[bugfix](scanmemory)modify scan memory profile (apache#41435)

1. memory usage in scan operator is empty.
2. peak memory usage is not in scan operator , it is in scanner.

---------

Co-authored-by: yiguolei <yiguolei@gmail.com>

Author: 2 people

be/src/pipeline/exec/operator.h

@@ -199,7 +199,7 @@ class PipelineXLocalStateBase {
     RuntimeProfile::Counter* _projection_timer = nullptr;
     RuntimeProfile::Counter* _exec_timer = nullptr;
     // Account for peak memory used by this node
-    RuntimeProfile::Counter* _peak_memory_usage_counter = nullptr;
+    RuntimeProfile::HighWaterMarkCounter* _peak_memory_usage_counter = nullptr;
     RuntimeProfile::Counter* _init_timer = nullptr;
     RuntimeProfile::Counter* _open_timer = nullptr;
     RuntimeProfile::Counter* _close_timer = nullptr;
@@ -371,7 +371,7 @@ class PipelineXSinkLocalStateBase {
     RuntimeProfile::Counter* _wait_for_finish_dependency_timer = nullptr;
     RuntimeProfile::Counter* _exec_timer = nullptr;
     RuntimeProfile::Counter* _memory_used_counter = nullptr;
-    RuntimeProfile::Counter* _peak_memory_usage_counter = nullptr;
+    RuntimeProfile::HighWaterMarkCounter* _peak_memory_usage_counter = nullptr;
 
     std::shared_ptr<QueryStatistics> _query_statistics = nullptr;
 };

be/src/pipeline/exec/scan_operator.cpp

@@ -1052,17 +1052,13 @@ Status ScanLocalState<Derived>::_init_profile() {
     _total_throughput_counter =
             profile()->add_rate_counter("TotalReadThroughput", _rows_read_counter);
     _num_scanners = ADD_COUNTER(_runtime_profile, "NumScanners", TUnit::UNIT);
+    _scanner_peak_memory_usage = _peak_memory_usage_counter;
+    //_runtime_profile->AddHighWaterMarkCounter("PeakMemoryUsage", TUnit::BYTES);
 
     // 2. counters for scanners
     _scanner_profile.reset(new RuntimeProfile("VScanner"));
     profile()->add_child(_scanner_profile.get(), true, nullptr);
 
-    _memory_usage_counter = ADD_LABEL_COUNTER_WITH_LEVEL(_scanner_profile, "MemoryUsage", 1);
-    _free_blocks_memory_usage =
-            _scanner_profile->AddHighWaterMarkCounter("FreeBlocks", TUnit::BYTES, "MemoryUsage", 1);
-    _scanner_peak_memory_usage =
-            _scanner_profile->AddHighWaterMarkCounter("PeakMemoryUsage", TUnit::BYTES);
-
     _newly_create_free_blocks_num =
             ADD_COUNTER(_scanner_profile, "NewlyCreateFreeBlocksNum", TUnit::UNIT);
     _scale_up_scanners_counter = ADD_COUNTER(_scanner_profile, "NumScaleUpScanners", TUnit::UNIT);

be/src/pipeline/exec/scan_operator.h

@@ -119,7 +119,6 @@ class ScanLocalStateBase : public PipelineXLocalState<>, public RuntimeFilterCon
     // time of filter output block from scanner
     RuntimeProfile::Counter* _filter_timer = nullptr;
     RuntimeProfile::Counter* _memory_usage_counter = nullptr;
-    RuntimeProfile::HighWaterMarkCounter* _free_blocks_memory_usage = nullptr;
     RuntimeProfile::Counter* _scale_up_scanners_counter = nullptr;
     // rows read from the scanner (including those discarded by (pre)filters)
     RuntimeProfile::Counter* _rows_read_counter = nullptr;

be/src/vec/exec/scan/scanner_context.cpp

@@ -133,6 +133,7 @@ Status ScannerContext::init() {
     _scanner_wait_batch_timer = _local_state->_scanner_wait_batch_timer;
     _scanner_ctx_sched_time = _local_state->_scanner_ctx_sched_time;
     _scale_up_scanners_counter = _local_state->_scale_up_scanners_counter;
+    _scanner_memory_used_counter = _local_state->_memory_used_counter;
 
 #ifndef BE_TEST
     // 3. get thread token
@@ -172,6 +173,7 @@ vectorized::BlockUPtr ScannerContext::get_free_block(bool force) {
     if (_free_blocks.try_dequeue(block)) {
         DCHECK(block->mem_reuse());
         _block_memory_usage -= block->allocated_bytes();
+        _scanner_memory_used_counter->set(_block_memory_usage);
         // A free block is reused, so the memory usage should be decreased
         // The caller of get_free_block will increase the memory usage
         update_peak_memory_usage(-block->allocated_bytes());
@@ -187,6 +189,7 @@ void ScannerContext::return_free_block(vectorized::BlockUPtr block) {
     if (block->mem_reuse() && _block_memory_usage < _max_bytes_in_queue) {
         size_t block_size_to_reuse = block->allocated_bytes();
         _block_memory_usage += block_size_to_reuse;
+        _scanner_memory_used_counter->set(_block_memory_usage);
         block->clear_column_data();
         if (_free_blocks.enqueue(std::move(block))) {
             update_peak_memory_usage(block_size_to_reuse);

be/src/vec/exec/scan/scanner_context.h

@@ -225,6 +225,8 @@ class ScannerContext : public std::enable_shared_from_this<ScannerContext>,
     const int _num_parallel_instances;
     std::shared_ptr<RuntimeProfile> _scanner_profile;
     RuntimeProfile::Counter* _scanner_sched_counter = nullptr;
+    // This counter refers to scan operator's local state
+    RuntimeProfile::Counter* _scanner_memory_used_counter = nullptr;
     RuntimeProfile::Counter* _newly_create_free_blocks_num = nullptr;
     RuntimeProfile::Counter* _scanner_wait_batch_timer = nullptr;
     RuntimeProfile::Counter* _scanner_ctx_sched_time = nullptr;