[Improvement](scheduler) Use a separate eager queue to execute cancel…

be/src/pipeline/pipeline_fragment_context.cpp

@@ -445,14 +445,14 @@ Status PipelineFragmentContext::_build_pipeline_tasks(const doris::TPipelineFrag
                 auto cur_task_id = _total_tasks++;
                 task_runtime_state->set_task_id(cur_task_id);
                 task_runtime_state->set_task_num(pipeline->num_tasks());
-                auto task = std::make_unique<PipelineTask>(pipeline, cur_task_id,
-                                                           task_runtime_state.get(), this,
-                                                           pipeline_id_to_profile[pip_idx].get(),
-                                                           get_local_exchange_state(pipeline), i);
+                auto task = std::shared_ptr<PipelineTask>(
+                        new PipelineTask(pipeline, cur_task_id, task_runtime_state.get(), this,
+                                         pipeline_id_to_profile[pip_idx].get(),
+                                         get_local_exchange_state(pipeline), i));
                 pipeline->incr_created_tasks(i, task.get());
                 task_runtime_state->set_task(task.get());
                 pipeline_id_to_task.insert({pipeline->id(), task.get()});
-                _tasks[i].emplace_back(std::move(task));
+                _tasks[i].emplace_back(task);
             }
         }
 
@@ -1679,7 +1679,7 @@ Status PipelineFragmentContext::submit() {
     auto* scheduler = _query_ctx->get_pipe_exec_scheduler();
     for (auto& task : _tasks) {
         for (auto& t : task) {
-            st = scheduler->schedule_task(t.get());
+            st = scheduler->schedule_task(t);
             if (!st) {
                 cancel(Status::InternalError("submit context to executor fail"));
                 std::lock_guard<std::mutex> l(_task_mutex);

be/src/pipeline/pipeline_fragment_context.h

@@ -226,7 +226,7 @@ class PipelineFragmentContext : public TaskExecutionContext {
 
     OperatorPtr _root_op = nullptr;
     // this is a [n * m] matrix. n is parallelism of pipeline engine and m is the number of pipelines.
-    std::vector<std::vector<std::unique_ptr<PipelineTask>>> _tasks;
+    std::vector<std::vector<std::shared_ptr<PipelineTask>>> _tasks;
 
     // TODO: remove the _sink and _multi_cast_stream_sink_senders to set both
     // of it in pipeline task not the fragment_context

be/src/pipeline/pipeline_task.cpp

@@ -394,7 +394,7 @@ Status PipelineTask::execute(bool* eos) {
         }
     }
 
-    RETURN_IF_ERROR(get_task_queue()->push_back(this));
+    RETURN_IF_ERROR(get_task_queue()->push_back(shared_from_this()));
     return Status::OK();
 }
 
@@ -554,10 +554,34 @@ std::string PipelineTask::debug_string() {
 
 void PipelineTask::wake_up() {
     // call by dependency
-    static_cast<void>(get_task_queue()->push_back(this));
+    static_cast<void>(get_task_queue()->push_back(shared_from_this()));
+}
+
+void PipelineTask::clear_blocking_state() {
+    _state->get_query_ctx()->get_execution_dependency()->set_always_ready();
+    // We use a lock to assure all dependencies are not deconstructed here.
+    std::unique_lock<std::mutex> lc(_dependency_lock);
+    if (!_finalized) {
+        _execution_dep->set_always_ready();
+        for (auto* dep : _filter_dependencies) {
+            dep->set_always_ready();
+        }
+        for (auto& deps : _read_dependencies) {
+            for (auto* dep : deps) {
+                dep->set_always_ready();
+            }
+        }
+        for (auto* dep : _write_dependencies) {
+            dep->set_always_ready();
+        }
+        for (auto* dep : _finish_dependencies) {
+            dep->set_always_ready();
+        }
+    }
 }
 
 QueryContext* PipelineTask::query_context() {
     return _fragment_context->get_query_ctx();
 }
+
 } // namespace doris::pipeline

be/src/pipeline/pipeline_task.h

@@ -45,7 +45,7 @@ class MultiCoreTaskQueue;
 class PriorityTaskQueue;
 class Dependency;
 
