Use shutdown_timeout=1

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

Author: revit13

test/coordinator/e2e/coordinator/configs_test.go

@@ -23,7 +23,11 @@ server:
   listen_addr: ":8080"
   read_timeout: 30s
   write_timeout: 120s
-  shutdown_timeout: 25s
+  # Recreated per spec behind a suite-lived Envoy; drain fast so a deleted
+  # coordinator stops serving immediately instead of lingering on a stale
+  # endpoint the gateway may still route to. 0s is avoided: it makes the
+  # server Shutdown context expire instantly and the process exit non-zero.
+  shutdown_timeout: 1s
 
 gateway:
   address: "http://envoy.${NAMESPACE}.svc:8081"

test/coordinator/e2e/coordinator/coordinator_test.go

@@ -31,16 +31,7 @@ import (
 	testutils "github.com/llm-d/llm-d-router/test/utils"
 )
 
-const (
-	// requestTimeout bounds text-only chat completions (prefill, decode).
-	requestTimeout = 60 * time.Second
-	// multimodalRequestTimeout bounds multimodal chat completions, which add the
-	// replace-media-urls, render, and encode stages: each image is fetched (e.g.
-	// from S3) and run through the encoder, so the round-trip runs well past the
-	// text-only budget. 600s matches the default client request timeout vLLM
-	// applies to its multimodal and chat-completion serving tests.
-	multimodalRequestTimeout = 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.
@@ -177,11 +168,7 @@ func runCoordinatorPipeline(body []byte, expectedSteps []string, expectedImages
 	gomega.Expect(err).ShouldNot(gomega.HaveOccurred())
 	req.Header.Set("Content-Type", "application/json")
 
-	timeout := requestTimeout
-	if expectedImages > 0 {
-		timeout = multimodalRequestTimeout
-	}
-	client := &http.Client{Timeout: timeout}
+	client := &http.Client{Timeout: requestTimeout}
 	resp, err := client.Do(req)
 	gomega.Expect(err).ShouldNot(gomega.HaveOccurred())
 	defer resp.Body.Close()

test/coordinator/e2e/coordinator/setup_test.go

@@ -187,14 +187,23 @@ func createCoordinator(config string) []string {
 	return objects
 }
 
-// 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).
+// waitForCoordinatorReady polls /readyz through Envoy and requires several
+// consecutive successes: right after the per-spec coordinator is recreated, a
+// single probe can succeed while Envoy is still converging on the new endpoint,
+// so the next request would hit a stale one and hang.
 func waitForCoordinatorReady() {
 	ginkgo.By("Waiting for coordinator to be reachable via gateway")
+	const requiredConsecutive = 3
+	consecutive := 0
 	gomega.Eventually(func() bool {
-		return pollReady(gatewayBaseURL() + "/readyz")
-	}, readyTimeout, defaultInterval).Should(gomega.BeTrue(), "coordinator should be reachable via gateway within the ready timeout")
+		if !pollReady(gatewayBaseURL() + "/readyz") {
+			consecutive = 0
+			return false
+		}
+		consecutive++
+		return consecutive >= requiredConsecutive
+	}, readyTimeout, defaultInterval).Should(gomega.BeTrue(),
+		"coordinator should be reachable via gateway within the ready timeout")
 }
 
 // pollReady reports whether a GET on url returns HTTP 200 within the client timeout.