fix(BREAKING): keep EPP names and labels within 63 characters

charts/modeldeployment/README.md

@@ -124,8 +124,9 @@ letting KEDA fail later when it creates the HPA:
 
 ## Naming conventions
 
-- InferencePool: `<name>-inferencepool`
-- EPP Deployment / Service / SA / Role / ConfigMap: `<name>-inferencepool-epp`
+- InferencePool, EPP Service and the EPP pods' `inferencepool` label: `<name>`
+  (these must fit 63 characters, so `name` is limited to 63 characters)
+- EPP Deployment / SA / Role / RoleBinding / ConfigMap: `<name>-inferencepool-epp`
 - HTTPRoute: `<name>-route`
 
 ## Example

charts/modeldeployment/templates/_helpers.tpl

@@ -24,17 +24,27 @@ Gateway that modelharness provisions in the workload namespace.
 {{- end }}
 
 {{/*
-Derived InferencePool name, matching KAITO's naming convention.
+InferencePool name. Equals the deployment name: Istio stamps it as the
+istio.io/inferencepool-name label value on the InferencePool's shadow
+Service, so it must fit the 63-character label value limit.
 */}}
 {{- define "modeldeployment.inferencePoolName" -}}
-{{ include "modeldeployment.name" . }}-inferencepool
+{{- include "modeldeployment.name" . -}}
 {{- end }}
 
 {{/*
-Derived EPP service name (used as the DestinationRule host).
+EPP Service name, also used as the `inferencepool` selector label value.
+Equals the deployment name so it fits the 63-character limit of both.
 */}}
 {{- define "modeldeployment.eppServiceName" -}}
-{{ include "modeldeployment.inferencePoolName" . }}-epp
+{{- include "modeldeployment.name" . -}}
+{{- end }}
+
+{{/*
+Name of the EPP Deployment / ServiceAccount / Role / RoleBinding / ConfigMap.
+*/}}
+{{- define "modeldeployment.eppName" -}}
+{{ include "modeldeployment.name" . }}-inferencepool-epp
 {{- end }}
 
 {{/*

charts/modeldeployment/templates/epp-configmap.yaml

@@ -1,7 +1,7 @@
 apiVersion: v1
 kind: ConfigMap
 metadata:
-  name: {{ include "modeldeployment.eppServiceName" . }}
+  name: {{ include "modeldeployment.eppName" . }}
   namespace: {{ include "modeldeployment.namespace" . }}
   labels:
     {{- include "modeldeployment.labels" . | nindent 4 }}

charts/modeldeployment/templates/epp-deployment.yaml

@@ -1,7 +1,7 @@
 apiVersion: apps/v1
 kind: Deployment
 metadata:
-  name: {{ include "modeldeployment.eppServiceName" . }}
+  name: {{ include "modeldeployment.eppName" . }}
   namespace: {{ include "modeldeployment.namespace" . }}
   labels:
     {{- include "modeldeployment.labels" . | nindent 4 }}
@@ -36,7 +36,7 @@ spec:
         checksum/epp-flow-control: {{ include "modeldeployment.eppFlowControlSettings" . | sha256sum }}
       {{- end }}
     spec:
-      serviceAccountName: {{ include "modeldeployment.eppServiceName" . }}
+      serviceAccountName: {{ include "modeldeployment.eppName" . }}
       terminationGracePeriodSeconds: 130
       {{- with .Values.epp.tolerations }}
       tolerations:
@@ -94,4 +94,4 @@ spec:
       volumes:
         - name: plugins-config-volume
           configMap:
-            name: {{ include "modeldeployment.eppServiceName" . }}
+            name: {{ include "modeldeployment.eppName" . }}

