duckdb 0.7.2-dev225.0 → 0.7.2-dev2320.0

package/src/duckdb/src/execution/aggregate_hashtable.cpp

@@ -3,9 +3,10 @@
 #include "duckdb/catalog/catalog_entry/aggregate_function_catalog_entry.hpp"
 #include "duckdb/common/algorithm.hpp"
 #include "duckdb/common/exception.hpp"
+#include "duckdb/common/radix_partitioning.hpp"
 #include "duckdb/common/row_operations/row_operations.hpp"
 #include "duckdb/common/types/null_value.hpp"
-#include "duckdb/common/types/row_data_collection.hpp"
+#include "duckdb/common/types/row/tuple_data_iterator.hpp"
 #include "duckdb/common/vector_operations/unary_executor.hpp"
 #include "duckdb/common/vector_operations/vector_operations.hpp"
 #include "duckdb/execution/expression_executor.hpp"
@@ -16,14 +17,14 @@
 
 namespace duckdb {
 
-using ValidityBytes = RowLayout::ValidityBytes;
+using ValidityBytes = TupleDataLayout::ValidityBytes;
 
 GroupedAggregateHashTable::GroupedAggregateHashTable(ClientContext &context, Allocator &allocator,
                                                      vector<LogicalType> group_types, vector<LogicalType> payload_types,
                                                      const vector<BoundAggregateExpression *> &bindings,
-                                                     HtEntryType entry_type)
+                                                     HtEntryType entry_type, idx_t initial_capacity)
     : GroupedAggregateHashTable(context, allocator, std::move(group_types), std::move(payload_types),
-                                AggregateObject::CreateAggregateObjects(bindings), entry_type) {
+                                AggregateObject::CreateAggregateObjects(bindings), entry_type, initial_capacity) {
 }
 
 GroupedAggregateHashTable::GroupedAggregateHashTable(ClientContext &context, Allocator &allocator,
@@ -31,40 +32,44 @@ GroupedAggregateHashTable::GroupedAggregateHashTable(ClientContext &context, All
     : GroupedAggregateHashTable(context, allocator, std::move(group_types), {}, vector<AggregateObject>()) {
 }
 
+AggregateHTAppendState::AggregateHTAppendState()
+    : ht_offsets(LogicalTypeId::BIGINT), hash_salts(LogicalTypeId::SMALLINT),
+      group_compare_vector(STANDARD_VECTOR_SIZE), no_match_vector(STANDARD_VECTOR_SIZE),
+      empty_vector(STANDARD_VECTOR_SIZE), new_groups(STANDARD_VECTOR_SIZE), addresses(LogicalType::POINTER),
+      chunk_state_initialized(false) {
+}
+
 GroupedAggregateHashTable::GroupedAggregateHashTable(ClientContext &context, Allocator &allocator,
                                                      vector<LogicalType> group_types_p,
                                                      vector<LogicalType> payload_types_p,
                                                      vector<AggregateObject> aggregate_objects_p,
-                                                     HtEntryType entry_type)
+                                                     HtEntryType entry_type, idx_t initial_capacity)
     : BaseAggregateHashTable(context, allocator, aggregate_objects_p, std::move(payload_types_p)),
-      entry_type(entry_type), capacity(0), entries(0), payload_page_offset(0), is_finalized(false),
-      ht_offsets(LogicalTypeId::BIGINT), hash_salts(LogicalTypeId::SMALLINT),
-      group_compare_vector(STANDARD_VECTOR_SIZE), no_match_vector(STANDARD_VECTOR_SIZE),
-      empty_vector(STANDARD_VECTOR_SIZE) {
-
+      entry_type(entry_type), capacity(0), is_finalized(false),
+      aggregate_allocator(make_shared<ArenaAllocator>(allocator)) {
 	// Append hash column to the end and initialise the row layout
 	group_types_p.emplace_back(LogicalType::HASH);
 	layout.Initialize(std::move(group_types_p), std::move(aggregate_objects_p));
+	tuple_size = layout.GetRowWidth();
+	tuples_per_block = Storage::BLOCK_SIZE / tuple_size;
 
 	// HT layout
 	hash_offset = layout.GetOffsets()[layout.ColumnCount() - 1];
+	data_collection = make_uniq<TupleDataCollection>(buffer_manager, layout);
+	data_collection->InitializeAppend(td_pin_state, TupleDataPinProperties::KEEP_EVERYTHING_PINNED);
 
-	tuple_size = layout.GetRowWidth();
-
-	D_ASSERT(tuple_size <= Storage::BLOCK_SIZE);
-	tuples_per_block = Storage::BLOCK_SIZE / tuple_size;
 	hashes_hdl = buffer_manager.Allocate(Storage::BLOCK_SIZE);
 	hashes_hdl_ptr = hashes_hdl.Ptr();
 
 	switch (entry_type) {
 	case HtEntryType::HT_WIDTH_64: {
 		hash_prefix_shift = (HASH_WIDTH - sizeof(aggr_ht_entry_64::salt)) * 8;
-		Resize<aggr_ht_entry_64>(STANDARD_VECTOR_SIZE * 2L);
+		Resize<aggr_ht_entry_64>(initial_capacity);
 		break;
 	}
 	case HtEntryType::HT_WIDTH_32: {
 		hash_prefix_shift = (HASH_WIDTH - sizeof(aggr_ht_entry_32::salt)) * 8;
-		Resize<aggr_ht_entry_32>(STANDARD_VECTOR_SIZE * 2L);
+		Resize<aggr_ht_entry_32>(initial_capacity);
 		break;
 	}
 	default:
@@ -72,44 +77,18 @@ GroupedAggregateHashTable::GroupedAggregateHashTable(ClientContext &context, All
 	}
 
 	predicates.resize(layout.ColumnCount() - 1, ExpressionType::COMPARE_EQUAL);
-	string_heap = make_unique<RowDataCollection>(buffer_manager, (idx_t)Storage::BLOCK_SIZE, 1, true);
 }
 
