Merge pull request #16932 from rsafonseca/version_skew

Fix version upgrade kubelet support

cmd/kops/update_cluster.go

@@ -27,11 +27,14 @@ import (
 
 	"github.com/spf13/cobra"
 	"github.com/spf13/viper"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 	"k8s.io/apimachinery/pkg/util/sets"
+	"k8s.io/client-go/kubernetes"
 	"k8s.io/client-go/tools/clientcmd"
 	"k8s.io/klog/v2"
 	"k8s.io/kops/cmd/kops/util"
 	"k8s.io/kops/pkg/apis/kops"
+	apisutil "k8s.io/kops/pkg/apis/kops/util"
 	"k8s.io/kops/pkg/assets"
 	"k8s.io/kops/pkg/commands/commandutils"
 	"k8s.io/kops/pkg/kubeconfig"
@@ -67,6 +70,10 @@ type UpdateClusterOptions struct {
 	SSHPublicKey       string
 	RunTasksOptions    fi.RunTasksOptions
 	AllowKopsDowngrade bool
+	// Bypasses kubelet vs control plane version skew checks,
+	// which by default prevent non-control plane instancegroups
+	// from being updated to a version greater than the control plane
+	IgnoreKubeletVersionSkew bool
 	// GetAssets is whether this is invoked from the CmdGetAssets.
 	GetAssets bool
 
@@ -103,6 +110,8 @@ func (o *UpdateClusterOptions) InitDefaults() {
 	o.Target = "direct"
 	o.SSHPublicKey = ""
 	o.OutDir = ""
+	// By default we enforce the version skew between control plane and worker nodes
+	o.IgnoreKubeletVersionSkew = false
 
 	// By default we export a kubecfg, but it doesn't have a static/eternal credential in it any more.
 	o.CreateKubecfg = true
@@ -163,6 +172,7 @@ func NewCmdUpdateCluster(f *util.Factory, out io.Writer) *cobra.Command {
 	cmd.RegisterFlagCompletionFunc("lifecycle-overrides", completeLifecycleOverrides)
 
 	cmd.Flags().BoolVar(&options.Prune, "prune", options.Prune, "Delete old revisions of cloud resources that were needed during an upgrade")
+	cmd.Flags().BoolVar(&options.IgnoreKubeletVersionSkew, "ignore-kubelet-version-skew", options.IgnoreKubeletVersionSkew, "Setting this to true will force updating the kubernetes version on all instance groups, regardles of which control plane version is running")
 
 	return cmd
 }
@@ -318,6 +328,15 @@ func RunUpdateCluster(ctx context.Context, f *util.Factory, out io.Writer, c *Up
 		return nil, err
 	}
 
+	minControlPlaneRunningVersion := cluster.Spec.KubernetesVersion
+	if !c.IgnoreKubeletVersionSkew {
+		minControlPlaneRunningVersion, err = checkControlPlaneRunningVersion(ctx, cluster.ObjectMeta.Name, minControlPlaneRunningVersion)
+		if err != nil {
+			klog.Warningf("error checking control plane running version, assuming no k8s upgrade in progress: %v", err)
+		} else {
+			klog.V(2).Infof("successfully checked control plane running version: %v", minControlPlaneRunningVersion)
+		}
+	}
 	applyCmd := &cloudup.ApplyClusterCmd{
 		Cloud:                      cloud,
 		Clientset:                  clientset,
@@ -332,6 +351,7 @@ func RunUpdateCluster(ctx context.Context, f *util.Factory, out io.Writer, c *Up
 		LifecycleOverrides:         lifecycleOverrideMap,
 		GetAssets:                  c.GetAssets,
 		DeletionProcessing:         deletionProcessing,
+		ControlPlaneRunningVersion: minControlPlaneRunningVersion,
 	}
 
 	applyResults, err := applyCmd.Run(ctx)
@@ -575,3 +595,38 @@ func matchInstanceGroupRoles(roles []string) predicates.Predicate[*kops.Instance
 		return false
 	}
 }
+
+// checkControlPlaneRunningVersion returns the minimum control plane running version
+func checkControlPlaneRunningVersion(ctx context.Context, clusterName string, version string) (string, error) {
+	configLoadingRules := clientcmd.NewDefaultClientConfigLoadingRules()
+	config, err := clientcmd.NewNonInteractiveDeferredLoadingClientConfig(
+		configLoadingRules,
+		&clientcmd.ConfigOverrides{CurrentContext: clusterName}).ClientConfig()
+	if err != nil {
+		return version, fmt.Errorf("cannot load kubecfg settings for %q: %v", clusterName, err)
+	}
+
+	k8sClient, err := kubernetes.NewForConfig(config)
+	if err != nil {
+		return version, fmt.Errorf("cannot build kubernetes api client for %q: %v", clusterName, err)
+	}
+
+	parsedVersion, err := apisutil.ParseKubernetesVersion(version)
+	if err != nil {
+		return version, fmt.Errorf("cannot parse kubernetes version %q: %v", clusterName, err)
+	}
+	nodeList, err := k8sClient.CoreV1().Nodes().List(ctx, metav1.ListOptions{
+		LabelSelector: "node-role.kubernetes.io/control-plane",
+	})
+	if err != nil {
+		return version, fmt.Errorf("cannot list nodes in cluster %q: %v", clusterName, err)
+	}
+	for _, node := range nodeList.Items {
+		if apisutil.IsKubernetesGTE(node.Status.NodeInfo.KubeletVersion, *parsedVersion) {
+			version = node.Status.NodeInfo.KubeletVersion
+			parsedVersion, _ = apisutil.ParseKubernetesVersion(version)
+		}
+
+	}
+	return strings.TrimPrefix(version, "v"), nil
+}

