[core] Fix drain state propagation race condition

src/ray/gcs/gcs_node_manager.cc

@@ -604,6 +604,10 @@ void GcsNodeManager::SetNodeDraining(
         << drain_request->DebugString();
     iter->second = drain_request;
   }
+
+  for (auto &listener : node_draining_listeners_) {
+    listener(node_id, true, drain_request->deadline_timestamp_ms());
+  }
 }
 
 std::shared_ptr<const rpc::GcsNodeInfo> GcsNodeManager::RemoveNode(

src/ray/gcs/gcs_node_manager.h

@@ -195,6 +195,17 @@ class GcsNodeManager : public rpc::NodeInfoGcsServiceHandler {
     node_added_listeners_.emplace_back(std::move(listener), io_context);
   }
 
+  /// Add listener to monitor when a node is set to draining.
+  /// The listener receives (node_id, is_draining, draining_deadline_timestamp_ms).
+  ///
+  /// \param listener The handler which processes the draining state change.
+  void AddNodeDrainingListener(
+      std::function<void(const NodeID &, bool, int64_t)> listener) {
+    absl::MutexLock lock(&mutex_);
+    RAY_CHECK(listener);
+    node_draining_listeners_.emplace_back(std::move(listener));
+  }
+
   /// Initialize with the gcs tables data synchronously.
   /// This should be called when GCS server restarts after a failure.
   ///
@@ -368,6 +379,12 @@ class GcsNodeManager : public rpc::NodeInfoGcsServiceHandler {
   std::vector<Postable<void(std::shared_ptr<const rpc::GcsNodeInfo>)>>
       node_removed_listeners_ ABSL_GUARDED_BY(mutex_);
 
+  /// Listeners which monitors when nodes are set to draining.
+  /// Uses std::function (not Postable) for synchronous invocation to ensure
+  /// scheduler sees draining state before HandleDrainNode returns.
+  std::vector<std::function<void(const NodeID &, bool, int64_t)>> node_draining_listeners_
+      ABSL_GUARDED_BY(mutex_);
+
   /// A publisher for publishing gcs messages.
   pubsub::GcsPublisher *gcs_publisher_;
   /// Storage for GCS tables.

src/ray/gcs/gcs_resource_manager.cc

@@ -295,6 +295,13 @@ void GcsResourceManager::OnNodeDead(const NodeID &node_id) {
   num_alive_nodes_--;
 }
 
+void GcsResourceManager::SetNodeDraining(const NodeID &node_id,
+                                         bool is_draining,
+                                         int64_t draining_deadline_timestamp_ms) {
+  cluster_resource_manager_.SetNodeDraining(
+      scheduling::NodeID(node_id.Binary()), is_draining, draining_deadline_timestamp_ms);
+}
+
 void GcsResourceManager::UpdatePlacementGroupLoad(
     const std::shared_ptr<rpc::PlacementGroupLoad> placement_group_load) {
   RAY_CHECK(placement_group_load != nullptr);

src/ray/gcs/gcs_resource_manager.h

@@ -104,6 +104,15 @@ class GcsResourceManager : public rpc::NodeResourceInfoGcsServiceHandler,
   /// \param node_id The specified node id.
   void OnNodeDead(const NodeID &node_id);
 
+  /// Set the draining state of a node.
+  ///
+  /// \param node_id The ID of the node.
+  /// \param is_draining Whether the node is draining.
+  /// \param draining_deadline_timestamp_ms The deadline for the drain operation.
+  void SetNodeDraining(const NodeID &node_id,
+                       bool is_draining,
+                       int64_t draining_deadline_timestamp_ms);
+
   /// Initialize with the gcs tables data synchronously.
   /// This should be called when GCS server restarts after a failure.
   ///

src/ray/gcs/gcs_server.cc

@@ -886,6 +886,11 @@ void GcsServer::InstallEventListeners() {
         gcs_autoscaler_state_manager_->OnNodeDead(node_id);
       },
       io_context_provider_.GetDefaultIOContext());
+  gcs_node_manager_->AddNodeDrainingListener(
+      [this](const NodeID &node_id, bool is_draining, int64_t deadline_timestamp_ms) {
+        gcs_resource_manager_->SetNodeDraining(
+            node_id, is_draining, deadline_timestamp_ms);
+      });
 
   // Install worker event listener.
   gcs_worker_manager_->AddWorkerDeadListener(

src/ray/gcs/tests/gcs_resource_manager_test.cc

@@ -227,4 +227,43 @@ TEST_F(GcsResourceManagerTest, TestGetDrainingNodes) {
   ASSERT_EQ(reply.draining_nodes(0).node_id(), node1->node_id());
 }
 
