cmd/k8s-operator,k8s-operator,go.mod: optionally create ServiceMonitor (#14248)

* cmd/k8s-operator,k8s-operator,go.mod: optionally create ServiceMonitor

Adds a new spec.metrics.serviceMonitor field to ProxyClass.
If that's set to true (and metrics are enabled), the operator
will create a Prometheus ServiceMonitor for each proxy to which
the ProxyClass applies.
Additionally, create a metrics Service for each proxy that has
metrics enabled.

Updates tailscale/tailscale#11292

Signed-off-by: Irbe Krumina <irbe@tailscale.com>

cmd/k8s-operator/connector.go

@@ -189,6 +189,7 @@ func (a *ConnectorReconciler) maybeProvisionConnector(ctx context.Context, logge
 			isExitNode: cn.Spec.ExitNode,
 		},
 		ProxyClassName: proxyClass,
+		proxyType:      proxyTypeConnector,
 	}
 
 	if cn.Spec.SubnetRouter != nil && len(cn.Spec.SubnetRouter.AdvertiseRoutes) > 0 {
@@ -253,7 +254,7 @@ func (a *ConnectorReconciler) maybeProvisionConnector(ctx context.Context, logge
 }
 
 func (a *ConnectorReconciler) maybeCleanupConnector(ctx context.Context, logger *zap.SugaredLogger, cn *tsapi.Connector) (bool, error) {
-	if done, err := a.ssr.Cleanup(ctx, logger, childResourceLabels(cn.Name, a.tsnamespace, "connector")); err != nil {
+	if done, err := a.ssr.Cleanup(ctx, logger, childResourceLabels(cn.Name, a.tsnamespace, "connector"), proxyTypeConnector); err != nil {
 		return false, fmt.Errorf("failed to cleanup Connector resources: %w", err)
 	} else if !done {
 		logger.Debugf("Connector cleanup not done yet, waiting for next reconcile")

cmd/k8s-operator/depaware.txt

@@ -378,7 +378,7 @@ tailscale.com/cmd/k8s-operator dependencies: (generated by github.com/tailscale/
         k8s.io/api/storage/v1beta1                                   from k8s.io/client-go/applyconfigurations/storage/v1beta1+
         k8s.io/api/storagemigration/v1alpha1                         from k8s.io/client-go/applyconfigurations/storagemigration/v1alpha1+
         k8s.io/apiextensions-apiserver/pkg/apis/apiextensions        from k8s.io/apiextensions-apiserver/pkg/apis/apiextensions/v1
-     💣 k8s.io/apiextensions-apiserver/pkg/apis/apiextensions/v1     from sigs.k8s.io/controller-runtime/pkg/webhook/conversion
+     💣 k8s.io/apiextensions-apiserver/pkg/apis/apiextensions/v1     from sigs.k8s.io/controller-runtime/pkg/webhook/conversion+
         k8s.io/apimachinery/pkg/api/equality                         from k8s.io/apiextensions-apiserver/pkg/apis/apiextensions/v1+
         k8s.io/apimachinery/pkg/api/errors                           from k8s.io/apimachinery/pkg/util/managedfields/internal+
         k8s.io/apimachinery/pkg/api/meta                             from k8s.io/apimachinery/pkg/api/validation+

cmd/k8s-operator/deploy/chart/templates/operator-rbac.yaml

@@ -30,6 +30,10 @@ rules:
 - apiGroups: ["tailscale.com"]
   resources: ["recorders", "recorders/status"]
   verbs: ["get", "list", "watch", "update"]
+- apiGroups: ["apiextensions.k8s.io"]
+  resources: ["customresourcedefinitions"]
+  verbs: ["get", "list", "watch"]
+  resourceNames: ["servicemonitors.monitoring.coreos.com"]
 ---
 apiVersion: rbac.authorization.k8s.io/v1
 kind: ClusterRoleBinding
@@ -65,6 +69,9 @@ rules:
 - apiGroups: ["rbac.authorization.k8s.io"]
   resources: ["roles", "rolebindings"]
   verbs: ["get", "create", "patch", "update", "list", "watch"]
+- apiGroups: ["monitoring.coreos.com"]
+  resources: ["servicemonitors"]
+  verbs: ["get", "list", "update", "create", "delete"]
 ---
 apiVersion: rbac.authorization.k8s.io/v1
 kind: RoleBinding

cmd/k8s-operator/deploy/crds/tailscale.com_proxyclasses.yaml

@@ -74,6 +74,8 @@ spec:
                       description: |-
                         Setting enable to true will make the proxy serve Tailscale metrics
                         at <pod-ip>:9002/metrics.
+                        A metrics Service named <proxy-statefulset>-metrics will also be created in the operator's namespace and will
+                        serve the metrics at <service-ip>:9002/metrics.
 
                         In 1.78.x and 1.80.x, this field also serves as the default value for
                         .spec.statefulSet.pod.tailscaleContainer.debug.enable. From 1.82.0, both
@@ -81,6 +83,25 @@ spec:
 
                         Defaults to false.
                       type: boolean
+                    serviceMonitor:
+                      description: |-
+                        Enable to create a Prometheus ServiceMonitor for scraping the proxy's Tailscale metrics.
+                        The ServiceMonitor will select the metrics Service that gets created when metrics are enabled.
+                        The ingested metrics for each Service monitor will have labels to identify the proxy:
+                        ts_proxy_type: ingress_service|ingress_resource|connector|proxygroup
+                        ts_proxy_parent_name: name of the parent resource (i.e name of the Connector, Tailscale Ingress, Tailscale Service or ProxyGroup)
+                        ts_proxy_parent_namespace: namespace of the parent resource (if the parent resource is not cluster scoped)
+                        job: ts_<proxy type>_[<parent namespace>]_<parent_name>
+                      type: object
+                      required:
+                        - enable
+                      properties:
+                        enable:
+                          description: If Enable is set to true, a Prometheus ServiceMonitor will be created. Enable can only be set to true if metrics are enabled.
+                          type: boolean
+                  x-kubernetes-validations:
+                    - rule: '!(has(self.serviceMonitor) && self.serviceMonitor.enable  && !self.enable)'
+                      message: ServiceMonitor can only be enabled if metrics are enabled
                 statefulSet:
                   description: |-
                     Configuration parameters for the proxy's StatefulSet. Tailscale

cmd/k8s-operator/deploy/manifests/operator.yaml

@@ -541,6 +541,8 @@ spec:
                                         description: |-
                                             Setting enable to true will make the proxy serve Tailscale metrics
                                             at <pod-ip>:9002/metrics.
+                                            A metrics Service named <proxy-statefulset>-metrics will also be created in the operator's namespace and will
+                                            serve the metrics at <service-ip>:9002/metrics.
 
                                             In 1.78.x and 1.80.x, this field also serves as the default value for
                                             .spec.statefulSet.pod.tailscaleContainer.debug.enable. From 1.82.0, both
@@ -548,9 +550,28 @@ spec:
 
                                             Defaults to false.
                                         type: boolean
+                                    serviceMonitor:
+                                        description: |-
+                                            Enable to create a Prometheus ServiceMonitor for scraping the proxy's Tailscale metrics.
+                                            The ServiceMonitor will select the metrics Service that gets created when metrics are enabled.
+                                            The ingested metrics for each Service monitor will have labels to identify the proxy:
+                                            ts_proxy_type: ingress_service|ingress_resource|connector|proxygroup
+                                            ts_proxy_parent_name: name of the parent resource (i.e name of the Connector, Tailscale Ingress, Tailscale Service or ProxyGroup)
+                                            ts_proxy_parent_namespace: namespace of the parent resource (if the parent resource is not cluster scoped)
+                                            job: ts_<proxy type>_[<parent namespace>]_<parent_name>
+                                        properties:
+                                            enable:
+                                                description: If Enable is set to true, a Prometheus ServiceMonitor will be created. Enable can only be set to true if metrics are enabled.
+                                                type: boolean
+                                        required:
+                                            - enable
+                                        type: object
                                 required:
                                     - enable
                                 type: object
+                                x-kubernetes-validations:
+                                    - message: ServiceMonitor can only be enabled if metrics are enabled
+                                      rule: '!(has(self.serviceMonitor) && self.serviceMonitor.enable  && !self.enable)'
                             statefulSet:
                                 description: |-
                                     Configuration parameters for the proxy's StatefulSet. Tailscale
@@ -4648,6 +4669,16 @@ rules:
         - list
         - watch
         - update
+    - apiGroups:
+        - apiextensions.k8s.io
+      resourceNames:
+        - servicemonitors.monitoring.coreos.com
+      resources:
+        - customresourcedefinitions
+      verbs:
+        - get
+        - list
+        - watch
 ---
 apiVersion: rbac.authorization.k8s.io/v1
 kind: ClusterRoleBinding
@@ -4728,6 +4759,16 @@ rules:
         - update
         - list
         - watch
+    - apiGroups:
+        - monitoring.coreos.com
+      resources:
+        - servicemonitors
+      verbs:
+        - get
+        - list
+        - update
+        - create
+        - delete
 ---
 apiVersion: rbac.authorization.k8s.io/v1
 kind: Role

cmd/k8s-operator/ingress.go

@@ -90,7 +90,7 @@ func (a *IngressReconciler) maybeCleanup(ctx context.Context, logger *zap.Sugare
 		return nil
 	}
 
