Merge pull request #31716 from coufon/explicitly_delete_pods_in_node_perf_test

Automatic merge from submit-queue

Explicitly delete pods in node performance tests

This PR explicitly deletes all created pods at the end in node e2e performance related tests.

The large number of pods may cause namespace cleanup times out (in #30878), therefore we explicitly delete all pods for cleaning up.

test/e2e_node/density_test.go

@@ -19,7 +19,6 @@ limitations under the License.
 package e2e_node
 
 import (
-	"errors"
 	"fmt"
 	"sort"
 	"strconv"
@@ -27,7 +26,6 @@ import (
 	"time"
 
 	"k8s.io/kubernetes/pkg/api"
-	apierrors "k8s.io/kubernetes/pkg/api/errors"
 	"k8s.io/kubernetes/pkg/api/unversioned"
 	"k8s.io/kubernetes/pkg/client/cache"
 	controllerframework "k8s.io/kubernetes/pkg/controller/framework"
@@ -67,15 +65,12 @@ var _ = framework.KubeDescribe("Density [Serial] [Slow]", func() {
 		ns = f.Namespace.Name
 		nodeName = framework.TestContext.NodeName
 		// Start a standalone cadvisor pod using 'createSync', the pod is running when it returns
-		createCadvisorPod(f)
+		f.PodClient().CreateSync(getCadvisorPod())
 		// Resource collector monitors fine-grain CPU/memory usage by a standalone Cadvisor with
 		// 1s housingkeeping interval
 		rc = NewResourceCollector(containerStatsPollingPeriod)
 	})
 
-	AfterEach(func() {
-	})
-
 	Context("create a batch of pods", func() {
 		// TODO(coufon): the values are generous, set more precise limits with benchmark data
 		// and add more tests
@@ -296,6 +291,8 @@ func runDensityBatchTest(f *framework.Framework, rc *ResourceCollector, testArg
 	time.Sleep(sleepBeforeCreatePods)
 
 	rc.Start()
+	// Explicitly delete pods to prevent namespace controller cleanning up timeout
+	defer deletePodsSync(f, append(pods, getCadvisorPod()))
 	defer rc.Stop()
 
 	By("Creating a batch of pods")
@@ -371,6 +368,8 @@ func runDensitySeqTest(f *framework.Framework, rc *ResourceCollector, testArg de
 	time.Sleep(sleepBeforeCreatePods)
 
 	rc.Start()
+	// Explicitly delete pods to prevent namespace controller cleanning up timeout
+	defer deletePodsSync(f, append(bgPods, append(testPods, getCadvisorPod())...))
 	defer rc.Stop()
 
 	// Create pods sequentially (back-to-back). e2eLags have been sorted.
@@ -394,16 +393,6 @@ func createBatchPodWithRateControl(f *framework.Framework, pods []*api.Pod, inte
 	return createTimes
 }
 
-// checkPodDeleted checks whether a pod has been successfully deleted
-func checkPodDeleted(f *framework.Framework, podName string) error {
-	ns := f.Namespace.Name
-	_, err := f.Client.Pods(ns).Get(podName)
-	if apierrors.IsNotFound(err) {
-		return nil
-	}
-	return errors.New("Pod Not Deleted")
-}
-
 // getPodStartLatency gets prometheus metric 'pod start latency' from kubelet
 func getPodStartLatency(node string) (framework.KubeletLatencyMetrics, error) {
 	latencyMetrics := framework.KubeletLatencyMetrics{}

test/e2e_node/resource_collector.go

@@ -292,8 +292,8 @@ func formatCPUSummary(summary framework.ContainersCPUSummary) string {
 }
 
 // createCadvisorPod creates a standalone cadvisor pod for fine-grain resource monitoring.
-func createCadvisorPod(f *framework.Framework) {
-	f.PodClient().CreateSync(&api.Pod{
+func getCadvisorPod() *api.Pod {
+	return &api.Pod{
 		ObjectMeta: api.ObjectMeta{
 			Name: cadvisorPodName,
 		},
@@ -363,24 +363,22 @@ func createCadvisorPod(f *framework.Framework) {
 				},
 			},
 		},
-	})
+	}
 }
 
-// deleteBatchPod deletes a batch of pods (synchronous).
-func deleteBatchPod(f *framework.Framework, pods []*api.Pod) {
-	ns := f.Namespace.Name
+// deletePodsSync deletes a list of pods and block until pods disappear.
+func deletePodsSync(f *framework.Framework, pods []*api.Pod) {
 	var wg sync.WaitGroup
 	for _, pod := range pods {
 		wg.Add(1)
 		go func(pod *api.Pod) {
 			defer wg.Done()
 
-			err := f.Client.Pods(ns).Delete(pod.ObjectMeta.Name, api.NewDeleteOptions(30))
+			err := f.PodClient().Delete(pod.ObjectMeta.Name, api.NewDeleteOptions(30))
 			Expect(err).NotTo(HaveOccurred())
 
-			Expect(framework.WaitForPodToDisappear(f.Client, ns, pod.ObjectMeta.Name, labels.Everything(),
-				30*time.Second, 10*time.Minute)).
-				NotTo(HaveOccurred())
+			Expect(framework.WaitForPodToDisappear(f.Client, f.Namespace.Name, pod.ObjectMeta.Name, labels.Everything(),
+				30*time.Second, 10*time.Minute)).NotTo(HaveOccurred())
 		}(pod)
 	}
 	wg.Wait()

test/e2e_node/resource_usage_test.go

@@ -52,7 +52,7 @@ var _ = framework.KubeDescribe("Resource-usage [Serial] [Slow]", func() {
 		// The Cadvsior of Kubelet has a housekeeping interval of 10s, which is too long to
 		// show the resource usage spikes. But changing its interval increases the overhead
 		// of kubelet. Hence we use a Cadvisor pod.
-		createCadvisorPod(f)
+		f.PodClient().CreateSync(getCadvisorPod())
 		rc = NewResourceCollector(containerStatsPollingPeriod)
 	})
 
@@ -135,12 +135,14 @@ func runResourceUsageTest(f *framework.Framework, rc *ResourceCollector, testArg
 		// sleep for an interval here to measure steady data
 		sleepAfterCreatePods = 10 * time.Second
 	)
+	pods := newTestPods(testArg.podsNr, ImageRegistry[pauseImage], "test_pod")
 
 	rc.Start()
+	// Explicitly delete pods to prevent namespace controller cleanning up timeout
+	defer deletePodsSync(f, append(pods, getCadvisorPod()))
 	defer rc.Stop()
 
 	By("Creating a batch of Pods")
-	pods := newTestPods(testArg.podsNr, ImageRegistry[pauseImage], "test_pod")
 	f.PodClient().CreateBatch(pods)
 
 	// wait for a while to let the node be steady