-class PipelineTask {
+class PipelineTask final : public std::enable_shared_from_this<PipelineTask> {
 public:
     PipelineTask(PipelinePtr& pipeline, uint32_t task_id, RuntimeState* state,
                  PipelineFragmentContext* fragment_context, RuntimeProfile* parent_profile,
@@ -137,28 +137,7 @@ class PipelineTask {
 
     void set_wake_up_early() { _wake_up_early = true; }
 
-    void clear_blocking_state() {
-        _state->get_query_ctx()->get_execution_dependency()->set_always_ready();
-        // We use a lock to assure all dependencies are not deconstructed here.
-        std::unique_lock<std::mutex> lc(_dependency_lock);
-        if (!_finalized) {
-            _execution_dep->set_always_ready();
-            for (auto* dep : _filter_dependencies) {
-                dep->set_always_ready();
-            }
-            for (auto& deps : _read_dependencies) {
-                for (auto* dep : deps) {
-                    dep->set_always_ready();
-                }
-            }
-            for (auto* dep : _write_dependencies) {
-                dep->set_always_ready();
-            }
-            for (auto* dep : _finish_dependencies) {
-                dep->set_always_ready();
-            }
-        }
-    }
+    void clear_blocking_state();
 
     void set_task_queue(MultiCoreTaskQueue* task_queue) { _task_queue = task_queue; }
     MultiCoreTaskQueue* get_task_queue() { return _task_queue; }

be/src/pipeline/task_queue.cpp

@@ -29,7 +29,7 @@
 namespace doris::pipeline {
 #include "common/compile_check_begin.h"
 
-PipelineTask* SubTaskQueue::try_take(bool is_steal) {
+std::shared_ptr<PipelineTask> SubTaskQueue::try_take(bool is_steal) {
     if (_queue.empty()) {
         return nullptr;
     }
@@ -54,7 +54,7 @@ void PriorityTaskQueue::close() {
     _wait_task.notify_all();
 }
 
-PipelineTask* PriorityTaskQueue::_try_take_unprotected(bool is_steal) {
+std::shared_ptr<PipelineTask> PriorityTaskQueue::_try_take_unprotected(bool is_steal) {
     if (_total_task_size == 0 || _closed) {
         return nullptr;
     }
@@ -90,13 +90,13 @@ int PriorityTaskQueue::_compute_level(uint64_t runtime) {
     return SUB_QUEUE_LEVEL - 1;
 }
 
-PipelineTask* PriorityTaskQueue::try_take(bool is_steal) {
+std::shared_ptr<PipelineTask> PriorityTaskQueue::try_take(bool is_steal) {
     // TODO other efficient lock? e.g. if get lock fail, return null_ptr
     std::unique_lock<std::mutex> lock(_work_size_mutex);
     return _try_take_unprotected(is_steal);
 }
 
-PipelineTask* PriorityTaskQueue::take(uint32_t timeout_ms) {
+std::shared_ptr<PipelineTask> PriorityTaskQueue::take(uint32_t timeout_ms) {
     std::unique_lock<std::mutex> lock(_work_size_mutex);
     auto task = _try_take_unprotected(false);
     if (task) {
@@ -111,7 +111,7 @@ PipelineTask* PriorityTaskQueue::take(uint32_t timeout_ms) {
     }
 }
 
-Status PriorityTaskQueue::push(PipelineTask* task) {
+Status PriorityTaskQueue::push(std::shared_ptr<PipelineTask>& task) {
     if (_closed) {
         return Status::InternalError("WorkTaskQueue closed");
     }
@@ -132,8 +132,12 @@ Status PriorityTaskQueue::push(PipelineTask* task) {
 
 MultiCoreTaskQueue::~MultiCoreTaskQueue() = default;
 
+static constexpr int NUM_EAGER_QUEUES = 1;
 MultiCoreTaskQueue::MultiCoreTaskQueue(int core_size)
-        : _prio_task_queues(core_size), _closed(false), _core_size(core_size) {}
+        : _prio_task_queues(core_size + NUM_EAGER_QUEUES),
+          _closed(false),
+          _core_size(core_size),
+          _urgent_queue_idx(core_size) {}
 
 void MultiCoreTaskQueue::close() {
     if (_closed) {
@@ -145,8 +149,8 @@ void MultiCoreTaskQueue::close() {
                           [](auto& prio_task_queue) { prio_task_queue.close(); });
 }
 
