duckdb 0.7.2-dev16.0 → 0.7.2-dev314.0

package/src/duckdb/src/main/client_context.cpp

@@ -813,12 +813,14 @@ unique_ptr<QueryResult> ClientContext::Query(const string &query, bool allow_str
 
 	unique_ptr<QueryResult> result;
 	QueryResult *last_result = nullptr;
+	bool last_had_result = false;
 	for (idx_t i = 0; i < statements.size(); i++) {
 		auto &statement = statements[i];
 		bool is_last_statement = i + 1 == statements.size();
 		PendingQueryParameters parameters;
 		parameters.allow_stream_result = allow_stream_result && is_last_statement;
 		auto pending_query = PendingQueryInternal(*lock, std::move(statement), parameters);
+		auto has_result = pending_query->properties.return_type == StatementReturnType::QUERY_RESULT;
 		unique_ptr<QueryResult> current_result;
 		if (pending_query->HasError()) {
 			current_result = make_unique<MaterializedQueryResult>(pending_query->GetErrorObject());
@@ -826,12 +828,17 @@ unique_ptr<QueryResult> ClientContext::Query(const string &query, bool allow_str
 			current_result = ExecutePendingQueryInternal(*lock, *pending_query);
 		}
 		// now append the result to the list of results
-		if (!last_result) {
+		if (!last_result || !last_had_result) {
 			// first result of the query
 			result = std::move(current_result);
 			last_result = result.get();
+			last_had_result = has_result;
 		} else {
 			// later results; attach to the result chain
+			// but only if there is a result
+			if (!has_result) {
+				continue;
+			}
 			last_result->next = std::move(current_result);
 			last_result = last_result->next.get();
 		}

package/src/duckdb/src/main/config.cpp

@@ -9,6 +9,9 @@
 #include "duckdb/common/thread.hpp"
 #endif
 
+#include <cstdio>
+#include <inttypes.h>
+
 namespace duckdb {
 
 #define DUCKDB_GLOBAL(_PARAM)                                                                                          \
@@ -214,9 +217,71 @@ void DBConfig::SetDefaultMaxMemory() {
 	}
 }
 
+idx_t CGroupBandwidthQuota(idx_t physical_cores, FileSystem &fs) {
+	static constexpr const char *CPU_MAX = "/sys/fs/cgroup/cpu.max";
+	static constexpr const char *CFS_QUOTA = "/sys/fs/cgroup/cpu/cpu.cfs_quota_us";
+	static constexpr const char *CFS_PERIOD = "/sys/fs/cgroup/cpu/cpu.cfs_period_us";
+
+	int64_t quota, period;
+	char byte_buffer[1000];
+	unique_ptr<FileHandle> handle;
+	int64_t read_bytes;
+
+	if (fs.FileExists(CPU_MAX)) {
+		// cgroup v2
+		// https://www.kernel.org/doc/html/latest/admin-guide/cgroup-v2.html
+		handle =
+		    fs.OpenFile(CPU_MAX, FileFlags::FILE_FLAGS_READ, FileSystem::DEFAULT_LOCK, FileSystem::DEFAULT_COMPRESSION);
+		read_bytes = fs.Read(*handle, (void *)byte_buffer, 999);
+		byte_buffer[read_bytes] = '\0';
+		if (std::sscanf(byte_buffer, "%" SCNd64 " %" SCNd64 "", &quota, &period) != 2) {
+			return physical_cores;
+		}
+	} else if (fs.FileExists(CFS_QUOTA) && fs.FileExists(CFS_PERIOD)) {
+		// cgroup v1
+		// https://www.kernel.org/doc/html/latest/scheduler/sched-bwc.html#management
+
+		// Read the quota, this indicates how many microseconds the CPU can be utilized by this cgroup per period
+		handle = fs.OpenFile(CFS_QUOTA, FileFlags::FILE_FLAGS_READ, FileSystem::DEFAULT_LOCK,
+		                     FileSystem::DEFAULT_COMPRESSION);
+		read_bytes = fs.Read(*handle, (void *)byte_buffer, 999);
+		byte_buffer[read_bytes] = '\0';
+		if (std::sscanf(byte_buffer, "%" SCNd64 "", &quota) != 1) {
+			return physical_cores;
+		}
+
+		// Read the time period, a cgroup can utilize the CPU up to quota microseconds every period
+		handle = fs.OpenFile(CFS_PERIOD, FileFlags::FILE_FLAGS_READ, FileSystem::DEFAULT_LOCK,
+		                     FileSystem::DEFAULT_COMPRESSION);
+		read_bytes = fs.Read(*handle, (void *)byte_buffer, 999);
+		byte_buffer[read_bytes] = '\0';
+		if (std::sscanf(byte_buffer, "%" SCNd64 "", &period) != 1) {
+			return physical_cores;
+		}
+	} else {
+		// No cgroup quota
+		return physical_cores;
+	}
+	if (quota > 0 && period > 0) {
+		return idx_t(std::ceil((double)quota / (double)period));
+	} else {
+		return physical_cores;
+	}
+}
+
+idx_t GetSystemMaxThreadsInternal(FileSystem &fs) {
+	idx_t physical_cores = std::thread::hardware_concurrency();
+#ifdef __linux__
+	auto cores_available_per_period = CGroupBandwidthQuota(physical_cores, fs);
+	return MaxValue<idx_t>(cores_available_per_period, 1);
+#else
+	return physical_cores;
+#endif
+}
+
 void DBConfig::SetDefaultMaxThreads() {
 #ifndef DUCKDB_NO_THREADS
-	options.maximum_threads = std::thread::hardware_concurrency();
+	options.maximum_threads = GetSystemMaxThreadsInternal(*file_system);
 #else
 	options.maximum_threads = 1;
 #endif

