Added allocatesMemoryInArena() method for aggregate functions.

Fixed runningAccumulate, now it works properly for complex functions.
More accurate threads handling in Aggregator.

dbms/include/DB/AggregateFunctions/AggregateFunctionGroupUniqArray.h

@@ -104,7 +104,10 @@ public:
 /// Generic implementation, it uses serialized representation as object descriptor.
 struct AggreagteFunctionGroupUniqArrayGenericData
 {
-	using Set = HashSetWithSavedHash<StringRef, StringRefHash, HashTableGrower<4>, HashTableAllocatorWithStackMemory<16>>;
+	static constexpr size_t INIT_ELEMS = 2; /// adjustable
+	static constexpr size_t ELEM_SIZE = sizeof(HashSetCellWithSavedHash<StringRef, StringRefHash>);
+	using Set = HashSetWithSavedHash<StringRef, StringRefHash, HashTableGrower<INIT_ELEMS>, HashTableAllocatorWithStackMemory<INIT_ELEMS * ELEM_SIZE>>;
+
 	Set value;
 };
 
@@ -136,6 +139,11 @@ public:
 		return std::make_shared<DataTypeArray>(input_data_type->clone());
 	}
 
+	bool allocatesMemoryInArena() const override
+	{
+		return true;
+	}
+
 	void serialize(ConstAggregateDataPtr place, WriteBuffer & buf) const override
 	{
 		auto & set = this->data(place).value;

dbms/include/DB/AggregateFunctions/IAggregateFunction.h

@@ -93,6 +93,12 @@ public:
 	/// Deserializes state. This function is called only for empty (just created) states.
 	virtual void deserialize(AggregateDataPtr place, ReadBuffer & buf, Arena * arena) const = 0;
 
+	/// Returns true if a function requires Arena to handle own states (see add(), merge(), deserialize()).
+	virtual bool allocatesMemoryInArena() const
+	{
+		return false;
+	}
+
 	/// Inserts results into a column.
 	virtual void insertResultInto(ConstAggregateDataPtr place, IColumn & to) const = 0;
 

dbms/include/DB/Columns/ColumnAggregateFunction.h

@@ -159,7 +159,8 @@ public:
 	/// Объединить состояние в последней строке с заданным
 	void insertMergeFrom(const IColumn & src, size_t n)
 	{
-		func->merge(getData().back(), static_cast<const ColumnAggregateFunction &>(src).getData()[n], &createOrGetArena());
+		Arena & arena = createOrGetArena();
+		func->merge(getData().back(), static_cast<const ColumnAggregateFunction &>(src).getData()[n], &arena);
 	}
 
 	Arena & createOrGetArena()

dbms/include/DB/Columns/ColumnString.h

@@ -12,6 +12,7 @@
 #include <DB/Common/SipHash.h>
 #include <DB/Common/memcpySmall.h>
 
+
 namespace DB
 {
 

dbms/include/DB/Functions/FunctionsMiscellaneous.h

@@ -25,6 +25,7 @@
 #include <DB/Common/UnicodeBar.h>
 #include <DB/Functions/IFunction.h>
 #include <DB/Functions/NumberTraits.h>
+#include <DB/Functions/ObjectPool.h>
 #include <DB/Interpreters/ExpressionActions.h>
 #include <ext/range.hpp>
 
@@ -1242,13 +1243,22 @@ public:
 		IColumn & result_column = *result_column_ptr;
 		result_column.reserve(column_with_states->size());
 
+		auto arena = (agg_func.allocatesMemoryInArena()) ?
+						arenas_pool.get(0, []{ return new Arena(); }) :
+						nullptr;
+
 		const auto & states = column_with_states->getData();
 		for (const auto & state_to_add : states)
 		{
-			agg_func.merge(place.get(), state_to_add, nullptr); /// Empty arena!
+			/// Will pass empty arena if agg_func does not allocate memory in arena
+			agg_func.merge(place.get(), state_to_add, arena.get());
 			agg_func.insertResultInto(place.get(), result_column);
 		}
 	}
+
+private:
+
+	ObjectPool<Arena, int> arenas_pool; /// Used only for complex functions
 };
 