-PipelineTask* MultiCoreTaskQueue::take(int core_id) {
-    PipelineTask* task = nullptr;
+std::shared_ptr<PipelineTask> MultiCoreTaskQueue::take(int core_id) {
+    std::shared_ptr<PipelineTask> task = nullptr;
     while (!_closed) {
         DCHECK(_prio_task_queues.size() > core_id)
                 << " list size: " << _prio_task_queues.size() << " core_id: " << core_id
@@ -156,9 +160,11 @@ PipelineTask* MultiCoreTaskQueue::take(int core_id) {
             task->set_core_id(core_id);
             break;
         }
-        task = _steal_take(core_id);
-        if (task) {
-            break;
+        if (core_id != _urgent_queue_idx) {
+            task = _steal_take(core_id);
+            if (task) {
+                break;
+            }
         }
         task = _prio_task_queues[core_id].take(WAIT_CORE_TASK_TIMEOUT_MS /* timeout_ms */);
         if (task) {
@@ -172,7 +178,7 @@ PipelineTask* MultiCoreTaskQueue::take(int core_id) {
     return task;
 }
 
-PipelineTask* MultiCoreTaskQueue::_steal_take(int core_id) {
+std::shared_ptr<PipelineTask> MultiCoreTaskQueue::_steal_take(int core_id) {
     DCHECK(core_id < _core_size);
     int next_id = core_id;
     for (int i = 1; i < _core_size; ++i) {
@@ -190,18 +196,19 @@ PipelineTask* MultiCoreTaskQueue::_steal_take(int core_id) {
     return nullptr;
 }
 
-Status MultiCoreTaskQueue::push_back(PipelineTask* task) {
+Status MultiCoreTaskQueue::push_back(std::shared_ptr<PipelineTask> task) {
     int core_id = task->get_previous_core_id();
     if (core_id < 0) {
         core_id = _next_core.fetch_add(1) % _core_size;
     }
     return push_back(task, core_id);
 }
 
-Status MultiCoreTaskQueue::push_back(PipelineTask* task, int core_id) {
-    DCHECK(core_id < _core_size);
+Status MultiCoreTaskQueue::push_back(std::shared_ptr<PipelineTask>& task, int core_id) {
+    DCHECK(core_id < _core_size || task->query_context()->is_cancelled());
     task->put_in_runnable_queue();
-    return _prio_task_queues[core_id].push(task);
+    return _prio_task_queues[task->query_context()->is_cancelled() ? _urgent_queue_idx : core_id]
+            .push(task);
 }
 
 void MultiCoreTaskQueue::update_statistics(PipelineTask* task, int64_t time_spent) {

be/src/pipeline/task_queue.h

@@ -38,9 +38,9 @@ class SubTaskQueue {
     friend class PriorityTaskQueue;
 
 public:
-    void push_back(PipelineTask* task) { _queue.emplace(task); }
+    void push_back(std::shared_ptr<PipelineTask>& task) { _queue.emplace(task); }
 
-    PipelineTask* try_take(bool is_steal);
+    std::shared_ptr<PipelineTask> try_take(bool is_steal);
 
     void set_level_factor(double level_factor) { _level_factor = level_factor; }
 
@@ -58,7 +58,7 @@ class SubTaskQueue {
     bool empty() { return _queue.empty(); }
 
 private:
-    std::queue<PipelineTask*> _queue;
+    std::queue<std::shared_ptr<PipelineTask>> _queue;
     // depends on LEVEL_QUEUE_TIME_FACTOR
     double _level_factor = 1;
 
@@ -72,18 +72,18 @@ class PriorityTaskQueue {
 
     void close();
 
-    PipelineTask* try_take(bool is_steal);
+    std::shared_ptr<PipelineTask> try_take(bool is_steal);
 
-    PipelineTask* take(uint32_t timeout_ms = 0);
+    std::shared_ptr<PipelineTask> take(uint32_t timeout_ms = 0);
 
-    Status push(PipelineTask* task);
+    Status push(std::shared_ptr<PipelineTask>& task);
 
     void inc_sub_queue_runtime(int level, uint64_t runtime) {
         _sub_queues[level].inc_runtime(runtime);
     }
 
 private:
-    PipelineTask* _try_take_unprotected(bool is_steal);
+    std::shared_ptr<PipelineTask> _try_take_unprotected(bool is_steal);
     static constexpr auto LEVEL_QUEUE_TIME_FACTOR = 2;
     static constexpr size_t SUB_QUEUE_LEVEL = 6;
     SubTaskQueue _sub_queues[SUB_QUEUE_LEVEL];
@@ -112,25 +112,26 @@ class MultiCoreTaskQueue {
     void close();
 
     // Get the task by core id.
-    PipelineTask* take(int core_id);
+    std::shared_ptr<PipelineTask> take(int core_id);
 