package/src/duckdb/src/main/database.cpp

@@ -191,8 +191,7 @@ void DatabaseInstance::Initialize(const char *database_path, DBConfig *user_conf
 	}
 
 	db_manager = make_unique<DatabaseManager>(*this);
-	buffer_manager =
-	    make_unique<BufferManager>(*this, config.options.temporary_directory, config.options.maximum_memory);
+	buffer_manager = make_unique<BufferManager>(*this, config.options.temporary_directory);
 	scheduler = make_unique<TaskScheduler>(*this);
 	object_cache = make_unique<ObjectCache>();
 	connection_manager = make_unique<ConnectionManager>();
@@ -271,6 +270,10 @@ BufferManager &DatabaseInstance::GetBufferManager() {
 	return *buffer_manager;
 }
 
+BufferPool &DatabaseInstance::GetBufferPool() {
+	return *config.buffer_pool;
+}
+
 DatabaseManager &DatabaseManager::Get(DatabaseInstance &db) {
 	return db.GetDatabaseManager();
 }
@@ -340,6 +343,11 @@ void DatabaseInstance::Configure(DBConfig &new_config) {
 	if (!config.default_allocator) {
 		config.default_allocator = Allocator::DefaultAllocatorReference();
 	}
+	if (new_config.buffer_pool) {
+		config.buffer_pool = std::move(new_config.buffer_pool);
+	} else {
+		config.buffer_pool = make_shared<BufferPool>(config.options.maximum_memory);
+	}
 }
 
 DBConfig &DBConfig::GetConfig(ClientContext &context) {

package/src/duckdb/src/optimizer/join_order/cardinality_estimator.cpp

@@ -6,24 +6,18 @@
 #include "duckdb/planner/operator/logical_comparison_join.hpp"
 #include "duckdb/planner/operator/logical_get.hpp"
 #include "duckdb/storage/data_table.hpp"
-#include "duckdb/storage/statistics/numeric_statistics.hpp"
 #include "duckdb/catalog/catalog_entry/table_catalog_entry.hpp"
 
 namespace duckdb {
 
 static TableCatalogEntry *GetCatalogTableEntry(LogicalOperator *op) {
-	if (op->type == LogicalOperatorType::LOGICAL_GET) {
-		auto get = (LogicalGet *)op;
-		TableCatalogEntry *entry = get->GetTable();
-		return entry;
+	if (!op) {
+		return nullptr;
 	}
-	for (auto &child : op->children) {
-		TableCatalogEntry *entry = GetCatalogTableEntry(child.get());
-		if (entry != nullptr) {
-			return entry;
-		}
-	}
-	return nullptr;
+	D_ASSERT(op->type == LogicalOperatorType::LOGICAL_GET);
+	auto get = (LogicalGet *)op;
+	TableCatalogEntry *entry = get->GetTable();
+	return entry;
 }
 