 GroupedAggregateHashTable::~GroupedAggregateHashTable() {
 	Destroy();
 }
 
-template <class FUNC>
-void GroupedAggregateHashTable::PayloadApply(FUNC fun) {
-	if (entries == 0) {
+void GroupedAggregateHashTable::Destroy() {
+	if (data_collection->Count() == 0) {
 		return;
 	}
-	idx_t apply_entries = entries;
-	idx_t page_nr = 0;
-	idx_t page_offset = 0;
-
-	for (auto &payload_chunk_ptr : payload_hds_ptrs) {
-		auto this_entries = MinValue(tuples_per_block, apply_entries);
-		page_offset = 0;
-		for (data_ptr_t ptr = payload_chunk_ptr, end = payload_chunk_ptr + this_entries * tuple_size; ptr < end;
-		     ptr += tuple_size) {
-			fun(page_nr, page_offset++, ptr);
-		}
-		apply_entries -= this_entries;
-		page_nr++;
-	}
-	D_ASSERT(apply_entries == 0);
-}
-
-void GroupedAggregateHashTable::NewBlock() {
-	auto pin = buffer_manager.Allocate(Storage::BLOCK_SIZE);
-	payload_hds.push_back(std::move(pin));
-	payload_hds_ptrs.push_back(payload_hds.back().Ptr());
-	payload_page_offset = 0;
-}
 
-void GroupedAggregateHashTable::Destroy() {
-	// check if there is a destructor
+	// Check if there is an aggregate with a destructor
 	bool has_destructor = false;
 	for (auto &aggr : layout.GetAggregates()) {
 		if (aggr.function.destructor) {
@@ -119,31 +98,25 @@ void GroupedAggregateHashTable::Destroy() {
 	if (!has_destructor) {
 		return;
 	}
-	// there are aggregates with destructors: loop over the hash table
-	// and call the destructor method for each of the aggregates
-	data_ptr_t data_pointers[STANDARD_VECTOR_SIZE];
-	Vector state_vector(LogicalType::POINTER, (data_ptr_t)data_pointers);
-	idx_t count = 0;
 
-	PayloadApply([&](idx_t page_nr, idx_t page_offset, data_ptr_t ptr) {
-		data_pointers[count++] = ptr;
-		if (count == STANDARD_VECTOR_SIZE) {
-			RowOperations::DestroyStates(layout, state_vector, count);
-			count = 0;
-		}
-	});
-	RowOperations::DestroyStates(layout, state_vector, count);
+	// There are aggregates with destructors: Call the destructor for each of the aggregates
+	RowOperationsState state(aggregate_allocator->GetAllocator());
+	TupleDataChunkIterator iterator(*data_collection, TupleDataPinProperties::DESTROY_AFTER_DONE, false);
+	auto &row_locations = iterator.GetChunkState().row_locations;
+	do {
+		RowOperations::DestroyStates(state, layout, row_locations, iterator.GetCurrentChunkCount());
+	} while (iterator.Next());
+	data_collection->Reset();
 }
 
 template <class ENTRY>
 void GroupedAggregateHashTable::VerifyInternal() {
 	auto hashes_ptr = (ENTRY *)hashes_hdl_ptr;
-	D_ASSERT(payload_hds.size() == payload_hds_ptrs.size());
 	idx_t count = 0;
 	for (idx_t i = 0; i < capacity; i++) {
 		if (hashes_ptr[i].page_nr > 0) {
 			D_ASSERT(hashes_ptr[i].page_offset < tuples_per_block);
-			D_ASSERT(hashes_ptr[i].page_nr <= payload_hds.size());
+			D_ASSERT(hashes_ptr[i].page_nr <= payload_hds_ptrs.size());
 			auto ptr = payload_hds_ptrs[hashes_ptr[i].page_nr - 1] + ((hashes_ptr[i].page_offset) * tuple_size);
 			auto hash = Load<hash_t>(ptr + hash_offset);
 			D_ASSERT((hashes_ptr[i].salt) == (hash >> hash_prefix_shift));
@@ -152,28 +125,37 @@ void GroupedAggregateHashTable::VerifyInternal() {
 		}
 	}
 	(void)count;
-	D_ASSERT(count == entries);
+	D_ASSERT(count == Count());
 }
 
-idx_t GroupedAggregateHashTable::MaxCapacity() {
-	idx_t max_pages = 0;
-	idx_t max_tuples = 0;
+idx_t GroupedAggregateHashTable::InitialCapacity() {
+	return STANDARD_VECTOR_SIZE * 2ULL;
+}
+
+idx_t GroupedAggregateHashTable::GetMaxCapacity(HtEntryType entry_type, idx_t tuple_size) {
+	idx_t max_pages;
+	idx_t max_tuples;
 
 	switch (entry_type) {
 	case HtEntryType::HT_WIDTH_32:
 		max_pages = NumericLimits<uint8_t>::Maximum();
 		max_tuples = NumericLimits<uint16_t>::Maximum();
 		break;
-	default:
-		D_ASSERT(entry_type == HtEntryType::HT_WIDTH_64);
+	case HtEntryType::HT_WIDTH_64:
 		max_pages = NumericLimits<uint32_t>::Maximum();
 		max_tuples = NumericLimits<uint16_t>::Maximum();
 		break;
+	default:
+		throw InternalException("Unsupported hash table width");
 	}
 
 	return max_pages * MinValue(max_tuples, (idx_t)Storage::BLOCK_SIZE / tuple_size);
 }
 
+idx_t GroupedAggregateHashTable::MaxCapacity() {
+	return GetMaxCapacity(entry_type, tuple_size);
+}
+
 void GroupedAggregateHashTable::Verify() {
 #ifdef DEBUG
 	switch (entry_type) {
@@ -189,53 +171,71 @@ void GroupedAggregateHashTable::Verify() {
 
 template <class ENTRY>
 void GroupedAggregateHashTable::Resize(idx_t size) {
-	Verify();
-
 	D_ASSERT(!is_finalized);
+	D_ASSERT(size >= STANDARD_VECTOR_SIZE);
+	D_ASSERT(IsPowerOfTwo(size));
 
-	if (size <= capacity) {
+	if (size < capacity) {
 		throw InternalException("Cannot downsize a hash table!");
 	}
-	D_ASSERT(size >= STANDARD_VECTOR_SIZE);
-
-	// size needs to be a power of 2
-	D_ASSERT((size & (size - 1)) == 0);
-	bitmask = size - 1;
+	capacity = size;
 