     // TODO combine these methods to `push_back(task, core_id = -1)`
-    Status push_back(PipelineTask* task);
+    Status push_back(std::shared_ptr<PipelineTask> task);
 
-    Status push_back(PipelineTask* task, int core_id);
+    Status push_back(std::shared_ptr<PipelineTask>& task, int core_id);
 
     void update_statistics(PipelineTask* task, int64_t time_spent);
 
-    int cores() const { return _core_size; }
+    int num_queues() const { return cast_set<int>(_prio_task_queues.size()); }
 
 private:
-    PipelineTask* _steal_take(int core_id);
+    std::shared_ptr<PipelineTask> _steal_take(int core_id);
 
     std::vector<PriorityTaskQueue> _prio_task_queues;
     std::atomic<uint32_t> _next_core = 0;
     std::atomic<bool> _closed;
 
-    int _core_size;
+    const int _core_size;
+    const int _urgent_queue_idx;
     static constexpr auto WAIT_CORE_TASK_TIMEOUT_MS = 100;
 };
 #include "common/compile_check_end.h"

be/src/pipeline/task_scheduler.cpp

@@ -51,27 +51,27 @@ TaskScheduler::~TaskScheduler() {
 }
 
 Status TaskScheduler::start() {
-    int cores = _task_queue.cores();
+    int num_queues = _task_queue.num_queues();
     RETURN_IF_ERROR(ThreadPoolBuilder(_name)
-                            .set_min_threads(cores)
-                            .set_max_threads(cores)
+                            .set_min_threads(num_queues)
+                            .set_max_threads(num_queues)
                             .set_max_queue_size(0)
                             .set_cgroup_cpu_ctl(_cgroup_cpu_ctl)
                             .build(&_fix_thread_pool));
-    LOG_INFO("TaskScheduler set cores").tag("size", cores);
-    _markers.resize(cores, true);
-    for (int i = 0; i < cores; ++i) {
+    LOG_INFO("TaskScheduler set cores").tag("size", num_queues);
+    _markers.resize(num_queues, true);
+    for (int i = 0; i < num_queues; ++i) {
         RETURN_IF_ERROR(_fix_thread_pool->submit_func([this, i] { _do_work(i); }));
     }
     return Status::OK();
 }
 
-Status TaskScheduler::schedule_task(PipelineTask* task) {
+Status TaskScheduler::schedule_task(std::shared_ptr<PipelineTask>& task) {
     return _task_queue.push_back(task);
 }
 
 // after _close_task, task maybe destructed.
-void _close_task(PipelineTask* task, Status exec_status) {
+void _close_task(std::shared_ptr<PipelineTask> task, Status exec_status) {
     // Has to attach memory tracker here, because the close task will also release some memory.
     // Should count the memory to the query or the query's memory will not decrease when part of
     // task finished.
@@ -95,11 +95,12 @@ void _close_task(PipelineTask* task, Status exec_status) {
     task->finalize();
     task->set_running(false);
     task->fragment_context()->close_a_pipeline(task->pipeline_id());
+    task.reset();
 }
 
 void TaskScheduler::_do_work(int index) {
     while (_markers[index]) {
-        auto* task = _task_queue.take(index);
+        auto task = _task_queue.take(index);
         if (!task) {
             continue;
         }
@@ -122,7 +123,7 @@ void TaskScheduler::_do_work(int index) {
             // If pipeline is canceled, it will report after pipeline closed, and will propagate
             // errors to downstream through exchange. So, here we needn't send_report.
             // fragment_ctx->send_report(true);
-            _close_task(task, fragment_ctx->get_query_ctx()->exec_status());
+            _close_task(std::move(task), fragment_ctx->get_query_ctx()->exec_status());
             continue;
         }
 
@@ -165,7 +166,7 @@ void TaskScheduler::_do_work(int index) {
             LOG(WARNING) << fmt::format("Pipeline task failed. query_id: {} reason: {}",
                                         print_id(task->query_context()->query_id()),
                                         status.to_string());
-            _close_task(task, status);
+            _close_task(std::move(task), status);
             continue;
         }
         fragment_ctx->trigger_report_if_necessary();
@@ -178,7 +179,7 @@ void TaskScheduler::_do_work(int index) {
                 task->set_running(false);
             } else {
                 Status exec_status = fragment_ctx->get_query_ctx()->exec_status();
-                _close_task(task, exec_status);
+                _close_task(std::move(task), exec_status);
             }
             continue;
         }