From 7de0bdde53953f419933e2d20a3e3b4da6570baa Mon Sep 17 00:00:00 2001
From: Chuang Wang <chuangw@google.com>
Date: Mon, 27 Jul 2026 00:54:50 -0700
Subject: [PATCH] Add a readyz probe to the benchmark glutton actor

ResumeActor returned as soon as the resume workflow finished, not when
glutton's port 80 listener was accepting connections. A ping issued right
after resume could land in that gap, get ECONNREFUSED, and come back from
the router as a 503.

ateom already has the gate: internal/readyz.WaitAll blocks RestoreWorkload
and RunWorkload until every container that declares a readyz probe returns
200. Containers without a probe are skipped entirely, so glutton was never
waited on. This adds /readyz to the HTTP mux -- on the same listener as
/ping, since the metrics server starts independently and would answer 200
before /ping is reachable -- and points the ActorTemplate at it.

Note that ActorTemplate.spec is immutable, so picking this up needs a
delete and recreate of the template, not an apply.
---
 benchmarking/workloads/manifests/workloads.yaml.tmpl |  9 +++++++++
 cmd/benchmarking/glutton/main.go                     | 12 ++++++++++++
 2 files changed, 21 insertions(+)

diff --git a/benchmarking/workloads/manifests/workloads.yaml.tmpl b/benchmarking/workloads/manifests/workloads.yaml.tmpl
index 5fa120487..84f2c062f 100644
--- a/benchmarking/workloads/manifests/workloads.yaml.tmpl
+++ b/benchmarking/workloads/manifests/workloads.yaml.tmpl
@@ -70,6 +70,15 @@ spec:
     - "--metrics-listen-addr=:9090"
     - "--data-dir=/tmp/glutton"
     - "--mode=http"
+    # Hold ResumeActor open until port 80 actually answers. Starting the
+    # process is not the same as it accepting connections, and without this
+    # gate a ping can arrive in the gap and be refused. Probing port 80 rather
+    # than the metrics port is deliberate: it is the port the router forwards
+    # to.
+    readyz:
+      httpGet:
+        path: /readyz
+        port: 80
   workerSelector:
     matchLabels:
       workload: benchmark-ateom
diff --git a/cmd/benchmarking/glutton/main.go b/cmd/benchmarking/glutton/main.go
index 90079bb81..085e92c0e 100644
--- a/cmd/benchmarking/glutton/main.go
+++ b/cmd/benchmarking/glutton/main.go
@@ -136,6 +136,18 @@ func main() {
 		// (re-exposable as additional routes if/when needed).
 		mux := http.NewServeMux()
 		mux.HandleFunc("/ping", httpPingHandler(svc))
+		// Readiness gate for the actor lifecycle. ateom blocks RestoreWorkload
+		// on this returning 200 (internal/readyz), so ResumeActor does not
+		// report success until the socket the router forwards to is genuinely
+		// serving. Without it a ping can beat the listener and get
+		// ECONNREFUSED, which the router turns into a 503.
+		//
+		// It must live on this listener rather than the metrics port: the
+		// metrics server starts independently, so a 200 there would say
+		// nothing about whether /ping is reachable yet.
+		mux.HandleFunc("/readyz", func(w http.ResponseWriter, r *http.Request) {
+			w.WriteHeader(http.StatusOK)
+		})
 		// otelhttp at the mux level + per-handler span follows
 		// docs/dev/best-practices/tracing.md: extract incoming context,
 		// then name the span after the operation in each handler.