[Pipeline](exec) Support shared scan in colo agg (apache#18457)

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

@@ -34,6 +34,7 @@ NewOlapScanNode::NewOlapScanNode(ObjectPool* pool, const TPlanNode& tnode,
                                  const DescriptorTbl& descs)
         : VScanNode(pool, tnode, descs), _olap_scan_node(tnode.olap_scan_node) {
     _output_tuple_id = tnode.olap_scan_node.tuple_id;
+    _col_distribute_ids = tnode.olap_scan_node.distribute_column_ids;
     if (_olap_scan_node.__isset.sort_info && _olap_scan_node.__isset.sort_limit) {
         _limit_per_scanner = _olap_scan_node.sort_limit;
     }

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

@@ -29,9 +29,11 @@ class PipScannerContext : public vectorized::ScannerContext {
                       const TupleDescriptor* input_tuple_desc,
                       const TupleDescriptor* output_tuple_desc,
                       const std::list<vectorized::VScanner*>& scanners, int64_t limit,
-                      int64_t max_bytes_in_blocks_queue)
+                      int64_t max_bytes_in_blocks_queue, const std::vector<int>& col_distribute_ids)
             : vectorized::ScannerContext(state, parent, input_tuple_desc, output_tuple_desc,
-                                         scanners, limit, max_bytes_in_blocks_queue) {}
+                                         scanners, limit, max_bytes_in_blocks_queue),
+              _col_distribute_ids(col_distribute_ids),
+              _need_colocate_distribute(!_col_distribute_ids.empty()) {}
 
     Status get_block_from_queue(RuntimeState* state, vectorized::BlockUPtr* block, bool* eos,
                                 int id, bool wait = false) override {
@@ -67,19 +69,54 @@ class PipScannerContext : public vectorized::ScannerContext {
         const int queue_size = _queue_mutexs.size();
         const int block_size = blocks.size();
         int64_t local_bytes = 0;
-        for (const auto& block : blocks) {
-            local_bytes += block->allocated_bytes();
-        }
 
-        for (int i = 0; i < queue_size && i < block_size; ++i) {
-            int queue = _next_queue_to_feed;
-            {
-                std::lock_guard<std::mutex> l(*_queue_mutexs[queue]);
-                for (int j = i; j < block_size; j += queue_size) {
-                    _blocks_queues[queue].emplace_back(std::move(blocks[j]));
+        if (_need_colocate_distribute) {
+            std::vector<uint64_t> hash_vals;
+            for (const auto& block : blocks) {
+                // vectorized calculate hash
+                int rows = block->rows();
+                const auto element_size = _max_queue_size;
+                hash_vals.resize(rows);
+                std::fill(hash_vals.begin(), hash_vals.end(), 0);
+                auto* __restrict hashes = hash_vals.data();
+
+                for (int j = 0; j < _col_distribute_ids.size(); ++j) {
+                    block->get_by_position(_col_distribute_ids[j])
+                            .column->update_crcs_with_value(
+                                    hash_vals, _output_tuple_desc->slots()[_col_distribute_ids[j]]
+                                                       ->type()
+                                                       .type);
+                }
+                for (int i = 0; i < rows; i++) {
+                    hashes[i] = hashes[i] % element_size;
+                }
+
+                std::vector<int> channel2rows[element_size];
+                for (int i = 0; i < rows; i++) {
+                    channel2rows[hashes[i]].emplace_back(i);
+                }
+
+                for (int i = 0; i < element_size; ++i) {
+                    if (!channel2rows[i].empty()) {
+                        _add_rows_colocate_blocks(block.get(), i, channel2rows[i]);
+                    }
                 }
             }
-            _next_queue_to_feed = queue + 1 < queue_size ? queue + 1 : 0;
+        } else {
+            for (const auto& block : blocks) {
+                local_bytes += block->allocated_bytes();
+            }
+
+            for (int i = 0; i < queue_size && i < block_size; ++i) {
+                int queue = _next_queue_to_feed;
+                {
+                    std::lock_guard<std::mutex> l(*_queue_mutexs[queue]);
+                    for (int j = i; j < block_size; j += queue_size) {
+                        _blocks_queues[queue].emplace_back(std::move(blocks[j]));
+                    }
+                }
+                _next_queue_to_feed = queue + 1 < queue_size ? queue + 1 : 0;
+            }
         }
         _current_used_bytes += local_bytes;
     }
@@ -95,18 +132,86 @@ class PipScannerContext : public vectorized::ScannerContext {
             _queue_mutexs.emplace_back(new std::mutex);
             _blocks_queues.emplace_back(std::list<vectorized::BlockUPtr>());
         }
+        if (_need_colocate_distribute) {
+            int real_block_size =
+                    limit == -1 ? _batch_size : std::min(static_cast<int64_t>(_batch_size), limit);
+            int64_t free_blocks_memory_usage = 0;
+            for (int i = 0; i < _max_queue_size; ++i) {
+                auto block = std::make_unique<vectorized::Block>(_output_tuple_desc->slots(),
+                                                                 real_block_size,
+                                                                 true /*ignore invalid slots*/);
+                free_blocks_memory_usage += block->allocated_bytes();
+                _colocate_mutable_blocks.emplace_back(new vectorized::MutableBlock(block.get()));
+                _colocate_blocks.emplace_back(std::move(block));
+                _colocate_block_mutexs.emplace_back(new std::mutex);
+            }
+            _free_blocks_memory_usage->add(free_blocks_memory_usage);
+        }
     }
 
