ate-setup no longer requires the GKE PodMonitoring CRD

cmd/ate-setup/internal/kube/apply_test.go

@@ -27,6 +27,10 @@ import (
 	"k8s.io/apimachinery/pkg/apis/meta/v1/unstructured"
 	"k8s.io/apimachinery/pkg/runtime/schema"
 	"k8s.io/apimachinery/pkg/util/wait"
+	"k8s.io/client-go/discovery/cached/memory"
+	fakediscovery "k8s.io/client-go/discovery/fake"
+	"k8s.io/client-go/restmapper"
+	clienttesting "k8s.io/client-go/testing"
 )
 
 // Teardown walks a fixed list of manifests covering every install shape, so a
@@ -158,3 +162,33 @@ func TestApplyWithRetryStopsOnCancellation(t *testing.T) {
 		t.Errorf("apply attempted %d times, want 1", ri.attempts)
 	}
 }
+
+// ApplyTolerant skips a kind whose CRD the cluster does not serve and never
+// reaches the dynamic client for it: the zero-value Dynamic panics if it does.
+func TestApplyTolerantSkipsUnservedKind(t *testing.T) {
+	// A cluster serves at least the core group; only the Managed Prometheus
+	// group is missing.
+	served := &fakediscovery.FakeDiscovery{Fake: &clienttesting.Fake{Resources: []*metav1.APIResourceList{{
+		GroupVersion: "v1",
+		APIResources: []metav1.APIResource{{Name: "configmaps", Kind: "ConfigMap", Namespaced: true}},
+	}}}}
+	disco := memory.NewMemCacheClient(served)
+	c := &Client{discovery: disco, mapper: restmapper.NewDeferredDiscoveryRESTMapper(disco)}
+
+	pm := &unstructured.Unstructured{}
+	pm.SetAPIVersion("monitoring.googleapis.com/v1")
+	pm.SetKind("PodMonitoring")
+	pm.SetNamespace("ate-system")
+	pm.SetName("atenet-router-envoy")
+
+	var skipped []string
+	err := c.ApplyTolerant(t.Context(), []*unstructured.Unstructured{pm}, func(obj *unstructured.Unstructured, _ error) {
+		skipped = append(skipped, Describe(obj))
+	})
+	if err != nil {
+		t.Fatalf("ApplyTolerant() = %v, want the unserved kind skipped", err)
+	}
+	if len(skipped) != 1 {
+		t.Errorf("skipped %v, want the PodMonitoring reported once", skipped)
+	}
+}

cmd/ate-setup/internal/steps/deploy.go

@@ -164,6 +164,9 @@ func (e *Env) DeployAteSystem(ctx context.Context, opts DeployOptions) error {
 	if err := e.applyWithAPIServerEnvHash(ctx, manifests); err != nil {
 		return err
 	}
+	if err := e.applyRouterMonitoring(ctx); err != nil {
+		return err
+	}
 
 	// After the bundle, which resets the pod template to the sidecar-free base.
 	if err := e.reconcileCloudSQLProxySidecar(ctx); err != nil {
@@ -218,6 +221,27 @@ func (e *Env) DeployAteSystem(ctx context.Context, opts DeployOptions) error {
 	return e.applyOtelEndpointOverride(ctx)
 }
 
+// applyRouterMonitoring applies the Managed Prometheus scrape of the router's
+// Envoy. Its PodMonitoring kind exists only where Google Managed Prometheus
+// does, so the object is skipped on a cluster that does not serve it, not
+// allowed to fail the install. Kind scrapes through its own collector.
+//
+// This is a temporary, GKE-specific skip, not an extension point. Do not extend
+// it to other in-tree provider CRDs: provider-specific CRDs belong out of tree,
+// and this one should move there longer term.
+func (e *Env) applyRouterMonitoring(ctx context.Context) error {
+	if e.Cfg.Kind {
+		return nil
+	}
+	objs, err := kube.LoadPath(e.Cfg.Manifest("atenet-router-monitoring.yaml"))
+	if err != nil {
+		return err
+	}
+	return e.Kube.ApplyTolerant(ctx, objs, func(obj *unstructured.Unstructured, _ error) {
+		log.Warnf("skipping %s: its CRD is not installed on this cluster", kube.Describe(obj))
+	})
+}
+
 // DeployPodCertificateController creates any missing signer CA pools, applies
 // the podcertificate controller, and waits until it is rolling and has
 // published its trust bundles.

