)
+			}
+		}
+
+		// Check that the condition has the required status
+		if nodeCondition != nil {
+			if nodeCondition.Status == s.conditionStatus {
+				nodes.Items = append(nodes.Items, node)
+			} else {
+				glog.V(4).Infof("Ignoring node %v with condition status %v", node.Name, nodeCondition.Status)
+			}
+		} else {
+			glog.V(2).Infof("Node %s doesn't have conditions of type %v", node.Name, s.conditionType)
+		}
+	}
+	return
+}
+
 // TODO Move this back to scheduler as a helper function that takes a Store,
 // rather than a method of StoreToNodeLister.
 // GetNodeInfo returns cached data for the minion 'id'.

plugin/pkg/scheduler/factory/factory.go

@@ -64,9 +64,10 @@ func NewConfigFactory(client *client.Client) *ConfigFactory {
 		Client:             client,
 		PodQueue:           cache.NewFIFO(cache.MetaNamespaceKeyFunc),
 		ScheduledPodLister: &cache.StoreToPodLister{},
-		NodeLister:         &cache.StoreToNodeLister{cache.NewStore(cache.MetaNamespaceKeyFunc)},
+		// Only nodes in the "Ready" condition with status == "True" are schedulable
-		ServiceLister:      &cache.StoreToServiceLister{cache.NewStore(cache.MetaNamespaceKeyFunc)},
+		NodeLister:     &cache.StoreToNodeLister{cache.NewStore(cache.MetaNamespaceKeyFunc)},
-		StopEverything:     make(chan struct{}),
+		ServiceLister:  &cache.StoreToServiceLister{cache.NewStore(cache.MetaNamespaceKeyFunc)},
+		StopEverything: make(chan struct{}),
 	}
 	modeler := scheduler.NewSimpleModeler(&cache.StoreToPodLister{c.PodQueue}, c.ScheduledPodLister)
 	c.modeler = modeler
@@ -150,8 +151,9 @@ func (f *ConfigFactory) CreateFromKeys(predicateKeys, priorityKeys util.StringSe
 	pluginArgs := PluginFactoryArgs{
 		PodLister:     f.PodLister,
 		ServiceLister: f.ServiceLister,
-		NodeLister:    f.NodeLister,
+		// All fit predicates only need to consider schedulable nodes.
-		NodeInfo:      f.NodeLister,
+		NodeLister: f.NodeLister.NodeCondition(api.NodeReady, api.ConditionTrue),
+		NodeInfo:   f.NodeLister,
 	}
 	predicateFuncs, err := getFitPredicateFunctions(predicateKeys, pluginArgs)
 	if err != nil {
@@ -191,8 +193,9 @@ func (f *ConfigFactory) CreateFromKeys(predicateKeys, priorityKeys util.StringSe
 	}
 
 	return &scheduler.Config{
-		Modeler:      f.modeler,
+		Modeler: f.modeler,
-		MinionLister: f.NodeLister,
+		// The scheduler only needs to consider schedulable nodes.
+		MinionLister: f.NodeLister.NodeCondition(api.NodeReady, api.ConditionTrue),
 		Algorithm:    algo,
 		Binder:       &binder{f.Client},
 		NextPod: func() *api.Pod {

test/integration/scheduler_test.go

@@ -21,6 +21,7 @@ package integration
 // This file tests the scheduler.
 
 import (
+	"fmt"
 	"net/http"
 	"net/http/httptest"
 	"testing"
@@ -31,9 +32,11 @@ import (
 	"github.com/GoogleCloudPlatform/kubernetes/pkg/api/testapi"
 	"github.com/GoogleCloudPlatform/kubernetes/pkg/apiserver"
 	"github.com/GoogleCloudPlatform/kubernetes/pkg/client"
+	"github.com/GoogleCloudPlatform/kubernetes/pkg/client/cache"
 	"github.com/GoogleCloudPlatform/kubernetes/pkg/client/record"
 	"github.com/GoogleCloudPlatform/kubernetes/pkg/master"
 	"github.com/GoogleCloudPlatform/kubernetes/pkg/tools/etcdtest"
+	"github.com/GoogleCloudPlatform/kubernetes/pkg/util"
 	"github.com/GoogleCloudPlatform/kubernetes/pkg/util/wait"
 	"github.com/GoogleCloudPlatform/kubernetes/plugin/pkg/admission/admit"
 	"github.com/GoogleCloudPlatform/kubernetes/plugin/pkg/scheduler"
@@ -45,6 +48,13 @@ func init() {
 	requireEtcd()
 }
 