     bool has_enough_space_in_blocks_queue() const override {
         return _current_used_bytes < _max_bytes_in_queue / 2 * _max_queue_size;
     }
 
+    virtual void _dispose_coloate_blocks_not_in_queue() override {
+        if (_need_colocate_distribute) {
+            for (int i = 0; i < _max_queue_size; ++i) {
+                std::scoped_lock s(*_colocate_block_mutexs[i], *_queue_mutexs[i]);
+                if (_colocate_blocks[i] && !_colocate_blocks[i]->empty()) {
+                    _current_used_bytes += _colocate_blocks[i]->allocated_bytes();
+                    _blocks_queues[i].emplace_back(std::move(_colocate_blocks[i]));
+                    _colocate_mutable_blocks[i]->clear();
+                }
+            }
+        }
+    }
+
 private:
     int _max_queue_size = 1;
     int _next_queue_to_feed = 0;
     std::vector<std::unique_ptr<std::mutex>> _queue_mutexs;
     std::vector<std::list<vectorized::BlockUPtr>> _blocks_queues;
     std::atomic_int64_t _current_used_bytes = 0;
+
+    const std::vector<int>& _col_distribute_ids;
+    const bool _need_colocate_distribute;
+    std::vector<vectorized::BlockUPtr> _colocate_blocks;
+    std::vector<std::unique_ptr<vectorized::MutableBlock>> _colocate_mutable_blocks;
+    std::vector<std::unique_ptr<std::mutex>> _colocate_block_mutexs;
+
+    void _add_rows_colocate_blocks(vectorized::Block* block, int loc,
+                                   const std::vector<int>& rows) {
+        int row_wait_add = rows.size();
+        const int batch_size = _batch_size;
+        const int* begin = &rows[0];
+        std::lock_guard<std::mutex> l(*_colocate_block_mutexs[loc]);
+
+        while (row_wait_add > 0) {
+            int row_add = 0;
+            int max_add = batch_size - _colocate_mutable_blocks[loc]->rows();
+            if (row_wait_add >= max_add) {
+                row_add = max_add;
+            } else {
+                row_add = row_wait_add;
+            }
+
+            _colocate_mutable_blocks[loc]->add_rows(block, begin, begin + row_add);
+            row_wait_add -= row_add;
+            begin += row_add;
+
+            if (row_add == max_add) {
+                _current_used_bytes += _colocate_blocks[loc]->allocated_bytes();
+                {
+                    std::lock_guard<std::mutex> queue_l(*_queue_mutexs[loc]);
+                    _blocks_queues[loc].emplace_back(std::move(_colocate_blocks[loc]));
+                }
+                bool get_block_not_empty = true;
+                _colocate_blocks[loc] = get_free_block(&get_block_not_empty, get_block_not_empty);
+                _colocate_mutable_blocks[loc]->set_muatable_columns(
+                        _colocate_blocks[loc]->mutate_columns());
+            }
+        }
+    }
 };
 } // namespace pipeline
 } // namespace doris

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

@@ -114,14 +114,17 @@ Status ScannerContext::init() {
     return Status::OK();
 }
 
