cpp_fix_v3.diff

diff --git a/python/ray/_private/node.py b/python/ray/_private/node.py
index afd481353e..ac5853a858 100644
--- a/python/ray/_private/node.py
+++ b/python/ray/_private/node.py
@@ -1852,6 +1852,30 @@ class Node:
                 wait=wait,
             )
 
+    def _raylet_graceful_shutdown_timeout(self, process):
+        # Written atomically by this raylet while it owns an Nsight worker/launcher.
+        # Use the raylet's effective config (which may come from another head node).
+        marker = os.path.join(
+            self.get_session_dir_path(), f"raylet-profiler-{process.pid}"
+        )
+        try:
+            with open(marker) as file:
+                return 5 + max(0, int(file.read())) / 1000
+        except (OSError, ValueError):
+            pass
+
+        # An explicit positive flush budget also opts into a longer graceful wait,
+        # e.g. when the marker cannot be published. Otherwise retain the 1s default.
+        flush_ms = max(
+            0, int(self._resolve_ray_config("worker_profiler_flush_timeout_ms", 0))
+        )
+        if flush_ms == 0:
+            return 1
+        worker_ms = max(
+            0, int(self._resolve_ray_config("kill_worker_timeout_milliseconds", 5000))
+        )
+        return 5 + (worker_ms + flush_ms) / 1000
+
     def _kill_process_impl(
         self, process_type, allow_graceful=False, check_alive=True, wait=False
     ):
@@ -1861,7 +1885,7 @@ class Node:
         process_infos = self.all_processes[process_type]
         if process_type != ray_constants.PROCESS_TYPE_REDIS_SERVER:
             assert len(process_infos) == 1
-        wait_timeout_seconds = 1
+        reap_timeout_seconds = 1
         for process_info in process_infos:
             process = process_info.process
             # Handle the case where the process has already exited.
@@ -1898,12 +1922,27 @@ class Node:
                 time.sleep(0.1)
 
             if allow_graceful:
+                graceful_start = time.monotonic()
+                graceful_timeout_seconds = (
+                    self._raylet_graceful_shutdown_timeout(process)
+                    if process_type == ray_constants.PROCESS_TYPE_RAYLET
+                    else 1
+                )
                 process.terminate()
-                # Allow the process one second to exit gracefully.
                 try:
-                    process.wait(timeout=wait_timeout_seconds)
+                    process.wait(timeout=graceful_timeout_seconds)
                 except subprocess.TimeoutExpired:
-                    pass
+                    # A worker can register between the initial marker read and
+                    # SIGTERM. Recheck once, without restarting the graceful budget.
+                    if process_type == ray_constants.PROCESS_TYPE_RAYLET:
+                        remaining = self._raylet_graceful_shutdown_timeout(process) - (
+                            time.monotonic() - graceful_start
+                        )
+                        if remaining > 0:
+                            try:
+                                process.wait(timeout=remaining)
+                            except subprocess.TimeoutExpired:
+                                pass
 
             # If the process did not exit, force kill it.
             if process.poll() is None:
@@ -1924,7 +1963,7 @@ class Node:
                 timeout = (
                     KILLED_PROCESS_REAP_TIMEOUT_SECONDS
                     if wait
-                    else wait_timeout_seconds
+                    else reap_timeout_seconds
                 )
                 try:
                     process.wait(timeout=timeout)
diff --git a/src/ray/common/ray_config_def.h b/src/ray/common/ray_config_def.h
index ed640d1d4f..65b449fe4b 100644
--- a/src/ray/common/ray_config_def.h
+++ b/src/ray/common/ray_config_def.h
@@ -342,6 +342,11 @@ RAY_CONFIG(int64_t, raylet_client_connect_timeout_milliseconds, 1000)
 /// the worker SIGKILL.
 RAY_CONFIG(int64_t, kill_worker_timeout_milliseconds, 5000)
 
+/// Nsight launcher flush budget, starting when raylet observes its worker exit.
+/// Non-positive values disable only the post-worker-exit wait. Supported with
+/// process-group cleanup on Linux; force kill and registration failure are excluded.
+RAY_CONFIG(int64_t, worker_profiler_flush_timeout_ms, 10000)
+
 /// Timeout for graceful actor shutdown (e.g. when actor goes out of scope).
 /// If an actor does not gracefully shut down within this timeout, it will be force
 /// killed. Set to -1 for infinite timeout to prevent the actor from being force killed
diff --git a/src/ray/raylet/BUILD.bazel b/src/ray/raylet/BUILD.bazel
index b0cd4c825b..0dc8928ab7 100644
--- a/src/ray/raylet/BUILD.bazel
+++ b/src/ray/raylet/BUILD.bazel
@@ -132,6 +132,7 @@ ray_cc_library(
         "@com_google_absl//absl/random",
         "@com_google_absl//absl/random:bit_gen_ref",
         "@com_google_absl//absl/strings",
+        "@nlohmann_json",
     ],
 )
 
@@ -357,6 +358,16 @@ ray_cc_library(
     ],
 )
 