charts/modeldeployment/templates/epp-rbac.yaml

@@ -1,15 +1,15 @@
 apiVersion: v1
 kind: ServiceAccount
 metadata:
-  name: {{ include "modeldeployment.eppServiceName" . }}
+  name: {{ include "modeldeployment.eppName" . }}
   namespace: {{ include "modeldeployment.namespace" . }}
   labels:
     {{- include "modeldeployment.labels" . | nindent 4 }}
 ---
 apiVersion: rbac.authorization.k8s.io/v1
 kind: Role
 metadata:
-  name: {{ include "modeldeployment.eppServiceName" . }}
+  name: {{ include "modeldeployment.eppName" . }}
   namespace: {{ include "modeldeployment.namespace" . }}
   labels:
     {{- include "modeldeployment.labels" . | nindent 4 }}
@@ -27,15 +27,15 @@ rules:
 apiVersion: rbac.authorization.k8s.io/v1
 kind: RoleBinding
 metadata:
-  name: {{ include "modeldeployment.eppServiceName" . }}
+  name: {{ include "modeldeployment.eppName" . }}
   namespace: {{ include "modeldeployment.namespace" . }}
   labels:
     {{- include "modeldeployment.labels" . | nindent 4 }}
 subjects:
   - kind: ServiceAccount
-    name: {{ include "modeldeployment.eppServiceName" . }}
+    name: {{ include "modeldeployment.eppName" . }}
     namespace: {{ include "modeldeployment.namespace" . }}
 roleRef:
   apiGroup: rbac.authorization.k8s.io
   kind: Role
-  name: {{ include "modeldeployment.eppServiceName" . }}
+  name: {{ include "modeldeployment.eppName" . }}

charts/modeldeployment/values.schema.json

@@ -64,7 +64,8 @@
   "properties": {
     "name": {
       "type": "string",
-      "description": "Deployment / InferenceSet name. Empty falls back to .Release.Name."
+      "maxLength": 63,
+      "description": "Deployment / InferenceSet name, also used as the InferencePool and EPP Service name and the EPP `inferencepool` selector label value, hence at most 63 characters. Empty falls back to .Release.Name."
     },
     "namespace": {
       "type": "string",

test/e2e/README.md

@@ -32,7 +32,7 @@ the production App Routing provisioning contract.
 - [`setup.go`](utils/setup.go) — `EnsureNamespace` (provisions the workload namespace + modelharness), `DeleteNamespace`, `SetupInferenceSetsWithRouting`, `TeardownInferenceSetsWithRouting`.
 - [`http.go`](utils/http.go) — authenticated `SendGatewayRequest` with `SendChat` / `SendModels` shortcuts against a `deploy.GatewayEndpoint`, taking `RequestOption`s (`WithPrompt`, `WithAuth`, `WithoutAuth`, `WithHeader`, `WithMethod`, `WithTransportRetry`, `WithOriginPath`, `WithTimeout`).
 - [`deployer.go`](utils/deployer.go) — `SetDeployer` / `CurrentDeployer` plus the `InstallModelDeployment`, `UpgradeModelDeployment`, `UninstallModelDeployment`, `InstallModelHarness`, `UninstallModelHarness`, `NamespaceAuthHeaders` helpers that delegate to the active backend (see [Deployment backends](#deployment-backends)).
-- [`inference.go`](utils/inference.go) — `WaitForInferenceSetReady`, `EPPServiceName`, snapshot/diff helpers.
+- [`inference.go`](utils/inference.go) — `WaitForInferenceSetReady`, `EPPDeploymentName`, `EPPSelector`, snapshot/diff helpers.
 - [`metrics.go`](utils/metrics.go), [`cluster.go`](utils/cluster.go), [`dynamic.go`](utils/dynamic.go), [`ginkgo.go`](utils/ginkgo.go).
 