 

dbms/include/DB/Functions/ObjectPool.h

@@ -1,3 +1,4 @@
+#pragma once
 #include <map>
 #include <memory>
 #include <stack>

dbms/include/DB/Interpreters/Aggregator.h

@@ -384,7 +384,7 @@ struct AggregationMethodConcat
 		{
 			/** Исправление, если все ключи - пустые массивы. Для них в хэш-таблицу записывается StringRef нулевой длины, но с ненулевым указателем.
 			  * Но при вставке в хэш-таблицу, такой StringRef оказывается равен другому ключу нулевой длины,
-			  * у которого указатель на данные может быть любым мусором и использовать его нельзя.
+			  *  у которого указатель на данные может быть любым мусором и использовать его нельзя.
 			  */
 			for (size_t i = 0; i < keys_size; ++i)
 				key_columns[i]->insertDefault();
@@ -1174,7 +1174,7 @@ protected:
 
 	template <typename Method>
 	void mergeBucketImpl(
-		ManyAggregatedDataVariants & data, Int32 bucket) const;
+		ManyAggregatedDataVariants & data, Int32 bucket, Arena * arena) const;
 
 	template <typename Method>
 	void convertBlockToTwoLevelImpl(

dbms/src/Interpreters/Aggregator.cpp

@@ -1480,7 +1480,7 @@ void NO_INLINE Aggregator::mergeSingleLevelDataImpl(
 
 template <typename Method>
 void NO_INLINE Aggregator::mergeBucketImpl(
-	ManyAggregatedDataVariants & data, Int32 bucket) const
+	ManyAggregatedDataVariants & data, Int32 bucket, Arena * arena) const
 {
 	/// Все результаты агрегации соединяем с первым.
 	AggregatedDataVariantsPtr & res = data[0];
@@ -1488,9 +1488,6 @@ void NO_INLINE Aggregator::mergeBucketImpl(
 	{
 		AggregatedDataVariants & current = *data[i];
 
-		/// Select Arena to avoid race conditions
-		Arena * arena = res->aggregates_pools.at(static_cast<size_t>(bucket) % size).get();
-
 		mergeDataImpl<Method>(
 			getDataVariant<Method>(*res).data.impls[bucket],
 			getDataVariant<Method>(current).data.impls[bucket],
@@ -1511,7 +1508,16 @@ public:
 	  *  которые все либо являются одноуровневыми, либо являются двухуровневыми.
 	  */
 	MergingAndConvertingBlockInputStream(const Aggregator & aggregator_, ManyAggregatedDataVariants & data_, bool final_, size_t threads_)
-		: aggregator(aggregator_), data(data_), final(final_), threads(threads_) {}
+		: aggregator(aggregator_), data(data_), final(final_), threads(threads_)
+	{
+		/// At least we need one arena in first data item per thread
+		if (!data.empty() && threads > data[0]->aggregates_pools.size())
+		{
+			Arenas & first_pool = data[0]->aggregates_pools;
+			for (size_t j = first_pool.size(); j < threads; j++)
+				first_pool.emplace_back(std::make_shared<Arena>());
+		}
+	}
 
 	String getName() const override { return "MergingAndConverting"; }
 
@@ -1653,17 +1659,21 @@ private:
 
 		try
 		{
-			/// TODO Возможно, поддержать no_more_keys
+			/// TODO: add no_more_keys support maybe
 
 			auto & merged_data = *data[0];
 			auto method = merged_data.type;
 			Block block;
 
+			/// Select Arena to avoid race conditions
+			size_t thread_number = static_cast<size_t>(bucket_num) % threads;
+			Arena * arena = merged_data.aggregates_pools.at(thread_number).get();
+
 			if (false) {}
 		#define M(NAME) \
 			else if (method == AggregatedDataVariants::Type::NAME) \
 			{ \
-				aggregator.mergeBucketImpl<decltype(merged_data.NAME)::element_type>(data, bucket_num); \
+				aggregator.mergeBucketImpl<decltype(merged_data.NAME)::element_type>(data, bucket_num, arena); \
 				block = aggregator.convertOneBucketToBlock(merged_data, *merged_data.NAME, final, bucket_num); \
 			}