+ray_cc_library(
+    name = "shutdown",
+    hdrs = ["shutdown.h"],
+    visibility = [":__subpackages__"],
+    deps = [
+        "//src/ray/asio:instrumented_io_context",
+        "//src/ray/protobuf:gcs_cc_proto",
+    ],
+)
+
 ray_cc_binary(
     name = "raylet",
     srcs = ["main.cc"],
@@ -370,6 +381,7 @@ ray_cc_binary(
         ":local_object_manager_interface",
         ":metrics",
         ":raylet_lib",
+        ":shutdown",
         ":worker_pool",
         "//src/ray/asio:instrumented_io_context",
         "//src/ray/asio:periodical_runner",
@@ -394,6 +406,7 @@ ray_cc_binary(
         "//src/ray/util:cmd_line_utils",
         "//src/ray/util:event",
         "//src/ray/util:process",
+        "//src/ray/util:process_utils",
         "//src/ray/util:raii",
         "//src/ray/util:stream_redirection",
         "//src/ray/util:stream_redirection_options",
diff --git a/src/ray/raylet/main.cc b/src/ray/raylet/main.cc
index e500e097f5..7d34eff183 100644
--- a/src/ray/raylet/main.cc
+++ b/src/ray/raylet/main.cc
@@ -48,6 +48,7 @@
 #include "ray/raylet/local_object_manager_interface.h"
 #include "ray/raylet/metrics.h"
 #include "ray/raylet/node_manager.h"
+#include "ray/raylet/shutdown.h"
 #include "ray/raylet/worker_pool.h"
 #include "ray/raylet_ipc_client/client_connection.h"
 #include "ray/raylet_rpc_client/raylet_client.h"
@@ -58,6 +59,7 @@
 #include "ray/util/event.h"
 #include "ray/util/network_util.h"
 #include "ray/util/process.h"
+#include "ray/util/process_utils.h"
 #include "ray/util/raii.h"
 #include "ray/util/stream_redirection.h"
 #include "ray/util/stream_redirection_options.h"
@@ -461,9 +463,10 @@ int main(int argc, char *argv[]) {
   ray::NodeID raylet_node_id = ray::NodeID::FromHex(node_id);
 
   std::atomic_bool shutting_down = false;
-  // Shut down raylet gracefully, in a synchronous fashion.
-  // This can be run by the signal handler or on the main io service.
-  auto shutdown_raylet_gracefully =
+  // Agent monitors and signal handlers share this entry point. Post all teardown
+  // to the main executor, which must keep serving worker disconnect replies.
+  auto shutdown_raylet_gracefully = ray::raylet::MakeRayletShutdownCallback(
+      main_service,
       [raylet_node_id,
        &shutting_down,
        &node_manager,
@@ -511,9 +514,15 @@ int main(int argc, char *argv[]) {
           remove(raylet_socket_name.c_str());
         };
 
-        gcs_client->Nodes().UnregisterSelf(
-            raylet_node_id, node_death_info, std::move(unregister_done_callback));
-      };
+        node_manager->PrepareForShutdown(
+            [&gcs_client,
+             raylet_node_id,
+             node_death_info,
+             done = std::move(unregister_done_callback)]() mutable {
+              gcs_client->Nodes().UnregisterSelf(
+                  raylet_node_id, node_death_info, std::move(done));
+            });
+      });
 
   gcs_client->InternalKV().AsyncGetInternalConfig([&](::ray::Status status,
                                                       const std::optional<std::string>
@@ -753,7 +762,8 @@ int main(int argc, char *argv[]) {
         node_manager_config.ray_debugger_external,
         /*clock=*/clock,
         worker_pool_metrics,
-        std::move(add_process_to_workers_cgroup_hook));
+        std::move(add_process_to_workers_cgroup_hook),
+        session_dir + "/raylet-profiler-" + std::to_string(ray::GetPID()));
 
     client_call_manager = std::make_unique<ray::rpc::ClientCallManager>(
         main_service, /*record_stats=*/true, node_ip_address);
diff --git a/src/ray/raylet/node_manager.cc b/src/ray/raylet/node_manager.cc
index ff9e598c9c..4514162f4b 100644
--- a/src/ray/raylet/node_manager.cc
+++ b/src/ray/raylet/node_manager.cc
@@ -1630,7 +1630,9 @@ void NodeManager::DisconnectClient(const std::shared_ptr<ClientConnection> &clie
           << "using process groups for worker cleanup. "
           << "Subreaper is deprecated and will be removed in a future release.";
     }
-    if (pg_enabled) {
+    const bool profiler_cleanup =
+        worker_pool_.DeferProfilerCleanup(worker->WorkerId(), graceful);
+    if (pg_enabled && !profiler_cleanup) {
       auto saved = worker->GetSavedProcessGroupId();
       if (saved.has_value()) {
         const auto wid = worker->WorkerId();
@@ -3042,7 +3044,12 @@ void NodeManager::HandleGetAgentPIDs(rpc::GetAgentPIDsRequest request,
   send_reply_callback(Status::OK(), /* success */ nullptr, /* failure */ nullptr);
 }
 
+void NodeManager::PrepareForShutdown(std::function<void()> done) {
+  worker_pool_.PrepareProfilerShutdown(std::move(done));
+}
+
 void NodeManager::Stop() {
+  worker_pool_.DrainProfilerProcesses();
   store_client_->Disconnect();
 #if !defined(_WIN32)
   // Best-effort process-group cleanup for any remaining workers before shutdown.
diff --git a/src/ray/raylet/node_manager.h b/src/ray/raylet/node_manager.h
index dc42ad231a..ca6722355e 100644
--- a/src/ray/raylet/node_manager.h
+++ b/src/ray/raylet/node_manager.h
@@ -251,6 +251,9 @@ class NodeManager : public rpc::NodeManagerServiceHandler,
   /// or object ids can be freed up across the cluster.
   void SetShouldGlobalGC();
 
