diff --git a/api/v1alpha1/condition_consts.go b/api/v1alpha1/condition_consts.go
index bbf96b8..508e0e9 100644
--- a/api/v1alpha1/condition_consts.go
+++ b/api/v1alpha1/condition_consts.go
@@ -51,4 +51,10 @@ const (
 
 	// MicrovmDeploymentProvisionFailedReason indicates that the microvm deployment failed to provision.
 	MicrovmDeploymentProvisionFailedReason = "MicrovmDeploymentProvisionFailed"
+
+	// MicrovmDeploymentUpdatingReason indicates the microvm deployment is in a pending state.
+	MicrovmDeploymentUpdatingReason = "MicrovmDeploymentUpdating"
+
+	// MicrovmDeploymentUpdateFailed indicates the microvm deployment is in a pending state.
+	MicrovmDeploymentUpdateFailedReason = "MicrovmDeploymentUpdateFailed"
 )
diff --git a/controllers/helpers_test.go b/controllers/helpers_test.go
index 68af332..00c681d 100644
--- a/controllers/helpers_test.go
+++ b/controllers/helpers_test.go
@@ -114,6 +114,18 @@ func reconcileMicrovmDeployment(client client.Client) (ctrl.Result, error) {
 	return mvmDepController.Reconcile(context.TODO(), request)
 }
 
+func reconcileMicrovmDeploymentNTimes(g *WithT, client client.Client, count int, r, rr int32) error {
+	for count > 0 {
+		ensureMicrovmReplicaSetState(g, client, r, rr)
+		if _, err := reconcileMicrovmDeployment(client); err != nil {
+			return err
+		}
+		count--
+	}
+
+	return nil
+}
+
 func getMicrovm(c client.Client, name, namespace string) (*infrav1.Microvm, error) {
 	key := client.ObjectKey{
 		Name:      name,
diff --git a/controllers/microvmdeployment_controller.go b/controllers/microvmdeployment_controller.go
index 828416b..f7b5590 100644
--- a/controllers/microvmdeployment_controller.go
+++ b/controllers/microvmdeployment_controller.go
@@ -170,6 +170,25 @@ func (r *MicrovmDeploymentReconciler) reconcileNormal(
 		}
 
 		mvmDeploymentScope.SetNotReady(infrav1.MicrovmDeploymentIncompleteReason, "Info", "")
+	// if we are here then a scale down has been requested.
+	// we delete the first found until the numbers balance out.
+	// TODO the way this works is very naive and often ends up deleting everything
+	// if the timing is wrong/right, find a better way https://github.com/weaveworks-liquidmetal/microvm-operator/issues/17
+	case createdSets > mvmDeploymentScope.RequiredSets():
+		mvmDeploymentScope.Info("MicrovmDeployment updating: delete microvmreplicaset")
+		mvmDeploymentScope.SetNotReady(infrav1.MicrovmDeploymentUpdatingReason, "Info", "")
+
+		rs := rsList[0]
+		if !rs.DeletionTimestamp.IsZero() {
+			return ctrl.Result{}, nil
+		}
+
+		if err := r.Delete(ctx, &rs); err != nil {
+			mvmDeploymentScope.Error(err, "failed deleting microvmreplicaset")
+			mvmDeploymentScope.SetNotReady(infrav1.MicrovmDeploymentUpdateFailedReason, "Error", "")
+
+			return ctrl.Result{}, err
+		}
 	// if all desired objects have been created, but are not quite ready yet,
 	// set the condition and requeue
 	default:
diff --git a/controllers/microvmdeployment_controller_test.go b/controllers/microvmdeployment_controller_test.go
index 0743e94..0d104d5 100644
--- a/controllers/microvmdeployment_controller_test.go
+++ b/controllers/microvmdeployment_controller_test.go
@@ -1,9 +1,11 @@
 package controllers_test
 
 import (
+	"context"
 	"testing"
 