 // The filter was made on top of a logical sample or other projection,
@@ -304,23 +298,36 @@ static bool IsLogicalFilter(LogicalOperator *op) {
 	return op->type == LogicalOperatorType::LOGICAL_FILTER;
 }
 
-static LogicalGet *GetLogicalGet(LogicalOperator *op) {
+static LogicalGet *GetLogicalGet(LogicalOperator *op, idx_t table_index = DConstants::INVALID_INDEX) {
 	LogicalGet *get = nullptr;
 	switch (op->type) {
 	case LogicalOperatorType::LOGICAL_GET:
 		get = (LogicalGet *)op;
 		break;
 	case LogicalOperatorType::LOGICAL_FILTER:
-		get = GetLogicalGet(op->children.at(0).get());
+		get = GetLogicalGet(op->children.at(0).get(), table_index);
 		break;
 	case LogicalOperatorType::LOGICAL_PROJECTION:
-		get = GetLogicalGet(op->children.at(0).get());
+		get = GetLogicalGet(op->children.at(0).get(), table_index);
 		break;
 	case LogicalOperatorType::LOGICAL_COMPARISON_JOIN: {
 		LogicalComparisonJoin *join = (LogicalComparisonJoin *)op;
+		// We should never be calling GetLogicalGet without a valid table_index.
+		// We are attempting to get the catalog table for a relation (for statistics/cardinality estimation)
+		// A logical join means there is a non-reorderable relation in the join plan. This means we need
+		// to know the exact table index to return.
+		D_ASSERT(table_index != DConstants::INVALID_INDEX);
 		if (join->join_type == JoinType::MARK || join->join_type == JoinType::LEFT) {
 			auto child = join->children.at(0).get();
-			get = GetLogicalGet(child);
+			get = GetLogicalGet(child, table_index);
+			if (get && get->table_index == table_index) {
+				return get;
+			}
+			child = join->children.at(1).get();
+			get = GetLogicalGet(child, table_index);
+			if (get && get->table_index == table_index) {
+				return get;
+			}
 		}
 		break;
 	}
@@ -370,16 +377,17 @@ void CardinalityEstimator::InitCardinalityEstimatorProps(vector<NodeOp> *node_op
 		if (op->type == LogicalOperatorType::LOGICAL_COMPARISON_JOIN) {
 			auto &join = (LogicalComparisonJoin &)*op;
 			if (join.join_type == JoinType::LEFT) {
-				// TODO: inspect child operators to get a more accurate cost
-				// and cardinality estimation. If an base op is a Logical Comparison join
-				// it is probably a left join, so cost of the larger table is a fine
-				// estimate
-				// No need to update a mark join cost because I say so.
+				// If a base op is a Logical Comparison join it is probably a left join,
+				// so the cost of the larger table is a fine estimate.
+				// TODO: provide better estimates for cost of mark joins
+				// MARK joins are used for anti and semi joins, so the cost can conceivably be
+				// less than the base table cardinality.
 				join_node->SetCost(join_node->GetBaseTableCardinality());
 			}
 		}
-		// update cardinality with filters
+		// Total domains can be affected by filters. So we update base table cardinality first
 		EstimateBaseTableCardinality(join_node, op);
+		// Then update total domains.
 		UpdateTotalDomains(join_node, op);
 	}
 
@@ -390,74 +398,84 @@ void CardinalityEstimator::InitCardinalityEstimatorProps(vector<NodeOp> *node_op
 void CardinalityEstimator::UpdateTotalDomains(JoinNode *node, LogicalOperator *op) {
 	auto relation_id = node->set->relations[0];
 	relation_attributes[relation_id].cardinality = node->GetCardinality<double>();
+	//! Initialize the distinct count for all columns used in joins with the current relation.
+	idx_t distinct_count = node->GetBaseTableCardinality();
 	TableCatalogEntry *catalog_table = nullptr;
-	auto get = GetLogicalGet(op);
-	if (get) {
-		catalog_table = GetCatalogTableEntry(get);
-	}
-
-	//! Initialize the tdoms for all columns the relation uses in join conditions.
-	unordered_set<idx_t>::iterator ite;
-	idx_t count = node->GetBaseTableCardinality();
 