-	if done, err := a.ssr.Cleanup(ctx, logger, childResourceLabels(ing.Name, ing.Namespace, "ingress")); err != nil {
+	if done, err := a.ssr.Cleanup(ctx, logger, childResourceLabels(ing.Name, ing.Namespace, "ingress"), proxyTypeIngressResource); err != nil {
 		return fmt.Errorf("failed to cleanup: %w", err)
 	} else if !done {
 		logger.Debugf("cleanup not done yet, waiting for next reconcile")
@@ -268,6 +268,7 @@ func (a *IngressReconciler) maybeProvision(ctx context.Context, logger *zap.Suga
 		Tags:                tags,
 		ChildResourceLabels: crl,
 		ProxyClassName:      proxyClass,
+		proxyType:           proxyTypeIngressResource,
 	}
 
 	if val := ing.GetAnnotations()[AnnotationExperimentalForwardClusterTrafficViaL7IngresProxy]; val == "true" {

cmd/k8s-operator/ingress_test.go

@@ -12,6 +12,7 @@
 	appsv1 "k8s.io/api/apps/v1"
 	corev1 "k8s.io/api/core/v1"
 	networkingv1 "k8s.io/api/networking/v1"
+	apiextensionsv1 "k8s.io/apiextensions-apiserver/pkg/apis/apiextensions/v1"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 	"k8s.io/apimachinery/pkg/types"
 	"sigs.k8s.io/controller-runtime/pkg/client/fake"
@@ -271,3 +272,124 @@ func TestTailscaleIngressWithProxyClass(t *testing.T) {
 	opts.proxyClass = ""
 	expectEqual(t, fc, expectedSTSUserspace(t, fc, opts), removeHashAnnotation)
 }
+
+func TestTailscaleIngressWithServiceMonitor(t *testing.T) {
+	pc := &tsapi.ProxyClass{
+		ObjectMeta: metav1.ObjectMeta{Name: "metrics", Generation: 1},
+		Spec: tsapi.ProxyClassSpec{
+			Metrics: &tsapi.Metrics{
+				Enable:         true,
+				ServiceMonitor: &tsapi.ServiceMonitor{Enable: true},
+			},
+		},
+		Status: tsapi.ProxyClassStatus{
+			Conditions: []metav1.Condition{{
+				Status:             metav1.ConditionTrue,
+				Type:               string(tsapi.ProxyClassReady),
+				ObservedGeneration: 1,
+			}}},
+	}
+	crd := &apiextensionsv1.CustomResourceDefinition{ObjectMeta: metav1.ObjectMeta{Name: serviceMonitorCRD}}
+	tsIngressClass := &networkingv1.IngressClass{ObjectMeta: metav1.ObjectMeta{Name: "tailscale"}, Spec: networkingv1.IngressClassSpec{Controller: "tailscale.com/ts-ingress"}}
+	fc := fake.NewClientBuilder().
+		WithScheme(tsapi.GlobalScheme).
+		WithObjects(pc, tsIngressClass).
+		WithStatusSubresource(pc).
+		Build()
+	ft := &fakeTSClient{}
+	fakeTsnetServer := &fakeTSNetServer{certDomains: []string{"foo.com"}}
+	zl, err := zap.NewDevelopment()
+	if err != nil {
+		t.Fatal(err)
+	}
+	ingR := &IngressReconciler{
+		Client: fc,
+		ssr: &tailscaleSTSReconciler{
+			Client:            fc,
+			tsClient:          ft,
+			tsnetServer:       fakeTsnetServer,
+			defaultTags:       []string{"tag:k8s"},
+			operatorNamespace: "operator-ns",
+			proxyImage:        "tailscale/tailscale",
+		},
+		logger: zl.Sugar(),
+	}
+	// 1. Enable metrics- expect metrics Service to be created
+	ing := &networkingv1.Ingress{
+		TypeMeta: metav1.TypeMeta{Kind: "Ingress", APIVersion: "networking.k8s.io/v1"},
+		ObjectMeta: metav1.ObjectMeta{
+			Name:      "test",
+			Namespace: "default",
+			// The apiserver is supposed to set the UID, but the fake client
+			// doesn't. So, set it explicitly because other code later depends
+			// on it being set.
+			UID: types.UID("1234-UID"),
+			Labels: map[string]string{
+				"tailscale.com/proxy-class": "metrics",
+			},
+		},
+		Spec: networkingv1.IngressSpec{
+			IngressClassName: ptr.To("tailscale"),
+			DefaultBackend: &networkingv1.IngressBackend{
+				Service: &networkingv1.IngressServiceBackend{
+					Name: "test",
+					Port: networkingv1.ServiceBackendPort{
+						Number: 8080,
+					},
+				},
+			},
+			TLS: []networkingv1.IngressTLS{
+				{Hosts: []string{"default-test"}},
+			},
+		},
+	}
+	mustCreate(t, fc, ing)
+	mustCreate(t, fc, &corev1.Service{
+		ObjectMeta: metav1.ObjectMeta{
+			Name:      "test",
+			Namespace: "default",
+		},
+		Spec: corev1.ServiceSpec{
+			ClusterIP: "1.2.3.4",
+			Ports: []corev1.ServicePort{{
+				Port: 8080,
+				Name: "http"},
+			},
+		},
+	})
+
+	expectReconciled(t, ingR, "default", "test")
+
+	fullName, shortName := findGenName(t, fc, "default", "test", "ingress")
+	opts := configOpts{
+		stsName:            shortName,
+		secretName:         fullName,
+		namespace:          "default",
+		tailscaleNamespace: "operator-ns",
+		parentType:         "ingress",
+		hostname:           "default-test",
+		app:                kubetypes.AppIngressResource,
+		enableMetrics:      true,
+		namespaced:         true,
+		proxyType:          proxyTypeIngressResource,
+	}
+	serveConfig := &ipn.ServeConfig{
+		TCP: map[uint16]*ipn.TCPPortHandler{443: {HTTPS: true}},
+		Web: map[ipn.HostPort]*ipn.WebServerConfig{"${TS_CERT_DOMAIN}:443": {Handlers: map[string]*ipn.HTTPHandler{"/": {Proxy: "http://1.2.3.4:8080/"}}}},
+	}
+	opts.serveConfig = serveConfig
+
+	expectEqual(t, fc, expectedSecret(t, fc, opts), nil)
+	expectEqual(t, fc, expectedHeadlessService(shortName, "ingress"), nil)
+	expectEqual(t, fc, expectedMetricsService(opts), nil)
+	expectEqual(t, fc, expectedSTSUserspace(t, fc, opts), removeHashAnnotation)
+	// 2. Enable ServiceMonitor - should not error when there is no ServiceMonitor CRD in cluster
+	mustUpdate(t, fc, "", "metrics", func(pc *tsapi.ProxyClass) {
+		pc.Spec.Metrics.ServiceMonitor = &tsapi.ServiceMonitor{Enable: true}
+	})
+	expectReconciled(t, ingR, "default", "test")
+	// 3. Create ServiceMonitor CRD and reconcile- ServiceMonitor should get created
+	mustCreate(t, fc, crd)
+	expectReconciled(t, ingR, "default", "test")
+	expectEqualUnstructured(t, fc, expectedServiceMonitor(t, opts))
+}