+type nodeMutationFunc func(t *testing.T, n *api.Node, nodeStore cache.Store, c *client.Client)
+
+type nodeStateManager struct {
+	makeSchedulable   nodeMutationFunc
+	makeUnSchedulable nodeMutationFunc
+}
+
 func TestUnschedulableNodes(t *testing.T) {
 	helper, err := master.NewEtcdHelper(newEtcdClient(), testapi.Version(), etcdtest.PathPrefix())
 	if err != nil {
@@ -69,21 +79,21 @@ func TestUnschedulableNodes(t *testing.T) {
 		AdmissionControl:  admit.NewAlwaysAdmit(),
 	})
 
-	client := client.NewOrDie(&client.Config{Host: s.URL, Version: testapi.Version()})
+	restClient := client.NewOrDie(&client.Config{Host: s.URL, Version: testapi.Version()})
 
-	schedulerConfigFactory := factory.NewConfigFactory(client)
+	schedulerConfigFactory := factory.NewConfigFactory(restClient)
 	schedulerConfig, err := schedulerConfigFactory.Create()
 	if err != nil {
 		t.Fatalf("Couldn't create scheduler config: %v", err)
 	}
 	eventBroadcaster := record.NewBroadcaster()
 	schedulerConfig.Recorder = eventBroadcaster.NewRecorder(api.EventSource{Component: "scheduler"})
-	eventBroadcaster.StartRecordingToSink(client.Events(""))
+	eventBroadcaster.StartRecordingToSink(restClient.Events(""))
 	scheduler.New(schedulerConfig).Run()
 
 	defer close(schedulerConfig.StopEverything)
 
-	DoTestUnschedulableNodes(t, client)
+	DoTestUnschedulableNodes(t, restClient, schedulerConfigFactory.NodeLister.Store)
 }
 
 func podScheduled(c *client.Client, podNamespace, podName string) wait.ConditionFunc {
@@ -103,50 +113,170 @@ func podScheduled(c *client.Client, podNamespace, podName string) wait.Condition
 	}
 }
 
-func DoTestUnschedulableNodes(t *testing.T, client *client.Client) {
+// Wait till the passFunc confirms that the object it expects to see is in the store.
-	node := &api.Node{
+// Used to observe reflected events.
-		ObjectMeta: api.ObjectMeta{Name: "node"},
+func waitForReflection(s cache.Store, key string, passFunc func(n interface{}) bool) error {
-		Spec:       api.NodeSpec{Unschedulable: true},
+	return wait.Poll(time.Millisecond*10, time.Second*20, func() (bool, error) {
-	}
+		if n, _, err := s.GetByKey(key); err == nil && passFunc(n) {
-	if _, err := client.Nodes().Create(node); err != nil {
+			return true, nil
-		t.Fatalf("Failed to create node: %v", err)
+		}
-	}
+		return false, nil
+	})
+}
 