+  /// Complete profiled worker shutdown while the executor serves disconnect replies.
+  void PrepareForShutdown(std::function<void()> done);
+
   /// Stop this node manager.
   void Stop();
 
diff --git a/src/ray/raylet/shutdown.h b/src/ray/raylet/shutdown.h
new file mode 100644
index 0000000000..953d80a294
--- /dev/null
+++ b/src/ray/raylet/shutdown.h
@@ -0,0 +1,37 @@
+// Copyright 2017 The Ray Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//  http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <functional>
+#include <utility>
+
+#include "ray/asio/instrumented_io_context.h"
+#include "src/ray/protobuf/gcs.pb.h"
+
+namespace ray::raylet {
+
+/// Agent monitor threads may invoke shutdown. Copy the death info and always post
+/// the callback so teardown never runs inline on a monitor thread.
+inline std::function<void(const rpc::NodeDeathInfo &)> MakeRayletShutdownCallback(
+    instrumented_io_context &io_service,
+    std::function<void(const rpc::NodeDeathInfo &)> shutdown) {
+  return [&io_service, shutdown = std::move(shutdown)](
+             const rpc::NodeDeathInfo &death_info) {
+    io_service.post([shutdown, death_info] { shutdown(death_info); },
+                    "Raylet.Shutdown");
+  };
+}
+
+}  // namespace ray::raylet
diff --git a/src/ray/raylet/tests/BUILD.bazel b/src/ray/raylet/tests/BUILD.bazel
index 8640c518de..61b55b124b 100644
--- a/src/ray/raylet/tests/BUILD.bazel
+++ b/src/ray/raylet/tests/BUILD.bazel
@@ -39,6 +39,7 @@ ray_cc_test(
         "//:ray_mock",
         "//src/ray/asio:periodical_runner",
         "//src/ray/core_worker_rpc_client:fake_core_worker_client",
+        "//src/ray/raylet:shutdown",
         "//src/ray/raylet:worker",
         "//src/ray/raylet:worker_pool",
         "//src/ray/util:clock",
diff --git a/src/ray/raylet/tests/worker_pool_test.cc b/src/ray/raylet/tests/worker_pool_test.cc
index bb288cc657..c02eb025fb 100644
--- a/src/ray/raylet/tests/worker_pool_test.cc
+++ b/src/ray/raylet/tests/worker_pool_test.cc
@@ -18,11 +18,14 @@
 #include <gtest/gtest.h>
 
 #include <algorithm>
+#include <cerrno>
+#include <csignal>
 #include <iostream>
 #include <list>
 #include <memory>
 #include <random>
 #include <string>
+#include <thread>
 #include <unordered_map>
 #include <unordered_set>
 #include <utility>
@@ -40,6 +43,7 @@
 #include "ray/core_worker_rpc_client/fake_core_worker_client.h"
 #include "ray/observability/fake_metric.h"
 #include "ray/raylet/runtime_env_agent_client.h"
+#include "ray/raylet/shutdown.h"
 #include "ray/raylet/worker.h"
 #include "ray/util/clock.h"
 #include "ray/util/fake_process.h"
@@ -49,6 +53,11 @@
 #include "ray/util/raii.h"
 #include "src/ray/protobuf/runtime_env_agent.pb.h"
 
+#if defined(__linux__)
+#include <sys/wait.h>
+#include <unistd.h>
+#endif
+
 using json = nlohmann::json;
 
 namespace ray::raylet {
@@ -2742,4 +2751,216 @@ TEST_F(WorkerPoolPortRangeTest, AssignsUniquePortsFromTheConfiguredRange) {
   }
 }
 