 `cases.go`:

test/e2e/cases.go

@@ -225,9 +225,10 @@ const (
 // unique across the entire table so Helm releases never collide.
 //
 // Conventions:
-//   - Name: deployment / Helm release name. AND the value carried in the
-//     `model` field of OpenAI-style requests (matched by the HTTPRoute as
-//     X-Gateway-Model-Name).
+//   - Name: deployment name (the Helm backend uses it as the release name,
+//     shortened when it exceeds Helm's 53-character limit). AND the value
+//     carried in the `model` field of OpenAI-style requests (matched by the
+//     HTTPRoute as X-Gateway-Model-Name).
 //   - Model: inference preset name, written to the InferenceSet's
 //     spec.template.inference.preset.name.
 //   - Namespace: per-case namespace — the suite installs into this

test/e2e/control_plane_error_spec.go

@@ -60,7 +60,7 @@ var _ = Describe("Control-plane Error Reporter",
 		})
 
 		It("emits inferencesetEPPNotReady when the InferenceSet EPP is scaled to zero", func() {
-			eppName := utils.EPPServiceName(modelName)
+			eppName := utils.EPPDeploymentName(modelName)
 			desired, _, err := utils.GetDeploymentReplicas(ctx, caseNS, eppName)
 			Expect(err).NotTo(HaveOccurred())
 			baseline, err := utils.ReporterEventBaseline(ctx, "inferencesetEPPNotReady", caseNS)

test/e2e/deploy/helm/chart_test.go

@@ -17,15 +17,22 @@ limitations under the License.
 package helm
 
 import (
+	"encoding/json"
+	"maps"
 	"os/exec"
 	"path/filepath"
+	"regexp"
+	"slices"
 	"strings"
 	"testing"
 
 	metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
+	"k8s.io/apimachinery/pkg/util/validation"
 	"sigs.k8s.io/yaml"
 )
 
+var maxLengthErrorRe = regexp.MustCompile(`name: String length must be less than or equal to 63|'/name': maxLength: got \d+, want 63`)
+
 func TestModelDeploymentScalingAnnotations(t *testing.T) {
 	if _, err := exec.LookPath("helm"); err != nil {
 		t.Skip("helm is required for chart rendering tests")
@@ -188,6 +195,89 @@ func TestModelDeploymentScalingValidation(t *testing.T) {
 	}
 }
 
+// The deployment name feeds the EPP Service name, the `inferencepool`
+// selector label value and the InferencePool name (which Istio copies into a
+// label on the pool's shadow Service), all capped at 63 characters. A
+// 63-character name must render objects that satisfy those limits, and a
+// 64-character name must be rejected by the chart.
+func TestModelDeploymentNameLength(t *testing.T) {
+	if _, err := exec.LookPath("helm"); err != nil {
+		t.Skip("helm is required for chart rendering tests")
+	}
+	chart := filepath.Join("..", "..", "..", "..", "charts", "modeldeployment")
+	render := func(name string) ([]byte, error) {
+		return exec.Command("helm", "template", "model", chart,
+			"--set", "model=test-model", "--set", "namespace=ns", "--set", "name="+name).CombinedOutput()
+	}
+
+	// Helm's JSON schema validator words the error differently across releases:
+	// "name: String length must be less than or equal to 63" before v3.18,
+	// "at '/name': maxLength: got 64, want 63" from v3.18 on.
+	output, err := render(strings.Repeat("a", 64))
+	if err == nil || !maxLengthErrorRe.Match(output) {
+		t.Errorf("expected a 64-character name to be rejected, got %v\n%s", err, output)
+	}
+
+	output, err = render(strings.Repeat("a", 63))
+	if err != nil {
+		t.Fatalf("render with a 63-character name: %v\n%s", err, output)
+	}
+	var kinds []string
+	for _, doc := range strings.Split(string(output), "\n---") {
+		var obj struct {
+			Kind     string            `json:"kind"`
+			Metadata metav1.ObjectMeta `json:"metadata"`
+			Spec     struct {
+				Selector json.RawMessage `json:"selector"`
+				Template struct {
+					Metadata metav1.ObjectMeta `json:"metadata"`
+				} `json:"template"`
+			} `json:"spec"`
+		}
+		if err := yaml.Unmarshal([]byte(doc), &obj); err != nil {
+			t.Fatalf("parse rendered object: %v\n%s", err, doc)
+		}
+		if obj.Kind == "" {
+			continue
+		}
+		kinds = append(kinds, obj.Kind)
+		labels := map[string]string{}
+		maps.Copy(labels, obj.Metadata.Labels)
+		maps.Copy(labels, obj.Spec.Template.Metadata.Labels)
+		switch obj.Kind {
+		case "Service":
+			if errs := validation.IsDNS1035Label(obj.Metadata.Name); len(errs) > 0 {
+				t.Errorf("Service name %q: %v", obj.Metadata.Name, errs)
+			}
+			var selector map[string]string
+			if err := json.Unmarshal(obj.Spec.Selector, &selector); err != nil {
+				t.Fatalf("parse Service selector: %v", err)
+			}
+			maps.Copy(labels, selector)
+		case "Deployment":
+			var selector metav1.LabelSelector
+			if err := json.Unmarshal(obj.Spec.Selector, &selector); err != nil {
+				t.Fatalf("parse Deployment selector: %v", err)
+			}
+			maps.Copy(labels, selector.MatchLabels)
+		case "InferencePool":
+			if len(obj.Metadata.Name) > 63 {
+				t.Errorf("InferencePool name %q exceeds 63 characters", obj.Metadata.Name)
+			}
+		}
+		for key, value := range labels {
+			if len(value) > 63 {
+				t.Errorf("%s %s: label %s=%q exceeds 63 characters", obj.Kind, obj.Metadata.Name, key, value)
+			}
+		}
+	}
+	for _, kind := range []string{"Service", "Deployment", "InferencePool"} {
+		if !slices.Contains(kinds, kind) {
+			t.Errorf("rendered chart has no %s; rendered kinds: %v", kind, kinds)
+		}
+	}
+}
+
 func renderModelDeployment(t *testing.T, values ...string) ([]byte, error) {
 	t.Helper()
 	chart := filepath.Join("..", "..", "..", "..", "charts", "modeldeployment")