-vectorized::BlockUPtr ScannerContext::get_free_block(bool* has_free_block) {
+vectorized::BlockUPtr ScannerContext::get_free_block(bool* has_free_block,
+                                                     bool get_block_not_empty) {
     {
         std::lock_guard l(_free_blocks_lock);
         if (!_free_blocks.empty()) {
-            auto block = std::move(_free_blocks.back());
-            _free_blocks.pop_back();
-            _free_blocks_memory_usage->add(-block->allocated_bytes());
-            return block;
+            if (!get_block_not_empty || _free_blocks.back()->mem_reuse()) {
+                auto block = std::move(_free_blocks.back());
+                _free_blocks.pop_back();
+                _free_blocks_memory_usage->add(-block->allocated_bytes());
+                return block;
+            }
         }
     }
     *has_free_block = false;
@@ -319,6 +322,7 @@ void ScannerContext::push_back_scanner_and_reschedule(VScanner* scanner) {
     // same scanner.
     if (scanner->need_to_close() && scanner->set_counted_down() &&
         (--_num_unfinished_scanners) == 0) {
+        _dispose_coloate_blocks_not_in_queue();
         _is_finished = true;
         _blocks_queue_added_cv.notify_one();
     }

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

@@ -55,7 +55,7 @@ class ScannerContext {
     virtual ~ScannerContext() = default;
     Status init();
 
-    vectorized::BlockUPtr get_free_block(bool* has_free_block);
+    vectorized::BlockUPtr get_free_block(bool* has_free_block, bool get_not_empty_block = false);
     void return_free_block(std::unique_ptr<vectorized::Block> block);
 
     // Append blocks from scanners to the blocks queue.
@@ -140,6 +140,8 @@ class ScannerContext {
     Status _close_and_clear_scanners(VScanNode* node, RuntimeState* state);
 
 protected:
+    virtual void _dispose_coloate_blocks_not_in_queue() {}
+
     RuntimeState* _state;
     VScanNode* _parent;
 

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

@@ -261,9 +261,9 @@ Status VScanNode::_init_profile() {
 
 Status VScanNode::_start_scanners(const std::list<VScanner*>& scanners) {
     if (_is_pipeline_scan) {
-        _scanner_ctx.reset(new pipeline::PipScannerContext(_state, this, _input_tuple_desc,
-                                                           _output_tuple_desc, scanners, limit(),
-                                                           _state->query_options().mem_limit / 20));
+        _scanner_ctx.reset(new pipeline::PipScannerContext(
+                _state, this, _input_tuple_desc, _output_tuple_desc, scanners, limit(),
+                _state->query_options().mem_limit / 20, _col_distribute_ids));
     } else {
         _scanner_ctx.reset(new ScannerContext(_state, this, _input_tuple_desc, _output_tuple_desc,
                                               scanners, limit(),

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

@@ -312,6 +312,7 @@ class VScanNode : public ExecNode {
     RuntimeProfile::HighWaterMarkCounter* _free_blocks_memory_usage;
 
     std::unordered_map<std::string, int> _colname_to_slot_id;
+    std::vector<int> _col_distribute_ids;
 
 private:
     // Register and get all runtime filters at Init phase.

fe/fe-core/src/main/java/org/apache/doris/planner/DistributedPlanner.java

@@ -932,9 +932,11 @@ private PlanFragment createAggregationFragment(
             return createPhase2DistinctAggregationFragment(node, childFragment, fragments);
         } else {
             if (canColocateAgg(node.getAggInfo(), childFragment.getDataPartition())
-                    && childFragment.getPlanRoot().shouldColoAgg()) {
+                    || childFragment.getPlanRoot().shouldColoAgg(node.getAggInfo())) {
+                childFragment.getPlanRoot().setShouldColoScan();
                 childFragment.addPlanRoot(node);
-                childFragment.setHasColocatePlanNode(true);
+                // pipeline here should use shared scan to improve performance
+                childFragment.setHasColocatePlanNode(!ConnectContext.get().getSessionVariable().enablePipelineEngine());
                 return childFragment;
             } else {
                 return createMergeAggregationFragment(node, childFragment);

fe/fe-core/src/main/java/org/apache/doris/planner/OlapScanNode.java

@@ -17,6 +17,7 @@
 
 package org.apache.doris.planner;
 
+import org.apache.doris.analysis.AggregateInfo;
 import org.apache.doris.analysis.Analyzer;
 import org.apache.doris.analysis.BaseTableRef;
 import org.apache.doris.analysis.BinaryPredicate;
@@ -185,15 +186,28 @@ public class OlapScanNode extends ScanNode {
     private Map<SlotRef, Expr> pointQueryEqualPredicats;
     private DescriptorTable descTable;
 