-	auto byte_size = size * sizeof(ENTRY);
+	bitmask = capacity - 1;
+	const auto byte_size = capacity * sizeof(ENTRY);
 	if (byte_size > (idx_t)Storage::BLOCK_SIZE) {
 		hashes_hdl = buffer_manager.Allocate(byte_size);
 		hashes_hdl_ptr = hashes_hdl.Ptr();
 	}
 	memset(hashes_hdl_ptr, 0, byte_size);
-	hashes_end_ptr = hashes_hdl_ptr + byte_size;
-	capacity = size;
 
-	auto hashes_arr = (ENTRY *)hashes_hdl_ptr;
+	if (Count() != 0) {
+		D_ASSERT(!payload_hds_ptrs.empty());
+		auto hashes_arr = (ENTRY *)hashes_hdl_ptr;
+
+		idx_t block_id = 0;
+		auto block_pointer = payload_hds_ptrs[block_id];
+		auto block_end = block_pointer + tuples_per_block * tuple_size;
+
+		TupleDataChunkIterator iterator(*data_collection, TupleDataPinProperties::ALREADY_PINNED, false);
+		const auto row_locations = iterator.GetRowLocations();
+		do {
+			for (idx_t i = 0; i < iterator.GetCurrentChunkCount(); i++) {
+				const auto &row_location = row_locations[i];
+				if (row_location > block_end || row_location < block_pointer) {
+					block_id++;
+					D_ASSERT(block_id < payload_hds_ptrs.size());
+					block_pointer = payload_hds_ptrs[block_id];
+					block_end = block_pointer + tuples_per_block * tuple_size;
+				}
+				D_ASSERT(row_location >= block_pointer && row_location < block_end);
+				D_ASSERT((row_location - block_pointer) % tuple_size == 0);
+
+				const auto hash = Load<hash_t>(row_location + hash_offset);
+				D_ASSERT((hash & bitmask) == (hash % capacity));
+				D_ASSERT(hash >> hash_prefix_shift <= NumericLimits<uint16_t>::Maximum());
+
+				auto entry_idx = (idx_t)hash & bitmask;
+				while (hashes_arr[entry_idx].page_nr > 0) {
+					entry_idx++;
+					if (entry_idx >= capacity) {
+						entry_idx = 0;
+					}
+				}
 
-	PayloadApply([&](idx_t page_nr, idx_t page_offset, data_ptr_t ptr) {
-		auto hash = Load<hash_t>(ptr + hash_offset);
-		D_ASSERT((hash & bitmask) == (hash % capacity));
-		auto entry_idx = (idx_t)hash & bitmask;
-		while (hashes_arr[entry_idx].page_nr > 0) {
-			entry_idx++;
-			if (entry_idx >= capacity) {
-				entry_idx = 0;
+				auto &ht_entry = hashes_arr[entry_idx];
+				D_ASSERT(!ht_entry.page_nr);
+				ht_entry.salt = hash >> hash_prefix_shift;
+				ht_entry.page_nr = block_id + 1;
+				ht_entry.page_offset = (row_location - block_pointer) / tuple_size;
 			}
-		}
-
-		D_ASSERT(!hashes_arr[entry_idx].page_nr);
-		D_ASSERT(hash >> hash_prefix_shift <= NumericLimits<uint16_t>::Maximum());
-
-		hashes_arr[entry_idx].salt = hash >> hash_prefix_shift;
-		hashes_arr[entry_idx].page_nr = page_nr + 1;
-		hashes_arr[entry_idx].page_offset = page_offset;
-	});
+		} while (iterator.Next());
+	}
 
 	Verify();
 }
 
-idx_t GroupedAggregateHashTable::AddChunk(DataChunk &groups, DataChunk &payload, AggregateType filter) {
+idx_t GroupedAggregateHashTable::AddChunk(AggregateHTAppendState &state, DataChunk &groups, DataChunk &payload,
+                                          AggregateType filter) {
 	vector<idx_t> aggregate_filter;
 
 	auto &aggregates = layout.GetAggregates();
@@ -245,60 +245,59 @@ idx_t GroupedAggregateHashTable::AddChunk(DataChunk &groups, DataChunk &payload,
 			aggregate_filter.push_back(i);
 		}
 	}
-	return AddChunk(groups, payload, aggregate_filter);
+	return AddChunk(state, groups, payload, aggregate_filter);
 }
 
-idx_t GroupedAggregateHashTable::AddChunk(DataChunk &groups, DataChunk &payload, const vector<idx_t> &filter) {
+idx_t GroupedAggregateHashTable::AddChunk(AggregateHTAppendState &state, DataChunk &groups, DataChunk &payload,
+                                          const vector<idx_t> &filter) {
 	Vector hashes(LogicalType::HASH);
 	groups.Hash(hashes);
 
-	return AddChunk(groups, hashes, payload, filter);
+	return AddChunk(state, groups, hashes, payload, filter);
 }
 
-idx_t GroupedAggregateHashTable::AddChunk(DataChunk &groups, Vector &group_hashes, DataChunk &payload,
-                                          const vector<idx_t> &filter) {
+idx_t GroupedAggregateHashTable::AddChunk(AggregateHTAppendState &state, DataChunk &groups, Vector &group_hashes,
+                                          DataChunk &payload, const vector<idx_t> &filter) {
 	D_ASSERT(!is_finalized);
-
 	if (groups.size() == 0) {
 		return 0;
 	}
-	// dummy
-	SelectionVector new_groups(STANDARD_VECTOR_SIZE);
 
+#ifdef DEBUG
 	D_ASSERT(groups.ColumnCount() + 1 == layout.ColumnCount());
 	for (idx_t i = 0; i < groups.ColumnCount(); i++) {
 		D_ASSERT(groups.GetTypes()[i] == layout.GetTypes()[i]);
 	}
+#endif
 
-	Vector addresses(LogicalType::POINTER);
-	auto new_group_count = FindOrCreateGroups(groups, group_hashes, addresses, new_groups);
-	VectorOperations::AddInPlace(addresses, layout.GetAggrOffset(), payload.size());
-
-	// now every cell has an entry
-	// update the aggregates
-	idx_t payload_idx = 0;
+	auto new_group_count = FindOrCreateGroups(state, groups, group_hashes, state.addresses, state.new_groups);
+	VectorOperations::AddInPlace(state.addresses, layout.GetAggrOffset(), payload.size());
 