 	bool direct_filter = false;
+	LogicalGet *get = nullptr;
+	bool get_updated = true;
 	for (auto &column : relation_attributes[relation_id].columns) {
-		//! for every column in the relation, get the count via either HLL, or assume it to be
+		//! for every column used in a filter in the relation, get the distinct count via HLL, or assume it to be
 		//! the cardinality
 		ColumnBinding key = ColumnBinding(relation_id, column);
+		auto actual_binding = relation_column_to_original_column.find(key);
+		// each relation has columns that are either projected or used as filters
+		// In order to get column statistics we need to make sure the actual binding still
+		// refers to the same base table relation, as non-reorderable joins may involve 2+
+		// base table relations and therefore the columns may also refer to 2 different
+		// base table relations
+		if (actual_binding != relation_column_to_original_column.end() &&
+		    (!get || get->table_index != actual_binding->second.table_index)) {
+			get = GetLogicalGet(op, actual_binding->second.table_index);
+			get_updated = true;
+		} else {
+			get_updated = false;
+		}
 
-		if (catalog_table) {
-			relation_attributes[relation_id].original_name = catalog_table->name;
-			// Get HLL stats here
-			auto actual_binding = relation_column_to_original_column[key];
+		if (get_updated) {
+			catalog_table = GetCatalogTableEntry(get);
+		}
 
-			auto base_stats = catalog_table->GetStatistics(context, actual_binding.column_index);
+		if (catalog_table && actual_binding != relation_column_to_original_column.end()) {
+			// Get HLL stats here
+			auto base_stats = catalog_table->GetStatistics(context, actual_binding->second.column_index);
 			if (base_stats) {
-				count = base_stats->GetDistinctCount();
+				distinct_count = base_stats->GetDistinctCount();
 			}
 
-			// means you have a direct filter on a column. The count/total domain for the column
+			// means you have a direct filter on a column. The distinct_count/total domain for the column
 			// should be decreased to match the predicted total domain matching the filter.
 			// We decrease the total domain for all columns in the equivalence set because filter pushdown
 			// will mean all columns are affected.
 			if (direct_filter) {
-				count = node->GetCardinality<idx_t>();
+				distinct_count = node->GetCardinality<idx_t>();
 			}
 
-			// HLL has estimation error, count can't be greater than cardinality of the table before filters
-			if (count > node->GetBaseTableCardinality()) {
-				count = node->GetBaseTableCardinality();
+			// HLL has estimation error, distinct_count can't be greater than cardinality of the table before filters
+			if (distinct_count > node->GetBaseTableCardinality()) {
+				distinct_count = node->GetBaseTableCardinality();
 			}
 		} else {
-			// No HLL. So if we know there is a direct filter, reduce count to cardinality with filter
-			// otherwise assume the total domain is still the cardinality
+			// No HLL. So if we know there is a direct filter, reduce the distinct count to the cardinality
+			// with filter effects. Otherwise assume the distinct count is still the cardinality
 			if (direct_filter) {
-				count = node->GetCardinality<idx_t>();
+				distinct_count = node->GetCardinality<idx_t>();
 			} else {
-				count = node->GetBaseTableCardinality();
+				distinct_count = node->GetBaseTableCardinality();
 			}
 		}
-
+		// Update the relation_to_tdom set with the estimated distinct count (or tdom) calculated above
 		for (auto &relation_to_tdom : relations_to_tdoms) {
 			column_binding_set_t i_set = relation_to_tdom.equivalent_relations;
 			if (i_set.count(key) != 1) {
 				continue;
 			}
 			if (catalog_table) {
-				if (relation_to_tdom.tdom_hll < count) {
-					relation_to_tdom.tdom_hll = count;
+				if (relation_to_tdom.tdom_hll < distinct_count) {
+					relation_to_tdom.tdom_hll = distinct_count;
 					relation_to_tdom.has_tdom_hll = true;
 				}
-				if (relation_to_tdom.tdom_no_hll > count) {
-					relation_to_tdom.tdom_no_hll = count;
+				if (relation_to_tdom.tdom_no_hll > distinct_count) {
+					relation_to_tdom.tdom_no_hll = distinct_count;
 				}
 			} else {
 				// Here we don't have catalog statistics, and the following is how we determine
 				// the tdom
 				// 1. If there is any hll data in the equivalence set, use that
 				// 2. Otherwise, use the table with the smallest cardinality
-				if (relation_to_tdom.tdom_no_hll > count && !relation_to_tdom.has_tdom_hll) {
-					relation_to_tdom.tdom_no_hll = count;
+				if (relation_to_tdom.tdom_no_hll > distinct_count && !relation_to_tdom.has_tdom_hll) {
+					relation_to_tdom.tdom_no_hll = distinct_count;
 				}
 			}
 			break;
@@ -465,9 +483,8 @@ void CardinalityEstimator::UpdateTotalDomains(JoinNode *node, LogicalOperator *o
 	}
 }
 