cmd/k8s-operator/metrics_resources.go

@@ -0,0 +1,272 @@
+// Copyright (c) Tailscale Inc & AUTHORS
+// SPDX-License-Identifier: BSD-3-Clause
+
+//go:build !plan9
+
+package main
+
+import (
+	"context"
+	"fmt"
+
+	"go.uber.org/zap"
+	corev1 "k8s.io/api/core/v1"
+	apierrors "k8s.io/apimachinery/pkg/api/errors"
+	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/apis/meta/v1/unstructured"
+	"k8s.io/apimachinery/pkg/runtime"
+	"k8s.io/apimachinery/pkg/types"
+	"sigs.k8s.io/controller-runtime/pkg/client"
+	tsapi "tailscale.com/k8s-operator/apis/v1alpha1"
+)
+
+const (
+	labelMetricsTarget = "tailscale.com/metrics-target"
+
+	// These labels get transferred from the metrics Service to the ingested Prometheus metrics.
+	labelPromProxyType            = "ts_proxy_type"
+	labelPromProxyParentName      = "ts_proxy_parent_name"
+	labelPromProxyParentNamespace = "ts_proxy_parent_namespace"
+	labelPromJob                  = "ts_prom_job"
+
+	serviceMonitorCRD = "servicemonitors.monitoring.coreos.com"
+)
+
+// ServiceMonitor contains a subset of fields of servicemonitors.monitoring.coreos.com Custom Resource Definition.
+// Duplicating it here allows us to avoid importing prometheus-operator library.
+// https://github.com/prometheus-operator/prometheus-operator/blob/bb4514e0d5d69f20270e29cfd4ad39b87865ccdf/pkg/apis/monitoring/v1/servicemonitor_types.go#L40
+type ServiceMonitor struct {
+	metav1.TypeMeta   `json:",inline"`
+	metav1.ObjectMeta `json:"metadata"`
+	Spec              ServiceMonitorSpec `json:"spec"`
+}
+
+// https://github.com/prometheus-operator/prometheus-operator/blob/bb4514e0d5d69f20270e29cfd4ad39b87865ccdf/pkg/apis/monitoring/v1/servicemonitor_types.go#L55
+type ServiceMonitorSpec struct {
+	// Endpoints defines the endpoints to be scraped on the selected Service(s).
+	// https://github.com/prometheus-operator/prometheus-operator/blob/bb4514e0d5d69f20270e29cfd4ad39b87865ccdf/pkg/apis/monitoring/v1/servicemonitor_types.go#L82
+	Endpoints []ServiceMonitorEndpoint `json:"endpoints"`
+	// JobLabel is the label on the Service whose value will become the value of the Prometheus job label for the metrics ingested via this ServiceMonitor.
+	// https://github.com/prometheus-operator/prometheus-operator/blob/bb4514e0d5d69f20270e29cfd4ad39b87865ccdf/pkg/apis/monitoring/v1/servicemonitor_types.go#L66
+	JobLabel string `json:"jobLabel"`
+	// NamespaceSelector selects the namespace of Service(s) that this ServiceMonitor allows to scrape.
+	// https://github.com/prometheus-operator/prometheus-operator/blob/bb4514e0d5d69f20270e29cfd4ad39b87865ccdf/pkg/apis/monitoring/v1/servicemonitor_types.go#L88
+	NamespaceSelector ServiceMonitorNamespaceSelector `json:"namespaceSelector,omitempty"`
+	// Selector is the label selector for Service(s) that this ServiceMonitor allows to scrape.
+	// https://github.com/prometheus-operator/prometheus-operator/blob/bb4514e0d5d69f20270e29cfd4ad39b87865ccdf/pkg/apis/monitoring/v1/servicemonitor_types.go#L85
+	Selector metav1.LabelSelector `json:"selector"`
+	// TargetLabels are labels on the selected Service that should be applied as Prometheus labels to the ingested metrics.
+	// https://github.com/prometheus-operator/prometheus-operator/blob/bb4514e0d5d69f20270e29cfd4ad39b87865ccdf/pkg/apis/monitoring/v1/servicemonitor_types.go#L72
+	TargetLabels []string `json:"targetLabels"`
+}
+
+// ServiceMonitorNamespaceSelector selects namespaces in which Prometheus operator will attempt to find Services for
+// this ServiceMonitor.
+// https://github.com/prometheus-operator/prometheus-operator/blob/bb4514e0d5d69f20270e29cfd4ad39b87865ccdf/pkg/apis/monitoring/v1/servicemonitor_types.go#L88
+type ServiceMonitorNamespaceSelector struct {
+	MatchNames []string `json:"matchNames,omitempty"`
+}
+
+// ServiceMonitorEndpoint defines an endpoint of Service to scrape. We only define port here. Prometheus by default
+// scrapes /metrics path, which is what we want.
+type ServiceMonitorEndpoint struct {
+	// Port is the name of the Service port that Prometheus will scrape.
+	Port string `json:"port,omitempty"`
+}
+
+func reconcileMetricsResources(ctx context.Context, logger *zap.SugaredLogger, opts *metricsOpts, pc *tsapi.ProxyClass, cl client.Client) error {
+	if opts.proxyType == proxyTypeEgress {
+		// Metrics are currently not being enabled for standalone egress proxies.
+		return nil
+	}
+	if pc == nil || pc.Spec.Metrics == nil || !pc.Spec.Metrics.Enable {
+		return maybeCleanupMetricsResources(ctx, opts, cl)
+	}
+	metricsSvc := &corev1.Service{
+		ObjectMeta: metav1.ObjectMeta{
+			Name:      metricsResourceName(opts.proxyStsName),
+			Namespace: opts.tsNamespace,
+			Labels:    metricsResourceLabels(opts),
+		},
+		Spec: corev1.ServiceSpec{
+			Selector: opts.proxyLabels,
+			Type:     corev1.ServiceTypeClusterIP,
+			Ports:    []corev1.ServicePort{{Protocol: "TCP", Port: 9002, Name: "metrics"}},
+		},
+	}
+	var err error
+	metricsSvc, err = createOrUpdate(ctx, cl, opts.tsNamespace, metricsSvc, func(svc *corev1.Service) {
+		svc.Spec.Ports = metricsSvc.Spec.Ports
+		svc.Spec.Selector = metricsSvc.Spec.Selector
+	})
+	if err != nil {
+		return fmt.Errorf("error ensuring metrics Service: %w", err)
+	}
+
+	crdExists, err := hasServiceMonitorCRD(ctx, cl)
+	if err != nil {
+		return fmt.Errorf("error verifying that %q CRD exists: %w", serviceMonitorCRD, err)
+	}
+	if !crdExists {
+		return nil
+	}
+
+	if pc.Spec.Metrics.ServiceMonitor == nil || !pc.Spec.Metrics.ServiceMonitor.Enable {
+		return maybeCleanupServiceMonitor(ctx, cl, opts.proxyStsName, opts.tsNamespace)
+	}
+
+	logger.Info("ensuring ServiceMonitor for metrics Service %s/%s", metricsSvc.Namespace, metricsSvc.Name)
+	svcMonitor, err := newServiceMonitor(metricsSvc)
+	if err != nil {
+		return fmt.Errorf("error creating ServiceMonitor: %w", err)
+	}
+	// We don't use createOrUpdate here because that does not work with unstructured types. We also do not update
+	// the ServiceMonitor because it is not expected that any of its fields would change. Currently this is good
+	// enough, but in future we might want to add logic to create-or-update unstructured types.
+	err = cl.Get(ctx, client.ObjectKeyFromObject(metricsSvc), svcMonitor.DeepCopy())
+	if apierrors.IsNotFound(err) {
+		if err := cl.Create(ctx, svcMonitor); err != nil {
+			return fmt.Errorf("error creating ServiceMonitor: %w", err)
+		}
+		return nil
+	}
+	if err != nil {
+		return fmt.Errorf("error getting ServiceMonitor: %w", err)
+	}
+	return nil
+}
+
+// maybeCleanupMetricsResources ensures that any metrics resources created for a proxy are deleted. Only metrics Service
+// gets deleted explicitly because the ServiceMonitor has Service's owner reference, so gets garbage collected
+// automatically.
+func maybeCleanupMetricsResources(ctx context.Context, opts *metricsOpts, cl client.Client) error {
+	sel := metricsSvcSelector(opts.proxyLabels, opts.proxyType)
+	return cl.DeleteAllOf(ctx, &corev1.Service{}, client.InNamespace(opts.tsNamespace), client.MatchingLabels(sel))
+}
+
+// maybeCleanupServiceMonitor cleans up any ServiceMonitor created for the named proxy StatefulSet.
+func maybeCleanupServiceMonitor(ctx context.Context, cl client.Client, stsName, ns string) error {
+	smName := metricsResourceName(stsName)
+	sm := serviceMonitorTemplate(smName, ns)
+	u, err := serviceMonitorToUnstructured(sm)
+	if err != nil {
+		return fmt.Errorf("error building ServiceMonitor: %w", err)
+	}
+	err = cl.Get(ctx, types.NamespacedName{Name: smName, Namespace: ns}, u)
+	if apierrors.IsNotFound(err) {
+		return nil // nothing to do
+	}
+	if err != nil {
+		return fmt.Errorf("error verifying if ServiceMonitor %s/%s exists: %w", ns, stsName, err)
+	}
+	return cl.Delete(ctx, u)
+}
+
+// newServiceMonitor takes a metrics Service created for a proxy and constructs and returns a ServiceMonitor for that
+// proxy that can be applied to the kube API server.
+// The ServiceMonitor is returned as Unstructured type - this allows us to avoid importing prometheus-operator API server client/schema.
+func newServiceMonitor(metricsSvc *corev1.Service) (*unstructured.Unstructured, error) {
+	sm := serviceMonitorTemplate(metricsSvc.Name, metricsSvc.Namespace)
+	sm.ObjectMeta.Labels = metricsSvc.Labels
+	sm.ObjectMeta.OwnerReferences = []metav1.OwnerReference{*metav1.NewControllerRef(metricsSvc, corev1.SchemeGroupVersion.WithKind("Service"))}
+	sm.Spec = ServiceMonitorSpec{
+		Selector: metav1.LabelSelector{MatchLabels: metricsSvc.Labels},
+		Endpoints: []ServiceMonitorEndpoint{{
+			Port: "metrics",
+		}},
+		NamespaceSelector: ServiceMonitorNamespaceSelector{
+			MatchNames: []string{metricsSvc.Namespace},
+		},
+		JobLabel: labelPromJob,
+		TargetLabels: []string{
+			labelPromProxyParentName,
+			labelPromProxyParentNamespace,
+			labelPromProxyType,
+		},
+	}
+	return serviceMonitorToUnstructured(sm)
+}
+
+// serviceMonitorToUnstructured takes a ServiceMonitor and converts it to Unstructured type that can be used by the c/r
+// client in Kubernetes API server calls.
+func serviceMonitorToUnstructured(sm *ServiceMonitor) (*unstructured.Unstructured, error) {
+	contents, err := runtime.DefaultUnstructuredConverter.ToUnstructured(sm)
+	if err != nil {
+		return nil, fmt.Errorf("error converting ServiceMonitor to Unstructured: %w", err)
+	}
+	u := &unstructured.Unstructured{}
+	u.SetUnstructuredContent(contents)
+	u.SetGroupVersionKind(sm.GroupVersionKind())
+	return u, nil
+}
+
+// metricsResourceName returns name for metrics Service and ServiceMonitor for a proxy StatefulSet.
+func metricsResourceName(stsName string) string {
+	// Maximum length of StatefulSet name if 52 chars, so this is fine.
+	return fmt.Sprintf("%s-metrics", stsName)
+}
+
+// metricsResourceLabels constructs labels that will be applied to metrics Service and metrics ServiceMonitor for a
+// proxy.
+func metricsResourceLabels(opts *metricsOpts) map[string]string {
+	lbls := map[string]string{
+		LabelManaged:             "true",
+		labelMetricsTarget:       opts.proxyStsName,
+		labelPromProxyType:       opts.proxyType,
+		labelPromProxyParentName: opts.proxyLabels[LabelParentName],
+	}
+	// Include namespace label for proxies created for a namespaced type.
+	if isNamespacedProxyType(opts.proxyType) {
+		lbls[labelPromProxyParentNamespace] = opts.proxyLabels[LabelParentNamespace]
+	}
+	lbls[labelPromJob] = promJobName(opts)
+	return lbls
+}
+
+// promJobName constructs the value of the Prometheus job label that will apply to all metrics for a ServiceMonitor.
+func promJobName(opts *metricsOpts) string {
+	// Include parent resource namespace for proxies created for namespaced types.
+	if opts.proxyType == proxyTypeIngressResource || opts.proxyType == proxyTypeIngressService {
+		return fmt.Sprintf("ts_%s_%s_%s", opts.proxyType, opts.proxyLabels[LabelParentNamespace], opts.proxyLabels[LabelParentName])
+	}
+	return fmt.Sprintf("ts_%s_%s", opts.proxyType, opts.proxyLabels[LabelParentName])
+}
+
+// metricsSvcSelector returns the minimum label set to uniquely identify a metrics Service for a proxy.
+func metricsSvcSelector(proxyLabels map[string]string, proxyType string) map[string]string {
+	sel := map[string]string{
+		labelPromProxyType:       proxyType,
+		labelPromProxyParentName: proxyLabels[LabelParentName],
+	}
+	// Include namespace label for proxies created for a namespaced type.
+	if isNamespacedProxyType(proxyType) {
+		sel[labelPromProxyParentNamespace] = proxyLabels[LabelParentNamespace]
+	}
+	return sel
+}
+
+// serviceMonitorTemplate returns a base ServiceMonitor type that, when converted to Unstructured, is a valid type that
+// can be used in kube API server calls via the c/r client.
+func serviceMonitorTemplate(name, ns string) *ServiceMonitor {
+	return &ServiceMonitor{
+		TypeMeta: metav1.TypeMeta{
+			Kind:       "ServiceMonitor",
+			APIVersion: "monitoring.coreos.com/v1",
+		},
+		ObjectMeta: metav1.ObjectMeta{
+			Name:      name,
+			Namespace: ns,
+		},
+	}
+}
+
+type metricsOpts struct {
+	proxyStsName string            // name of StatefulSet for proxy
+	tsNamespace  string            // namespace in which Tailscale is installed
+	proxyLabels  map[string]string // labels of the proxy StatefulSet
+	proxyType    string
+}
+
+func isNamespacedProxyType(typ string) bool {
+	return typ == proxyTypeIngressResource || typ == proxyTypeIngressService
+}

