|
| 1 | +package tnf |
| 2 | + |
| 3 | +import ( |
| 4 | + "context" |
| 5 | + "fmt" |
| 6 | + "math/rand" |
| 7 | + "slices" |
| 8 | + "time" |
| 9 | + |
| 10 | + g "github.com/onsi/ginkgo/v2" |
| 11 | + o "github.com/onsi/gomega" |
| 12 | + v1 "github.com/openshift/api/config/v1" |
| 13 | + "github.com/openshift/origin/test/extended/etcd/helpers" |
| 14 | + exutil "github.com/openshift/origin/test/extended/util" |
| 15 | + "github.com/pkg/errors" |
| 16 | + "go.etcd.io/etcd/api/v3/etcdserverpb" |
| 17 | + corev1 "k8s.io/api/core/v1" |
| 18 | + metav1 "k8s.io/apimachinery/pkg/apis/meta/v1" |
| 19 | +) |
| 20 | + |
| 21 | +var _ = g.Describe("[sig-node][apigroup:config.openshift.io] Two Nodes OCP with fencing recovery", func() { |
| 22 | + defer g.GinkgoRecover() |
| 23 | + |
| 24 | + var ( |
| 25 | + oc = exutil.NewCLIWithoutNamespace("").AsAdmin() |
| 26 | + etcdClientFactory *helpers.EtcdClientFactoryImpl |
| 27 | + nodeA, nodeB corev1.Node |
| 28 | + ) |
| 29 | + |
| 30 | + g.BeforeEach(func() { |
| 31 | + infraStatus := getInfraStatus(oc) |
| 32 | + if infraStatus.ControlPlaneTopology != v1.DualReplicaTopologyMode { |
| 33 | + g.Skip("Cluster is not in DualReplicaTopologyMode skipping test") |
| 34 | + } |
| 35 | + |
| 36 | + nodes, err := oc.AdminKubeClient().CoreV1().Nodes().List(context.Background(), metav1.ListOptions{}) |
| 37 | + o.Expect(err).ShouldNot(o.HaveOccurred(), "Expected to retrieve nodes without error") |
| 38 | + o.Expect(len(nodes.Items)).To(o.BeNumerically("==", 2), "Expected to find 2 Nodes only") |
| 39 | + |
| 40 | + // Select the first index randomly |
| 41 | + randomIndex := rand.Intn(len(nodes.Items)) |
| 42 | + nodeA = nodes.Items[randomIndex] |
| 43 | + // Select the remaining index |
| 44 | + nodeB = nodes.Items[(randomIndex+1)%len(nodes.Items)] |
| 45 | + g.GinkgoT().Printf("Randomly selected %s to be gracefully shut down and %s to take the lead\n", nodeB.Name, nodeA.Name) |
| 46 | + |
| 47 | + kubeClient := oc.KubeClient() |
| 48 | + etcdClientFactory = helpers.NewEtcdClientFactory(kubeClient) |
| 49 | + |
| 50 | + g.GinkgoT().Printf("Ensure both nodes are healthy before starting the test\n") |
| 51 | + o.Eventually(func() error { |
| 52 | + return helpers.EnsureHealthyMember(g.GinkgoT(), etcdClientFactory, nodeA.Name) |
| 53 | + }, time.Minute, 5*time.Second).ShouldNot(o.HaveOccurred(), "expect to ensure Node A healthy without error") |
| 54 | + |
| 55 | + o.Eventually(func() error { |
| 56 | + return helpers.EnsureHealthyMember(g.GinkgoT(), etcdClientFactory, nodeB.Name) |
| 57 | + }, time.Minute, 5*time.Second).ShouldNot(o.HaveOccurred(), "expect to ensure Node B healthy without error") |
| 58 | + }) |
| 59 | + |
| 60 | + g.It("Should support a graceful node shutdown", func() { |
| 61 | + msg := fmt.Sprintf("Shutting down %s gracefully in 1 minute", nodeB.Name) |
| 62 | + g.By(msg) |
| 63 | + // NOTE: Using `shutdown` alone would cause the node to be permanently removed from the cluster. |
| 64 | + // To prevent this, we use the `--reboot` flag, which ensures a graceful shutdown and allows the |
| 65 | + // node to rejoin the cluster upon restart. A one-minute delay is added to give the debug node |
| 66 | + // sufficient time to cleanly exit before the shutdown process completes. |
| 67 | + _, _, err := runOnNodeNS(oc, nodeB.Name, "openshift-etcd", "shutdown --reboot +1") |
| 68 | + o.Expect(err).To(o.BeNil(), "Expected to gracefully shutdown the node without errors") |
| 69 | + time.Sleep(time.Minute) |
| 70 | + |
| 71 | + msg = fmt.Sprintf("Ensuring %s leaves the member list", nodeB.Name) |
| 72 | + g.By(msg) |
| 73 | + o.Eventually(func() error { |
| 74 | + return helpers.EnsureMemberRemoved(g.GinkgoT(), etcdClientFactory, nodeB.Name) |
| 75 | + }, 5*time.Minute, 30*time.Second).ShouldNot(o.HaveOccurred()) |
| 76 | + |
| 77 | + msg = fmt.Sprintf("Ensuring that %s is a healthy voting member and adds %s back as learner", nodeA.Name, nodeB.Name) |
| 78 | + g.By(msg) |
| 79 | + o.Eventually(func() error { |
| 80 | + members, err := getMembers(etcdClientFactory) |
| 81 | + if err != nil { |
| 82 | + return err |
| 83 | + } |
| 84 | + if len(members) != 2 { |
| 85 | + return fmt.Errorf("Not enough members") |
| 86 | + } |
| 87 | + |
| 88 | + if started, learner, err := getMemberState(&nodeA, members); err != nil { |
| 89 | + return err |
| 90 | + } else if !started || learner { |
| 91 | + return fmt.Errorf("Expected node: %s to be a started and voting member. Membership: %+v", nodeA.Name, members) |
| 92 | + } |
| 93 | + |