 	. "github.com/onsi/gomega"
+	"github.com/weaveworks-liquidmetal/controller-pkg/types/microvm"
 	infrav1 "github.com/weaveworks-liquidmetal/microvm-operator/api/v1alpha1"
 	"k8s.io/apimachinery/pkg/runtime"
 )
@@ -79,3 +81,64 @@ func TestMicrovmDep_ReconcileNormal_CreateSucceeds(t *testing.T) {
 	g.Expect(microvmReplicaSetsCreated(g, client)).To(Equal(expectedReplicaSets), "Expected all Microvms to have been created after two reconciliations")
 	assertOneSetPerHost(g, reconciled, client)
 }
+
+func TestMicrovmDep_ReconcileNormal_UpdateSucceeds(t *testing.T) {
+	g := NewWithT(t)
+
+	// updating a replicaset with 2 replicas
+	var (
+		initialReplicaSetCount int   = 2
+		scaledReplicaSetCount  int32 = 1
+		expectedReplicas       int32 = 2
+		initialReplicaCount    int32 = 4
+		scaledReplicaCount     int32 = 2
+	)
+
+	mvmD := createMicrovmDeployment(expectedReplicas, initialReplicaSetCount)
+	objects := []runtime.Object{mvmD}
+	client := createFakeClient(g, objects)
+
+	// create
+	g.Expect(reconcileMicrovmDeploymentNTimes(g, client, initialReplicaSetCount+4, expectedReplicas, expectedReplicas)).To(Succeed())
+
+	reconciled, err := getMicrovmDeployment(client, testMicrovmDeploymentName, testNamespace)
+	g.Expect(err).NotTo(HaveOccurred(), "Getting microvm should not fail")
+
+	assertMDFinalizer(g, reconciled)
+	assertConditionTrue(g, reconciled, infrav1.MicrovmDeploymentReadyCondition)
+	g.Expect(reconciled.Status.Ready).To(BeTrue(), "MicrovmDeployment should be ready now")
+	g.Expect(reconciled.Status.Replicas).To(Equal(initialReplicaCount), "Expected the record to contain 4 replicas")
+	g.Expect(reconciled.Status.ReadyReplicas).To(Equal(initialReplicaCount), "Expected all replicas to be ready")
+	g.Expect(microvmReplicaSetsCreated(g, client)).To(Equal(initialReplicaSetCount), "Expected 2 replicasets to exist")
+
+	// update, scale down to 1
+	reconciled.Spec.Hosts = []microvm.Host{{Endpoint: "1.2.3.4:9090"}}
+	g.Expect(client.Update(context.TODO(), reconciled)).To(Succeed())
+
+	// first reconciliation
+	result, err := reconcileMicrovmDeployment(client)
+	g.Expect(err).NotTo(HaveOccurred(), "Reconciling microvmdeployment the first time should not error")
+	g.Expect(result.IsZero()).To(BeFalse(), "Expect requeue to be requested after update")
+
+	reconciled, err = getMicrovmDeployment(client, testMicrovmDeploymentName, testNamespace)
+	g.Expect(err).NotTo(HaveOccurred(), "Getting microvmdeployment should not fail")
+
+	assertConditionFalse(g, reconciled, infrav1.MicrovmDeploymentReadyCondition, infrav1.MicrovmDeploymentUpdatingReason)
+	g.Expect(reconciled.Status.Ready).To(BeFalse(), "MicrovmDeployment should not be ready")
+	g.Expect(reconciled.Status.Replicas).To(Equal(initialReplicaCount), "Expected the record to contain 4 replicas")
+	g.Expect(reconciled.Status.ReadyReplicas).To(Equal(initialReplicaCount), "Expected all replicas to be ready")
+
+	// second reconciliation
+	result, err = reconcileMicrovmDeployment(client)
+	g.Expect(err).NotTo(HaveOccurred(), "Reconciling microvmdeployment the second time should not error")
+	g.Expect(result.IsZero()).To(BeTrue(), "Expect requeue to not be requested after reconcile")
+
+	reconciled, err = getMicrovmDeployment(client, testMicrovmDeploymentName, testNamespace)
+	g.Expect(err).NotTo(HaveOccurred(), "Getting microvmdeployment should not fail")
+
+	assertConditionTrue(g, reconciled, infrav1.MicrovmDeploymentReadyCondition)
+	g.Expect(reconciled.Status.Ready).To(BeTrue(), "MicrovmDeployment should be ready again")
+	g.Expect(reconciled.Status.Replicas).To(Equal(scaledReplicaCount), "Expected the record to contain 2 replicas")
+	g.Expect(reconciled.Status.ReadyReplicas).To(Equal(scaledReplicaCount), "Expected all replicas to be ready")
+	g.Expect(microvmReplicaSetsCreated(g, client)).To(Equal(int(scaledReplicaSetCount)), "Expected replicasets to have been scaled down after two reconciliations")
+}