test/e2e/deploy/helm/helm.go

@@ -20,7 +20,9 @@ package helm
 
 import (
 	"context"
+	"crypto/sha256"
 	"encoding/base64"
+	"encoding/hex"
 	"encoding/json"
 	"fmt"
 	"os"
@@ -287,8 +289,9 @@ func isNotFound(out []byte) bool {
 
 // InstallModelDeployment runs `helm upgrade --install` for the modeldeployment
 // chart with the supplied values. It is idempotent — re-running it on an
-// existing release reconciles to the new values. The release name equals
-// values.Name and the release namespace equals values.Namespace.
+// existing release reconciles to the new values. The release name is
+// modelDeploymentReleaseName(values.Name) and the release namespace equals
+// values.Namespace.
 func (d *Deployer) InstallModelDeployment(ctx context.Context, values deploy.ModelDeploymentValues) error {
 	if err := values.Validate(); err != nil {
 		return err
@@ -304,7 +307,7 @@ func (d *Deployer) InstallModelDeployment(ctx context.Context, values deploy.Mod
 	}
 
 	args := []string{
-		"upgrade", "--install", values.Name, chart,
+		"upgrade", "--install", modelDeploymentReleaseName(values.Name), chart,
 		"--namespace", values.Namespace,
 		"--create-namespace",
 	}
@@ -335,7 +338,7 @@ func (d *Deployer) UpgradeModelDeployment(ctx context.Context, values deploy.Mod
 		return err
 	}
 
-	args := []string{"upgrade", values.Name, chart, "--namespace", values.Namespace}
+	args := []string{"upgrade", modelDeploymentReleaseName(values.Name), chart, "--namespace", values.Namespace}
 	args = append(args, setArgs(values)...)
 	args = append(args, gatewaySetArgs(values.Gateway)...)
 
