auto pass through hashagg

dbms/src/Common/HashTable/HashTableKeyHolder.h

@@ -27,7 +27,7 @@
   * make a persistent copy of the key in each of the following cases:
   * 1) the aggregation method doesn't use temporary keys, so they're persistent
   *    from the start;
-  * 1) the key is already present in the hash table;
+  * 2) the key is already present in the hash table;
   * 3) that particular key is stored by value, e.g. a short StringRef key in
   *    StringHashMap.
   *

dbms/src/DataStreams/AutoPassThroughAggregatingBlockInputStream.cpp

@@ -0,0 +1,47 @@
+// Copyright 2024 PingCAP, Inc.
+//
+// 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.
+
+#include <DataStreams/AutoPassThroughAggregatingBlockInputStream.h>
+
+namespace DB
+{
+
+template <bool force_streaming>
+Block AutoPassThroughAggregatingBlockInputStream<force_streaming>::readImpl()
+{
+    while (!build_done)
+    {
+        Block block = children[0]->read();
+        if (block)
+        {
+            auto_pass_through_context->onBlock<force_streaming>(block);
+        }
+        else
+        {
+            build_done = true;
+            break;
+        }
+
+        if (!auto_pass_through_context->passThroughBufferEmpty())
+            return auto_pass_through_context->popPassThroughBuffer();
+    }
+
+    assert(build_done);
+    if (!auto_pass_through_context->passThroughBufferEmpty())
+        return auto_pass_through_context->popPassThroughBuffer();
+
+    return auto_pass_through_context->getData();
+}
+
+} // namespace DB

dbms/src/DataStreams/AutoPassThroughAggregatingBlockInputStream.h

@@ -0,0 +1,60 @@
+// Copyright 2024 PingCAP, Inc.
+//
+// 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 <DataStreams/IProfilingBlockInputStream.h>
+#include <Interpreters/Aggregator.h>
+#include <Operators/AutoPassThroughHashAggContext.h>
+
+namespace DB
+{
+static constexpr std::string_view autoPassThroughAggregatingExtraInfo = "auto pass through";
+
+template <bool force_streaming>
+class AutoPassThroughAggregatingBlockInputStream : public IProfilingBlockInputStream
+{
+    static constexpr auto NAME = "Aggregating";
+
+public:
+    AutoPassThroughAggregatingBlockInputStream(
+        const BlockInputStreamPtr & input_,
+        const Aggregator::Params & params_,
+        const String & req_id,
+        UInt64 row_limit_unit)
+    {
+        children.push_back(input_);
+        auto_pass_through_context = std::make_unique<AutoPassThroughHashAggContext>(
+            children[0]->getHeader(),
+            params_,
+            [&]() { return this->isCancelled(); },
+            req_id,
+            row_limit_unit);
+    }
+
+    String getName() const override { return NAME; }
+
+    Block getHeader() const override { return auto_pass_through_context->getHeader(); }
+
+protected:
+    Block readImpl() override;
+
+private:
+    AutoPassThroughHashAggContextPtr auto_pass_through_context;
+    bool build_done = false;
+};
+
+template class AutoPassThroughAggregatingBlockInputStream<true>;
+template class AutoPassThroughAggregatingBlockInputStream<false>;
+} // namespace DB

dbms/src/DataStreams/ExchangeSenderBlockInputStream.h

@@ -38,6 +38,8 @@ class ExchangeSenderBlockInputStream : public IProfilingBlockInputStream
     String getName() const override { return name; }
     Block getHeader() const override { return children.back()->getHeader(); }
 
+    bool canHandleSelectiveBlock() const override { return true; }
+
 protected:
     Block readImpl() override;
     void readPrefixImpl() override { writer->prepare(getHeader()); }

dbms/src/DataStreams/ExpressionBlockInputStream.cpp

@@ -39,7 +39,22 @@ Block ExpressionBlockInputStream::readImpl()
     Block res = children.back()->read();
     if (!res)
         return res;
-    expression->execute(res);
+
+    if (res.info.selective)
+    {
+        const auto ori_rows = res.rows();
+        auto ori_info = res.info;
+        expression->execute(res);
+        res.info = ori_info;
+        // When block.info.selective is not null, the expression action should be cast/project.
+        // So the rows should not change.
+        RUNTIME_CHECK(ori_rows == res.rows());
+    }
+    else
+    {
+        expression->execute(res);
+    }
     return res;
 }
+
 } // namespace DB

dbms/src/DataStreams/ExpressionBlockInputStream.h

@@ -41,6 +41,8 @@ class ExpressionBlockInputStream : public IProfilingBlockInputStream
     String getName() const override { return NAME; }
     Block getHeader() const override;
 
