Merge remote-tracking branch 'origin/4.18' into 4.19

Author: yadvr

plugins/integrations/kubernetes-service/src/main/java/com/cloud/kubernetes/cluster/KubernetesClusterManagerImpl.java

@@ -1956,6 +1956,7 @@ public ConfigKey<?>[] getConfigKeys() {
             KubernetesClusterStartTimeout,
             KubernetesClusterScaleTimeout,
             KubernetesClusterUpgradeTimeout,
+            KubernetesClusterUpgradeRetries,
             KubernetesClusterExperimentalFeaturesEnabled,
             KubernetesMaxClusterSize
         };

plugins/integrations/kubernetes-service/src/main/java/com/cloud/kubernetes/cluster/KubernetesClusterService.java

@@ -72,6 +72,12 @@ public interface KubernetesClusterService extends PluggableService, Configurable
             "Timeout interval (in seconds) in which upgrade operation for a Kubernetes cluster should be completed. Not strictly obeyed while upgrade is in progress on a node",
             true,
             KubernetesServiceEnabled.key());
+    static final ConfigKey<Integer> KubernetesClusterUpgradeRetries = new ConfigKey<Integer>("Advanced", Integer.class,
+            "cloud.kubernetes.cluster.upgrade.retries",
+            "3",
+            "The number of retries if fail to upgrade kubernetes cluster due to some reasons (e.g. drain node, etcdserver leader changed)",
+            true,
+            KubernetesServiceEnabled.key());
     static final ConfigKey<Boolean> KubernetesClusterExperimentalFeaturesEnabled = new ConfigKey<Boolean>("Advanced", Boolean.class,
             "cloud.kubernetes.cluster.experimental.features.enabled",
             "false",

plugins/integrations/kubernetes-service/src/main/java/com/cloud/kubernetes/cluster/actionworkers/KubernetesClusterUpgradeWorker.java

@@ -77,39 +77,62 @@ private Pair<Boolean, String> runInstallScriptOnVM(final UserVm vm, final int in
     }
 