cmd/k8s-operator/operator.go

@@ -24,8 +24,11 @@
 	discoveryv1 "k8s.io/api/discovery/v1"
 	networkingv1 "k8s.io/api/networking/v1"
 	rbacv1 "k8s.io/api/rbac/v1"
+	apiextensionsv1 "k8s.io/apiextensions-apiserver/pkg/apis/apiextensions/v1"
+	"k8s.io/apimachinery/pkg/fields"
 	"k8s.io/apimachinery/pkg/types"
 	"k8s.io/client-go/rest"
+	toolscache "k8s.io/client-go/tools/cache"
 	"sigs.k8s.io/controller-runtime/pkg/builder"
 	"sigs.k8s.io/controller-runtime/pkg/cache"
 	"sigs.k8s.io/controller-runtime/pkg/client"
@@ -239,21 +242,29 @@ func runReconcilers(opts reconcilerOpts) {
 	nsFilter := cache.ByObject{
 		Field: client.InNamespace(opts.tailscaleNamespace).AsSelector(),
 	}
+	// We watch the ServiceMonitor CRD to ensure that reconcilers are re-triggered if user's workflows result in the
+	// ServiceMonitor CRD applied after some of our resources that define ServiceMonitor creation. This selector
+	// ensures that we only watch the ServiceMonitor CRD and that we don't cache full contents of it.
+	serviceMonitorSelector := cache.ByObject{
+		Field:     fields.SelectorFromSet(fields.Set{"metadata.name": serviceMonitorCRD}),
+		Transform: crdTransformer(startlog),
+	}
 	mgrOpts := manager.Options{
 		// TODO (irbekrm): stricter filtering what we watch/cache/call
 		// reconcilers on. c/r by default starts a watch on any
 		// resources that we GET via the controller manager's client.
 		Cache: cache.Options{
 			ByObject: map[client.Object]cache.ByObject{
-				&corev1.Secret{}:             nsFilter,
+				&corev1.Secret{}:                            nsFilter,
-				&corev1.ServiceAccount{}:     nsFilter,
+				&corev1.ServiceAccount{}:                    nsFilter,
-				&corev1.Pod{}:                nsFilter,
+				&corev1.Pod{}:                               nsFilter,
-				&corev1.ConfigMap{}:          nsFilter,
+				&corev1.ConfigMap{}:                         nsFilter,
-				&appsv1.StatefulSet{}:        nsFilter,
+				&appsv1.StatefulSet{}:                       nsFilter,
-				&appsv1.Deployment{}:         nsFilter,
+				&appsv1.Deployment{}:                        nsFilter,
-				&discoveryv1.EndpointSlice{}: nsFilter,
+				&discoveryv1.EndpointSlice{}:                nsFilter,
-				&rbacv1.Role{}:               nsFilter,
+				&rbacv1.Role{}:                              nsFilter,
-				&rbacv1.RoleBinding{}:        nsFilter,
+				&rbacv1.RoleBinding{}:                       nsFilter,
+				&apiextensionsv1.CustomResourceDefinition{}: serviceMonitorSelector,
 			},
 		},
 		Scheme: tsapi.GlobalScheme,
@@ -422,8 +433,13 @@ func runReconcilers(opts reconcilerOpts) {
 		startlog.Fatalf("could not create egress EndpointSlices reconciler: %v", err)
 	}
 
+	// ProxyClass reconciler gets triggered on ServiceMonitor CRD changes to ensure that any ProxyClasses, that
+	// define that a ServiceMonitor should be created, were set to invalid because the CRD did not exist get
+	// reconciled if the CRD is applied at a later point.
+	serviceMonitorFilter := handler.EnqueueRequestsFromMapFunc(proxyClassesWithServiceMonitor(mgr.GetClient(), opts.log))
 	err = builder.ControllerManagedBy(mgr).
 		For(&tsapi.ProxyClass{}).
+		Watches(&apiextensionsv1.CustomResourceDefinition{}, serviceMonitorFilter).
 		Complete(&ProxyClassReconciler{
 			Client:   mgr.GetClient(),
 			recorder: eventRecorder,
@@ -1018,6 +1034,49 @@ func epsFromExternalNameService(cl client.Client, logger *zap.SugaredLogger, ns
 	}
 }
 