+    bool canHandleSelectiveBlock() const override { return true; }
+
 protected:
     Block readImpl() override;
 

dbms/src/DataStreams/IBlockInputStream.h

@@ -188,6 +188,8 @@ class IBlockInputStream : private boost::noncopyable
 
     virtual void appendInfo(FmtBuffer & /*buffer*/) const {};
 
+    virtual bool canHandleSelectiveBlock() const { return false; }
+
 protected:
     virtual uint64_t collectCPUTimeNsImpl(bool /*is_thread_runner*/) { return 0; }
 

dbms/src/DataStreams/IProfilingBlockInputStream.cpp

@@ -95,6 +95,15 @@ Block IProfilingBlockInputStream::read(FilterPtr & res_filter, bool return_filte
 
         if (quota != nullptr)
             checkQuota(res);
+
+        // Check if child can handle selective block.
+        for (const auto & child : children)
+        {
+            RUNTIME_CHECK_MSG(
+                child->canHandleSelectiveBlock() || !res.info.selective,
+                "{} cannot handle selective block",
+                getName());
+        }
     }
     else
     {
@@ -228,7 +237,7 @@ static bool handleOverflowMode(OverflowMode mode, const String & message, int co
     default:
         throw Exception("Logical error: unknown overflow mode", ErrorCodes::LOGICAL_ERROR);
     }
-};
+}
 
 
 bool IProfilingBlockInputStream::checkTimeLimit() const

dbms/src/DataStreams/ParallelAggregatingBlockInputStream.cpp

@@ -66,7 +66,6 @@ void ParallelAggregatingBlockInputStream::cancel(bool kill)
         processor.cancel(kill);
 }
 
-
 Block ParallelAggregatingBlockInputStream::readImpl()
 {
     if (!executed)

dbms/src/Debug/MockExecutor/AggregationBinder.cpp

@@ -34,6 +34,8 @@ bool AggregationBinder::toTiPBExecutor(
     tipb_executor->set_executor_id(name);
     tipb_executor->set_fine_grained_shuffle_stream_count(fine_grained_shuffle_stream_count);
     auto * agg = tipb_executor->mutable_aggregation();
+    if (switcher)
+        agg->set_pre_agg_mode(switcher->mode);
     buildAggExpr(agg, collator_id, context);
     buildGroupBy(agg, collator_id, context);
     auto * child_executor = agg->mutable_child();
@@ -87,6 +89,7 @@ void AggregationBinder::toMPPSubPlan(
     // todo support avg
     if (has_uniq_raw_res)
         throw Exception("uniq raw res not supported in mpp query");
+    // Partial agg cannot be fine grained shuffle. So set fine_grained_shuffle_stream_count as 0.
     std::shared_ptr<AggregationBinder> partial_agg = std::make_shared<AggregationBinder>(
         executor_index,
         output_schema_for_partial_agg,
@@ -95,7 +98,8 @@ void AggregationBinder::toMPPSubPlan(
         std::move(agg_exprs),
         std::move(gby_exprs),
         false,
-        fine_grained_shuffle_stream_count);
+        /*fine_grained_shuffle_stream_count*/ 0,
+        switcher);
     partial_agg->children.push_back(children[0]);
     std::vector<size_t> partition_keys;
     size_t agg_func_num = partial_agg->agg_exprs.size();
@@ -136,6 +140,8 @@ void AggregationBinder::toMPPSubPlan(
         gby_exprs.push_back(std::make_shared<ASTIdentifier>(output_schema_for_partial_agg[agg_func_num + i].first));
     }
     children[0] = exchange_receiver;
+    // Because this aggregation is 2nd agg, so reset auto_pass_through flag.
+    switcher = nullptr;
 }
 
 bool AggregationBinder::needAppendProject() const
@@ -235,7 +241,8 @@ ExecutorBinderPtr compileAggregation(
     size_t & executor_index,
     ASTPtr agg_funcs,
     ASTPtr group_by_exprs,
-    uint64_t fine_grained_shuffle_stream_count)
+    uint64_t fine_grained_shuffle_stream_count,
+    std::shared_ptr<AutoPassThroughSwitcher> switcher)
 {
     std::vector<ASTPtr> agg_exprs;
     std::vector<ASTPtr> gby_exprs;
@@ -308,7 +315,8 @@ ExecutorBinderPtr compileAggregation(
         std::move(agg_exprs),
         std::move(gby_exprs),
         true,
-        fine_grained_shuffle_stream_count);
+        fine_grained_shuffle_stream_count,
+        switcher);
     aggregation->children.push_back(input);
     return aggregation;
 }

dbms/src/Debug/MockExecutor/AggregationBinder.h

@@ -15,6 +15,7 @@
 #pragma once
 
 #include <Debug/MockExecutor/ExecutorBinder.h>
+#include <Operators/AutoPassThroughHashAggContext.h>
 #include <Parsers/ASTFunction.h>
 