     private void upgradeKubernetesClusterNodes() {
-        Pair<Boolean, String> result = null;
         for (int i = 0; i < clusterVMs.size(); ++i) {
             UserVm vm = clusterVMs.get(i);
             String hostName = vm.getHostName();
             if (StringUtils.isNotEmpty(hostName)) {
                 hostName = hostName.toLowerCase();
             }
-            result = null;
+            Pair<Boolean, String> result;
             if (LOGGER.isInfoEnabled()) {
                 LOGGER.info(String.format("Upgrading node on VM %s in Kubernetes cluster %s with Kubernetes version(%s) ID: %s",
                         vm.getDisplayName(), kubernetesCluster.getName(), upgradeVersion.getSemanticVersion(), upgradeVersion.getUuid()));
             }
-            try {
-                result = SshHelper.sshExecute(publicIpAddress, sshPort, getControlNodeLoginUser(), sshKeyFile, null,
-                        String.format("sudo /opt/bin/kubectl drain %s --ignore-daemonsets --delete-emptydir-data", hostName),
-                        10000, 10000, 60000);
-            } catch (Exception e) {
-                logTransitStateDetachIsoAndThrow(Level.ERROR, String.format("Failed to upgrade Kubernetes cluster : %s, unable to drain Kubernetes node on VM : %s", kubernetesCluster.getName(), vm.getDisplayName()), kubernetesCluster, clusterVMs, KubernetesCluster.Event.OperationFailed, e);
-            }
-            if (!result.first()) {
-                logTransitStateDetachIsoAndThrow(Level.ERROR, String.format("Failed to upgrade Kubernetes cluster : %s, unable to drain Kubernetes node on VM : %s", kubernetesCluster.getName(), vm.getDisplayName()), kubernetesCluster, clusterVMs, KubernetesCluster.Event.OperationFailed, null);
+            String errorMessage = String.format("Failed to upgrade Kubernetes cluster : %s, unable to drain Kubernetes node on VM : %s", kubernetesCluster.getName(), vm.getDisplayName());
+            for (int retry = KubernetesClusterService.KubernetesClusterUpgradeRetries.value(); retry >= 0; retry--) {
+                try {
+                    result = SshHelper.sshExecute(publicIpAddress, sshPort, getControlNodeLoginUser(), sshKeyFile, null,
+                            String.format("sudo /opt/bin/kubectl drain %s --ignore-daemonsets --delete-emptydir-data", hostName),
+                            10000, 10000, 60000);
+                    if (result.first()) {
+                        break;
+                    }
+                    if (retry > 0) {
+                        LOGGER.error(String.format("%s, retries left: %s", errorMessage, retry));
+                    } else {
+                        logTransitStateDetachIsoAndThrow(Level.ERROR, errorMessage, kubernetesCluster, clusterVMs, KubernetesCluster.Event.OperationFailed, null);
+                    }
+                } catch (Exception e) {
+                    if (retry > 0) {
+                        LOGGER.error(String.format("%s due to %s, retries left: %s", errorMessage, e, retry));
+                    } else {
+                        logTransitStateDetachIsoAndThrow(Level.ERROR, errorMessage, kubernetesCluster, clusterVMs, KubernetesCluster.Event.OperationFailed, e);
+                    }
+                }
             }
             if (System.currentTimeMillis() > upgradeTimeoutTime) {
                 logTransitStateDetachIsoAndThrow(Level.ERROR, String.format("Failed to upgrade Kubernetes cluster : %s, upgrade action timed out", kubernetesCluster.getName()), kubernetesCluster, clusterVMs, KubernetesCluster.Event.OperationFailed, null);
             }
-            try {
-                deployProvider();
-                result = runInstallScriptOnVM(vm, i);
-            } catch (Exception e) {
-                logTransitStateDetachIsoAndThrow(Level.ERROR, String.format("Failed to upgrade Kubernetes cluster : %s, unable to upgrade Kubernetes node on VM : %s", kubernetesCluster.getName(), vm.getDisplayName()), kubernetesCluster, clusterVMs, KubernetesCluster.Event.OperationFailed, e);
-            }
-            if (!result.first()) {
-                logTransitStateDetachIsoAndThrow(Level.ERROR, String.format("Failed to upgrade Kubernetes cluster : %s, unable to upgrade Kubernetes node on VM : %s", kubernetesCluster.getName(), vm.getDisplayName()), kubernetesCluster, clusterVMs, KubernetesCluster.Event.OperationFailed, null);
+            errorMessage = String.format("Failed to upgrade Kubernetes cluster : %s, unable to upgrade Kubernetes node on VM : %s", kubernetesCluster.getName(), vm.getDisplayName());
+            for (int retry = KubernetesClusterService.KubernetesClusterUpgradeRetries.value(); retry >= 0; retry--) {
+                try {
+                    deployProvider();
+                    result = runInstallScriptOnVM(vm, i);
+                    if (result.first()) {
+                        break;
+                    }
+                    if (retry > 0) {
+                        LOGGER.error(String.format("%s, retries left: %s", errorMessage, retry));
+                    } else {
+                        logTransitStateDetachIsoAndThrow(Level.ERROR, errorMessage, kubernetesCluster, clusterVMs, KubernetesCluster.Event.OperationFailed, null);
+                    }
+                } catch (Exception e) {
+                    if (retry > 0) {
+                        LOGGER.error(String.format("%s due to %s, retries left: %s", errorMessage, e, retry));
+                    } else {
+                        logTransitStateDetachIsoAndThrow(Level.ERROR, errorMessage, kubernetesCluster, clusterVMs, KubernetesCluster.Event.OperationFailed, e);
+                    }
+                }
             }
             if (System.currentTimeMillis() > upgradeTimeoutTime) {
                 logTransitStateDetachIsoAndThrow(Level.ERROR, String.format("Failed to upgrade Kubernetes cluster : %s, upgrade action timed out", kubernetesCluster.getName()), kubernetesCluster, clusterVMs, KubernetesCluster.Event.OperationFailed, null);

test/integration/smoke/test_kubernetes_clusters.py

@@ -283,13 +283,15 @@ def deleteKubernetesSupportedVersion(cls, version_id):
         cls.apiclient.deleteKubernetesSupportedVersion(deleteKubernetesSupportedVersionCmd)
 