+// Verify SetNodeDraining() immediately updates ClusterResourceManager.
+TEST_F(GcsResourceManagerTest, DrainStateImmediatelyVisibleToScheduler) {
+  auto node = GenNodeInfo();
+  node->mutable_resources_total()->insert({"CPU", 10});
+  gcs_resource_manager_->OnNodeAdd(*node);
+
+  auto node_id = NodeID::FromBinary(node->node_id());
+  auto scheduling_node_id = scheduling::NodeID(node_id.Binary());
+
+  UpdateFromResourceViewSync(node_id,
+                             {{"CPU", 10}},
+                             {{"CPU", 10}},
+                             /*idle_ms=*/0,
+                             /*is_draining=*/false);
+
+  ASSERT_FALSE(cluster_resource_manager_.IsNodeDraining(scheduling_node_id));
+
+  gcs_node_manager_->AddNodeDrainingListener(
+      [this](const NodeID &node_id, bool is_draining, int64_t deadline_timestamp_ms) {
+        cluster_resource_manager_.SetNodeDraining(
+            scheduling::NodeID(node_id.Binary()), is_draining, deadline_timestamp_ms);
+      });
+
+  auto drain_request = std::make_shared<rpc::autoscaler::DrainNodeRequest>();
+  drain_request->set_node_id(node_id.Binary());
+  drain_request->set_reason(rpc::autoscaler::DRAIN_NODE_REASON_IDLE_TERMINATION);
+  drain_request->set_reason_message("idle termination");
+
+  gcs_node_manager_->AddNode(std::make_shared<rpc::GcsNodeInfo>(*node));
+  drain_request->set_deadline_timestamp_ms(12345);
+  gcs_node_manager_->SetNodeDraining(node_id, drain_request);
+
+  ASSERT_TRUE(cluster_resource_manager_.IsNodeDraining(scheduling_node_id));
+  const auto &node_resources =
+      cluster_resource_manager_.GetNodeResources(scheduling_node_id);
+  ASSERT_EQ(node_resources.draining_deadline_timestamp_ms,
+            drain_request->deadline_timestamp_ms());
+}
+
 }  // namespace ray

src/ray/raylet/scheduling/cluster_resource_manager.cc

@@ -101,9 +101,11 @@ bool ClusterResourceManager::UpdateNode(
   // Last update time to the local node resources view.
   local_view.last_resource_update_time = absl::Now();
 
-  local_view.is_draining = resource_view_sync_message.is_draining();
-  local_view.draining_deadline_timestamp_ms =
-      resource_view_sync_message.draining_deadline_timestamp_ms();
+  if (!local_view.is_draining) {
+    local_view.is_draining = resource_view_sync_message.is_draining();
+    local_view.draining_deadline_timestamp_ms =
+        resource_view_sync_message.draining_deadline_timestamp_ms();
+  }
 
   AddOrUpdateNode(node_id, local_view);
   received_node_resources_[node_id] = std::move(local_view);
@@ -115,6 +117,26 @@ bool ClusterResourceManager::RemoveNode(scheduling::NodeID node_id) {
   return nodes_.erase(node_id) != 0;
 }
 
+bool ClusterResourceManager::SetNodeDraining(const scheduling::NodeID &node_id,
+                                             bool is_draining,
+                                             int64_t draining_deadline_timestamp_ms) {
+  auto it = nodes_.find(node_id);
+  if (it == nodes_.end()) {
+    return false;
+  }
+
+  auto *local_view = it->second.GetMutableLocalView();
+  local_view->is_draining = is_draining;
+  local_view->draining_deadline_timestamp_ms = draining_deadline_timestamp_ms;
+  auto rnr_it = received_node_resources_.find(node_id);
+  if (rnr_it != received_node_resources_.end()) {
+    rnr_it->second.is_draining = is_draining;
+    rnr_it->second.draining_deadline_timestamp_ms = draining_deadline_timestamp_ms;
+  }
+
+  return true;
+}
+
 bool ClusterResourceManager::GetNodeResources(scheduling::NodeID node_id,
                                               NodeResources *ret_resources) const {
   auto it = nodes_.find(node_id);

src/ray/raylet/scheduling/cluster_resource_manager.h

@@ -132,6 +132,18 @@ class ClusterResourceManager {
     return node.GetLocalView().is_draining;
   }
 
+  /// Set the draining state of a node.
+  /// This is called when the autoscaler commits to draining a node, ensuring
+  /// the scheduler immediately sees the node as unavailable for scheduling.
+  ///
+  /// \param node_id The ID of the node.
+  /// \param is_draining Whether the node is draining.
+  /// \param draining_deadline_timestamp_ms The deadline for the drain operation.
+  /// \return true if the node exists and was updated, false otherwise.
+  bool SetNodeDraining(const scheduling::NodeID &node_id,
+                       bool is_draining,
+                       int64_t draining_deadline_timestamp_ms);
+
   /// @param max_num_nodes_to_include Max number of nodes to include in the debug string.
   ///   If not specified, all nodes will be included.
   std::string DebugString(