+#if defined(__linux__)
+class CountingProfilerProcess : public FakeProcess {
+ public:
+  using FakeProcess::FakeProcess;
+  bool IsAlive() const override {
+    ++probes;
+    return FakeProcess::IsAlive();
+  }
+  pid_t GetId() const override {
+    ++pid_reads;
+    return FakeProcess::GetId();
+  }
+  void Kill() override {
+    ++kills;
+    FakeProcess::Kill();
+  }
+  mutable int probes = 0;
+  mutable int pid_reads = 0;
+  int kills = 0;
+};
+
+class WorkerPoolProfilerTest : public WorkerPoolTest {
+ public:
+  void SetUp() override {
+    WorkerPoolTest::SetUp();
+    io_service_.stop();
+    thread_io_service_->join();
+    io_service_.restart();
+    RayConfig::instance().initialize(
+        R"({"process_group_cleanup_enabled":true,
+            "kill_worker_timeout_milliseconds":100,
+            "worker_profiler_flush_timeout_ms":100})");
+  }
+  void TearDown() override {
+    for (auto &[_, profiler] : worker_pool_->profiler_processes_) {
+      worker_pool_->TryFinishProfiler(profiler, true);
+    }
+    worker_pool_->DrainProfilerProcesses();
+    worker_pool_.reset();
+    io_service_.stop();
+    runtime_env_reference.clear();
+    if (child_pid_ > 0) {
+      kill(child_pid_, SIGKILL);
+      while (waitpid(child_pid_, nullptr, 0) < 0 && errno == EINTR) {
+      }
+    }
+  }
+
+  struct Tracked {
+    std::shared_ptr<WorkerInterface> worker;
+    CountingProfilerProcess *process;
+    std::shared_ptr<CountingProfilerProcess> launcher;
+  };
+
+  Tracked Track(bool alive = false, pid_t pid = -1) {
+    const pid_t pgid = Process::PID_MAX_LIMIT + 100 + next_pid_++;
+    auto process =
+        std::make_unique<CountingProfilerProcess>(pid > 0 ? pid : pgid + 1000);
+    auto *ptr = process.get();
+    ptr->SetAlive(alive);
+    auto worker = worker_pool_->CreateWorker(WorkerID::FromRandom(), std::move(process));
+    auto launcher = std::make_shared<CountingProfilerProcess>(pgid);
+    worker->SetSavedProcessGroupId(pgid);
+    auto &state = worker_pool_->states_by_lang_[Language::PYTHON];
+    state.registered_workers.insert(worker);
+    auto &record = state.worker_processes[worker->WorkerId()];
+    record.worker_type = rpc::WorkerType::WORKER;
+    record.is_pending_registration = false;
+    record.proc = launcher;
+    record.profiler_cleanup_owned = true;
+    worker_pool_->profiler_processes_.emplace(
+        worker->WorkerId(), WorkerPool::ProfilerProcess{launcher, worker, pgid});
+    return {worker, ptr, launcher};
+  }
+
+  void Poll() { worker_pool_->PollProfilerProcesses(); }
+  auto &Entry(const Tracked &t) {
+    return worker_pool_->profiler_processes_.at(t.worker->WorkerId());
+  }
+  bool Owns(const Tracked &t) {
+    return worker_pool_->profiler_processes_.contains(t.worker->WorkerId());
+  }
+  bool FinishAgain(const Tracked &t) {
+    return worker_pool_->TryFinishProfiler(Entry(t), true);
+  }
+
+ protected:
+  pid_t child_pid_ = -1;
+
+ private:
+  int next_pid_ = 0;
+};
+
+TEST_F(WorkerPoolProfilerTest, ExitIsIrreversibleAndShutdownPreservesDeadline) {
+  auto t = Track();
+  EXPECT_TRUE(worker_pool_->DeferProfilerCleanup(t.worker->WorkerId(), true));
+  Poll();
+  auto deadline = Entry(t).flush_deadline;
+  ASSERT_TRUE(deadline.has_value());
+  const int probes = t.process->probes;
+  const int reads = t.process->pid_reads;
+  t.process->SetAlive(true);  // The old numeric W PID has been reused.
+  fake_clock_.AdvanceTime(absl::Milliseconds(90));
+  bool done = false;
+  worker_pool_->PrepareProfilerShutdown([&] { done = true; });
+  EXPECT_EQ(Entry(t).flush_deadline, deadline);
+  EXPECT_FALSE(done);
+  fake_clock_.AdvanceTime(absl::Milliseconds(10));
+  Poll();
+  EXPECT_TRUE(done);
+  EXPECT_FALSE(Owns(t));
+  EXPECT_EQ(t.process->probes, probes);
+  EXPECT_EQ(t.process->pid_reads, reads);
+  EXPECT_EQ(t.launcher->kills, 1);
+  worker_pool_->DrainProfilerProcesses();
+  EXPECT_EQ(t.launcher->kills, 1);
+}
+
+TEST_F(WorkerPoolProfilerTest, ConsumedGroupSurvivesLateDisconnect) {
+  auto t = Track();
+  EXPECT_TRUE(worker_pool_->DeferProfilerCleanup(t.worker->WorkerId(), true));
+  Poll();
+  EXPECT_TRUE(FinishAgain(t));
+  EXPECT_EQ(Entry(t).pgid, -1);
+  t.launcher->SetAlive(true);  // Simulate reuse after releasing the group.
+  EXPECT_TRUE(FinishAgain(t));
+  Poll();
+  EXPECT_FALSE(Owns(t));
+  EXPECT_TRUE(t.worker->IsDead());
+  EXPECT_TRUE(worker_pool_->GetAllRegisteredWorkers(true, false).empty());
+  ASSERT_EQ(worker_pool_->GetRegisteredWorker(t.worker->Connection()), t.worker);
+  // DisconnectClient queries this before removing the still-registered worker.
+  // Both EOF and graceful disconnect must skip its ordinary group cleanup.
+  EXPECT_TRUE(worker_pool_->DeferProfilerCleanup(t.worker->WorkerId(), false));
+  EXPECT_TRUE(worker_pool_->DeferProfilerCleanup(t.worker->WorkerId(), true));
+  worker_pool_->DisconnectWorker(t.worker, rpc::WorkerExitType::SYSTEM_ERROR);
+  EXPECT_EQ(worker_pool_->GetRegisteredWorker(t.worker->Connection()), nullptr);
+  worker_pool_->DrainProfilerProcesses();
+  EXPECT_EQ(t.launcher->kills, 1);
+}
+
+TEST_F(WorkerPoolProfilerTest, AgentShutdownYieldsAndForceKillsEscapedWorker) {
+  int ready[2];
+  ASSERT_EQ(pipe(ready), 0);
+  child_pid_ = fork();
+  if (child_pid_ == 0) {
+    close(ready[0]);
+    if (setsid() < 0) {
+      _exit(1);
+    }
+    signal(SIGTERM, SIG_IGN);
+    alarm(10);  // A missing force kill fails by SIGALRM instead of hanging the test.
+    const char ok = 1;
+    if (write(ready[1], &ok, 1) != 1) {
+      _exit(2);
+    }
+    close(ready[1]);
+    for (;;) {
+      pause();
+    }
+  }
+  close(ready[1]);
+  char ok = 0;
+  const auto bytes = child_pid_ > 0 ? read(ready[0], &ok, 1) : -1;
+  close(ready[0]);
+  ASSERT_GT(child_pid_, 0);
+  ASSERT_EQ(bytes, 1);
+  auto t = Track(true, child_pid_);
+  ASSERT_EQ(getpgid(child_pid_), child_pid_);
+  ASSERT_NE(getpgid(child_pid_), *t.worker->GetSavedProcessGroupId());
+
+  bool prepared = false, done = false, replied = false;
+  auto shutdown = MakeRayletShutdownCallback(
+      io_service_, [&](const rpc::NodeDeathInfo &info) {
+        EXPECT_TRUE(io_service_.get_executor().running_in_this_thread());
+        EXPECT_EQ(info.reason_message(), "agent failed");
+        prepared = true;
+        worker_pool_->PrepareProfilerShutdown([&] { done = true; });
+      });
+  std::thread agent([&] {
+    rpc::NodeDeathInfo info;
+    info.set_reason_message("agent failed");
+    shutdown(info);
+    info.set_reason_message("caller reused its message");
+  });
+  agent.join();
+  EXPECT_FALSE(prepared);
+  io_service_.post([&] { replied = true; }, "test.disconnect_reply");
+  io_service_.poll();
+  EXPECT_TRUE(prepared);
+  EXPECT_TRUE(replied);
+  EXPECT_FALSE(done);
+  EXPECT_EQ(kill(child_pid_, 0), 0);
+
+  // DestroyWorker disconnects first, then calls KillAsync, which MarkDead suppresses.
+  EXPECT_TRUE(worker_pool_->DeferProfilerCleanup(t.worker->WorkerId(), false));
+  t.worker->KillAsync(io_service_, true);
+  int status = 0;
+  pid_t reaped;
+  do {
+    reaped = waitpid(child_pid_, &status, 0);
+  } while (reaped < 0 && errno == EINTR);
+  ASSERT_EQ(reaped, child_pid_);
+  child_pid_ = -1;
+  ASSERT_TRUE(WIFSIGNALED(status));
+  EXPECT_EQ(WTERMSIG(status), SIGKILL);
+  Poll();
+  EXPECT_TRUE(done);
+  EXPECT_EQ(t.launcher->kills, 1);
+}
+#endif
+
 }  // namespace ray::raylet
diff --git a/src/ray/raylet/worker_pool.cc b/src/ray/raylet/worker_pool.cc
index 93323ba7b2..679dd5defa 100644
--- a/src/ray/raylet/worker_pool.cc
+++ b/src/ray/raylet/worker_pool.cc
@@ -16,12 +16,17 @@
 