-	pod := &api.Pod{
+func DoTestUnschedulableNodes(t *testing.T, restClient *client.Client, nodeStore cache.Store) {
-		ObjectMeta: api.ObjectMeta{Name: "my-pod"},
+	goodCondition := api.NodeCondition{
-		Spec: api.PodSpec{
+		Type:              api.NodeReady,
-			Containers: []api.Container{{Name: "container", Image: "kubernetes/pause:go"}},
+		Status:            api.ConditionTrue,
+		Reason:            fmt.Sprintf("schedulable condition"),
+		LastHeartbeatTime: util.Time{time.Now()},
+	}
+	badCondition := api.NodeCondition{
+		Type:              api.NodeReady,
+		Status:            api.ConditionUnknown,
+		Reason:            fmt.Sprintf("unschedulable condition"),
+		LastHeartbeatTime: util.Time{time.Now()},
+	}
+	// Create a new schedulable node, since we're first going to apply
+	// the unschedulable condition and verify that pods aren't scheduled.
+	node := &api.Node{
+		ObjectMeta: api.ObjectMeta{Name: "node-scheduling-test-node"},
+		Spec:       api.NodeSpec{Unschedulable: false},
+		Status: api.NodeStatus{
+			Conditions: []api.NodeCondition{goodCondition},
 		},
 	}
-	myPod, err := client.Pods(api.NamespaceDefault).Create(pod)
+	nodeKey, err := cache.MetaNamespaceKeyFunc(node)
 	if err != nil {
-		t.Fatalf("Failed to create pod: %v", err)
+		t.Fatalf("Couldn't retrieve key for node %v", node.Name)
-	}
-	// There are no schedulable nodes - the pod shouldn't be scheduled.
-	err = wait.Poll(time.Second, time.Second*10, podScheduled(client, myPod.Namespace, myPod.Name))
-	if err == nil {
-		t.Errorf("Pod scheduled successfully on unschedulable nodes")
-	}
-	if err != wait.ErrWaitTimeout {
-		t.Errorf("Failed while waiting for scheduled pod: %v", err)
 	}
 
-	// Make the node schedulable and wait until the pod is scheduled.
+	// The test does the following for each nodeStateManager in this list:
-	newNode, err := client.Nodes().Get(node.Name)
+	//	1. Create a new node
-	if err != nil {
+	//	2. Apply the makeUnSchedulable function
-		t.Fatalf("Failed to get node: %v", err)
+	//	3. Create a new pod
-	}
+	//  4. Check that the pod doesn't get assigned to the node
-	newNode.Spec.Unschedulable = false
+	//  5. Apply the schedulable function
-	if _, err = client.Nodes().Update(newNode); err != nil {
+	//  6. Check that the pod *does* get assigned to the node
-		t.Fatalf("Failed to update node: %v", err)
+	//  7. Delete the pod and node.
-	}
+
-	err = wait.Poll(time.Second, time.Second*10, podScheduled(client, myPod.Namespace, myPod.Name))
+	nodeModifications := []nodeStateManager{
-	if err != nil {
+		// Test node.Spec.Unschedulable=true/false
-		t.Errorf("Failed to schedule a pod: %v", err)
+		{
+			makeUnSchedulable: func(t *testing.T, n *api.Node, s cache.Store, c *client.Client) {
+				n.Spec.Unschedulable = true
+				if _, err := c.Nodes().Update(n); err != nil {
+					t.Fatalf("Failed to update node with unschedulable=true: %v", err)
+				}
+				err = waitForReflection(s, nodeKey, func(node interface{}) bool {
+					// An unschedulable node should get deleted from the store
+					return node == nil
+				})
+				if err != nil {
+					t.Fatalf("Failed to observe reflected update for setting unschedulable=true: %v", err)
+				}
+			},
+			makeSchedulable: func(t *testing.T, n *api.Node, s cache.Store, c *client.Client) {
+				n.Spec.Unschedulable = false
+				if _, err := c.Nodes().Update(n); err != nil {
+					t.Fatalf("Failed to update node with unschedulable=false: %v", err)
+				}
+				err = waitForReflection(s, nodeKey, func(node interface{}) bool {
+					return node != nil && node.(*api.Node).Spec.Unschedulable == false
+				})
+				if err != nil {
+					t.Fatalf("Failed to observe reflected update for setting unschedulable=false: %v", err)
+				}
+			},
+		},
+		// Test node.Status.Conditions=ConditionTrue/Unknown
+		{
+			makeUnSchedulable: func(t *testing.T, n *api.Node, s cache.Store, c *client.Client) {
+				n.Status = api.NodeStatus{
+					Conditions: []api.NodeCondition{badCondition},
+				}
+				if _, err = c.Nodes().UpdateStatus(n); err != nil {
+					t.Fatalf("Failed to update node with bad status condition: %v", err)
+				}
+				err = waitForReflection(s, nodeKey, func(node interface{}) bool {
+					return node != nil && node.(*api.Node).Status.Conditions[0].Status == api.ConditionUnknown
+				})
+				if err != nil {
+					t.Fatalf("Failed to observe reflected update for status condition update: %v", err)
+				}
+			},
+			makeSchedulable: func(t *testing.T, n *api.Node, s cache.Store, c *client.Client) {
+				n.Status = api.NodeStatus{
+					Conditions: []api.NodeCondition{goodCondition},
+				}
+				if _, err = c.Nodes().UpdateStatus(n); err != nil {
+					t.Fatalf("Failed to update node with healthy status condition: %v", err)
+				}
+				waitForReflection(s, nodeKey, func(node interface{}) bool {
+					return node != nil && node.(*api.Node).Status.Conditions[0].Status == api.ConditionTrue
+				})
+				if err != nil {
+					t.Fatalf("Failed to observe reflected update for status condition update: %v", err)
+				}
+			},
+		},
 	}
 
