From 719ca8d5cb00ee86df101dcade4f93f5278b9f3c Mon Sep 17 00:00:00 2001
From: itsnothuy <huytrngqu@gmail.com>
Date: Tue, 24 Feb 2026 11:51:59 +0700
Subject: [PATCH 1/2] Check runtime backend readiness in IsReady() for strict
 mode
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Under strict_readiness_=true, IsReady() only checked lifecycle state
(ModelStates()) but not runtime backend readiness. This meant that
/v2/health/ready returned 200 even when a model's backend had crashed
(e.g., Python backend stub process died), because the lifecycle still
showed READY.

Add a runtime readiness check via ModelIsReady() for models in READY
lifecycle state. ModelIsReady() calls model->IsReady() which invokes
TRITONBACKEND_ModelInstanceReady — in the Python backend this calls
StubActive() -> waitpid(WNOHANG), a non-blocking microsecond-level
check.

This ensures /v2/health/ready correctly reflects backend health,
enabling orchestrators (K8s, etc.) to detect and restart unhealthy
pods.

Fixes: triton-inference-server/server#8604
---
 src/server.cc | 20 ++++++++++++++++++++
 1 file changed, 20 insertions(+)

diff --git a/src/server.cc b/src/server.cc
index b1cc4eafe..d8f634b62 100644
--- a/src/server.cc
+++ b/src/server.cc
@@ -445,6 +445,26 @@ InferenceServer::IsReady(bool* ready)
           *ready = false;
           goto strict_done;
         }
+
+        // For models in READY lifecycle state, also check runtime
+        // backend instance readiness. This catches cases where the
+        // model was loaded successfully (lifecycle = READY) but the
+        // backend has become unhealthy at runtime (e.g., Python backend
+        // stub process died).
+        if (vs.second.first == ModelReadyState::READY) {
+          bool model_ready = false;
+          Status status =
+              ModelIsReady(mv.first.name_, vs.first, &model_ready);
+          if (!status.IsOk() || !model_ready) {
+            LOG_VERBOSE(1) << "Model '" << mv.first.name_ << "' version "
+                           << vs.first
+                           << " is in READY lifecycle state but failed "
+                              "runtime readiness check during server "
+                              "readiness evaluation";
+            *ready = false;
+            goto strict_done;
+          }
+        }
       }
     }
   strict_done:;

From 1fd66859e48e4c30de991d0d6853e01b4b1f01fd Mon Sep 17 00:00:00 2001
From: Huy Tran <158990013+itsnothuy@users.noreply.github.com>
Date: Wed, 25 Feb 2026 09:44:53 +0700
Subject: [PATCH 2/2] Update src/server.cc

Co-authored-by: J Wyman <jeremy.wyman@outlook.com>
---
 src/server.cc | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/src/server.cc b/src/server.cc
index d8f634b62..b9ed47212 100644
--- a/src/server.cc
+++ b/src/server.cc
@@ -453,8 +453,7 @@ InferenceServer::IsReady(bool* ready)
         // stub process died).
         if (vs.second.first == ModelReadyState::READY) {
           bool model_ready = false;
-          Status status =
-              ModelIsReady(mv.first.name_, vs.first, &model_ready);
+          Status status = ModelIsReady(mv.first.name_, vs.first, &model_ready);
           if (!status.IsOk() || !model_ready) {
             LOG_VERBOSE(1) << "Model '" << mv.first.name_ << "' version "
                            << vs.first