+	// Now every cell has an entry, update the aggregates
 	auto &aggregates = layout.GetAggregates();
 	idx_t filter_idx = 0;
+	idx_t payload_idx = 0;
+	RowOperationsState row_state(aggregate_allocator->GetAllocator());
 	for (idx_t i = 0; i < aggregates.size(); i++) {
 		auto &aggr = aggregates[i];
 		if (filter_idx >= filter.size() || i < filter[filter_idx]) {
 			// Skip all the aggregates that are not in the filter
 			payload_idx += aggr.child_count;
-			VectorOperations::AddInPlace(addresses, aggr.payload_size, payload.size());
+			VectorOperations::AddInPlace(state.addresses, aggr.payload_size, payload.size());
 			continue;
 		}
 		D_ASSERT(i == filter[filter_idx]);
 
 		if (aggr.aggr_type != AggregateType::DISTINCT && aggr.filter) {
-			RowOperations::UpdateFilteredStates(filter_set.GetFilterData(i), aggr, addresses, payload, payload_idx);
+			RowOperations::UpdateFilteredStates(row_state, filter_set.GetFilterData(i), aggr, state.addresses, payload,
+			                                    payload_idx);
 		} else {
-			RowOperations::UpdateStates(aggr, addresses, payload, payload_idx, payload.size());
+			RowOperations::UpdateStates(row_state, aggr, state.addresses, payload, payload_idx, payload.size());
 		}
 
-		// move to the next aggregate
+		// Move to the next aggregate
 		payload_idx += aggr.child_count;
-		VectorOperations::AddInPlace(addresses, aggr.payload_size, payload.size());
+		VectorOperations::AddInPlace(state.addresses, aggr.payload_size, payload.size());
 		filter_idx++;
 	}
 
@@ -316,209 +315,258 @@ void GroupedAggregateHashTable::FetchAggregates(DataChunk &groups, DataChunk &re
 	if (groups.size() == 0) {
 		return;
 	}
+
 	// find the groups associated with the addresses
 	// FIXME: this should not use the FindOrCreateGroups, creating them is unnecessary
+	AggregateHTAppendState append_state;
 	Vector addresses(LogicalType::POINTER);
-	FindOrCreateGroups(groups, addresses);
+	FindOrCreateGroups(append_state, groups, addresses);
 	// now fetch the aggregates
-	RowOperations::FinalizeStates(layout, addresses, result, 0);
+	RowOperationsState row_state(aggregate_allocator->GetAllocator());
+	RowOperations::FinalizeStates(row_state, layout, addresses, result, 0);
+}
+
+idx_t GroupedAggregateHashTable::ResizeThreshold() {
+	return capacity / LOAD_FACTOR;
 }
 
 template <class ENTRY>
-idx_t GroupedAggregateHashTable::FindOrCreateGroupsInternal(DataChunk &groups, Vector &group_hashes, Vector &addresses,
+idx_t GroupedAggregateHashTable::FindOrCreateGroupsInternal(AggregateHTAppendState &state, DataChunk &groups,
+                                                            Vector &group_hashes_v, Vector &addresses_v,
                                                             SelectionVector &new_groups_out) {
 	D_ASSERT(!is_finalized);
+	D_ASSERT(groups.ColumnCount() + 1 == layout.ColumnCount());
+	D_ASSERT(group_hashes_v.GetType() == LogicalType::HASH);
+	D_ASSERT(state.ht_offsets.GetVectorType() == VectorType::FLAT_VECTOR);
+	D_ASSERT(state.ht_offsets.GetType() == LogicalType::BIGINT);
+	D_ASSERT(addresses_v.GetType() == LogicalType::POINTER);
+	D_ASSERT(state.hash_salts.GetType() == LogicalType::SMALLINT);
 
-	if (entries + groups.size() > MaxCapacity()) {
+	if (Count() + groups.size() > MaxCapacity()) {
 		throw InternalException("Hash table capacity reached");
 	}
 
-	// resize at 50% capacity, also need to fit the entire vector
-	if (capacity - entries <= groups.size() || entries > capacity / LOAD_FACTOR) {
+	// Resize at 50% capacity, also need to fit the entire vector
+	if (capacity - Count() <= groups.size() || Count() > ResizeThreshold()) {
+		Verify();
 		Resize<ENTRY>(capacity * 2);
 	}
+	D_ASSERT(capacity - Count() >= groups.size()); // we need to be able to fit at least one vector of data
 
-	D_ASSERT(capacity - entries >= groups.size());
-	D_ASSERT(groups.ColumnCount() + 1 == layout.ColumnCount());
-	// we need to be able to fit at least one vector of data
-	D_ASSERT(capacity - entries >= groups.size());
-	D_ASSERT(group_hashes.GetType() == LogicalType::HASH);
+	group_hashes_v.Flatten(groups.size());
+	auto group_hashes = FlatVector::GetData<hash_t>(group_hashes_v);
 
-	group_hashes.Flatten(groups.size());
-	auto group_hashes_ptr = FlatVector::GetData<hash_t>(group_hashes);
+	addresses_v.Flatten(groups.size());
+	auto addresses = FlatVector::GetData<data_ptr_t>(addresses_v);
 
-	D_ASSERT(ht_offsets.GetVectorType() == VectorType::FLAT_VECTOR);
-	D_ASSERT(ht_offsets.GetType() == LogicalType::BIGINT);
-
-	D_ASSERT(addresses.GetType() == LogicalType::POINTER);
-	addresses.Flatten(groups.size());
-	auto addresses_ptr = FlatVector::GetData<data_ptr_t>(addresses);
-
-	// now compute the entry in the table based on the hash using a modulo
-	UnaryExecutor::Execute<hash_t, uint64_t>(group_hashes, ht_offsets, groups.size(), [&](hash_t element) {
+	// Compute the entry in the table based on the hash using a modulo,
+	// and precompute the hash salts for faster comparison below
+	auto ht_offsets_ptr = FlatVector::GetData<uint64_t>(state.ht_offsets);
+	auto hash_salts_ptr = FlatVector::GetData<uint16_t>(state.hash_salts);
+	for (idx_t r = 0; r < groups.size(); r++) {
+		auto element = group_hashes[r];
 		D_ASSERT((element & bitmask) == (element % capacity));
-		return (element & bitmask);
-	});
-	auto ht_offsets_ptr = FlatVector::GetData<uint64_t>(ht_offsets);
-
-	// precompute the hash salts for faster comparison below
-	D_ASSERT(hash_salts.GetType() == LogicalType::SMALLINT);
-	UnaryExecutor::Execute<hash_t, uint16_t>(group_hashes, hash_salts, groups.size(),
-	                                         [&](hash_t element) { return (element >> hash_prefix_shift); });
-	auto hash_salts_ptr = FlatVector::GetData<uint16_t>(hash_salts);
-
+		ht_offsets_ptr[r] = element & bitmask;
+		hash_salts_ptr[r] = element >> hash_prefix_shift;
+	}
 	// we start out with all entries [0, 1, 2, ..., groups.size()]
 	const SelectionVector *sel_vector = FlatVector::IncrementalSelectionVector();
 