 namespace DB::mock
@@ -33,14 +34,16 @@ class AggregationBinder : public ExecutorBinder
         ASTs && agg_exprs_,
         ASTs && gby_exprs_,
         bool is_final_mode_,
-        uint64_t fine_grained_shuffle_stream_count_)
+        uint64_t fine_grained_shuffle_stream_count_,
+        std::shared_ptr<AutoPassThroughSwitcher> switcher_)
         : ExecutorBinder(index_, "aggregation_" + std::to_string(index_), output_schema_)
         , has_uniq_raw_res(has_uniq_raw_res_)
         , need_append_project(need_append_project_)
         , agg_exprs(std::move(agg_exprs_))
         , gby_exprs(std::move(gby_exprs_))
         , is_final_mode(is_final_mode_)
         , fine_grained_shuffle_stream_count(fine_grained_shuffle_stream_count_)
+        , switcher(switcher_)
     {}
 
     bool toTiPBExecutor(
@@ -73,6 +76,7 @@ class AggregationBinder : public ExecutorBinder
     bool is_final_mode;
     DAGSchema output_schema_for_partial_agg;
     uint64_t fine_grained_shuffle_stream_count;
+    std::shared_ptr<AutoPassThroughSwitcher> switcher;
 
 private:
     void buildGroupBy(tipb::Aggregation * agg, int32_t collator_id, const Context & context) const;
@@ -85,6 +89,7 @@ ExecutorBinderPtr compileAggregation(
     size_t & executor_index,
     ASTPtr agg_funcs,
     ASTPtr group_by_exprs,
-    uint64_t fine_grained_shuffle_stream_count = 0);
+    uint64_t fine_grained_shuffle_stream_count = 0,
+    std::shared_ptr<AutoPassThroughSwitcher> switcher = nullptr);
 
 } // namespace DB::mock

dbms/src/Flash/Coprocessor/FineGrainedShuffle.h

@@ -23,7 +23,7 @@ static constexpr std::string_view enableFineGrainedShuffleExtraInfo = "enable fi
 
 static constexpr size_t maxFineGrainedStreamCount = 1024;
 
-inline bool enableFineGrainedShuffle(uint64_t stream_count)
+inline bool fineGrainedShuffleEnabled(uint64_t stream_count)
 {
     return stream_count > 0;
 }
@@ -40,7 +40,7 @@ struct FineGrainedShuffle
         , batch_size(executor ? executor->fine_grained_shuffle_batch_size() : 0)
     {}
 
-    bool enable() const { return enableFineGrainedShuffle(stream_count); }
+    bool enabled() const { return fineGrainedShuffleEnabled(stream_count); }
 
     const UInt64 stream_count;
     const UInt64 batch_size;

dbms/src/Flash/Mpp/ExchangeReceiver.cpp

@@ -322,7 +322,7 @@ ExchangeReceiverBase<RPCContext>::ExchangeReceiverBase(
     : exc_log(Logger::get(req_id, executor_id))
     , rpc_context(std::move(rpc_context_))
     , source_num(source_num_)
-    , enable_fine_grained_shuffle_flag(enableFineGrainedShuffle(fine_grained_shuffle_stream_count_))
+    , enable_fine_grained_shuffle_flag(fineGrainedShuffleEnabled(fine_grained_shuffle_stream_count_))
     , output_stream_count(
           enable_fine_grained_shuffle_flag ? std::min(max_streams_, fine_grained_shuffle_stream_count_) : max_streams_)
     , max_buffer_size(getMaxBufferSize(source_num, settings.recv_queue_size))

dbms/src/Flash/Pipeline/Pipeline.cpp

@@ -121,7 +121,7 @@ void Pipeline::addPlanNode(const PhysicalPlanNodePtr & plan_node)
 {
     assert(plan_node);
     /// For fine grained mode, all plan node should enable fine grained shuffle.
-    if (!plan_node->getFineGrainedShuffle().enable())
+    if (!plan_node->getFineGrainedShuffle().enabled())
         is_fine_grained_mode = false;
     plan_nodes.push_back(plan_node);
 }

dbms/src/Flash/Planner/PhysicalPlan.cpp

@@ -93,6 +93,7 @@ void PhysicalPlan::build(const tipb::Executor * executor)
 {
     RUNTIME_CHECK(executor);
     RUNTIME_CHECK(executor->has_executor_id());
+
     const auto & executor_id = executor->executor_id();
     switch (executor->tp())
     {

dbms/src/Flash/Planner/PhysicalPlanNode.cpp

@@ -145,7 +145,7 @@ void PhysicalPlanNode::buildPipeline(
 
 EventPtr PhysicalPlanNode::sinkComplete(PipelineExecutorContext & exec_context)
 {
-    if (getFineGrainedShuffle().enable())
+    if (getFineGrainedShuffle().enabled())
         return nullptr;
     return doSinkComplete(exec_context);
 }