+// proxyClassesWithServiceMonitor returns an event handler that, given that the event is for the Prometheus
+// ServiceMonitor CRD, returns all ProxyClasses that define that a ServiceMonitor should be created.
+func proxyClassesWithServiceMonitor(cl client.Client, logger *zap.SugaredLogger) handler.MapFunc {
+	return func(ctx context.Context, o client.Object) []reconcile.Request {
+		crd, ok := o.(*apiextensionsv1.CustomResourceDefinition)
+		if !ok {
+			logger.Debugf("[unexpected] ServiceMonitor CRD handler received an object that is not a CustomResourceDefinition")
+			return nil
+		}
+		if crd.Name != serviceMonitorCRD {
+			logger.Debugf("[unexpected] ServiceMonitor CRD handler received an unexpected CRD %q", crd.Name)
+			return nil
+		}
+		pcl := &tsapi.ProxyClassList{}
+		if err := cl.List(ctx, pcl); err != nil {
+			logger.Debugf("[unexpected] error listing ProxyClasses: %v", err)
+			return nil
+		}
+		reqs := make([]reconcile.Request, 0)
+		for _, pc := range pcl.Items {
+			if pc.Spec.Metrics != nil && pc.Spec.Metrics.ServiceMonitor != nil && pc.Spec.Metrics.ServiceMonitor.Enable {
+				reqs = append(reqs, reconcile.Request{
+					NamespacedName: types.NamespacedName{Namespace: pc.Namespace, Name: pc.Name},
+				})
+			}
+		}
+		return reqs
+	}
+}
+
+// crdTransformer gets called before a CRD is stored to c/r cache, it removes the CRD spec to reduce memory consumption.
+func crdTransformer(log *zap.SugaredLogger) toolscache.TransformFunc {
+	return func(o any) (any, error) {
+		crd, ok := o.(*apiextensionsv1.CustomResourceDefinition)
+		if !ok {
+			log.Infof("[unexpected] CRD transformer called for a non-CRD type")
+			return crd, nil
+		}
+		crd.Spec = apiextensionsv1.CustomResourceDefinitionSpec{}
+		return crd, nil
+	}
+}
+
 // indexEgressServices adds a local index to a cached Tailscale egress Services meant to be exposed on a ProxyGroup. The
 // index is used a list filter.
 func indexEgressServices(o client.Object) []string {

cmd/k8s-operator/proxyclass.go

@@ -15,6 +15,7 @@
 	dockerref "github.com/distribution/reference"
 	"go.uber.org/zap"
 	corev1 "k8s.io/api/core/v1"
+	apiextensionsv1 "k8s.io/apiextensions-apiserver/pkg/apis/apiextensions/v1"
 	apiequality "k8s.io/apimachinery/pkg/api/equality"
 	apierrors "k8s.io/apimachinery/pkg/api/errors"
 	apivalidation "k8s.io/apimachinery/pkg/api/validation"
@@ -95,7 +96,7 @@ func (pcr *ProxyClassReconciler) Reconcile(ctx context.Context, req reconcile.Re
 	pcr.mu.Unlock()
 
 	oldPCStatus := pc.Status.DeepCopy()
-	if errs := pcr.validate(pc); errs != nil {
+	if errs := pcr.validate(ctx, pc); errs != nil {
 		msg := fmt.Sprintf(messageProxyClassInvalid, errs.ToAggregate().Error())
 		pcr.recorder.Event(pc, corev1.EventTypeWarning, reasonProxyClassInvalid, msg)
 		tsoperator.SetProxyClassCondition(pc, tsapi.ProxyClassReady, metav1.ConditionFalse, reasonProxyClassInvalid, msg, pc.Generation, pcr.clock, logger)
@@ -111,7 +112,7 @@ func (pcr *ProxyClassReconciler) Reconcile(ctx context.Context, req reconcile.Re
 	return reconcile.Result{}, nil
 }
 
-func (pcr *ProxyClassReconciler) validate(pc *tsapi.ProxyClass) (violations field.ErrorList) {
+func (pcr *ProxyClassReconciler) validate(ctx context.Context, pc *tsapi.ProxyClass) (violations field.ErrorList) {
 	if sts := pc.Spec.StatefulSet; sts != nil {
 		if len(sts.Labels) > 0 {
 			if errs := metavalidation.ValidateLabels(sts.Labels, field.NewPath(".spec.statefulSet.labels")); errs != nil {
@@ -167,6 +168,16 @@ func (pcr *ProxyClassReconciler) validate(pc *tsapi.ProxyClass) (violations fiel
 			}
 		}
 	}
+	if pc.Spec.Metrics != nil && pc.Spec.Metrics.ServiceMonitor != nil && pc.Spec.Metrics.ServiceMonitor.Enable {
+		found, err := hasServiceMonitorCRD(ctx, pcr.Client)
+		if err != nil {
+			pcr.logger.Infof("[unexpected]: error retrieving %q CRD: %v", serviceMonitorCRD, err)
+			// best effort validation - don't error out here
+		} else if !found {
+			msg := fmt.Sprintf("ProxyClass defines that a ServiceMonitor custom resource should be created, but %q CRD was not found", serviceMonitorCRD)
+			violations = append(violations, field.TypeInvalid(field.NewPath("spec", "metrics", "serviceMonitor"), "enable", msg))
+		}
+	}
 	// We do not validate embedded fields (security context, resource
 	// requirements etc) as we inherit upstream validation for those fields.
 	// Invalid values would get rejected by upstream validations at apply
@@ -174,6 +185,16 @@ func (pcr *ProxyClassReconciler) validate(pc *tsapi.ProxyClass) (violations fiel
 	return violations
 }
 
+func hasServiceMonitorCRD(ctx context.Context, cl client.Client) (bool, error) {
+	sm := &apiextensionsv1.CustomResourceDefinition{}
+	if err := cl.Get(ctx, types.NamespacedName{Name: serviceMonitorCRD}, sm); apierrors.IsNotFound(err) {
+		return false, nil
+	} else if err != nil {
+		return false, err
+	}
+	return true, nil
+}
+
 // maybeCleanup removes tailscale.com finalizer and ensures that the ProxyClass
 // is no longer counted towards k8s_proxyclass_resources.
 func (pcr *ProxyClassReconciler) maybeCleanup(ctx context.Context, logger *zap.SugaredLogger, pc *tsapi.ProxyClass) error {

cmd/k8s-operator/proxyclass_test.go

@@ -8,10 +8,12 @@
 package main
 
 import (
+	"context"
 	"testing"
 	"time"
 
 	"go.uber.org/zap"
+	apiextensionsv1 "k8s.io/apiextensions-apiserver/pkg/apis/apiextensions/v1"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 	"k8s.io/apimachinery/pkg/types"
 	"k8s.io/client-go/tools/record"
@@ -134,6 +136,25 @@ func TestProxyClass(t *testing.T) {
 		"Warning CustomTSEnvVar ProxyClass overrides the default value for EXPERIMENTAL_ALLOW_PROXYING_CLUSTER_TRAFFIC_VIA_INGRESS env var for tailscale container. Running with custom values for Tailscale env vars is not recommended and might break in the future."}
 	expectReconciled(t, pcr, "", "test")
 	expectEvents(t, fr, expectedEvents)
+
+	// 6. A ProxyClass with ServiceMonitor enabled and in a cluster that has not ServiceMonitor CRD is invalid
+	pc.Spec.Metrics = &tsapi.Metrics{Enable: true, ServiceMonitor: &tsapi.ServiceMonitor{Enable: true}}
+	mustUpdate(t, fc, "", "test", func(proxyClass *tsapi.ProxyClass) {
+		proxyClass.Spec = pc.Spec
+	})
+	expectReconciled(t, pcr, "", "test")
+	msg = `ProxyClass is not valid: spec.metrics.serviceMonitor: Invalid value: "enable": ProxyClass defines that a ServiceMonitor custom resource should be created, but "servicemonitors.monitoring.coreos.com" CRD was not found`
+	tsoperator.SetProxyClassCondition(pc, tsapi.ProxyClassReady, metav1.ConditionFalse, reasonProxyClassInvalid, msg, 0, cl, zl.Sugar())
+	expectEqual(t, fc, pc, nil)
+	expectedEvent = "Warning ProxyClassInvalid " + msg
+	expectEvents(t, fr, []string{expectedEvent})
+
+	// 7. A ProxyClass with ServiceMonitor enabled and in a cluster that does have the ServiceMonitor CRD is valid
+	crd := &apiextensionsv1.CustomResourceDefinition{ObjectMeta: metav1.ObjectMeta{Name: serviceMonitorCRD}}
+	mustCreate(t, fc, crd)
+	expectReconciled(t, pcr, "", "test")
+	tsoperator.SetProxyClassCondition(pc, tsapi.ProxyClassReady, metav1.ConditionTrue, reasonProxyClassValid, reasonProxyClassValid, 0, cl, zl.Sugar())
+	expectEqual(t, fc, pc, nil)
 }
 
 func TestValidateProxyClass(t *testing.T) {
@@ -180,7 +201,7 @@ func TestValidateProxyClass(t *testing.T) {
 	} {
 		t.Run(name, func(t *testing.T) {
 			pcr := &ProxyClassReconciler{}
-			err := pcr.validate(tc.pc)
+			err := pcr.validate(context.Background(), tc.pc)
 			valid := err == nil
 			if valid != tc.valid {
 				t.Errorf("expected valid=%v, got valid=%v, err=%v", tc.valid, valid, err)

cmd/k8s-operator/proxygroup.go

@@ -259,6 +259,15 @@ func (r *ProxyGroupReconciler) maybeProvision(ctx context.Context, pg *tsapi.Pro
 	}); err != nil {
 		return fmt.Errorf("error provisioning StatefulSet: %w", err)
 	}
+	mo := &metricsOpts{
+		tsNamespace:  r.tsNamespace,
+		proxyStsName: pg.Name,
+		proxyLabels:  pgLabels(pg.Name, nil),
+		proxyType:    "proxygroup",
+	}
+	if err := reconcileMetricsResources(ctx, logger, mo, proxyClass, r.Client); err != nil {
+		return fmt.Errorf("error reconciling metrics resources: %w", err)
+	}
 
 	if err := r.cleanupDanglingResources(ctx, pg); err != nil {
 		return fmt.Errorf("error cleaning up dangling resources: %w", err)
@@ -327,6 +336,14 @@ func (r *ProxyGroupReconciler) maybeCleanup(ctx context.Context, pg *tsapi.Proxy
 		}
 	}
 