-	idx_t remaining_entries = groups.size();
-
-	// make a chunk that references the groups and the hashes
-	DataChunk group_chunk;
-	group_chunk.InitializeEmpty(layout.GetTypes());
+	// Make a chunk that references the groups and the hashes and convert to unified format
+	if (state.group_chunk.ColumnCount() == 0) {
+		state.group_chunk.InitializeEmpty(layout.GetTypes());
+	}
+	D_ASSERT(state.group_chunk.ColumnCount() == layout.GetTypes().size());
 	for (idx_t grp_idx = 0; grp_idx < groups.ColumnCount(); grp_idx++) {
-		group_chunk.data[grp_idx].Reference(groups.data[grp_idx]);
+		state.group_chunk.data[grp_idx].Reference(groups.data[grp_idx]);
 	}
-	group_chunk.data[groups.ColumnCount()].Reference(group_hashes);
-	group_chunk.SetCardinality(groups);
+	state.group_chunk.data[groups.ColumnCount()].Reference(group_hashes_v);
+	state.group_chunk.SetCardinality(groups);
 
 	// convert all vectors to unified format
-	auto group_data = group_chunk.ToUnifiedFormat();
+	if (!state.chunk_state_initialized) {
+		data_collection->InitializeAppend(state.chunk_state);
+		state.chunk_state_initialized = true;
+	}
+	TupleDataCollection::ToUnifiedFormat(state.chunk_state, state.group_chunk);
+	if (!state.group_data) {
+		state.group_data = unique_ptr<UnifiedVectorFormat[]>(new UnifiedVectorFormat[state.group_chunk.ColumnCount()]);
+	}
+	TupleDataCollection::GetVectorData(state.chunk_state, state.group_data.get());
 
 	idx_t new_group_count = 0;
+	idx_t remaining_entries = groups.size();
 	while (remaining_entries > 0) {
 		idx_t new_entry_count = 0;
 		idx_t need_compare_count = 0;
 		idx_t no_match_count = 0;
 
-		// first figure out for each remaining whether or not it belongs to a full or empty group
+		// For each remaining entry, figure out whether or not it belongs to a full or empty group
 		for (idx_t i = 0; i < remaining_entries; i++) {
 			const idx_t index = sel_vector->get_index(i);
-			const auto ht_entry_ptr = ((ENTRY *)this->hashes_hdl_ptr) + ht_offsets_ptr[index];
-			if (ht_entry_ptr->page_nr == 0) { // we use page number 0 as a "unused marker"
-				// cell is empty; setup the new entry
-				if (payload_page_offset == tuples_per_block || payload_hds.empty()) {
-					NewBlock();
-				}
-
-				auto entry_payload_ptr = payload_hds_ptrs.back() + (payload_page_offset * tuple_size);
-
-				D_ASSERT(group_hashes_ptr[index] >> hash_prefix_shift <= NumericLimits<uint16_t>::Maximum());
-				D_ASSERT(payload_page_offset < tuples_per_block);
-				D_ASSERT(payload_hds.size() < NumericLimits<uint32_t>::Maximum());
-				D_ASSERT(payload_page_offset + 1 < NumericLimits<uint16_t>::Maximum());
-
-				ht_entry_ptr->salt = group_hashes_ptr[index] >> hash_prefix_shift;
+			auto &ht_entry = *(((ENTRY *)this->hashes_hdl_ptr) + ht_offsets_ptr[index]);
+			if (ht_entry.page_nr == 0) { // Cell is unoccupied (we use page number 0 as a "unused marker")
+				D_ASSERT(group_hashes[index] >> hash_prefix_shift <= NumericLimits<uint16_t>::Maximum());
+				D_ASSERT(payload_hds_ptrs.size() < NumericLimits<uint32_t>::Maximum());
 
-				// page numbers start at one so we can use 0 as empty flag
-				// GetPtr undoes this
-				ht_entry_ptr->page_nr = payload_hds.size();
-				ht_entry_ptr->page_offset = payload_page_offset++;
+				// Set page nr to 1 for now to mark it as occupied (will be corrected later) and set the salt
+				ht_entry.page_nr = 1;
+				ht_entry.salt = group_hashes[index] >> hash_prefix_shift;
 
-				// update selection lists for outer loops
-				empty_vector.set_index(new_entry_count++, index);
+				// Update selection lists for outer loops
+				state.empty_vector.set_index(new_entry_count++, index);
 				new_groups_out.set_index(new_group_count++, index);
-				entries++;
-
-				addresses_ptr[index] = entry_payload_ptr;
-
-			} else {
-				// cell is occupied: add to check list
-				// only need to check if hash salt in ptr == prefix of hash in payload
-				if (ht_entry_ptr->salt == hash_salts_ptr[index]) {
-					group_compare_vector.set_index(need_compare_count++, index);
-
-					auto page_ptr = payload_hds_ptrs[ht_entry_ptr->page_nr - 1];
-					auto page_offset = ht_entry_ptr->page_offset * tuple_size;
-					addresses_ptr[index] = page_ptr + page_offset;
-
+			} else { // Cell is occupied: Compare salts
+				if (ht_entry.salt == hash_salts_ptr[index]) {
+					state.group_compare_vector.set_index(need_compare_count++, index);
 				} else {
-					no_match_vector.set_index(no_match_count++, index);
+					state.no_match_vector.set_index(no_match_count++, index);
 				}
 			}
 		}
 