 #include <algorithm>
 #include <boost/date_time/posix_time/posix_time.hpp>
+#include <cerrno>
+#include <chrono>
+#include <csignal>
+#include <cstdio>
 #include <deque>
 #include <fstream>
 #include <iostream>
 #include <memory>
 #include <optional>
 #include <string>
+#include <thread>
 #include <tuple>
 #include <unordered_set>
 #include <utility>
@@ -30,6 +35,7 @@
 #include "absl/random/random.h"
 #include "absl/strings/str_format.h"
 #include "absl/strings/str_split.h"
+#include "nlohmann/json.hpp"
 #include "ray/common/constants.h"
 #include "ray/common/lease/lease_spec.h"
 #include "ray/common/protobuf_utils.h"
@@ -139,8 +145,10 @@ WorkerPool::WorkerPool(instrumented_io_context &io_service,
                        int ray_debugger_external,
                        ClockInterface &clock,
                        WorkerPoolMetrics &worker_pool_metrics,
-                       AddProcessToCgroupHook add_to_cgroup_hook)
+                       AddProcessToCgroupHook add_to_cgroup_hook,
+                       std::string profiler_shutdown_marker_path)
     : clock_(clock),
+      profiler_shutdown_marker_path_(std::move(profiler_shutdown_marker_path)),
       io_service_(&io_service),
       node_id_(node_id),
       node_address_(std::move(node_address)),
@@ -195,13 +203,20 @@ WorkerPool::WorkerPool(instrumented_io_context &io_service,
     RAY_LOG(INFO) << "Initialized the worker port pool with " << ports.size()
                   << " shuffled ports.";
   }
+#if defined(__linux__)
+  UpdateProfilerShutdownMarker();
+#endif
 }
 
 WorkerPool::~WorkerPool() {
-  absl::flat_hash_map<pid_t, std::unique_ptr<ProcessInterface>> procs_to_kill;
+  DrainProfilerProcesses();
+  absl::flat_hash_map<pid_t, std::shared_ptr<ProcessInterface>> procs_to_kill;
   for (auto &entry : states_by_lang_) {
     // Kill all the worker processes.
     for (auto &worker_process : entry.second.worker_processes) {
+      if (worker_process.second.profiler_cleanup_owned) {
+        continue;
+      }
       auto pid = worker_process.second.proc->GetId();
       procs_to_kill.try_emplace(pid, std::move(worker_process.second.proc));
     }
@@ -212,6 +227,223 @@ WorkerPool::~WorkerPool() {
   }
 }
 