+	mo := &metricsOpts{
+		proxyLabels: pgLabels(pg.Name, nil),
+		tsNamespace: r.tsNamespace,
+		proxyType:   "proxygroup"}
+	if err := maybeCleanupMetricsResources(ctx, mo, r.Client); err != nil {
+		return false, fmt.Errorf("error cleaning up metrics resources: %w", err)
+	}
+
 	logger.Infof("cleaned up ProxyGroup resources")
 	r.mu.Lock()
 	r.proxyGroups.Remove(pg.UID)

cmd/k8s-operator/proxygroup_test.go

@@ -17,6 +17,7 @@
 	appsv1 "k8s.io/api/apps/v1"
 	corev1 "k8s.io/api/core/v1"
 	rbacv1 "k8s.io/api/rbac/v1"
+	apiextensionsv1 "k8s.io/apiextensions-apiserver/pkg/apis/apiextensions/v1"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 	"k8s.io/client-go/tools/record"
 	"sigs.k8s.io/controller-runtime/pkg/client"
@@ -76,6 +77,13 @@ func TestProxyGroup(t *testing.T) {
 		l:        zl.Sugar(),
 		clock:    cl,
 	}
+	crd := &apiextensionsv1.CustomResourceDefinition{ObjectMeta: metav1.ObjectMeta{Name: serviceMonitorCRD}}
+	opts := configOpts{
+		proxyType:          "proxygroup",
+		stsName:            pg.Name,
+		parentType:         "proxygroup",
+		tailscaleNamespace: "tailscale",
+	}
 
 	t.Run("proxyclass_not_ready", func(t *testing.T) {
 		expectReconciled(t, reconciler, "", pg.Name)
@@ -190,6 +198,27 @@ func TestProxyGroup(t *testing.T) {
 		expectProxyGroupResources(t, fc, pg, true, "518a86e9fae64f270f8e0ec2a2ea6ca06c10f725035d3d6caca132cd61e42a74")
 	})
 
+	t.Run("enable_metrics", func(t *testing.T) {
+		pc.Spec.Metrics = &tsapi.Metrics{Enable: true}
+		mustUpdate(t, fc, "", pc.Name, func(p *tsapi.ProxyClass) {
+			p.Spec = pc.Spec
+		})
+		expectReconciled(t, reconciler, "", pg.Name)
+		expectEqual(t, fc, expectedMetricsService(opts), nil)
+	})
+	t.Run("enable_service_monitor_no_crd", func(t *testing.T) {
+		pc.Spec.Metrics.ServiceMonitor = &tsapi.ServiceMonitor{Enable: true}
+		mustUpdate(t, fc, "", pc.Name, func(p *tsapi.ProxyClass) {
+			p.Spec.Metrics = pc.Spec.Metrics
+		})
+		expectReconciled(t, reconciler, "", pg.Name)
+	})
+	t.Run("create_crd_expect_service_monitor", func(t *testing.T) {
+		mustCreate(t, fc, crd)
+		expectReconciled(t, reconciler, "", pg.Name)
+		expectEqualUnstructured(t, fc, expectedServiceMonitor(t, opts))
+	})
+
 	t.Run("delete_and_cleanup", func(t *testing.T) {
 		if err := fc.Delete(context.Background(), pg); err != nil {
 			t.Fatal(err)
@@ -197,7 +226,7 @@ func TestProxyGroup(t *testing.T) {
 
 		expectReconciled(t, reconciler, "", pg.Name)
 
-		expectMissing[tsapi.Recorder](t, fc, "", pg.Name)
+		expectMissing[tsapi.ProxyGroup](t, fc, "", pg.Name)
 		if expected := 0; reconciler.proxyGroups.Len() != expected {
 			t.Fatalf("expected %d ProxyGroups, got %d", expected, reconciler.proxyGroups.Len())
 		}
@@ -206,6 +235,7 @@ func TestProxyGroup(t *testing.T) {
 		if diff := cmp.Diff(tsClient.deleted, []string{"nodeid-1", "nodeid-2", "nodeid-0"}); diff != "" {
 			t.Fatalf("unexpected deleted devices (-got +want):\n%s", diff)
 		}
+		expectMissing[corev1.Service](t, reconciler, "tailscale", metricsResourceName(pg.Name))
 		// The fake client does not clean up objects whose owner has been
 		// deleted, so we can't test for the owned resources getting deleted.
 	})

cmd/k8s-operator/sts.go

@@ -94,6 +94,12 @@
 	podAnnotationLastSetTailnetTargetFQDN = "tailscale.com/operator-last-set-ts-tailnet-target-fqdn"
 	// podAnnotationLastSetConfigFileHash is sha256 hash of the current tailscaled configuration contents.
 	podAnnotationLastSetConfigFileHash = "tailscale.com/operator-last-set-config-file-hash"
+
+	proxyTypeEgress          = "egress_service"
+	proxyTypeIngressService  = "ingress_service"
+	proxyTypeIngressResource = "ingress_resource"
+	proxyTypeConnector       = "connector"
+	proxyTypeProxyGroup      = "proxygroup"
 )
 
 var (
@@ -122,6 +128,8 @@ type tailscaleSTSConfig struct {
 	Hostname string
 	Tags     []string // if empty, use defaultTags
 
+	proxyType string
+
 	// Connector specifies a configuration of a Connector instance if that's
 	// what this StatefulSet should be created for.
 	Connector *connector
@@ -197,14 +205,22 @@ func (a *tailscaleSTSReconciler) Provision(ctx context.Context, logger *zap.Suga
 	if err != nil {
 		return nil, fmt.Errorf("failed to reconcile statefulset: %w", err)
 	}
-
+	mo := &metricsOpts{
+		proxyStsName: hsvc.Name,
+		tsNamespace:  hsvc.Namespace,
+		proxyLabels:  hsvc.Labels,
+		proxyType:    sts.proxyType,
+	}
+	if err = reconcileMetricsResources(ctx, logger, mo, sts.ProxyClass, a.Client); err != nil {
+		return nil, fmt.Errorf("failed to ensure metrics resources: %w", err)
+	}
 	return hsvc, nil
 }
 
 // Cleanup removes all resources associated that were created by Provision with
 // the given labels. It returns true when all resources have been removed,
 // otherwise it returns false and the caller should retry later.
-func (a *tailscaleSTSReconciler) Cleanup(ctx context.Context, logger *zap.SugaredLogger, labels map[string]string) (done bool, _ error) {
+func (a *tailscaleSTSReconciler) Cleanup(ctx context.Context, logger *zap.SugaredLogger, labels map[string]string, typ string) (done bool, _ error) {
 	// Need to delete the StatefulSet first, and delete it with foreground
 	// cascading deletion. That way, the pod that's writing to the Secret will
 	// stop running before we start looking at the Secret's contents, and
@@ -257,6 +273,14 @@ func (a *tailscaleSTSReconciler) Cleanup(ctx context.Context, logger *zap.Sugare
 			return false, err
 		}
 	}
+	mo := &metricsOpts{
+		proxyLabels: labels,
+		tsNamespace: a.operatorNamespace,
+		proxyType:   typ,
+	}
+	if err := maybeCleanupMetricsResources(ctx, mo, a.Client); err != nil {
+		return false, fmt.Errorf("error cleaning up metrics resources: %w", err)
+	}
 	return true, nil
 }
 

cmd/k8s-operator/svc.go

@@ -152,7 +152,12 @@ func (a *ServiceReconciler) maybeCleanup(ctx context.Context, logger *zap.Sugare
 		return nil
 	}
 