     @classmethod
-    def listKubernetesCluster(cls, cluster_id = None):
+    def listKubernetesCluster(cls, cluster_id = None, cluster_name = None):
         listKubernetesClustersCmd = listKubernetesClusters.listKubernetesClustersCmd()
         listKubernetesClustersCmd.listall = True
         if cluster_id != None:
             listKubernetesClustersCmd.id = cluster_id
+        if cluster_name != None:
+            listKubernetesClustersCmd.name = cluster_name
         clusterResponse = cls.apiclient.listKubernetesClusters(listKubernetesClustersCmd)
-        if cluster_id != None and clusterResponse != None:
+        if (cluster_id != None or cluster_name != None) and clusterResponse != None:
             return clusterResponse[0]
         return clusterResponse
 
@@ -528,24 +530,6 @@ def test_06_delete_kubernetes_cluster(self):
 
         return
 
-    @attr(tags=["advanced", "smoke"], required_hardware="true")
-    @skipTestIf("hypervisorNotSupported")
-    def test_07_deploy_kubernetes_ha_cluster(self):
-        """Test to deploy a new HA Kubernetes cluster
-
-        # Validate the following:
-        # 1. createKubernetesCluster should return valid info for new cluster
-        # 2. The Cloud Database contains the valid information
-        """
-        if self.setup_failed == True:
-            self.fail("Setup incomplete")
-        if self.default_network:
-            self.skipTest("HA cluster on shared network requires external ip address, skipping it")
-        global k8s_cluster
-        k8s_cluster = self.getValidKubernetesCluster(1, 3)
-        self.debug("HA Kubernetes cluster with ID: %s successfully deployed" % k8s_cluster.id)
-        return
-
     @attr(tags=["advanced", "smoke"], required_hardware="true")
     @skipTestIf("hypervisorNotSupported")
     def test_08_upgrade_kubernetes_ha_cluster(self):
@@ -573,24 +557,6 @@ def test_08_upgrade_kubernetes_ha_cluster(self):
         self.debug("Kubernetes cluster with ID: %s successfully upgraded" % k8s_cluster.id)
         return
 
-    @attr(tags=["advanced", "smoke"], required_hardware="true")
-    @skipTestIf("hypervisorNotSupported")
-    def test_09_delete_kubernetes_ha_cluster(self):
-        """Test to delete a HA Kubernetes cluster
-
-        # Validate the following:
-        # 1. deleteKubernetesCluster should delete an existing HA Kubernetes cluster
-        """
-        if self.setup_failed == True:
-            self.fail("Setup incomplete")
-        if self.default_network:
-            self.skipTest("HA cluster on shared network requires external ip address, skipping it")
-        global k8s_cluster
-        k8s_cluster = self.getValidKubernetesCluster(1, 3)
-
-        self.debug("Deleting Kubernetes cluster with ID: %s" % k8s_cluster.id)
-        return
-
     @attr(tags=["advanced", "smoke"], required_hardware="true")
     @skipTestIf("hypervisorNotSupported")
     def test_10_vpc_tier_kubernetes_cluster(self):
@@ -818,8 +784,14 @@ def createNewKubernetesCluster(self, version, size, control_nodes) :
             cluster = self.createKubernetesCluster(name, version.id, size, control_nodes)
             self.verifyKubernetesCluster(cluster, name, version.id, size, control_nodes)
         except Exception as ex:
+            cluster = self.listKubernetesCluster(cluster_name = name)
+            if cluster != None:
+                self.deleteKubernetesClusterAndVerify(cluster.id, False, True)
             self.fail("Kubernetes cluster deployment failed: %s" % ex)
         except AssertionError as err:
+            cluster = self.listKubernetesCluster(cluster_name = name)
+            if cluster != None:
+                self.deleteKubernetesClusterAndVerify(cluster.id, False, True)
             self.fail("Kubernetes cluster deployment failed during cluster verification: %s" % err)
         return cluster