-	err = client.Pods(api.NamespaceDefault).Delete(myPod.Name, nil)
+	for i, mod := range nodeModifications {
-	if err != nil {
+		unSchedNode, err := restClient.Nodes().Create(node)
-		t.Errorf("Failed to delete pod: %v", err)
+		if err != nil {
+			t.Fatalf("Failed to create node: %v", err)
+		}
+
+		// Apply the unschedulable modification to the node, and wait for the reflection
+		mod.makeUnSchedulable(t, unSchedNode, nodeStore, restClient)
+
+		// Create the new pod, note that this needs to happen post unschedulable
+		// modification or we have a race in the test.
+		pod := &api.Pod{
+			ObjectMeta: api.ObjectMeta{Name: "node-scheduling-test-pod"},
+			Spec: api.PodSpec{
+				Containers: []api.Container{{Name: "container", Image: "kubernetes/pause:go"}},
+			},
+		}
+		myPod, err := restClient.Pods(api.NamespaceDefault).Create(pod)
+		if err != nil {
+			t.Fatalf("Failed to create pod: %v", err)
+		}
+
+		// There are no schedulable nodes - the pod shouldn't be scheduled.
+		err = wait.Poll(time.Second, time.Second*10, podScheduled(restClient, myPod.Namespace, myPod.Name))
+		if err == nil {
+			t.Errorf("Pod scheduled successfully on unschedulable nodes")
+		}
+		if err != wait.ErrWaitTimeout {
+			t.Errorf("Test %d: failed while trying to confirm the pod does not get scheduled on the node: %v", err)
+		} else {
+			t.Logf("Test %d: Pod did not get scheduled on an unschedulable node", i)
+		}
+
+		// Apply the schedulable modification to the node, and wait for the reflection
+		schedNode, err := restClient.Nodes().Get(unSchedNode.Name)
+		if err != nil {
+			t.Fatalf("Failed to get node: %v", err)
+		}
+		mod.makeSchedulable(t, schedNode, nodeStore, restClient)
+
+		// Wait until the pod is scheduled.
+		err = wait.Poll(time.Second, time.Second*10, podScheduled(restClient, myPod.Namespace, myPod.Name))
+		if err != nil {
+			t.Errorf("Test %d: failed to schedule a pod: %v", err)
+		} else {
+			t.Logf("Test %d: Pod got scheduled on a schedulable node", i)
+		}
+
+		err = restClient.Pods(api.NamespaceDefault).Delete(myPod.Name, nil)
+		if err != nil {
+			t.Errorf("Failed to delete pod: %v", err)
+		}
+		err = restClient.Nodes().Delete(schedNode.Name)
+		if err != nil {
+			t.Errorf("Failed to delete node: %v", err)
+		}
 	}
 }