-	if done, err := a.ssr.Cleanup(ctx, logger, childResourceLabels(svc.Name, svc.Namespace, "svc")); err != nil {
+	proxyTyp := proxyTypeEgress
+	if a.shouldExpose(svc) {
+		proxyTyp = proxyTypeIngressService
+	}
+
+	if done, err := a.ssr.Cleanup(ctx, logger, childResourceLabels(svc.Name, svc.Namespace, "svc"), proxyTyp); err != nil {
 		return fmt.Errorf("failed to cleanup: %w", err)
 	} else if !done {
 		logger.Debugf("cleanup not done yet, waiting for next reconcile")
@@ -256,6 +261,10 @@ func (a *ServiceReconciler) maybeProvision(ctx context.Context, logger *zap.Suga
 		ChildResourceLabels: crl,
 		ProxyClassName:      proxyClass,
 	}
+	sts.proxyType = proxyTypeEgress
+	if a.shouldExpose(svc) {
+		sts.proxyType = proxyTypeIngressService
+	}
 
 	a.mu.Lock()
 	if a.shouldExposeClusterIP(svc) {

cmd/k8s-operator/testutils_test.go

@@ -8,6 +8,7 @@
 import (
 	"context"
 	"encoding/json"
+	"fmt"
 	"net/netip"
 	"reflect"
 	"strings"
@@ -21,6 +22,7 @@
 	corev1 "k8s.io/api/core/v1"
 	apierrors "k8s.io/apimachinery/pkg/api/errors"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/apis/meta/v1/unstructured"
 	"k8s.io/apimachinery/pkg/types"
 	"k8s.io/client-go/tools/record"
 	"sigs.k8s.io/controller-runtime/pkg/client"
@@ -39,7 +41,10 @@ type configOpts struct {
 	secretName                                     string
 	hostname                                       string
 	namespace                                      string
+	tailscaleNamespace                             string
+	namespaced                                     bool
 	parentType                                     string
+	proxyType                                      string
 	priorityClassName                              string
 	firewallMode                                   string
 	tailnetTargetIP                                string
@@ -56,6 +61,7 @@ type configOpts struct {
 	app                                            string
 	shouldRemoveAuthKey                            bool
 	secretExtraData                                map[string][]byte
+	enableMetrics                                  bool
 }
 
 func expectedSTS(t *testing.T, cl client.Client, opts configOpts) *appsv1.StatefulSet {
@@ -150,6 +156,29 @@ func expectedSTS(t *testing.T, cl client.Client, opts configOpts) *appsv1.Statef
 		Name:  "TS_INTERNAL_APP",
 		Value: opts.app,
 	})
+	if opts.enableMetrics {
+		tsContainer.Env = append(tsContainer.Env,
+			corev1.EnvVar{
+				Name:  "TS_DEBUG_ADDR_PORT",
+				Value: "$(POD_IP):9001"},
+			corev1.EnvVar{
+				Name:  "TS_TAILSCALED_EXTRA_ARGS",
+				Value: "--debug=$(TS_DEBUG_ADDR_PORT)",
+			},
+			corev1.EnvVar{
+				Name:  "TS_LOCAL_ADDR_PORT",
+				Value: "$(POD_IP):9002",
+			},
+			corev1.EnvVar{
+				Name:  "TS_ENABLE_METRICS",
+				Value: "true",
+			},
+		)
+		tsContainer.Ports = append(tsContainer.Ports,
+			corev1.ContainerPort{Name: "debug", ContainerPort: 9001, Protocol: "TCP"},
+			corev1.ContainerPort{Name: "metrics", ContainerPort: 9002, Protocol: "TCP"},
+		)
+	}
 	ss := &appsv1.StatefulSet{
 		TypeMeta: metav1.TypeMeta{
 			Kind:       "StatefulSet",
@@ -241,6 +270,29 @@ func expectedSTSUserspace(t *testing.T, cl client.Client, opts configOpts) *apps
 			{Name: "serve-config", ReadOnly: true, MountPath: "/etc/tailscaled"},
 		},
 	}
+	if opts.enableMetrics {
+		tsContainer.Env = append(tsContainer.Env,
+			corev1.EnvVar{
+				Name:  "TS_DEBUG_ADDR_PORT",
+				Value: "$(POD_IP):9001"},
+			corev1.EnvVar{
+				Name:  "TS_TAILSCALED_EXTRA_ARGS",
+				Value: "--debug=$(TS_DEBUG_ADDR_PORT)",
+			},
+			corev1.EnvVar{
+				Name:  "TS_LOCAL_ADDR_PORT",
+				Value: "$(POD_IP):9002",
+			},
+			corev1.EnvVar{
+				Name:  "TS_ENABLE_METRICS",
+				Value: "true",
+			},
+		)
+		tsContainer.Ports = append(tsContainer.Ports, corev1.ContainerPort{
+			Name: "debug", ContainerPort: 9001, Protocol: "TCP"},
+			corev1.ContainerPort{Name: "metrics", ContainerPort: 9002, Protocol: "TCP"},
+		)
+	}
 	volumes := []corev1.Volume{
 		{
 			Name: "tailscaledconfig",
@@ -335,6 +387,87 @@ func expectedHeadlessService(name string, parentType string) *corev1.Service {
 	}
 }
 
+func expectedMetricsService(opts configOpts) *corev1.Service {
+	labels := metricsLabels(opts)
+	selector := map[string]string{
+		"tailscale.com/managed":              "true",
+		"tailscale.com/parent-resource":      "test",
+		"tailscale.com/parent-resource-type": opts.parentType,
+	}
+	if opts.namespaced {
+		selector["tailscale.com/parent-resource-ns"] = opts.namespace
+	}
+	return &corev1.Service{
+		ObjectMeta: metav1.ObjectMeta{
+			Name:      metricsResourceName(opts.stsName),
+			Namespace: opts.tailscaleNamespace,
+			Labels:    labels,
+		},
+		Spec: corev1.ServiceSpec{
+			Selector: selector,
+			Type:     corev1.ServiceTypeClusterIP,
+			Ports:    []corev1.ServicePort{{Protocol: "TCP", Port: 9002, Name: "metrics"}},
+		},
+	}
+}
+
+func metricsLabels(opts configOpts) map[string]string {
+	promJob := fmt.Sprintf("ts_%s_default_test", opts.proxyType)
+	if !opts.namespaced {
+		promJob = fmt.Sprintf("ts_%s_test", opts.proxyType)
+	}
+	labels := map[string]string{
+		"tailscale.com/managed":        "true",
+		"tailscale.com/metrics-target": opts.stsName,
+		"ts_prom_job":                  promJob,
+		"ts_proxy_type":                opts.proxyType,
+		"ts_proxy_parent_name":         "test",
+	}
+	if opts.namespaced {
+		labels["ts_proxy_parent_namespace"] = "default"
+	}
+	return labels
+}
+
+func expectedServiceMonitor(t *testing.T, opts configOpts) *unstructured.Unstructured {
+	t.Helper()
+	labels := metricsLabels(opts)
+	name := metricsResourceName(opts.stsName)
+	sm := &ServiceMonitor{
+		ObjectMeta: metav1.ObjectMeta{
+			Name:            name,
+			Namespace:       opts.tailscaleNamespace,
+			Labels:          labels,
+			ResourceVersion: "1",
+			OwnerReferences: []metav1.OwnerReference{{APIVersion: "v1", Kind: "Service", Name: name, BlockOwnerDeletion: ptr.To(true), Controller: ptr.To(true)}},
+		},
+		TypeMeta: metav1.TypeMeta{
+			Kind:       "ServiceMonitor",
+			APIVersion: "monitoring.coreos.com/v1",
+		},
+		Spec: ServiceMonitorSpec{
+			Selector: metav1.LabelSelector{MatchLabels: labels},
+			Endpoints: []ServiceMonitorEndpoint{{
+				Port: "metrics",
+			}},
+			NamespaceSelector: ServiceMonitorNamespaceSelector{
+				MatchNames: []string{opts.tailscaleNamespace},
+			},
+			JobLabel: "ts_prom_job",
+			TargetLabels: []string{
+				"ts_proxy_parent_name",
+				"ts_proxy_parent_namespace",
+				"ts_proxy_type",
+			},
+		},
+	}
+	u, err := serviceMonitorToUnstructured(sm)
+	if err != nil {
+		t.Fatalf("error converting ServiceMonitor to unstructured: %v", err)
+	}
+	return u
+}
+
 func expectedSecret(t *testing.T, cl client.Client, opts configOpts) *corev1.Secret {
 	t.Helper()
 	s := &corev1.Secret{
@@ -502,6 +635,21 @@ func expectEqual[T any, O ptrObject[T]](t *testing.T, client client.Client, want
 	}
 }
 