+    private Set<Integer> distributionColumnIds;
+
+    private boolean shouldColoScan = false;
+
     // Constructs node to scan given data files of table 'tbl'.
     public OlapScanNode(PlanNodeId id, TupleDescriptor desc, String planNodeName) {
         super(id, desc, planNodeName, StatisticalType.OLAP_SCAN_NODE);
         olapTable = (OlapTable) desc.getTable();
+        distributionColumnIds = Sets.newTreeSet();
+
+        Set<String> distColumnName = olapTable != null
+                ? olapTable.getDistributionColumnNames() : Sets.newTreeSet();
+        int columnId = 0;
         // use for Nereids to generate uniqueId set for inverted index to avoid scan unnecessary big size column
         for (SlotDescriptor slotDescriptor : desc.getSlots()) {
             if (slotDescriptor.getColumn() != null) {
                 outputColumnUniqueIds.add(slotDescriptor.getColumn().getUniqueId());
+                if (distColumnName.contains(slotDescriptor.getColumn().getName().toLowerCase())) {
+                    distributionColumnIds.add(columnId);
+                }
             }
+            columnId++;
         }
     }
 
@@ -1166,17 +1180,48 @@ public int getNumInstances() {
     }
 
     @Override
-    public boolean shouldColoAgg() {
-        // In pipeline exec engine, the instance num is parallel instance. we should disable colo agg
-        // in parallelInstance >= tablet_num * 2 to use more thread to speed up the query
-        if (ConnectContext.get().getSessionVariable().enablePipelineEngine()) {
-            int parallelInstance = ConnectContext.get().getSessionVariable().getParallelExecInstanceNum();
-            return parallelInstance < result.size() * 2;
+    public boolean shouldColoAgg(AggregateInfo aggregateInfo) {
+        distributionColumnIds.clear();
+        if (ConnectContext.get().getSessionVariable().enablePipelineEngine()
+                && ConnectContext.get().getSessionVariable().enableColocateScan()) {
+            List<Expr> aggPartitionExprs = aggregateInfo.getInputPartitionExprs();
+            List<SlotDescriptor> slots = desc.getSlots();
+            for (Expr aggExpr : aggPartitionExprs) {
+                if (aggExpr instanceof SlotRef) {
+                    SlotDescriptor slotDesc = ((SlotRef) aggExpr).getDesc();
+                    int columnId = 0;
+                    for (SlotDescriptor slotDescriptor : slots) {
+                        if (slotDescriptor.equals(slotDesc)) {
+                            if (slotDescriptor.getType().isFixedLengthType()
+                                    || slotDescriptor.getType().isStringType()) {
+                                distributionColumnIds.add(columnId);
+                            } else {
+                                return false;
+                            }
+                        }
+                        columnId++;
+                    }
+                }
+            }
+
+            for (int i = 0; i < slots.size(); i++) {
+                if (!distributionColumnIds.contains(i) && (!slots.get(i).getType().isFixedLengthType()
+                        || slots.get(i).getType().isStringType())) {
+                    return false;
+                }
+            }
+
+            return !distributionColumnIds.isEmpty();
         } else {
-            return true;
+            return false;
         }
     }
 
+    @Override
+    public void setShouldColoScan() {
+        shouldColoScan = true;
+    }
+
     @Override
     protected void toThrift(TPlanNode msg) {
         List<String> keyColumnNames = new ArrayList<String>();
@@ -1233,6 +1278,10 @@ protected void toThrift(TPlanNode msg) {
         if (outputColumnUniqueIds != null) {
             msg.olap_scan_node.setOutputColumnUniqueIds(outputColumnUniqueIds);
         }
+
+        if (shouldColoScan) {
+            msg.olap_scan_node.setDistributeColumnIds(new ArrayList<>(distributionColumnIds));
+        }
     }
 
     // export some tablets

fe/fe-core/src/main/java/org/apache/doris/planner/PlanNode.java

@@ -20,6 +20,7 @@
 
 package org.apache.doris.planner;
 
+import org.apache.doris.analysis.AggregateInfo;
 import org.apache.doris.analysis.Analyzer;
 import org.apache.doris.analysis.BitmapFilterPredicate;
 import org.apache.doris.analysis.CompoundPredicate;
@@ -835,10 +836,12 @@ public int getNumInstances() {
         return numInstances;
     }
 
-    public boolean shouldColoAgg() {
-        return true;
+    public boolean shouldColoAgg(AggregateInfo aggregateInfo) {
+        return false;
     }
 
+    public void setShouldColoScan() {}
+
     public void setNumInstances(int numInstances) {
         this.numInstances = numInstances;
     }