Merge pull request #26692 from wojtek-t/fix_scalability_test

test/e2e/density.go

@@ -158,6 +158,11 @@ var _ = framework.KubeDescribe("Density", func() {
 		c = f.Client
 		ns = f.Namespace.Name
 
+		// In large clusters we may get to this point but still have a bunch
+		// of nodes without Routes created. Since this would make a node
+		// unschedulable, we need to wait until all of them are schedulable.
+		framework.ExpectNoError(framework.WaitForAllNodesSchedulable(c))
+
 		nodes := framework.GetReadySchedulableNodesOrDie(c)
 		nodeCount = len(nodes.Items)
 		Expect(nodeCount).NotTo(BeZero())

test/e2e/framework/util.go

@@ -2596,6 +2596,17 @@ func waitListSchedulableNodesOrDie(c *client.Client) *api.NodeList {
 	return nodes
 }
 
+// Node is schedulable if:
+// 1) doesn't have "unschedulable" field set
+// 2) it's Ready condition is set to true
+// 3) doesn't have NetworkUnavailable condition set to true
+func isNodeSchedulable(node *api.Node) bool {
+	nodeReady := IsNodeConditionSetAsExpected(node, api.NodeReady, true)
+	networkReady := IsNodeConditionUnset(node, api.NodeNetworkUnavailable) ||
+		IsNodeConditionSetAsExpected(node, api.NodeNetworkUnavailable, false)
+	return !node.Spec.Unschedulable && nodeReady && networkReady
+}
+
 // GetReadySchedulableNodesOrDie addresses the common use case of getting nodes you can do work on.
 // 1) Needs to be schedulable.
 // 2) Needs to be ready.
@@ -2605,14 +2616,32 @@ func GetReadySchedulableNodesOrDie(c *client.Client) (nodes *api.NodeList) {
 	// previous tests may have cause failures of some nodes. Let's skip
 	// 'Not Ready' nodes, just in case (there is no need to fail the test).
 	FilterNodes(nodes, func(node api.Node) bool {
-		nodeReady := IsNodeConditionSetAsExpected(&node, api.NodeReady, true)
-		networkReady := IsNodeConditionUnset(&node, api.NodeNetworkUnavailable) ||
-			IsNodeConditionSetAsExpected(&node, api.NodeNetworkUnavailable, false)
-		return !node.Spec.Unschedulable && nodeReady && networkReady
+		return isNodeSchedulable(&node)
 	})
 	return nodes
 }
 
+func WaitForAllNodesSchedulable(c *client.Client) error {
+	return wait.PollImmediate(30*time.Second, 2*time.Hour, func() (bool, error) {
+		opts := api.ListOptions{
+			ResourceVersion: "0",
+			FieldSelector:   fields.Set{"spec.unschedulable": "false"}.AsSelector(),
+		}
+		nodes, err := c.Nodes().List(opts)
+		if err != nil {
+			Logf("Unexpected error listing nodes: %v", err)
+			// Ignore the error here - it will be retried.
+			return false, nil
+		}
+		for _, node := range nodes.Items {
+			if !isNodeSchedulable(&node) {
+				return false, nil
+			}
+		}
+		return true, nil
+	})
+}
+
 func ScaleRC(c *client.Client, ns, name string, size uint, wait bool) error {
 	By(fmt.Sprintf("Scaling replication controller %s in namespace %s to %d", name, ns, size))
 	scaler, err := kubectl.ScalerFor(api.Kind("ReplicationController"), c)

test/e2e/load.go

@@ -82,6 +82,11 @@ var _ = framework.KubeDescribe("Load capacity", func() {
 	BeforeEach(func() {
 		c = f.Client
 
+		// In large clusters we may get to this point but still have a bunch
+		// of nodes without Routes created. Since this would make a node
+		// unschedulable, we need to wait until all of them are schedulable.
+		framework.ExpectNoError(framework.WaitForAllNodesSchedulable(c))
+
 		ns = f.Namespace.Name
 		nodes := framework.GetReadySchedulableNodesOrDie(c)
 		nodeCount = len(nodes.Items)