+func expectEqualUnstructured(t *testing.T, client client.Client, want *unstructured.Unstructured) {
+	t.Helper()
+	got := &unstructured.Unstructured{}
+	got.SetGroupVersionKind(want.GroupVersionKind())
+	if err := client.Get(context.Background(), types.NamespacedName{
+		Name:      want.GetName(),
+		Namespace: want.GetNamespace(),
+	}, got); err != nil {
+		t.Fatalf("getting %q: %v", want.GetName(), err)
+	}
+	if diff := cmp.Diff(got, want); diff != "" {
+		t.Fatalf("unexpected contents of Unstructured (-got +want):\n%s", diff)
+	}
+}
+
 func expectMissing[T any, O ptrObject[T]](t *testing.T, client client.Client, ns, name string) {
 	t.Helper()
 	obj := O(new(T))

go.mod

@@ -396,7 +396,7 @@ require (
 	gopkg.in/yaml.v2 v2.4.0 // indirect
 	gopkg.in/yaml.v3 v3.0.1
 	howett.net/plist v1.0.0 // indirect
-	k8s.io/apiextensions-apiserver v0.30.3 // indirect
+	k8s.io/apiextensions-apiserver v0.30.3
 	k8s.io/klog/v2 v2.130.1 // indirect
 	k8s.io/kube-openapi v0.0.0-20240228011516-70dd3763d340 // indirect
 	k8s.io/utils v0.0.0-20240711033017-18e509b52bc8

k8s-operator/api.md

@@ -326,7 +326,8 @@ _Appears in:_
 
 | Field | Description | Default | Validation |
 | --- | --- | --- | --- |
-| `enable` _boolean_ | Setting enable to true will make the proxy serve Tailscale metrics<br />at <pod-ip>:9002/metrics.<br />In 1.78.x and 1.80.x, this field also serves as the default value for<br />.spec.statefulSet.pod.tailscaleContainer.debug.enable. From 1.82.0, both<br />fields will independently default to false.<br />Defaults to false. |  |  |
+| `enable` _boolean_ | Setting enable to true will make the proxy serve Tailscale metrics<br />at <pod-ip>:9002/metrics.<br />A metrics Service named <proxy-statefulset>-metrics will also be created in the operator's namespace and will<br />serve the metrics at <service-ip>:9002/metrics.<br />In 1.78.x and 1.80.x, this field also serves as the default value for<br />.spec.statefulSet.pod.tailscaleContainer.debug.enable. From 1.82.0, both<br />fields will independently default to false.<br />Defaults to false. |  |  |
+| `serviceMonitor` _[ServiceMonitor](#servicemonitor)_ | Enable to create a Prometheus ServiceMonitor for scraping the proxy's Tailscale metrics.<br />The ServiceMonitor will select the metrics Service that gets created when metrics are enabled.<br />The ingested metrics for each Service monitor will have labels to identify the proxy:<br />ts_proxy_type: ingress_service\|ingress_resource\|connector\|proxygroup<br />ts_proxy_parent_name: name of the parent resource (i.e name of the Connector, Tailscale Ingress, Tailscale Service or ProxyGroup)<br />ts_proxy_parent_namespace: namespace of the parent resource (if the parent resource is not cluster scoped)<br />job: ts_<proxy type>_[<parent namespace>]_<parent_name> |  |  |
 
 
 #### Name
@@ -836,6 +837,22 @@ _Appears in:_
 | `name` _string_ | The name of a Kubernetes Secret in the operator's namespace that contains<br />credentials for writing to the configured bucket. Each key-value pair<br />from the secret's data will be mounted as an environment variable. It<br />should include keys for AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY if<br />using a static access key. |  |  |
 
 
+#### ServiceMonitor
+
+
+
+
+
+
+
+_Appears in:_
+- [Metrics](#metrics)
+
+| Field | Description | Default | Validation |
+| --- | --- | --- | --- |
+| `enable` _boolean_ | If Enable is set to true, a Prometheus ServiceMonitor will be created. Enable can only be set to true if metrics are enabled. |  |  |
+
+
 #### StatefulSet
 
 

k8s-operator/apis/v1alpha1/register.go

@@ -10,6 +10,7 @@
 
 	"tailscale.com/k8s-operator/apis"
 
+	apiextensionsv1 "k8s.io/apiextensions-apiserver/pkg/apis/apiextensions/v1"
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
 	"k8s.io/apimachinery/pkg/runtime"
 	"k8s.io/apimachinery/pkg/runtime/schema"
@@ -39,12 +40,18 @@ func init() {
 	localSchemeBuilder.Register(addKnownTypes)
 
 	GlobalScheme = runtime.NewScheme()
+	// Add core types
 	if err := scheme.AddToScheme(GlobalScheme); err != nil {
 		panic(fmt.Sprintf("failed to add k8s.io scheme: %s", err))
 	}
+	// Add tailscale.com types
 	if err := AddToScheme(GlobalScheme); err != nil {
 		panic(fmt.Sprintf("failed to add tailscale.com scheme: %s", err))
 	}
+	// Add apiextensions types (CustomResourceDefinitions/CustomResourceDefinitionLists)
+	if err := apiextensionsv1.AddToScheme(GlobalScheme); err != nil {
+		panic(fmt.Sprintf("failed to add apiextensions.k8s.io scheme: %s", err))
+	}
 }
 
 // Adds the list of known types to api.Scheme.

k8s-operator/apis/v1alpha1/types_proxyclass.go

@@ -161,9 +161,12 @@ type Pod struct {
 	TopologySpreadConstraints []corev1.TopologySpreadConstraint `json:"topologySpreadConstraints,omitempty"`
 }
 
+// +kubebuilder:validation:XValidation:rule="!(has(self.serviceMonitor) && self.serviceMonitor.enable  && !self.enable)",message="ServiceMonitor can only be enabled if metrics are enabled"
 type Metrics struct {
 	// Setting enable to true will make the proxy serve Tailscale metrics
 	// at <pod-ip>:9002/metrics.
+	// A metrics Service named <proxy-statefulset>-metrics will also be created in the operator's namespace and will
+	// serve the metrics at <service-ip>:9002/metrics.
 	//
 	// In 1.78.x and 1.80.x, this field also serves as the default value for
 	// .spec.statefulSet.pod.tailscaleContainer.debug.enable. From 1.82.0, both
@@ -171,6 +174,20 @@ type Metrics struct {
 	//
 	// Defaults to false.
 	Enable bool `json:"enable"`
+	// Enable to create a Prometheus ServiceMonitor for scraping the proxy's Tailscale metrics.
+	// The ServiceMonitor will select the metrics Service that gets created when metrics are enabled.
+	// The ingested metrics for each Service monitor will have labels to identify the proxy:
+	// ts_proxy_type: ingress_service|ingress_resource|connector|proxygroup
+	// ts_proxy_parent_name: name of the parent resource (i.e name of the Connector, Tailscale Ingress, Tailscale Service or ProxyGroup)
+	// ts_proxy_parent_namespace: namespace of the parent resource (if the parent resource is not cluster scoped)
+	// job: ts_<proxy type>_[<parent namespace>]_<parent_name>
+	// +optional
+	ServiceMonitor *ServiceMonitor `json:"serviceMonitor"`
+}
+
+type ServiceMonitor struct {
+	// If Enable is set to true, a Prometheus ServiceMonitor will be created. Enable can only be set to true if metrics are enabled.
+	Enable bool `json:"enable"`
 }
 
 type Container struct {

k8s-operator/apis/v1alpha1/zz_generated.deepcopy.go

@@ -319,6 +319,11 @@ func (in *Env) DeepCopy() *Env {
 // DeepCopyInto is an autogenerated deepcopy function, copying the receiver, writing into out. in must be non-nil.
 func (in *Metrics) DeepCopyInto(out *Metrics) {
 	*out = *in
+	if in.ServiceMonitor != nil {
+		in, out := &in.ServiceMonitor, &out.ServiceMonitor
+		*out = new(ServiceMonitor)
+		**out = **in
+	}
 }
 
 // DeepCopy is an autogenerated deepcopy function, copying the receiver, creating a new Metrics.
@@ -526,7 +531,7 @@ func (in *ProxyClassSpec) DeepCopyInto(out *ProxyClassSpec) {
 	if in.Metrics != nil {
 		in, out := &in.Metrics, &out.Metrics
 		*out = new(Metrics)
-		**out = **in
+		(*in).DeepCopyInto(*out)
 	}
 	if in.TailscaleConfig != nil {
 		in, out := &in.TailscaleConfig, &out.TailscaleConfig
@@ -991,6 +996,21 @@ func (in *S3Secret) DeepCopy() *S3Secret {
 	return out
 }
 
+// DeepCopyInto is an autogenerated deepcopy function, copying the receiver, writing into out. in must be non-nil.
+func (in *ServiceMonitor) DeepCopyInto(out *ServiceMonitor) {
+	*out = *in
+}
+
+// DeepCopy is an autogenerated deepcopy function, copying the receiver, creating a new ServiceMonitor.
+func (in *ServiceMonitor) DeepCopy() *ServiceMonitor {
+	if in == nil {
+		return nil
+	}
+	out := new(ServiceMonitor)
+	in.DeepCopyInto(out)
+	return out
+}
+
 // DeepCopyInto is an autogenerated deepcopy function, copying the receiver, writing into out. in must be non-nil.
 func (in *StatefulSet) DeepCopyInto(out *StatefulSet) {
 	*out = *in