+#if defined(__linux__)
+bool WorkerPool::TryFinishProfiler(ProfilerProcess &profiler, bool force) {
+  if (profiler.state == ProfilerState::finished) {
+    return true;
+  }
+  const auto now = clock_.SteadyNow();
+  if (!force) {
+    if (profiler.state == ProfilerState::waiting_for_worker_exit) {
+      if (profiler.worker->GetProcess().IsAlive()) {
+        return false;
+      }
+      profiler.state = ProfilerState::flushing;
+      profiler.flush_deadline =
+          now + std::chrono::milliseconds(std::max<int64_t>(
+                    0, RayConfig::instance().worker_profiler_flush_timeout_ms()));
+    }
+    // Once flushing, never probe or signal the old worker PID again.
+    if (now < *profiler.flush_deadline && profiler.launcher->IsAlive()) {
+      return false;
+    }
+  }
+  // W may have escaped the saved process group (e.g. via setsid). Terminate it
+  // directly before MarkDead suppresses DestroyWorker's subsequent KillAsync.
+  // Never touch W's old PID once its exit has been observed.
+  if (profiler.state == ProfilerState::waiting_for_worker_exit &&
+      !profiler.worker_kill_sent) {
+    profiler.worker_kill_sent = true;
+    kill(profiler.worker->GetProcess().GetId(), SIGKILL);
+  }
+  // Consume ownership before signaling the launcher/group. Stop filters dead
+  // workers; the process record retains ownership until DisconnectWorker.
+  profiler.state = ProfilerState::finished;
+  profiler.worker->MarkDead();
+  const auto pgid = std::exchange(profiler.pgid, -1);
+  if (profiler.launcher->IsAlive()) {
+    profiler.launcher->Kill();
+  }
+  RAY_LOG(INFO).WithField(profiler.worker->WorkerId())
+      << "Profiler cleanup: sending SIGKILL to pgid=" << pgid;
+  auto error = KillProcessGroup(pgid, SIGKILL);
+  if (error && *error && error->value() != ESRCH) {
+    RAY_LOG(WARNING) << "Profiler group cleanup failed: " << error->message();
+  }
+  return true;
+}
+
+void WorkerPool::UpdateProfilerShutdownMarker() {
+  if (profiler_shutdown_marker_path_.empty()) {
+    return;
+  }
+  if (profiler_processes_.empty()) {
+    std::remove(profiler_shutdown_marker_path_.c_str());
+    return;
+  }
+  // Publish the effective C++ budget, including on non-head nodes. Python only
+  // extends its graceful wait while this raylet owns a profiled worker/launcher.
+  const auto timeout_ms =
+      std::max<int64_t>(0, RayConfig::instance().kill_worker_timeout_milliseconds()) +
+      std::max<int64_t>(0, RayConfig::instance().worker_profiler_flush_timeout_ms());
+  const auto temporary = profiler_shutdown_marker_path_ + ".tmp";
+  std::ofstream marker(temporary);
+  marker << timeout_ms;
+  marker.close();
+  if (!marker || std::rename(temporary.c_str(), profiler_shutdown_marker_path_.c_str())) {
+    RAY_LOG(WARNING) << "Could not publish profiler shutdown budget";
+    std::remove(temporary.c_str());
+  }
+}
+
+void WorkerPool::ScheduleProfilerPoll() {
+  if (profiler_timer_pending_) {
+    return;
+  }
+  if (!profiler_timer_) {
+    profiler_timer_ = std::make_unique<boost::asio::steady_timer>(*io_service_);
+  }
+  profiler_timer_pending_ = true;
+  profiler_timer_->expires_after(std::chrono::milliseconds(10));
+  profiler_timer_->async_wait(
+      [this, lifetime = std::weak_ptr<int>(profiler_lifetime_)](
+          const boost::system::error_code &ec) {
+        if (ec || lifetime.expired()) {
+          return;
+        }
+        profiler_timer_pending_ = false;
+        PollProfilerProcesses();
+      });
+}
+
+void WorkerPool::PollProfilerProcesses() {
+  bool pending = false;
+  const auto now = clock_.SteadyNow();
+  for (auto it = profiler_processes_.begin(); it != profiler_processes_.end();) {
+    auto &profiler = it->second;
+    if (!profiler.cleanup_requested) {
+      ++it;
+      continue;
+    }
+    const bool force =
+        profiler_shutdown_deadline_ && now >= *profiler_shutdown_deadline_;
+    if (TryFinishProfiler(profiler, force)) {
+      profiler_processes_.erase(it++);
+      if (profiler_processes_.empty()) {
+        UpdateProfilerShutdownMarker();
+      }
+      continue;
+    }
+    if (profiler.state == ProfilerState::waiting_for_worker_exit &&
+        profiler_worker_exit_deadline_ && now >= *profiler_worker_exit_deadline_ &&
+        !profiler.worker_kill_sent) {
+      profiler.worker_kill_sent = true;
+      kill(profiler.worker->GetProcess().GetId(), SIGKILL);
+    }
+    pending = true;
+    ++it;
+  }
+  if (pending) {
+    ScheduleProfilerPoll();