docs/cli/kops_update_cluster.md

@@ -29,6 +29,7 @@ kops update cluster [CLUSTER] [flags]
       --allow-kops-downgrade           Allow an older version of kOps to update the cluster than last used
       --create-kube-config             Will control automatically creating the kube config file on your local filesystem (default true)
   -h, --help                           help for cluster
+      --ignore-kubelet-version-skew    Setting this to true will force updating the kubernetes version on all instance groups, regardles of which control plane version is running
       --instance-group strings         Instance groups to update (defaults to all if not specified)
       --instance-group-roles strings   Instance group roles to update (control-plane,apiserver,node,bastion)
       --internal                       Use the cluster's internal DNS name. Implies --create-kube-config

pkg/apis/kops/model/instance_group.go

@@ -34,17 +34,20 @@ type InstanceGroup interface {
 	// RawClusterSpec returns the cluster spec for the instance group.
 	// If possible, prefer abstracted methods over accessing this data directly.
 	RawClusterSpec() *kops.ClusterSpec
+
+	// ForceKubernetesVersion overrides the Kubernetes version for this instance group.
+	// (The default is to use the cluster-wide Kubernetes version, but this allows
+	// us to override it for the nodes to respect the node skew policy.)
+	ForceKubernetesVersion(version string) error
 }
 
 // ForInstanceGroup creates an InstanceGroup model for the given cluster and instance group.
 func ForInstanceGroup(cluster *kops.Cluster, ig *kops.InstanceGroup) (InstanceGroup, error) {
-	kubernetesVersionString := cluster.Spec.KubernetesVersion
-	kubernetesVersion, err := ParseKubernetesVersion(kubernetesVersionString)
-	if err != nil {
-		return nil, fmt.Errorf("error parsing Kubernetes version %q: %v", kubernetesVersionString, err)
+	m := &instanceGroupModel{cluster: cluster, ig: ig}
+	if err := m.ForceKubernetesVersion(cluster.Spec.KubernetesVersion); err != nil {
+		return nil, err
 	}
-
-	return &instanceGroupModel{cluster: cluster, ig: ig, kubernetesVersion: kubernetesVersion}, nil
+	return m, nil
 }
 
 // instanceGroupModel is a concrete implementation of InstanceGroup.
@@ -67,3 +70,12 @@ func (m *instanceGroupModel) GetCloudProvider() kops.CloudProviderID {
 func (m *instanceGroupModel) RawClusterSpec() *kops.ClusterSpec {
 	return &m.cluster.Spec
 }
+
+func (m *instanceGroupModel) ForceKubernetesVersion(kubernetesVersionString string) error {
+	kubernetesVersion, err := ParseKubernetesVersion(kubernetesVersionString)
+	if err != nil {
+		return fmt.Errorf("error parsing Kubernetes version %q: %v", kubernetesVersionString, err)
+	}
+	m.kubernetesVersion = kubernetesVersion
+	return nil
+}

pkg/assets/builder.go

@@ -51,6 +51,11 @@ type AssetBuilder struct {
 	// KubernetesVersion is the version of kubernetes we are installing
 	KubernetesVersion semver.Version
 
+	// KubeletSupportedVersion is the max version of kubelet that we are currently allowed to run on worker nodes.
+	// This is used to avoid violating the kubelet supported version skew policy,
+	// (we are not allowed to run a newer kubelet on a worker node than the control plane)
+	KubeletSupportedVersion string
+
 	// StaticManifests records manifests used by nodeup:
 	// * e.g. sidecar manifests for static pods run by kubelet
 	StaticManifests []*StaticManifest

pkg/nodemodel/nodeupconfigbuilder.go

@@ -224,6 +224,12 @@ func (n *nodeUpConfigBuilder) BuildConfig(ig *kops.InstanceGroup, wellKnownAddre
 		return nil, nil, fmt.Errorf("building instance group model: %w", err)
 	}
 
+	if !hasAPIServer && n.assetBuilder.KubeletSupportedVersion != "" {
+		if err := igModel.ForceKubernetesVersion(n.assetBuilder.KubeletSupportedVersion); err != nil {
+			return nil, nil, err
+		}
+	}
+
 	kubernetesAssets, err := BuildKubernetesFileAssets(igModel, n.assetBuilder)
 	if err != nil {
 		return nil, nil, err

upup/pkg/fi/cloudup/apply_cluster.go

@@ -137,6 +137,9 @@ type ApplyClusterCmd struct {
 
 	// InstanceGroupFilter is a predicate that restricts which instance groups we will update.
 	InstanceGroupFilter predicates.Predicate[*kops.InstanceGroup]
+
+	// The current oldest version of control plane nodes, defaults to version defined in cluster spec if IgnoreVersionSkew was set
+	ControlPlaneRunningVersion string
 }
 
 // ApplyResults holds information about an ApplyClusterCmd operation.
@@ -239,6 +242,9 @@ func (c *ApplyClusterCmd) Run(ctx context.Context) (*ApplyResults, error) {
 	}
 
 	assetBuilder := assets.NewAssetBuilder(c.Clientset.VFSContext(), c.Cluster.Spec.Assets, c.Cluster.Spec.KubernetesVersion, c.GetAssets)
+	if len(c.ControlPlaneRunningVersion) > 0 && c.ControlPlaneRunningVersion != c.Cluster.Spec.KubernetesVersion {
+		assetBuilder.KubeletSupportedVersion = c.ControlPlaneRunningVersion
+	}
 	err = c.upgradeSpecs(ctx, assetBuilder)
 	if err != nil {
 		return nil, err