-		// for each of the locations that are empty, serialize the group columns to the locations
-		RowOperations::Scatter(group_chunk, group_data.get(), layout, addresses, *string_heap, empty_vector,
-		                       new_entry_count);
-		RowOperations::InitializeStates(layout, addresses, empty_vector, new_entry_count);
+		if (new_entry_count != 0) {
+			// Append everything that belongs to an empty group
+			data_collection->AppendUnified(td_pin_state, state.chunk_state, state.group_chunk, state.empty_vector,
+			                               new_entry_count);
+			RowOperations::InitializeStates(layout, state.chunk_state.row_locations,
+			                                *FlatVector::IncrementalSelectionVector(), new_entry_count);
+
+			// Get the pointers to the (possibly) newly created blocks of the data collection
+			idx_t block_id = payload_hds_ptrs.empty() ? 0 : payload_hds_ptrs.size() - 1;
+			UpdateBlockPointers();
+			auto block_pointer = payload_hds_ptrs[block_id];
+			auto block_end = block_pointer + tuples_per_block * tuple_size;
+
+			// Set the page nrs/offsets in the 1st part of the HT now that the data has been appended
+			const auto row_locations = FlatVector::GetData<data_ptr_t>(state.chunk_state.row_locations);
+			for (idx_t new_entry_idx = 0; new_entry_idx < new_entry_count; new_entry_idx++) {
+				const auto &row_location = row_locations[new_entry_idx];
+				if (row_location > block_end || row_location < block_pointer) {
+					block_id++;
+					D_ASSERT(block_id < payload_hds_ptrs.size());
+					block_pointer = payload_hds_ptrs[block_id];
+					block_end = block_pointer + tuples_per_block * tuple_size;
+				}
+				D_ASSERT(row_location >= block_pointer && row_location < block_end);
+				D_ASSERT((row_location - block_pointer) % tuple_size == 0);
+				const auto index = state.empty_vector.get_index(new_entry_idx);
+				auto &ht_entry = *(((ENTRY *)this->hashes_hdl_ptr) + ht_offsets_ptr[index]);
+				ht_entry.page_nr = block_id + 1;
+				ht_entry.page_offset = (row_location - block_pointer) / tuple_size;
+				addresses[index] = row_location;
+			}
+		}
 
-		// now we have only the tuples remaining that might match to an existing group
-		// start performing comparisons with each of the groups
-		RowOperations::Match(group_chunk, group_data.get(), layout, addresses, predicates, group_compare_vector,
-		                     need_compare_count, &no_match_vector, no_match_count);
+		if (need_compare_count != 0) {
+			// Get the pointers to the rows that need to be compared
+			for (idx_t need_compare_idx = 0; need_compare_idx < need_compare_count; need_compare_idx++) {
+				const auto index = state.group_compare_vector.get_index(need_compare_idx);
+				const auto &ht_entry = *(((ENTRY *)this->hashes_hdl_ptr) + ht_offsets_ptr[index]);
+				auto page_ptr = payload_hds_ptrs[ht_entry.page_nr - 1];
+				auto page_offset = ht_entry.page_offset * tuple_size;
+				addresses[index] = page_ptr + page_offset;
+			}
 
-		// each of the entries that do not match we move them to the next entry in the HT
+			// Perform group comparisons
+			RowOperations::Match(state.group_chunk, state.group_data.get(), layout, addresses_v, predicates,
+			                     state.group_compare_vector, need_compare_count, &state.no_match_vector,
+			                     no_match_count);
+		}
+
+		// Linear probing: each of the entries that do not match move to the next entry in the HT
 		for (idx_t i = 0; i < no_match_count; i++) {
-			idx_t index = no_match_vector.get_index(i);
+			idx_t index = state.no_match_vector.get_index(i);
 			ht_offsets_ptr[index]++;
 			if (ht_offsets_ptr[index] >= capacity) {
 				ht_offsets_ptr[index] = 0;
 			}
 		}
-		sel_vector = &no_match_vector;
+		sel_vector = &state.no_match_vector;
 		remaining_entries = no_match_count;
 	}
 
 	return new_group_count;
 }
 
+void GroupedAggregateHashTable::UpdateBlockPointers() {
+	for (const auto &id_and_handle : td_pin_state.row_handles) {
+		const auto &id = id_and_handle.first;
+		const auto &handle = id_and_handle.second;
+		if (payload_hds_ptrs.empty() || id > payload_hds_ptrs.size() - 1) {
+			payload_hds_ptrs.resize(id + 1);
+		}
+		payload_hds_ptrs[id] = handle.Ptr();
+	}
+}
+
 // this is to support distinct aggregations where we need to record whether we
 // have already seen a value for a group
-idx_t GroupedAggregateHashTable::FindOrCreateGroups(DataChunk &groups, Vector &group_hashes, Vector &addresses_out,
+idx_t GroupedAggregateHashTable::FindOrCreateGroups(AggregateHTAppendState &state, DataChunk &groups,
+                                                    Vector &group_hashes, Vector &addresses_out,
                                                     SelectionVector &new_groups_out) {
 	switch (entry_type) {
 	case HtEntryType::HT_WIDTH_64:
-		return FindOrCreateGroupsInternal<aggr_ht_entry_64>(groups, group_hashes, addresses_out, new_groups_out);
+		return FindOrCreateGroupsInternal<aggr_ht_entry_64>(state, groups, group_hashes, addresses_out, new_groups_out);
 	case HtEntryType::HT_WIDTH_32:
-		return FindOrCreateGroupsInternal<aggr_ht_entry_32>(groups, group_hashes, addresses_out, new_groups_out);
+		return FindOrCreateGroupsInternal<aggr_ht_entry_32>(state, groups, group_hashes, addresses_out, new_groups_out);
 	default:
 		throw InternalException("Unknown HT entry width");
 	}
 }
 
-void GroupedAggregateHashTable::FindOrCreateGroups(DataChunk &groups, Vector &addresses) {
+void GroupedAggregateHashTable::FindOrCreateGroups(AggregateHTAppendState &state, DataChunk &groups,
+                                                   Vector &addresses) {
 	// create a dummy new_groups sel vector
-	SelectionVector new_groups(STANDARD_VECTOR_SIZE);
-	FindOrCreateGroups(groups, addresses, new_groups);
+	FindOrCreateGroups(state, groups, addresses, state.new_groups);
 }
 
