diff --git a/be/src/runtime/fragment_mgr.cpp b/be/src/runtime/fragment_mgr.cpp
index cae968fee7b94e..e81852f97f020c 100644
--- a/be/src/runtime/fragment_mgr.cpp
+++ b/be/src/runtime/fragment_mgr.cpp
@@ -336,10 +336,10 @@ void FragmentMgr::stop() {
         _cancel_thread->join();
     }
 
+    _thread_pool->shutdown();
     // Only me can delete
     _query_ctx_map.clear();
     _pipeline_map.clear();
-    _thread_pool->shutdown();
 }
 
 std::string FragmentMgr::to_http_path(const std::string& file_name) {
diff --git a/be/src/runtime/fragment_mgr.h b/be/src/runtime/fragment_mgr.h
index 4c17fff2acb795..4eab31ff3a24ab 100644
--- a/be/src/runtime/fragment_mgr.h
+++ b/be/src/runtime/fragment_mgr.h
@@ -171,6 +171,10 @@ class FragmentMgr : public RestMonitorIface {
 
     ThreadPool* get_thread_pool() { return _thread_pool.get(); }
 
+    // When fragment mgr is going to stop, the _stop_background_threads_latch is set to 0
+    // and other module that use fragment mgr's thread pool should get this signal and exit.
+    bool shutting_down() { return _stop_background_threads_latch.count() == 0; }
+
     int32_t running_query_num() { return _query_ctx_map.num_items(); }
 
     std::string dump_pipeline_tasks(int64_t duration = 0);
@@ -215,7 +219,7 @@ class FragmentMgr : public RestMonitorIface {
 
     CountDownLatch _stop_background_threads_latch;
     scoped_refptr<Thread> _cancel_thread;
-    // every job is a pool
+    // This pool is used as global async task pool
     std::unique_ptr<ThreadPool> _thread_pool;
 
     std::shared_ptr<MetricEntity> _entity;
diff --git a/be/src/vec/sink/writer/async_result_writer.cpp b/be/src/vec/sink/writer/async_result_writer.cpp
index ea623d2dcc304e..57f656bda39f44 100644
--- a/be/src/vec/sink/writer/async_result_writer.cpp
+++ b/be/src/vec/sink/writer/async_result_writer.cpp
@@ -144,10 +144,18 @@ void AsyncResultWriter::process_block(RuntimeState* state, RuntimeProfile* profi
         //1) wait scan operator write data
         {
             std::unique_lock l(_m);
-            while (!_eos && _data_queue.empty() && _writer_status.ok()) {
+            // When the query is cancelled, _writer_status may be set to error status in force_close method.
+            // When the BE process is exit gracefully, the fragment mgr's thread pool will be shutdown,
+            // and the async thread will be exit.
+            while (!_eos && _data_queue.empty() && _writer_status.ok() &&
+                   !ExecEnv::GetInstance()->fragment_mgr()->shutting_down()) {
                 // Add 1s to check to avoid lost signal
                 _cv.wait_for(l, std::chrono::seconds(1));
             }
+            // If writer status is not ok, then we should not change its status to avoid lost the actual error status.
+            if (ExecEnv::GetInstance()->fragment_mgr()->shutting_down() && _writer_status.ok()) {
+                _writer_status.update(Status::InternalError<false>("FragmentMgr is shutting down"));
+            }
 
             //check if eos or writer error
             if ((_eos && _data_queue.empty()) || !_writer_status.ok()) {