+  } else if (profiler_shutdown_done_) {
+    auto done = std::move(profiler_shutdown_done_);
+    done();  // May destroy the pool; do not access members afterwards.
+  }
+}
+#endif
+
+bool WorkerPool::DeferProfilerCleanup(const WorkerID &worker_id, bool graceful) {
+#if defined(__linux__)
+  auto it = profiler_processes_.find(worker_id);
+  if (it != profiler_processes_.end()) {
+    if (!graceful) {
+      TryFinishProfiler(it->second, true);
+      profiler_processes_.erase(it);
+      if (profiler_processes_.empty()) {
+        UpdateProfilerShutdownMarker();
+      }
+      return true;
+    }
+    it->second.cleanup_requested = true;
+    // Do not finish inline: DisconnectClient must send its reply first.
+    ScheduleProfilerPoll();
+    return true;
+  }
+  // Terminal cleanup can precede a late EOF while the worker is still registered.
+  // Its record outlives the registry entry, so disconnect must not reacquire PGID.
+  for (const auto &[_, state] : states_by_lang_) {
+    auto process = state.worker_processes.find(worker_id);
+    if (process != state.worker_processes.end() &&
+        process->second.profiler_cleanup_owned) {
+      return true;
+    }
+  }
+#endif
+  return false;
+}
+
+void WorkerPool::PrepareProfilerShutdown(std::function<void()> done) {
+  RAY_CHECK(!profiler_shutdown_started_);
+  profiler_shutdown_started_ = true;
+#if defined(__linux__)
+  const auto now = clock_.SteadyNow();
+  profiler_worker_exit_deadline_ =
+      now + std::chrono::milliseconds(std::max<int64_t>(
+                0, RayConfig::instance().kill_worker_timeout_milliseconds()));
+  profiler_shutdown_deadline_ =
+      *profiler_worker_exit_deadline_ +
+      std::chrono::milliseconds(std::max<int64_t>(
+          0, RayConfig::instance().worker_profiler_flush_timeout_ms()));
+  profiler_shutdown_done_ = std::move(done);
+  for (auto &[_, profiler] : profiler_processes_) {
+    profiler.cleanup_requested = true;
+    // Observe exit before signaling, preserving any existing flush deadline.
+    if (!TryFinishProfiler(profiler, false) &&
+        profiler.state == ProfilerState::waiting_for_worker_exit) {
+      kill(profiler.worker->GetProcess().GetId(), SIGTERM);
+    }
+  }
+  PollProfilerProcesses();
+#else
+  done();
+#endif
+}
+
+void WorkerPool::DrainProfilerProcesses() {
+#if defined(__linux__)
+  profiler_lifetime_.reset();
+  if (profiler_timer_) {
+    boost::system::error_code ignored;
+    profiler_timer_->cancel(ignored);
+  }
+  profiler_shutdown_done_ = nullptr;
+  // This is a fallback for Stop/destruction without PrepareProfilerShutdown.
+  // Never wait for a live W here: it may need a disconnect reply on this executor.
+  // Only launchers whose W has exited may consume their remaining flush budget.
+  while (!profiler_processes_.empty()) {
+    for (auto it = profiler_processes_.begin(); it != profiler_processes_.end();) {
+      auto &profiler = it->second;
+      bool finished = TryFinishProfiler(profiler, false);
+      if (!finished &&
+          (profiler.state == ProfilerState::waiting_for_worker_exit ||
+           (profiler_shutdown_deadline_ &&
+            clock_.SteadyNow() >= *profiler_shutdown_deadline_))) {
+        finished = TryFinishProfiler(profiler, true);
+      }
+      if (finished) {
+        profiler_processes_.erase(it++);
+      } else {
+        ++it;
+      }
+    }
+    if (!profiler_processes_.empty()) {
+      std::this_thread::sleep_for(std::chrono::milliseconds(10));
+    }
+  }
+  UpdateProfilerShutdownMarker();
+#endif
+}
+
 void WorkerPool::Start() {
   if (RayConfig::instance().kill_idle_workers_interval_ms() > 0) {
     periodical_runner_->RunFnPeriodically(
@@ -864,6 +1096,11 @@ Status WorkerPool::RegisterWorker(const std::shared_ptr<WorkerInterface> &worker
                                   pid_t pid,
                                   std::function<void(Status, int)> send_reply_callback) {
   RAY_CHECK(worker);
+  if (profiler_shutdown_started_) {
+    const auto status = Status::Invalid("Raylet is shutting down");
+    send_reply_callback(status, /*port=*/0);
+    return status;
+  }
   auto &state = GetStateForLanguage(worker->GetLanguage());
   const WorkerID &worker_id = worker->WorkerId();
   auto it = state.worker_processes.find(worker_id);
@@ -911,6 +1148,28 @@ Status WorkerPool::RegisterWorker(const std::shared_ptr<WorkerInterface> &worker
       << ", worker_type: " << rpc::WorkerType_Name(worker->GetWorkerType());
   worker->SetAssignedPort(port);
 