-idx_t GroupedAggregateHashTable::FindOrCreateGroups(DataChunk &groups, Vector &addresses_out,
-                                                    SelectionVector &new_groups_out) {
+idx_t GroupedAggregateHashTable::FindOrCreateGroups(AggregateHTAppendState &state, DataChunk &groups,
+                                                    Vector &addresses_out, SelectionVector &new_groups_out) {
 	Vector hashes(LogicalType::HASH);
 	groups.Hash(hashes);
-	return FindOrCreateGroups(groups, hashes, addresses_out, new_groups_out);
+	return FindOrCreateGroups(state, groups, hashes, addresses_out, new_groups_out);
 }
 
 struct FlushMoveState {
-	FlushMoveState(Allocator &allocator, RowLayout &layout)
-	    : new_groups(STANDARD_VECTOR_SIZE), group_addresses(LogicalType::POINTER),
+	explicit FlushMoveState(TupleDataCollection &collection_p)
+	    : collection(collection_p), hashes(LogicalType::HASH), group_addresses(LogicalType::POINTER),
 	      new_groups_sel(STANDARD_VECTOR_SIZE) {
-		vector<LogicalType> group_types(layout.GetTypes().begin(), layout.GetTypes().end() - 1);
-		groups.Initialize(allocator, group_types);
+		const auto &layout = collection.GetLayout();
+		vector<column_t> column_ids;
+		column_ids.reserve(layout.ColumnCount() - 1);
+		for (idx_t col_idx = 0; col_idx < layout.ColumnCount() - 1; col_idx++) {
+			column_ids.emplace_back(col_idx);
+		}
+		// FIXME DESTROY_AFTER_DONE if we make it possible to pass a selection vector to RowOperations::DestroyStates?
+		collection.InitializeScan(scan_state, column_ids, TupleDataPinProperties::UNPIN_AFTER_DONE);
+		collection.InitializeScanChunk(scan_state, groups);
+		hash_col_idx = layout.ColumnCount() - 1;
 	}
 
+	bool Scan();
+
+	TupleDataCollection &collection;
+	TupleDataScanState scan_state;
 	DataChunk groups;
-	SelectionVector new_groups;
+
+	idx_t hash_col_idx;
+	Vector hashes;
+
+	AggregateHTAppendState append_state;
 	Vector group_addresses;
 	SelectionVector new_groups_sel;
 };
 
-void GroupedAggregateHashTable::FlushMove(FlushMoveState &state, Vector &source_addresses, Vector &source_hashes,
-                                          idx_t count) {
-	D_ASSERT(source_addresses.GetType() == LogicalType::POINTER);
-	D_ASSERT(source_hashes.GetType() == LogicalType::HASH);
-
-	state.groups.Reset();
-	state.groups.SetCardinality(count);
-	for (idx_t col_no = 0; col_no < state.groups.ColumnCount(); col_no++) {
-		auto &column = state.groups.data[col_no];
-		RowOperations::Gather(source_addresses, *FlatVector::IncrementalSelectionVector(), column,
-		                      *FlatVector::IncrementalSelectionVector(), count, layout, col_no);
+bool FlushMoveState::Scan() {
+	if (collection.Scan(scan_state, groups)) {
+		collection.Gather(scan_state.chunk_state.row_locations, *FlatVector::IncrementalSelectionVector(),
+		                  groups.size(), hash_col_idx, hashes, *FlatVector::IncrementalSelectionVector());
+		return true;
 	}
 
-	FindOrCreateGroups(state.groups, source_hashes, state.group_addresses, state.new_groups_sel);
-
-	RowOperations::CombineStates(layout, source_addresses, state.group_addresses, count);
+	collection.FinalizePinState(scan_state.pin_state);
+	return false;
 }
 
 void GroupedAggregateHashTable::Combine(GroupedAggregateHashTable &other) {
@@ -527,126 +575,67 @@ void GroupedAggregateHashTable::Combine(GroupedAggregateHashTable &other) {
 	D_ASSERT(other.layout.GetAggrWidth() == layout.GetAggrWidth());
 	D_ASSERT(other.layout.GetDataWidth() == layout.GetDataWidth());
 	D_ASSERT(other.layout.GetRowWidth() == layout.GetRowWidth());
-	D_ASSERT(other.tuples_per_block == tuples_per_block);
 
-	if (other.entries == 0) {
+	if (other.Count() == 0) {
 		return;
 	}
 
-	Vector addresses(LogicalType::POINTER);
-	auto addresses_ptr = FlatVector::GetData<data_ptr_t>(addresses);
-
-	Vector hashes(LogicalType::HASH);
-	auto hashes_ptr = FlatVector::GetData<hash_t>(hashes);
-
-	idx_t group_idx = 0;
-
-	FlushMoveState state(allocator, layout);
-	other.PayloadApply([&](idx_t page_nr, idx_t page_offset, data_ptr_t ptr) {
-		auto hash = Load<hash_t>(ptr + hash_offset);
+	FlushMoveState state(*other.data_collection);
+	RowOperationsState row_state(aggregate_allocator->GetAllocator());
+	while (state.Scan()) {
+		FindOrCreateGroups(state.append_state, state.groups, state.hashes, state.group_addresses, state.new_groups_sel);
+		RowOperations::CombineStates(row_state, layout, state.scan_state.chunk_state.row_locations,
+		                             state.group_addresses, state.groups.size());
+	}
 
-		hashes_ptr[group_idx] = hash;
-		addresses_ptr[group_idx] = ptr;
-		group_idx++;
-		if (group_idx == STANDARD_VECTOR_SIZE) {
-			FlushMove(state, addresses, hashes, group_idx);
-			group_idx = 0;
-		}
-	});
-	FlushMove(state, addresses, hashes, group_idx);
-	string_heap->Merge(*other.string_heap);
 	Verify();
 }
 
