From 11cdcca35d6416546f3a7b1ced091db507565846 Mon Sep 17 00:00:00 2001
From: Yu-Ju Hong <yjhong@google.com>
Date: Mon, 10 Oct 2016 10:26:44 -0700
Subject: [PATCH] Ignore mirror pods with RestartPolicy == Never in restart
 tests

Kubelet does not sync the mirror pods once they have terminated. If, for some
reason, that such mirror pods get deleted once they have terminated (either by
the node controller or by users), kubelet will not attempt to recreate them.
However, when kubelet restarts, it will examine the static pods, sync once,
and create a mirror pod. This has led to unexpected pod counts in disruptive
tests where kubelet gets restarted on purpose (see #34003).

This change disregard such mirror pods when totaling the pods to fix the test
flake until we have time to implement a long-term solution.
---
 test/e2e/restart.go | 30 ++++++++++++++++++++++++++++--
 1 file changed, 28 insertions(+), 2 deletions(-)

diff --git a/test/e2e/restart.go b/test/e2e/restart.go
index 32f917738ec..1b0c27a064f 100644
--- a/test/e2e/restart.go
+++ b/test/e2e/restart.go
@@ -22,6 +22,7 @@ import (
 
 	"k8s.io/kubernetes/pkg/api"
 	"k8s.io/kubernetes/pkg/fields"
+	kubepod "k8s.io/kubernetes/pkg/kubelet/pod"
 	"k8s.io/kubernetes/pkg/labels"
 	"k8s.io/kubernetes/pkg/util/wait"
 	"k8s.io/kubernetes/test/e2e/framework"
@@ -30,6 +31,28 @@ import (
 	. "github.com/onsi/gomega"
 )
 
+func isRestartNeverMirrorPod(p *api.Pod) bool {
+	if !kubepod.IsMirrorPod(p) {
+		return false
+	}
+	return p.Spec.RestartPolicy == api.RestartPolicyNever
+}
+
+func filterIrrelevantPods(pods []*api.Pod) []*api.Pod {
+	var results []*api.Pod
+	for _, p := range pods {
+		if isRestartNeverMirrorPod(p) {
+			// Mirror pods with restart policy == Never will not get
+			// recreated if they are deleted after the pods have
+			// terminated. For now, we discount such pods.
+			// https://github.com/kubernetes/kubernetes/issues/34003
+			continue
+		}
+		results = append(results, p)
+	}
+	return results
+}
+
 var _ = framework.KubeDescribe("Restart [Disruptive]", func() {
 	f := framework.NewDefaultFramework("restart")
 	var ps *framework.PodStore
@@ -57,7 +80,9 @@ var _ = framework.KubeDescribe("Restart [Disruptive]", func() {
 		framework.Logf("Got the following nodes before restart: %v", nodeNamesBefore)
 
 		By("ensuring all pods are running and ready")
-		pods := ps.List()
+		allPods := ps.List()
+		pods := filterIrrelevantPods(allPods)
+
 		podNamesBefore := make([]string, len(pods))
 		for i, p := range pods {
 			podNamesBefore[i] = p.ObjectMeta.Name
@@ -105,7 +130,8 @@ func waitForNPods(ps *framework.PodStore, expect int, timeout time.Duration) ([]
 	var pods []*api.Pod
 	var errLast error
 	found := wait.Poll(framework.Poll, timeout, func() (bool, error) {
-		pods = ps.List()
+		allPods := ps.List()
+		pods := filterIrrelevantPods(allPods)
 		if len(pods) != expect {
 			errLast = fmt.Errorf("expected to find %d pods but found only %d", expect, len(pods))
 			framework.Logf("Error getting pods: %v", errLast)