diff --git a/cli/command/service/formatter.go b/cli/command/service/formatter.go
index 4c81406a4a..972129a7a3 100644
--- a/cli/command/service/formatter.go
+++ b/cli/command/service/formatter.go
@@ -49,6 +49,9 @@ Placement:
 {{- if .TaskPlacementPreferences }}
  Preferences:   {{ .TaskPlacementPreferences }}
 {{- end }}
+{{- if .MaxReplicas }}
+ Max Replicas Per Node:   {{ .MaxReplicas }}
+{{- end }}
 {{- if .HasUpdateConfig }}
 UpdateConfig:
  Parallelism:	{{ .UpdateParallelism }}
@@ -284,6 +287,13 @@ func (ctx *serviceInspectContext) TaskPlacementPreferences() []string {
 	return strings
 }
 
+func (ctx *serviceInspectContext) MaxReplicas() uint64 {
+	if ctx.Service.Spec.TaskTemplate.Placement != nil {
+		return ctx.Service.Spec.TaskTemplate.Placement.MaxReplicas
+	}
+	return 0
+}
+
 func (ctx *serviceInspectContext) HasUpdateConfig() bool {
 	return ctx.Service.Spec.UpdateConfig != nil
 }
diff --git a/cli/command/service/list.go b/cli/command/service/list.go
index 188ffb6e55..90fe330102 100644
--- a/cli/command/service/list.go
+++ b/cli/command/service/list.go
@@ -120,9 +120,16 @@ func GetServicesStatus(services []swarm.Service, nodes []swarm.Node, tasks []swa
 	for _, service := range services {
 		info[service.ID] = ListInfo{}
 		if service.Spec.Mode.Replicated != nil && service.Spec.Mode.Replicated.Replicas != nil {
-			info[service.ID] = ListInfo{
-				Mode:     "replicated",
-				Replicas: fmt.Sprintf("%d/%d", running[service.ID], *service.Spec.Mode.Replicated.Replicas),
+			if service.Spec.TaskTemplate.Placement != nil && service.Spec.TaskTemplate.Placement.MaxReplicas > 0 {
+				info[service.ID] = ListInfo{
+					Mode:     "replicated",
+					Replicas: fmt.Sprintf("%d/%d (max %d per node)", running[service.ID], *service.Spec.Mode.Replicated.Replicas, service.Spec.TaskTemplate.Placement.MaxReplicas),
+				}
+			} else {
+				info[service.ID] = ListInfo{
+					Mode:     "replicated",
+					Replicas: fmt.Sprintf("%d/%d", running[service.ID], *service.Spec.Mode.Replicated.Replicas),
+				}
 			}
 		} else if service.Spec.Mode.Global != nil {
 			info[service.ID] = ListInfo{
diff --git a/cli/command/service/opts.go b/cli/command/service/opts.go
index 2478ca8e89..ec4ea403d6 100644
--- a/cli/command/service/opts.go
+++ b/cli/command/service/opts.go
@@ -500,6 +500,7 @@ type serviceOptions struct {
 	restartPolicy  restartPolicyOptions
 	constraints    opts.ListOpts
 	placementPrefs placementPrefOpts
+	maxReplicas    uint64
 	update         updateOptions
 	rollback       updateOptions
 	networks       opts.NetworkOpt
@@ -541,6 +542,10 @@ func (options *serviceOptions) ToServiceMode() (swarm.ServiceMode, error) {
 			return serviceMode, errors.Errorf("replicas can only be used with replicated mode")
 		}
 
+		if options.maxReplicas > 0 {
+			return serviceMode, errors.New("replicas-max-per-node can only be used with replicated mode")
+		}
+
 		serviceMode.Global = &swarm.GlobalService{}
 	case "replicated":
 		serviceMode.Replicated = &swarm.ReplicatedService{
@@ -645,6 +650,7 @@ func (options *serviceOptions) ToService(ctx context.Context, apiClient client.N
 			Placement: &swarm.Placement{
 				Constraints: options.constraints.GetAll(),
 				Preferences: options.placementPrefs.prefs,
+				MaxReplicas: options.maxReplicas,
 			},
 			LogDriver: options.logDriver.toLogDriver(),
 		},
@@ -747,6 +753,8 @@ func addServiceFlags(flags *pflag.FlagSet, opts *serviceOptions, defaultFlagValu
 
 	flags.Var(&opts.stopGrace, flagStopGracePeriod, flagDesc(flagStopGracePeriod, "Time to wait before force killing a container (ns|us|ms|s|m|h)"))
 	flags.Var(&opts.replicas, flagReplicas, "Number of tasks")
+	flags.Uint64Var(&opts.maxReplicas, flagMaxReplicas, defaultFlagValues.getUint64(flagMaxReplicas), "Maximum number of tasks per node (default 0 = unlimited)")
+	flags.SetAnnotation(flagMaxReplicas, "version", []string{"1.40"})
 