-TableFilterSet *CardinalityEstimator::GetTableFilters(LogicalOperator *op) {
-	// First check table filters
-	auto get = GetLogicalGet(op);
+TableFilterSet *CardinalityEstimator::GetTableFilters(LogicalOperator *op, idx_t table_index) {
+	auto get = GetLogicalGet(op, table_index);
 	return get ? &get->table_filters : nullptr;
 }
 
@@ -529,9 +546,10 @@ idx_t CardinalityEstimator::InspectConjunctionOR(idx_t cardinality, idx_t column
 	return cardinality_after_filters;
 }
 
-idx_t CardinalityEstimator::InspectTableFilters(idx_t cardinality, LogicalOperator *op, TableFilterSet *table_filters) {
+idx_t CardinalityEstimator::InspectTableFilters(idx_t cardinality, LogicalOperator *op, TableFilterSet *table_filters,
+                                                idx_t table_index) {
 	idx_t cardinality_after_filters = cardinality;
-	auto get = GetLogicalGet(op);
+	auto get = GetLogicalGet(op, table_index);
 	unique_ptr<BaseStatistics> column_statistics;
 	for (auto &it : table_filters->filters) {
 		column_statistics = nullptr;
@@ -562,17 +580,30 @@ idx_t CardinalityEstimator::InspectTableFilters(idx_t cardinality, LogicalOperat
 
 void CardinalityEstimator::EstimateBaseTableCardinality(JoinNode *node, LogicalOperator *op) {
 	auto has_logical_filter = IsLogicalFilter(op);
-	auto table_filters = GetTableFilters(op);
+	D_ASSERT(node->set->count == 1);
+	auto relation_id = node->set->relations[0];
 
-	auto card_after_filters = node->GetBaseTableCardinality();
-	if (table_filters) {
-		double inspect_result = (double)InspectTableFilters(card_after_filters, op, table_filters);
-		card_after_filters = MinValue(inspect_result, (double)card_after_filters);
-	}
-	if (has_logical_filter) {
-		card_after_filters *= DEFAULT_SELECTIVITY;
+	double lowest_card_found = NumericLimits<double>::Maximum();
+	for (auto &column : relation_attributes[relation_id].columns) {
+		auto card_after_filters = node->GetBaseTableCardinality();
+		ColumnBinding key = ColumnBinding(relation_id, column);
+		TableFilterSet *table_filters = nullptr;
+		auto actual_binding = relation_column_to_original_column.find(key);
+		if (actual_binding != relation_column_to_original_column.end()) {
+			table_filters = GetTableFilters(op, actual_binding->second.table_index);
+		}
+
+		if (table_filters) {
+			double inspect_result =
+			    (double)InspectTableFilters(card_after_filters, op, table_filters, actual_binding->second.table_index);
+			card_after_filters = MinValue(inspect_result, (double)card_after_filters);
+		}
+		if (has_logical_filter) {
+			card_after_filters *= DEFAULT_SELECTIVITY;
+		}
+		lowest_card_found = MinValue(card_after_filters, lowest_card_found);
 	}
-	node->SetEstimatedCardinality(card_after_filters);
+	node->SetEstimatedCardinality(lowest_card_found);
 }
 
 } // namespace duckdb

package/src/duckdb/src/optimizer/join_order/join_order_optimizer.cpp

@@ -152,7 +152,7 @@ bool JoinOrderOptimizer::ExtractJoinRelations(LogicalOperator &input_op, vector<
 		// new NULL values in the right side, so pushing this condition through the join leads to incorrect results
 		// for this reason, we just start a new JoinOptimizer pass in each of the children of the join
 