-struct PartitionInfo {
-	PartitionInfo() : addresses(LogicalType::POINTER), hashes(LogicalType::HASH), group_count(0) {
-		addresses_ptr = FlatVector::GetData<data_ptr_t>(addresses);
-		hashes_ptr = FlatVector::GetData<hash_t>(hashes);
-	};
-	Vector addresses;
-	Vector hashes;
-	idx_t group_count;
-	data_ptr_t *addresses_ptr;
-	hash_t *hashes_ptr;
-};
-
-void GroupedAggregateHashTable::Partition(vector<GroupedAggregateHashTable *> &partition_hts, hash_t mask,
-                                          idx_t shift) {
-	D_ASSERT(partition_hts.size() > 1);
-	vector<PartitionInfo> partition_info(partition_hts.size());
-
-	FlushMoveState state(allocator, layout);
-	PayloadApply([&](idx_t page_nr, idx_t page_offset, data_ptr_t ptr) {
-		auto hash = Load<hash_t>(ptr + hash_offset);
-
-		idx_t partition = (hash & mask) >> shift;
-		D_ASSERT(partition < partition_hts.size());
-
-		auto &info = partition_info[partition];
-
-		info.hashes_ptr[info.group_count] = hash;
-		info.addresses_ptr[info.group_count] = ptr;
-		info.group_count++;
-		if (info.group_count == STANDARD_VECTOR_SIZE) {
-			D_ASSERT(partition_hts[partition]);
-			partition_hts[partition]->FlushMove(state, info.addresses, info.hashes, info.group_count);
-			info.group_count = 0;
-		}
-	});
-
-	idx_t info_idx = 0;
-	idx_t total_count = 0;
-	for (auto &partition_entry : partition_hts) {
-		auto &info = partition_info[info_idx++];
-		partition_entry->FlushMove(state, info.addresses, info.hashes, info.group_count);
-
-		partition_entry->string_heap->Merge(*string_heap);
-		partition_entry->Verify();
-		total_count += partition_entry->Size();
+void GroupedAggregateHashTable::Partition(vector<GroupedAggregateHashTable *> &partition_hts, idx_t radix_bits) {
+	const auto num_partitions = RadixPartitioning::NumberOfPartitions(radix_bits);
+	D_ASSERT(partition_hts.size() == num_partitions);
+
+	// Partition the data
+	auto partitioned_data =
+	    make_uniq<RadixPartitionedTupleData>(buffer_manager, layout, radix_bits, layout.ColumnCount() - 1);
+	partitioned_data->Partition(*data_collection, TupleDataPinProperties::KEEP_EVERYTHING_PINNED);
+	D_ASSERT(partitioned_data->GetPartitions().size() == num_partitions);
+
+	// Move the partitioned data collections to the partitioned hash tables and initialize the 1st part of the HT
+	auto &partitions = partitioned_data->GetPartitions();
+	for (idx_t partition_idx = 0; partition_idx < num_partitions; partition_idx++) {
+		auto &partition_ht = *partition_hts[partition_idx];
+		partition_ht.data_collection = std::move(partitions[partition_idx]);
+		partition_ht.aggregate_allocator = aggregate_allocator;
+		partition_ht.InitializeFirstPart();
+		partition_ht.Verify();
 	}
-	(void)total_count;
-	D_ASSERT(total_count == entries);
 }
 
-idx_t GroupedAggregateHashTable::Scan(AggregateHTScanState &scan_state, DataChunk &result) {
-	idx_t this_n;
-	Vector addresses(LogicalType::POINTER);
-	auto data_pointers = FlatVector::GetData<data_ptr_t>(addresses);
-	{
-		lock_guard<mutex> l(scan_state.lock);
-		if (scan_state.scan_position >= entries) {
-			return 0;
-		}
-		auto remaining = entries - scan_state.scan_position;
-		this_n = MinValue((idx_t)STANDARD_VECTOR_SIZE, remaining);
-
-		auto chunk_idx = scan_state.scan_position / tuples_per_block;
-		auto chunk_offset = (scan_state.scan_position % tuples_per_block) * tuple_size;
-		D_ASSERT(chunk_offset + tuple_size <= Storage::BLOCK_SIZE);
-
-		auto read_ptr = payload_hds_ptrs[chunk_idx++];
-		for (idx_t i = 0; i < this_n; i++) {
-			data_pointers[i] = read_ptr + chunk_offset;
-			chunk_offset += tuple_size;
-			if (chunk_offset >= tuples_per_block * tuple_size) {
-				read_ptr = payload_hds_ptrs[chunk_idx++];
-				chunk_offset = 0;
-			}
-		}
-		scan_state.scan_position += this_n;
+void GroupedAggregateHashTable::InitializeFirstPart() {
+	data_collection->GetBlockPointers(payload_hds_ptrs);
+	auto size = MaxValue<idx_t>(NextPowerOfTwo(Count() * 2L), capacity);
+	switch (entry_type) {
+	case HtEntryType::HT_WIDTH_64:
+		Resize<aggr_ht_entry_64>(size);
+		break;
+	case HtEntryType::HT_WIDTH_32:
+		Resize<aggr_ht_entry_32>(size);
+		break;
+	default:
+		throw InternalException("Unknown HT entry width");
 	}
+}
 
-	result.SetCardinality(this_n);
-	// fetch the group columns (ignoring the final hash column
+idx_t GroupedAggregateHashTable::Scan(TupleDataParallelScanState &gstate, TupleDataLocalScanState &lstate,
+                                      DataChunk &result) {
+	data_collection->Scan(gstate, lstate, result);
+
+	RowOperationsState row_state(aggregate_allocator->GetAllocator());
 	const auto group_cols = layout.ColumnCount() - 1;
-	for (idx_t col_no = 0; col_no < group_cols; col_no++) {
-		auto &column = result.data[col_no];
-		RowOperations::Gather(addresses, *FlatVector::IncrementalSelectionVector(), column,
-		                      *FlatVector::IncrementalSelectionVector(), result.size(), layout, col_no);
-	}
+	RowOperations::FinalizeStates(row_state, layout, lstate.chunk_state.row_locations, result, group_cols);
 
-	RowOperations::FinalizeStates(layout, addresses, result, group_cols);
-	return this_n;
+	return result.size();
 }
 
 void GroupedAggregateHashTable::Finalize() {
@@ -654,8 +643,11 @@ void GroupedAggregateHashTable::Finalize() {
 		return;
 	}
 
-	// early release hashes, not needed for partition/scan
+	// Early release hashes (not needed for partition/scan) and data collection (will be pinned again when scanning)
 	hashes_hdl.Destroy();
+	data_collection->FinalizePinState(td_pin_state);
+	data_collection->Unpin();
+
 	is_finalized = true;
 }