 	flags.StringVar(&opts.restartPolicy.condition, flagRestartCondition, "", flagDesc(flagRestartCondition, `Restart when condition is met ("none"|"on-failure"|"any")`))
 	flags.Var(&opts.restartPolicy.delay, flagRestartDelay, flagDesc(flagRestartDelay, "Delay between restart attempts (ns|us|ms|s|m|h)"))
@@ -853,6 +861,7 @@ const (
 	flagLabelAdd                = "label-add"
 	flagLimitCPU                = "limit-cpu"
 	flagLimitMemory             = "limit-memory"
+	flagMaxReplicas             = "replicas-max-per-node"
 	flagMode                    = "mode"
 	flagMount                   = "mount"
 	flagMountRemove             = "mount-rm"
diff --git a/cli/command/service/opts_test.go b/cli/command/service/opts_test.go
index 344893f552..a128e24152 100644
--- a/cli/command/service/opts_test.go
+++ b/cli/command/service/opts_test.go
@@ -224,3 +224,12 @@ func TestToServiceUpdateRollback(t *testing.T) {
 	assert.Check(t, is.DeepEqual(service.UpdateConfig, expected.UpdateConfig))
 	assert.Check(t, is.DeepEqual(service.RollbackConfig, expected.RollbackConfig))
 }
+
+func TestToServiceMaxReplicasGlobalModeConflict(t *testing.T) {
+	opt := serviceOptions{
+		mode:        "global",
+		maxReplicas: 1,
+	}
+	_, err := opt.ToServiceMode()
+	assert.Error(t, err, "replicas-max-per-node can only be used with replicated mode")
+}
diff --git a/cli/command/service/update.go b/cli/command/service/update.go
index 97a67576a8..12ac6ed118 100644
--- a/cli/command/service/update.go
+++ b/cli/command/service/update.go
@@ -387,6 +387,10 @@ func updateService(ctx context.Context, apiClient client.NetworkAPIClient, flags
 		return err
 	}
 
+	if anyChanged(flags, flagMaxReplicas) {
+		updateUint64(flagMaxReplicas, &task.Placement.MaxReplicas)
+	}
+
 	if anyChanged(flags, flagUpdateParallelism, flagUpdateDelay, flagUpdateMonitor, flagUpdateFailureAction, flagUpdateMaxFailureRatio, flagUpdateOrder) {
 		if spec.UpdateConfig == nil {
 			spec.UpdateConfig = updateConfigFromDefaults(defaults.Service.Update)
diff --git a/cli/command/service/update_test.go b/cli/command/service/update_test.go
index 5cfc3b85ed..6ece18195c 100644
--- a/cli/command/service/update_test.go
+++ b/cli/command/service/update_test.go
@@ -808,3 +808,23 @@ func TestUpdateNetworks(t *testing.T) {
 	assert.NilError(t, err)
 	assert.Check(t, is.DeepEqual([]swarm.NetworkAttachmentConfig{{Target: "id999"}}, svc.TaskTemplate.Networks))
 }
+
+func TestUpdateMaxReplicas(t *testing.T) {
+	ctx := context.Background()
+
+	svc := swarm.ServiceSpec{
+		TaskTemplate: swarm.TaskSpec{
+			ContainerSpec: &swarm.ContainerSpec{},
+			Placement: &swarm.Placement{
+				MaxReplicas: 1,
+			},
+		},
+	}
+
+	flags := newUpdateCommand(nil).Flags()
+	flags.Set(flagMaxReplicas, "2")
+	err := updateService(ctx, nil, flags, &svc)
+	assert.NilError(t, err)
+
+	assert.DeepEqual(t, svc.TaskTemplate.Placement, &swarm.Placement{MaxReplicas: uint64(2)})
+}
diff --git a/docs/reference/commandline/service_create.md b/docs/reference/commandline/service_create.md
index 1840b767da..daf7107540 100644
--- a/docs/reference/commandline/service_create.md
+++ b/docs/reference/commandline/service_create.md
@@ -61,6 +61,7 @@ Options:
   -q, --quiet                              Suppress progress output
       --read-only                          Mount the container's root filesystem as read only
       --replicas uint                      Number of tasks
+      --replicas-max-per-node uint         Maximum number of tasks per node (default 0 = unlimited)
       --reserve-cpu decimal                Reserve CPUs
       --reserve-memory bytes               Reserve Memory
       --restart-condition string           Restart when condition is met ("none"|"on-failure"|"any") (default "any")
@@ -757,6 +758,26 @@ appends a new placement preference after all existing placement preferences.
 `--placement-pref-rm` removes an existing placement preference that matches the
 argument.
 
+### Specify maximum replicas per node (--replicas-max-per-node)
+
+Use the `--replicas-max-per-node` flag to set the maximum number of replica tasks that can run on a node.
+The following command creates a nginx service with 2 replica tasks but only one replica task per node.
+
+One example where this can be useful is to balance tasks over a set of data centers together with `--placement-pref`
+and let `--replicas-max-per-node` setting make sure that replicas are not migrated to another datacenter during
+maintenance or datacenter failure.
+
+The example below illustrates this:
+
+```bash
+$ docker service create \
+  --name nginx \
+  --replicas 2 \
+  --replicas-max-per-node 1 \
+  --placement-pref 'spread=node.labels.datacenter' \
+  nginx
+```
+
 ### Attach a service to an existing network (--network)
 
 You can use overlay networks to connect one or more services within the swarm.