-		// Keep track of all of the filter bindings the new join order optimizer makes
+		// Keep track of all filter bindings the new join order optimizer makes
 		vector<column_binding_map_t<ColumnBinding>> child_binding_maps;
 		idx_t child_bindings_it = 0;
 		for (auto &child : op->children) {
@@ -222,12 +222,25 @@ bool JoinOrderOptimizer::ExtractJoinRelations(LogicalOperator &input_op, vector<
 	}
 	case LogicalOperatorType::LOGICAL_PROJECTION: {
 		auto proj = (LogicalProjection *)op;
-		// we run the join order optimizer witin the subquery as well
+		// we run the join order optimizer within the subquery as well
 		JoinOrderOptimizer optimizer(context);
 		op->children[0] = optimizer.Optimize(std::move(op->children[0]));
 		// projection, add to the set of relations
 		auto relation = make_unique<SingleJoinRelation>(&input_op, parent);
-		relation_mapping[proj->table_index] = relations.size();
+		auto relation_id = relations.size();
+		// push one child column binding map back.
+		vector<column_binding_map_t<ColumnBinding>> child_binding_maps;
+		child_binding_maps.emplace_back(column_binding_map_t<ColumnBinding>());
+		optimizer.cardinality_estimator.CopyRelationMap(child_binding_maps.at(0));
+		// This logical projection may sit on top of a logical comparison join that has been pushed down
+		// we want to copy the binding info of both tables
+		relation_mapping[proj->table_index] = relation_id;
+		for (auto &binding_info : child_binding_maps.at(0)) {
+			cardinality_estimator.AddRelationToColumnMapping(
+			    ColumnBinding(proj->table_index, binding_info.first.column_index), binding_info.second);
+			cardinality_estimator.AddColumnToRelationMap(binding_info.second.table_index,
+			                                             binding_info.second.column_index);
+		}
 		relations.push_back(std::move(relation));
 		return true;
 	}

package/src/duckdb/src/optimizer/statistics/expression/propagate_aggregate.cpp

@@ -5,15 +5,21 @@ namespace duckdb {
 
 unique_ptr<BaseStatistics> StatisticsPropagator::PropagateExpression(BoundAggregateExpression &aggr,
                                                                      unique_ptr<Expression> *expr_ptr) {
-	vector<unique_ptr<BaseStatistics>> stats;
+	vector<BaseStatistics> stats;
 	stats.reserve(aggr.children.size());
 	for (auto &child : aggr.children) {
-		stats.push_back(PropagateExpression(child));
+		auto stat = PropagateExpression(child);
+		if (!stat) {
+			stats.push_back(BaseStatistics::CreateUnknown(child->return_type));
+		} else {
+			stats.push_back(stat->Copy());
+		}
 	}
 	if (!aggr.function.statistics) {
 		return nullptr;
 	}
-	return aggr.function.statistics(context, aggr, aggr.bind_info.get(), stats, node_stats.get());
+	AggregateStatisticsInput input(aggr.bind_info.get(), stats, node_stats.get());
+	return aggr.function.statistics(context, aggr, input);
 }
 
 } // namespace duckdb

package/src/duckdb/src/optimizer/statistics/expression/propagate_and_compress.cpp

@@ -5,7 +5,6 @@
 #include "duckdb/planner/expression/bound_constant_expression.hpp"
 #include "duckdb/planner/expression/bound_function_expression.hpp"
 #include "duckdb/storage/statistics/base_statistics.hpp"
-#include "duckdb/storage/statistics/numeric_statistics.hpp"
 #include "duckdb/common/operator/subtract.hpp"
 
 namespace duckdb {
@@ -44,14 +43,14 @@ bool GetCastType(hugeint_t range, LogicalType &cast_type) {
 }
 
 template <class T>
-unique_ptr<Expression> TemplatedCastToSmallestType(unique_ptr<Expression> expr, NumericStatistics &num_stats) {
+unique_ptr<Expression> TemplatedCastToSmallestType(unique_ptr<Expression> expr, BaseStatistics &stats) {
 	// Compute range
-	if (num_stats.min.IsNull() || num_stats.max.IsNull()) {
+	if (!NumericStats::HasMinMax(stats)) {
 		return expr;
 	}
 