@@ -357,20 +360,34 @@ func gatewaySetArgs(values deploy.GatewayValues) []string {
 	return args
 }
 
-// UninstallModelDeployment runs `helm uninstall` for the named release.
-// Missing releases are treated as success (so cleanup is idempotent).
+// UninstallModelDeployment runs `helm uninstall` for the named model
+// deployment's release. Missing releases are treated as success (so cleanup
+// is idempotent).
 func (d *Deployer) UninstallModelDeployment(ctx context.Context, name, namespace string) error {
 	if name == "" {
 		return fmt.Errorf("modeldeployment: name is required")
 	}
-	args := []string{"uninstall", name, "--namespace", namespace, "--ignore-not-found", "--wait"}
+	args := []string{"uninstall", modelDeploymentReleaseName(name), "--namespace", namespace, "--ignore-not-found", "--wait"}
 
 	if out, err := d.run(ctx, args...); err != nil {
 		return fmt.Errorf("helm uninstall %s failed: %w\n%s", name, err, string(out))
 	}
 	return nil
 }
 
+// modelDeploymentReleaseName returns the Helm release name for a model
+// deployment. It is the deployment name, except that names longer than Helm's
+// 53-character release name limit (deployment names may reach 63 characters)
+// are truncated and suffixed with a hash of the full name so they stay unique.
+func modelDeploymentReleaseName(name string) string {
+	const maxReleaseNameLength, hashLength = 53, 8
+	if len(name) <= maxReleaseNameLength {
+		return name
+	}
+	sum := sha256.Sum256([]byte(name))
+	return name[:maxReleaseNameLength-hashLength-1] + "-" + hex.EncodeToString(sum[:])[:hashLength]
+}
+
 // setArgs builds the `--set` arguments for the modeldeployment chart.
 func setArgs(v deploy.ModelDeploymentValues) []string {
 	args := []string{

test/e2e/deploy/helm/helm_test.go

@@ -742,6 +742,47 @@ func assertNamespaceLabelled(t *testing.T, calls [][]string, namespace string) {
 
 // Upgrade must not double as create: a mistyped release name has to fail
 // rather than silently provision a second deployment.
+// Deployment names may reach 63 characters, but Helm release names are capped
+// at 53. Install, upgrade and uninstall must all target the same shortened
+// release, while the chart still receives the full deployment name.
+func TestLongDeploymentNameUsesShortenedRelease(t *testing.T) {
+	d, calls := newTestDeployer(t, nil)
+
+	name := strings.Repeat("a", 63)
+	values := deploy.ModelDeploymentValues{
+		Name: name, Namespace: "e2e-ns", Model: "phi-4-mini-instruct", Replicas: 1,
+	}
+	if err := d.InstallModelDeployment(context.Background(), values); err != nil {
+		t.Fatalf("InstallModelDeployment: %v", err)
+	}
+	if err := d.UpgradeModelDeployment(context.Background(), values); err != nil {
+		t.Fatalf("UpgradeModelDeployment: %v", err)
+	}
+	if err := d.UninstallModelDeployment(context.Background(), name, "e2e-ns"); err != nil {
+		t.Fatalf("UninstallModelDeployment: %v", err)
+	}
+
+	release := modelDeploymentReleaseName(name)
+	if len(release) != 53 || !strings.HasPrefix(release, strings.Repeat("a", 44)+"-") {
+		t.Fatalf("release name %q is not a 53-character hash-suffixed prefix", release)
+	}
+	if got := modelDeploymentReleaseName(strings.Repeat("a", 62) + "b"); got == release {
+		t.Errorf("names differing past the truncation point share release %q", got)
+	}
+	if got := modelDeploymentReleaseName("phi"); got != "phi" {
+		t.Errorf("short name release = %q, want %q", got, "phi")
+	}
+
+	install, upgrade, uninstall := (*calls)[0], (*calls)[1], (*calls)[2]
+	if install[2] != release || upgrade[1] != release || uninstall[1] != release {
+		t.Errorf("release mismatch: install %q, upgrade %q, uninstall %q, want %q",
+			install[2], upgrade[1], uninstall[1], release)
+	}
+	if !hasArg(install, "name="+name) || !hasArg(upgrade, "name="+name) {
+		t.Errorf("chart must receive the full deployment name: install %v, upgrade %v", install, upgrade)
+	}
+}
+
 func TestUpgradeModelDeploymentDoesNotInstall(t *testing.T) {
 	d, calls := newTestDeployer(t, nil)