cmd/ate-setup/internal/steps/deploy_test.go

@@ -117,6 +117,61 @@ func TestSetPodcertWorkersPerSignerRejectsMissingDeployment(t *testing.T) {
 	}
 }
 
+// A stock cluster has no Managed Prometheus CRD, so no system bundle may carry
+// one: applying it would fail the whole install.
+func TestSystemBundlesNeedNoManagedPrometheusCRD(t *testing.T) {
+	for _, tc := range []struct {
+		name string
+		cfg  config.Config
+	}{
+		{"base", config.Config{Router: config.RouterEnvoy}},
+		{"agentgateway", config.Config{Router: config.RouterAgentgateway}},
+		{"kind", config.Config{Router: config.RouterEnvoy, Kind: true}},
+		{"kind agentgateway", config.Config{Router: config.RouterAgentgateway, Kind: true}},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			cfg := tc.cfg
+			cfg.Root = repoRoot(t)
+			e := &Env{Cfg: &cfg}
+			manifest, err := e.render(cfg.Path(SystemOverlay(&cfg)))
+			if err != nil {
+				t.Fatalf("render: %v", err)
+			}
+			objs, err := kube.DecodeManifestBytes(manifest)
+			if err != nil {
+				t.Fatalf("decoding the bundle: %v", err)
+			}
+			for _, obj := range objs {
+				if obj.GroupVersionKind().Group == "monitoring.googleapis.com" {
+					t.Errorf("bundle carries %s, which needs a Managed Prometheus CRD", kube.Describe(obj))
+				}
+			}
+		})
+	}
+}
+
+// Kind has no Managed Prometheus, so it never reaches the cluster for the
+// router monitoring. The nil Kube client panics if it does.
+func TestApplyRouterMonitoringSkipsKind(t *testing.T) {
+	e := &Env{Cfg: &config.Config{Root: repoRoot(t), Kind: true}}
+	if err := e.applyRouterMonitoring(t.Context()); err != nil {
+		t.Errorf("applyRouterMonitoring() = %v, want nil", err)
+	}
+}
+
+// The monitoring leaves the bundles, so the file the installer applies on its
+// own has to be the router's PodMonitoring.
+func TestRouterMonitoringManifest(t *testing.T) {
+	e := &Env{Cfg: &config.Config{Root: repoRoot(t)}}
+	objs, err := kube.LoadPath(e.Cfg.Manifest("atenet-router-monitoring.yaml"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(objs) != 1 || objs[0].GetKind() != "PodMonitoring" || objs[0].GetNamespace() != "ate-system" {
+		t.Errorf("manifest holds %v, want one PodMonitoring in ate-system", objs)
+	}
+}
+
 // Upgrading from a release that predates a signer pool: deploying the
 // controller alone has to create the missing pool before the new pod needs it.
 // The manifest root is empty, so the deploy stops at rendering, after the pools.

manifests/ate-install/base/kustomization.yaml

@@ -26,12 +26,15 @@ kind: Kustomization
 # both the overlay's flags and the env override. The same goes for the
 # sandboxconfig and atenet-egress files: cmd/ate-setup applies the variant it
 # selects, so they are not listed here either.
+#
+# atenet-router-monitoring.yaml is absent too: its PodMonitoring kind exists
+# only where Google Managed Prometheus does, so cmd/ate-setup applies it
+# separately and skips it on a cluster without that CRD.
 resources:
   - ../ate-api-server.yaml
   - ../ate-idp-server.yaml
   - ../ate-controller.yaml
   - ../atelet.yaml
   - ../atenet-router.yaml
-  - ../atenet-router-monitoring.yaml
   - ../ate-otel-config.yaml
   - ../ate-system-networkpolicies.yaml