Merge pull request #8110 from ykakarap/node-label-in-place-propagation_md-ms

⚠️ in-place propagation from MD to MS

Author: k8s-ci-robot

api/v1beta1/machinedeployment_types.go

@@ -54,8 +54,18 @@ const (
 	// proportions in case the deployment has surge replicas.
 	MaxReplicasAnnotation = "machinedeployment.clusters.x-k8s.io/max-replicas"
 
-	// MachineDeploymentUniqueLabel is the label applied to Machines
-	// in a MachineDeployment containing the hash of the template.
+	// MachineDeploymentUniqueLabel is used to uniquely identify the Machines of a MachineSet.
+	// The MachineDeployment controller will set this label on a MachineSet when it is created.
+	// The label is also applied to the Machines of the MachineSet and used in the MachineSet selector.
+	// Note: For the lifetime of the MachineSet the label's value has to stay the same, otherwise the
+	// MachineSet selector would no longer match its Machines.
+	// Note: In previous Cluster API versions (< v1.4.0), the label value was the hash of the full machine template.
+	// With the introduction of in-place mutation the machine template of the MachineSet can change.
+	// Because of that it is impossible that the label's value to always be the hash of the full machine template.
+	// (Because the hash changes when the machine template changes).
+	// As a result, we use the hash of the machine template while ignoring all in-place mutable fields, i.e. the
+	// machine template with only fields that could trigger a rollout for the machine-template-hash, making it
+	// independent of the changes to any in-place mutable fields.
 	MachineDeploymentUniqueLabel = "machine-template-hash"
 )
 

internal/controllers/machinedeployment/machinedeployment_controller.go

@@ -37,6 +37,7 @@ import (
 
 	clusterv1 "sigs.k8s.io/cluster-api/api/v1beta1"
 	"sigs.k8s.io/cluster-api/controllers/external"
+	"sigs.k8s.io/cluster-api/internal/util/ssa"
 	"sigs.k8s.io/cluster-api/util"
 	"sigs.k8s.io/cluster-api/util/annotations"
 	"sigs.k8s.io/cluster-api/util/conditions"
@@ -50,6 +51,10 @@ var (
 	machineDeploymentKind = clusterv1.GroupVersion.WithKind("MachineDeployment")
 )
 
+// machineDeploymentManagerName is the manager name used for Server-Side-Apply (SSA) operations
+// in the MachineDeployment controller.
+const machineDeploymentManagerName = "capi-machinedeployment"
+
 // +kubebuilder:rbac:groups=core,resources=events,verbs=get;list;watch;create;patch
 // +kubebuilder:rbac:groups=core,resources=secrets,verbs=get;list;watch
 // +kubebuilder:rbac:groups=core,resources=nodes,verbs=get;list;watch;create;update;patch;delete
@@ -248,6 +253,19 @@ func (r *Reconciler) reconcile(ctx context.Context, cluster *clusterv1.Cluster,
 		}
 	}
 
+	// Loop over all MachineSets and cleanup managed fields.
+	// We do this so that MachineSets that were created/patched before (< v1.4.0) the controller adopted
+	// Server-Side-Apply (SSA) can also work with SSA. Otherwise, fields would be co-owned by our "old" "manager" and
+	// "capi-machinedeployment" and then we would not be able to e.g. drop labels and annotations.
+	// Note: We are cleaning up managed fields for all MachineSets, so we're able to remove this code in a few
+	// Cluster API releases. If we do this only for selected MachineSets, we would have to keep this code forever.
+	for idx := range msList {
+		machineSet := msList[idx]
+		if err := ssa.CleanUpManagedFieldsForSSAAdoption(ctx, machineSet, machineDeploymentManagerName, r.Client); err != nil {
+			return ctrl.Result{}, errors.Wrapf(err, "failed to clean up managedFields of MachineSet %s", klog.KObj(machineSet))
+		}
+	}
+
 	if d.Spec.Paused {
 		return ctrl.Result{}, r.sync(ctx, d, msList)
 	}
@@ -302,7 +320,7 @@ func (r *Reconciler) getMachineSetsForDeployment(ctx context.Context, d *cluster
 			continue
 		}
 
-		// Attempt to adopt machine if it meets previous conditions and it has no controller references.
+		// Attempt to adopt MachineSet if it meets previous conditions and it has no controller references.
 		if metav1.GetControllerOf(ms) == nil {
 			if err := r.adoptOrphan(ctx, d, ms); err != nil {
 				log.Error(err, "Failed to adopt MachineSet into MachineDeployment")