Add Stable-Service fix

Signed-off-by: Revital Sur <eres@il.ibm.com>

Author: revit13

test/coordinator/e2e/coordinator/coordinator_test.go

@@ -31,11 +31,7 @@ import (
 	testutils "github.com/llm-d/llm-d-router/test/utils"
 )
 
-// requestTimeout is intentionally long (applies to all specs, text-only and
-// multimodal) so the request completes even on the intermittent slow/black-holed
-// path, letting the end-of-spec coordinator log dump show whether and when the
-// request actually reached the coordinator.
-const requestTimeout = 600 * time.Second
+const requestTimeout = 60 * time.Second
 
 // testImageURL and testImageURL2 are publicly accessible images used to
 // exercise multimodal requests that trigger the encode stage.

test/coordinator/e2e/coordinator/e2e_suite_test.go

@@ -97,9 +97,10 @@ var (
 
 	readyTimeout = env.GetEnvDuration("READY_TIMEOUT", defaultReadyTimeout, ginkgo.GinkgoLogr)
 
-	portForwardSessions []*gexec.Session
-	rendererObjects     []string
-	createdNameSpace    bool
+	portForwardSessions  []*gexec.Session
+	rendererObjects      []string
+	stableServiceObjects []string
+	createdNameSpace     bool
 )
 
 func TestCoordinatorE2E(t *testing.T) {
@@ -131,6 +132,10 @@ var _ = ginkgo.BeforeSuite(func() {
 	}
 
 	rendererObjects = createRenderer()
+
+	// Coordinator and EPP Services are created once and kept stable across specs
+	// so Envoy's STRICT_DNS clusters never have to re-resolve a rotated ClusterIP.
+	stableServiceObjects = createStableServices()
 })
 
 var _ = ginkgo.ReportAfterSuite("cleanup", func(report ginkgo.Report) {
@@ -148,6 +153,9 @@ var _ = ginkgo.ReportAfterSuite("cleanup", func(report ginkgo.Report) {
 	if len(rendererObjects) > 0 {
 		testutils.DeleteObjects(testConfig, rendererObjects, nsName)
 	}
+	if len(stableServiceObjects) > 0 {
+		testutils.DeleteObjects(testConfig, stableServiceObjects, nsName)
+	}
 	for _, session := range portForwardSessions {
 		session.Terminate()
 	}

test/coordinator/e2e/coordinator/setup_test.go

@@ -96,7 +96,9 @@ func createEndPointPicker(phase, config string) []string {
 
 	objects := make([]string, 1, 8)
 	objects[0] = "ConfigMap/" + cmName
-	objects = append(objects, applyManifest(manifest, eppSubstitutions())...)
+	// The Service is created once by createStableServices; recreate only the
+	// rest (ServiceAccount, RoleBinding, Deployment) per spec.
+	objects = append(objects, applyManifest(manifest, eppSubstitutions(), "Service")...)
 	podsInDeploymentsReady(objects)
 	return objects
 }
@@ -177,7 +179,9 @@ func createCoordinator(config string) []string {
 	objects[0] = "ConfigMap/llm-d-coordinator-config"
 
 	docs := e2eutil.RunKustomize(coordinatorComponentDir)
-	docs = e2eutil.FilterKinds(docs, "ConfigMap")
+	// The Service is created once by createStableServices so its ClusterIP is
+	// stable across specs; recreate only the Deployment (and SA) per spec.
+	docs = e2eutil.FilterKinds(docs, "ConfigMap", "Service")
 	docs = e2eutil.SubstituteMany(docs, coordinatorSubstitutions())
 	docs = e2eutil.RemoveEmptyArgs(docs)
 	objects = append(objects, testutils.CreateObjsFromYaml(testConfig, docs, nsName)...)
@@ -188,8 +192,11 @@ func createCoordinator(config string) []string {
 }
 
 // waitForCoordinatorReady polls /readyz through Envoy until it returns 200,
-// catching Envoy's STRICT_DNS resolution lagging behind the per-test Service
-// (podsInDeploymentsReady already confirms the coordinator pod itself is ready).
+// confirming the freshly recreated coordinator pod is reachable through the
+// gateway before the test sends its request. The gateway Service is stable
+// across specs (see createStableServices), so this waits for the new pod to
+// appear behind it, not for Envoy to re-resolve a rotated ClusterIP.
+// (podsInDeploymentsReady already confirms the coordinator pod itself is ready.)
 func waitForCoordinatorReady() {
 	ginkgo.By("Waiting for coordinator to be reachable via gateway")
 	gomega.Eventually(func() bool {
@@ -222,13 +229,39 @@ func createEPPConfigMap(name, content string) {
 	}
 }
 
-func applyManifest(path string, subs map[string]string) []string {
+func applyManifest(path string, subs map[string]string, excludeKinds ...string) []string {
 	docs := testutils.ReadYaml(path)
 	docs = e2eutil.SubstituteMany(docs, subs)
 	docs = e2eutil.RemoveEmptyArgs(docs)
+	if len(excludeKinds) > 0 {
+		docs = e2eutil.FilterKinds(docs, excludeKinds...)
+	}
 	return testutils.CreateObjsFromYaml(testConfig, docs, getNamespace())
 }
 
+// createStableServices creates the coordinator and per-phase EPP Services once,
+// up front, and returns their ids for suite teardown. Envoy fronts these via
+// STRICT_DNS clusters and outlives the per-spec workload; recreating the
+// Services each spec would rotate their ClusterIPs and force Envoy to
+// re-resolve, so only the Deployments behind them churn per spec. Mirrors the
+// non-coordinator e2e, which creates the EPP Services once in its per-container
+// setup and recreates only the Deployment per test.
+func createStableServices() []string {
+	var objects []string
+
+	docs := e2eutil.RunKustomize(coordinatorComponentDir)
+	docs = e2eutil.FilterKinds(docs, "ConfigMap", "Deployment", "ServiceAccount")
+	docs = e2eutil.SubstituteMany(docs, coordinatorSubstitutions())
+	docs = e2eutil.RemoveEmptyArgs(docs)
+	objects = append(objects, testutils.CreateObjsFromYaml(testConfig, docs, getNamespace())...)
+
+	for _, manifest := range []string{encodeEPPManifest, prefillEPPManifest, decodeEPPManifest} {
+		objects = append(objects,
+			applyManifest(manifest, eppSubstitutions(), "ServiceAccount", "RoleBinding", "Deployment")...)
+	}
+	return objects
+}
+
 func eppSubstitutions() map[string]string {
 	return map[string]string{
 		"${EPP_NAME}":              eppName,