+#if defined(__linux__)
+  // Nsight's launcher exports after the registered Python process exits.
+  // P != W alone is insufficient (e.g. py_executable wrappers or rocprof-sys).
+  if (RayConfig::instance().process_group_cleanup_enabled() &&
+      worker->GetLanguage() == Language::PYTHON && it->second.proc->GetId() != pid) {
+    const auto env = nlohmann::json::parse(
+        it->second.runtime_env_info.serialized_runtime_env(), nullptr, false);
+    const auto saved = worker->GetSavedProcessGroupId();
+    if (env.is_object() && env.contains("_nsight") &&
+        (env["_nsight"] == "default" ||
+         (env["_nsight"].is_object() && !env["_nsight"].empty())) &&
+        saved && *saved > 1 && *saved != getpgrp()) {
+      const bool first = profiler_processes_.empty();
+      profiler_processes_.emplace(
+          worker_id, ProfilerProcess{it->second.proc, worker, *saved});
+      it->second.profiler_cleanup_owned = true;
+      if (first) {
+        UpdateProfilerShutdownMarker();
+      }
+    }
+  }
+#endif
   state.registered_workers.insert(worker);
 
   // Send the reply immediately for worker registrations.
diff --git a/src/ray/raylet/worker_pool.h b/src/ray/raylet/worker_pool.h
index b6a0a3447d..3439f3b4df 100644
--- a/src/ray/raylet/worker_pool.h
+++ b/src/ray/raylet/worker_pool.h
@@ -18,6 +18,7 @@
 
 #include <algorithm>
 #include <boost/asio/io_service.hpp>
+#include <boost/asio/steady_timer.hpp>
 #include <boost/functional/hash.hpp>
 #include <deque>
 #include <list>
@@ -197,6 +198,13 @@ class IOWorkerPoolInterface {
 /// Used for new scheduler unit tests.
 class WorkerPoolInterface : public IOWorkerPoolInterface {
  public:
+  /// Own group cleanup for a registered Nsight worker; defer only graceful exits.
+  virtual bool DeferProfilerCleanup(const WorkerID &, bool graceful) { return false; }
+  /// Run before Stop, with the main event loop still serving disconnect replies.
+  virtual void PrepareProfilerShutdown(std::function<void()> done) { done(); }
+  /// Drain exited workers' launchers; force-finish live workers without an IPC wait.
+  virtual void DrainProfilerProcesses() {}
+
   /// Pop an idle worker from the pool. The caller is responsible for pushing
   /// the worker back onto the pool once the worker has completed its work.
   ///
@@ -322,6 +330,10 @@ inline std::ostream &operator<<(std::ostream &os,
 /// is a container for a unit of work.
 class WorkerPool : public WorkerPoolInterface {
  public:
+  bool DeferProfilerCleanup(const WorkerID &worker_id, bool graceful) override;
+  void PrepareProfilerShutdown(std::function<void()> done) override;
+  void DrainProfilerProcesses() override;
+
   /// Create a pool and asynchronously start at least the specified number of workers per
   /// language.
   /// Once each worker process has registered with an external server, the
@@ -371,7 +383,8 @@ class WorkerPool : public WorkerPoolInterface {
       int ray_debugger_external,
       ClockInterface &clock,
       WorkerPoolMetrics &worker_pool_metrics,
-      AddProcessToCgroupHook add_to_cgroup_hook = [](const std::string &) {});
+      AddProcessToCgroupHook add_to_cgroup_hook = [](const std::string &) {},
+      std::string profiler_shutdown_marker_path = "");
 
   /// Destructor responsible for freeing a set of workers owned by this class.
   ~WorkerPool() override;
@@ -682,8 +695,8 @@ class WorkerPool : public WorkerPoolInterface {
     bool is_pending_registration = true;
     /// The type of the worker.
     rpc::WorkerType worker_type;
-    /// The worker process instance.
-    std::unique_ptr<ProcessInterface> proc;
+    /// Retained by profiler_processes_ until post-worker-exit export completes.
+    std::shared_ptr<ProcessInterface> proc;
     /// The worker process start time (monotonic, for measuring startup latency).
     SteadyTimePoint start_time;
     /// The runtime env Info.
@@ -692,6 +705,8 @@ class WorkerPool : public WorkerPoolInterface {
     std::vector<std::string> dynamic_options;
     /// The duration to keep the newly created worker alive before it's assigned a lease.
     std::optional<absl::Duration> worker_startup_keep_alive_duration;
+    /// Profiler cleanup owns this group, even after completion, until disconnect.
+    bool profiler_cleanup_owned = false;
   };
 
   /// An internal data structure that maintains the pool state per language.
@@ -746,6 +761,35 @@ class WorkerPool : public WorkerPoolInterface {
   std::list<IdleWorkerEntry> idle_of_all_languages_;
 
  private:
+#if defined(__linux__)
+  enum class ProfilerState { waiting_for_worker_exit, flushing, finished };
+  struct ProfilerProcess {
+    std::shared_ptr<ProcessInterface> launcher;
+    std::shared_ptr<WorkerInterface> worker;
+    pid_t pgid;
+    ProfilerState state = ProfilerState::waiting_for_worker_exit;
+    bool cleanup_requested = false;
+    bool worker_kill_sent = false;
+    std::optional<SteadyTimePoint> flush_deadline = std::nullopt;
+  };
+  // All access, including destruction, is serialized with the main executor.
+  // Retains launcher ownership after DisconnectWorker removes the process record.
+  absl::flat_hash_map<WorkerID, ProfilerProcess> profiler_processes_;
+  std::unique_ptr<boost::asio::steady_timer> profiler_timer_;
+  // Cancellation alone does not invalidate an already queued successful callback.
+  std::shared_ptr<int> profiler_lifetime_ = std::make_shared<int>(0);
+  bool profiler_timer_pending_ = false;
+  std::optional<SteadyTimePoint> profiler_worker_exit_deadline_;
+  std::optional<SteadyTimePoint> profiler_shutdown_deadline_;
+  std::function<void()> profiler_shutdown_done_;
+  bool TryFinishProfiler(ProfilerProcess &profiler, bool force);
+  void PollProfilerProcesses();
+  void ScheduleProfilerPoll();
+  void UpdateProfilerShutdownMarker();
+#endif
+  bool profiler_shutdown_started_ = false;
+  const std::string profiler_shutdown_marker_path_;
+
   /// A helper function that returns the reference of the pool state
   /// for a given language.
   State &GetStateForLanguage(const Language &language);
@@ -975,6 +1019,7 @@ class WorkerPool : public WorkerPoolInterface {
   static inline const ProcessInterface &kNullProcess = Process();
 
   friend class WorkerPoolTest;
+  friend class WorkerPoolProfilerTest;
   friend class WorkerPoolDriverRegisteredTest;
 };