-	auto signed_min_val = num_stats.min.GetValue<T>();
-	auto signed_max_val = num_stats.max.GetValue<T>();
+	auto signed_min_val = NumericStats::Min(stats).GetValue<T>();
+	auto signed_max_val = NumericStats::Max(stats).GetValue<T>();
 	if (signed_max_val < signed_min_val) {
 		return expr;
 	}
@@ -82,7 +81,7 @@ unique_ptr<Expression> TemplatedCastToSmallestType(unique_ptr<Expression> expr,
 	return BoundCastExpression::AddDefaultCastToType(std::move(minus_expr), cast_type);
 }
 
-unique_ptr<Expression> CastToSmallestType(unique_ptr<Expression> expr, NumericStatistics &num_stats) {
+unique_ptr<Expression> CastToSmallestType(unique_ptr<Expression> expr, BaseStatistics &num_stats) {
 	auto physical_type = expr->return_type.InternalType();
 	switch (physical_type) {
 	case PhysicalType::UINT8:
@@ -111,7 +110,7 @@ void StatisticsPropagator::PropagateAndCompress(unique_ptr<Expression> &expr, un
 	stats = PropagateExpression(expr);
 	if (stats) {
 		if (expr->return_type.IsIntegral()) {
-			expr = CastToSmallestType(std::move(expr), (NumericStatistics &)*stats);
+			expr = CastToSmallestType(std::move(expr), *stats);
 		}
 	}
 }

package/src/duckdb/src/optimizer/statistics/expression/propagate_cast.cpp

@@ -1,24 +1,27 @@
 #include "duckdb/optimizer/statistics_propagator.hpp"
 #include "duckdb/planner/expression/bound_cast_expression.hpp"
-#include "duckdb/storage/statistics/numeric_statistics.hpp"
 
 namespace duckdb {
 
-static unique_ptr<BaseStatistics> StatisticsOperationsNumericNumericCast(const BaseStatistics *input_p,
+static unique_ptr<BaseStatistics> StatisticsOperationsNumericNumericCast(const BaseStatistics &input,
                                                                          const LogicalType &target) {
-	auto &input = (NumericStatistics &)*input_p;
-
-	Value min = input.min, max = input.max;
+	if (!NumericStats::HasMinMax(input)) {
+		return nullptr;
+	}
+	Value min = NumericStats::Min(input);
+	Value max = NumericStats::Max(input);
 	if (!min.DefaultTryCastAs(target) || !max.DefaultTryCastAs(target)) {
 		// overflow in cast: bailout
 		return nullptr;
 	}
-	auto stats = make_unique<NumericStatistics>(target, std::move(min), std::move(max), input.stats_type);
-	stats->CopyBase(*input_p);
-	return std::move(stats);
+	auto result = NumericStats::CreateEmpty(target);
+	result.CopyBase(input);
+	NumericStats::SetMin(result, min);
+	NumericStats::SetMax(result, max);
+	return result.ToUnique();
 }
 
-static unique_ptr<BaseStatistics> StatisticsNumericCastSwitch(const BaseStatistics *input, const LogicalType &target) {
+static unique_ptr<BaseStatistics> StatisticsNumericCastSwitch(const BaseStatistics &input, const LogicalType &target) {
 	switch (target.InternalType()) {
 	case PhysicalType::INT8:
 	case PhysicalType::INT16:
@@ -48,13 +51,13 @@ unique_ptr<BaseStatistics> StatisticsPropagator::PropagateExpression(BoundCastEx
 	case PhysicalType::INT128:
 	case PhysicalType::FLOAT:
 	case PhysicalType::DOUBLE:
-		result_stats = StatisticsNumericCastSwitch(child_stats.get(), cast.return_type);
+		result_stats = StatisticsNumericCastSwitch(*child_stats, cast.return_type);
 		break;
 	default:
 		return nullptr;
 	}
 	if (cast.try_cast && result_stats) {
-		result_stats->validity_stats = make_unique<ValidityStatistics>(true, true);
+		result_stats->Set(StatsInfo::CAN_HAVE_NULL_VALUES);
 	}
 	return result_stats;
 }

package/src/duckdb/src/optimizer/statistics/expression/propagate_columnref.cpp

@@ -9,7 +9,7 @@ unique_ptr<BaseStatistics> StatisticsPropagator::PropagateExpression(BoundColumn
 	if (stats == statistics_map.end()) {
 		return nullptr;
 	}
-	return stats->second->Copy();
+	return stats->second->ToUnique();
 }
 
 } // namespace duckdb