AllNodesReady waits for all system pods to be running

test/e2e/framework/util.go

@@ -167,6 +167,13 @@ var (
 
 	// For parsing Kubectl version for version-skewed testing.
 	gitVersionRegexp = regexp.MustCompile("GitVersion:\"(v.+?)\"")
+
+	// Slice of regexps for names of pods that have to be running to consider a Node "healthy"
+	requiredPerNodePods = []*regexp.Regexp{
+		regexp.MustCompile(".*kube-proxy.*"),
+		regexp.MustCompile(".*fluentd-elasticsearch.*"),
+		regexp.MustCompile(".*node-problem-detector.*"),
+	}
 )
 
 // GetServerArchitecture fetches the architecture of the cluster's apiserver.
@@ -4003,7 +4010,10 @@ func WaitForNodeToBe(c *client.Client, name string, conditionType api.NodeCondit
 	return false
 }
 
-// checks whether all registered nodes are ready
+// Checks whether all registered nodes are ready.
+// TODO: we should change the AllNodesReady call in AfterEach to WaitForAllNodesHealthy,
+// and figure out how to do it in a configurable way, as we can't expect all setups to run
+// default test add-ons.
 func AllNodesReady(c *client.Client, timeout time.Duration) error {
 	Logf("Waiting up to %v for all nodes to be ready", timeout)
 
@@ -4033,6 +4043,71 @@ func AllNodesReady(c *client.Client, timeout time.Duration) error {
 	return nil
 }
 
+// checks whether all registered nodes are ready and all required Pods are running on them.
+func WaitForAllNodesHealthy(c *client.Client, timeout time.Duration) error {
+	Logf("Waiting up to %v for all nodes to be ready", timeout)
+
+	var notReady []api.Node
+	var missingPodsPerNode map[string][]string
+	err := wait.PollImmediate(Poll, timeout, func() (bool, error) {
+		notReady = nil
+		// It should be OK to list unschedulable Nodes here.
+		nodes, err := c.Nodes().List(api.ListOptions{ResourceVersion: "0"})
+		if err != nil {
+			return false, err
+		}
+		for _, node := range nodes.Items {
+			if !IsNodeConditionSetAsExpected(&node, api.NodeReady, true) {
+				notReady = append(notReady, node)
+			}
+		}
+		pods, err := c.Pods(api.NamespaceAll).List(api.ListOptions{ResourceVersion: "0"})
+		if err != nil {
+			return false, err
+		}
+
+		systemPodsPerNode := make(map[string][]string)
+		for _, pod := range pods.Items {
+			if pod.Namespace == api.NamespaceSystem && pod.Status.Phase == api.PodRunning {
+				if pod.Spec.NodeName != "" {
+					systemPodsPerNode[pod.Spec.NodeName] = append(systemPodsPerNode[pod.Spec.NodeName], pod.Name)
+				}
+			}
+		}
+		missingPodsPerNode = make(map[string][]string)
+		for _, node := range nodes.Items {
+			if !system.IsMasterNode(&node) {
+				for _, requiredPod := range requiredPerNodePods {
+					foundRequired := false
+					for _, presentPod := range systemPodsPerNode[node.Name] {
+						if requiredPod.MatchString(presentPod) {
+							foundRequired = true
+							break
+						}
+					}
+					if !foundRequired {
+						missingPodsPerNode[node.Name] = append(missingPodsPerNode[node.Name], requiredPod.String())
+					}
+				}
+			}
+		}
+		return len(notReady) == 0 && len(missingPodsPerNode) == 0, nil
+	})
+
+	if err != nil && err != wait.ErrWaitTimeout {
+		return err
+	}
+
+	if len(notReady) > 0 {
+		return fmt.Errorf("Not ready nodes: %v", notReady)
+	}
+	if len(missingPodsPerNode) > 0 {
+		return fmt.Errorf("Not running system Pods: %v", missingPodsPerNode)
+	}
+	return nil
+
+}
+
 // Filters nodes in NodeList in place, removing nodes that do not
 // satisfy the given condition
 // TODO: consider merging with pkg/client/cache.NodeLister

test/e2e/scheduler_predicates.go

@@ -198,6 +198,7 @@ var _ = framework.KubeDescribe("SchedulerPredicates [Serial]", func() {
 		ns = f.Namespace.Name
 		nodeList = &api.NodeList{}
 
+		framework.WaitForAllNodesHealthy(c, time.Minute)
 		masterNodes, nodeList = framework.GetMasterAndWorkerNodesOrDie(c)
 
 		err := framework.CheckTestingNSDeletedExcept(c, ns)