| 94 | + // Ensure GNS node is unstarted and a learner member (i.e. !learner) |
| 95 | + if started, learner, err := getMemberState(&nodeB, members); err != nil { |
| 96 | + return err |
| 97 | + } else if started || !learner { |
| 98 | + return fmt.Errorf("Expected node: %s to be a unstarted and learning member. Membership: %+v", nodeB.Name, members) |
| 99 | + } |
| 100 | + |
| 101 | + g.GinkgoT().Logf("membership: %+v", members) |
| 102 | + return nil |
| 103 | + }, 2*time.Minute, 15*time.Second).ShouldNot(o.HaveOccurred()) |
| 104 | + |
| 105 | + msg = fmt.Sprintf("Ensuring %s rejoins as learner", nodeB.Name) |
| 106 | + g.By(msg) |
| 107 | + o.Eventually(func() error { |
| 108 | + members, err := getMembers(etcdClientFactory) |
| 109 | + if err != nil { |
| 110 | + return err |
| 111 | + } |
| 112 | + if len(members) != 2 { |
| 113 | + return fmt.Errorf("Not enough members") |
| 114 | + } |
| 115 | + |
| 116 | + if started, learner, err := getMemberState(&nodeA, members); err != nil { |
| 117 | + return err |
| 118 | + } else if !started || learner { |
| 119 | + return fmt.Errorf("Expected node: %s to be a started and voting member. Membership: %+v", nodeA.Name, members) |
| 120 | + } |
| 121 | + |
| 122 | + if started, learner, err := getMemberState(&nodeB, members); err != nil { |
| 123 | + return err |
| 124 | + } else if !started || !learner { |
| 125 | + return fmt.Errorf("Expected node: %s to be a started and learner member. Membership: %+v", nodeB.Name, members) |
| 126 | + } |
| 127 | + |
| 128 | + g.GinkgoT().Logf("membership: %+v", members) |
| 129 | + return nil |
| 130 | + }, 10*time.Minute, 15*time.Second).ShouldNot(o.HaveOccurred()) |
| 131 | + |
| 132 | + msg = fmt.Sprintf("Ensuring %s node is promoted back as voting member", nodeB.Name) |
| 133 | + g.By(msg) |
| 134 | + o.Eventually(func() error { |
| 135 | + members, err := getMembers(etcdClientFactory) |
| 136 | + if err != nil { |
| 137 | + return err |
| 138 | + } |
| 139 | + if len(members) != 2 { |
| 140 | + return fmt.Errorf("Not enough members") |
| 141 | + } |
| 142 | + |
| 143 | + if started, learner, err := getMemberState(&nodeA, members); err != nil { |
| 144 | + return err |
| 145 | + } else if !started || learner { |
| 146 | + return fmt.Errorf("Expected node: %s to be a started and voting member. Membership: %+v", nodeA.Name, members) |
| 147 | + } |
| 148 | + |
| 149 | + if started, learner, err := getMemberState(&nodeB, members); err != nil { |
| 150 | + return err |
| 151 | + } else if !started || learner { |
| 152 | + return fmt.Errorf("Expected node: %s to be a started and voting member. Membership: %+v", nodeB.Name, members) |
| 153 | + } |
| 154 | + |
| 155 | + g.GinkgoT().Logf("membership: %+v", members) |
| 156 | + return nil |
| 157 | + }, 10*time.Minute, 15*time.Second).ShouldNot(o.HaveOccurred()) |
| 158 | + }) |
| 159 | +}) |
| 160 | + |
| 161 | +func getMembers(etcdClientFactory helpers.EtcdClientCreator) ([]*etcdserverpb.Member, error) { |
| 162 | + etcdClient, closeFn, err := etcdClientFactory.NewEtcdClient() |
| 163 | + if err != nil { |
| 164 | + return []*etcdserverpb.Member{}, errors.Wrap(err, "could not get a etcd client") |
| 165 | + } |
| 166 | + defer closeFn() |
| 167 | + |
| 168 | + ctx, cancel := context.WithTimeout(context.TODO(), 10*time.Second) |
| 169 | + defer cancel() |
| 170 | + m, err := etcdClient.MemberList(ctx) |
| 171 | + if err != nil { |
| 172 | + return []*etcdserverpb.Member{}, errors.Wrap(err, "could not get the member list") |
| 173 | + } |
| 174 | + return m.Members, nil |
| 175 | +} |
| 176 | + |
| 177 | +func getMemberState(node *corev1.Node, members []*etcdserverpb.Member) (started, learner bool, err error) { |
| 178 | + // Etcd members that have been added to the member list but haven't |
| 179 | + // joined yet will have an empty Name field. We can match them via Peer URL. |
| 180 | + peerURL := fmt.Sprintf("https://%s:2380", node.Status.Addresses[0].Address) |
| 181 | + var found bool |
| 182 | + for _, m := range members { |
| 183 | + if m.Name == node.Name { |
| 184 | + found = true |
| 185 | + started = true |
| 186 | + learner = m.IsLearner |
| 187 | + break |
| 188 | + } |
| 189 | + if slices.Contains(m.PeerURLs, peerURL) { |
| 190 | + found = true |
| 191 | + learner = m.IsLearner |
| 192 | + break |
| 193 | + } |
| 194 | + } |
| 195 | + if !found { |
| 196 | + return false, false, fmt.Errorf("could not find node %v", node.Name) |
| 197 | + } |
| 198 | + return started, learner, nil |
| 199 | +} |
0 commit comments