physical::PhysicalPtr InjectJoinFilters::pushDownFiltersInternal(
const physical::PhysicalPtr &probe_child,
const physical::PhysicalPtr &build_child,
const physical::FilterJoinPtr &filter_join) const {
switch (probe_child->getPhysicalType()) {
case P::PhysicalType::kAggregate: // Fall through
case P::PhysicalType::kHashJoin:
case P::PhysicalType::kSample:
case P::PhysicalType::kSelection:
case P::PhysicalType::kSort:
case P::PhysicalType::kWindowAggregate: {
DCHECK_GE(probe_child->getNumChildren(), 1u);
const P::PhysicalPtr child = probe_child->children()[0];
if (E::SubsetOfExpressions(filter_join->probe_attributes(),
child->getOutputAttributes())) {
const P::PhysicalPtr new_child =
pushDownFiltersInternal(child, build_child, filter_join);
if (new_child != child) {
std::vector<P::PhysicalPtr> new_children = probe_child->children();
new_children[0] = new_child;
return probe_child->copyWithNewChildren(new_children);
}
}
}
default:
break;
}
if (probe_child != filter_join->left()) {
// TODO(jianqiao): may need to update probe_attributes.
return P::FilterJoin::Create(probe_child,
build_child,
filter_join->probe_attributes(),
filter_join->build_attributes(),
E::ToNamedExpressions(probe_child->getOutputAttributes()),
filter_join->build_side_filter_predicate(),
filter_join->is_anti_join(),
filter_join->hasRepartition(),
filter_join->cloneOutputPartitionSchemeHeader());
} else {
return filter_join;
}
}
P::PhysicalPtr ReduceGroupByAttributes::applyToNode(const P::PhysicalPtr &input) {
P::TableReferencePtr table_reference;
if (P::SomeTableReference::MatchesWithConditionalCast(input, &table_reference)) {
// Collect the attributes-to-TableReference mapping info.
for (const auto &attr : table_reference->attribute_list()) {
source_.emplace(attr->id(), std::make_pair(table_reference, attr));
}
return input;
}
P::AggregatePtr aggregate;
if (!P::SomeAggregate::MatchesWithConditionalCast(input, &aggregate) ||
aggregate->grouping_expressions().size() <= 1u) {
return input;
}
// Divide the group-by attributes into groups based on their source table.
std::map<P::TableReferencePtr, std::vector<E::AttributeReferencePtr>> table_attributes;
for (const auto &expr : aggregate->grouping_expressions()) {
const auto source_it = source_.find(expr->id());
if (source_it != source_.end()) {
table_attributes[source_it->second.first].emplace_back(source_it->second.second);
}
}
std::unordered_set<E::ExprId> erased_grouping_attr_ids;
std::vector<std::pair<P::TableReferencePtr, E::AttributeReferencePtr>> hoisted_tables;
// For each group (i.e. each source table), if it is profitable then we pull
// the table up the aggregation.
for (const auto &pair : table_attributes) {
const P::TableReferencePtr table = pair.first;
const std::vector<E::AttributeReferencePtr> &attributes = pair.second;
// TODO(jianqiao): find a cost-based metic instead of hard-coding the threshold
// number of group-by attributes.
if (attributes.size() <= FLAGS_reduce_group_by_attributes_threshold) {
continue;
}
std::vector<AttributeInfo> attr_infos;
for (const auto &attr : attributes) {
attr_infos.emplace_back(attr,
cost_model_->impliesUniqueAttributes(table, {attr}),
!attr->getValueType().isVariableLength(),
attr->getValueType().maximumByteLength());
}
std::vector<const AttributeInfo *> attr_info_refs;
for (const auto &info : attr_infos) {
attr_info_refs.emplace_back(&info);
}
std::sort(attr_info_refs.begin(),
attr_info_refs.end(),
AttributeInfo::IsBetterThan);
const AttributeInfo &best_candidate = *attr_info_refs.front();
if (!best_candidate.is_unique) {
// Cannot find a key attribute. Give up pulling this table up.
continue;
}
const E::AttributeReferencePtr key_attribute = best_candidate.attribute;
hoisted_tables.emplace_back(table, key_attribute);
for (const auto &attr : attributes) {
if (attr->id() != key_attribute->id()) {
erased_grouping_attr_ids.emplace(attr->id());
}
}
}
if (erased_grouping_attr_ids.empty()) {
return input;
}
// Reconstuct the Aggregate node with reduced group-by attributes and then
// construct HashJoin nodes on top of the Aggregate.
std::vector<E::NamedExpressionPtr> reduced_grouping_expressions;
for (const auto &expr : aggregate->grouping_expressions()) {
if (erased_grouping_attr_ids.find(expr->id()) == erased_grouping_attr_ids.end()) {
reduced_grouping_expressions.emplace_back(expr);
}
}
const P::AggregatePtr new_aggregate =
P::Aggregate::Create(aggregate->input(),
reduced_grouping_expressions,
aggregate->aggregate_expressions(),
aggregate->filter_predicate());
P::PhysicalPtr output = new_aggregate;
std::vector<E::NamedExpressionPtr> project_expressions =
E::ToNamedExpressions(output->getOutputAttributes());
for (const auto &pair : hoisted_tables) {
const P::TableReferencePtr &source_table = pair.first;
const E::AttributeReferencePtr &probe_attribute = pair.second;
E::AttributeReferencePtr build_attribute;
std::vector<E::AttributeReferencePtr> new_attribute_list;
for (const auto &attr : source_table->attribute_list()) {
if (attr->id() == probe_attribute->id()) {
build_attribute =
E::AttributeReference::Create(optimizer_context_->nextExprId(),
attr->attribute_name(),
attr->attribute_alias(),
attr->relation_name(),
attr->getValueType(),
E::AttributeReferenceScope::kLocal);
new_attribute_list.emplace_back(build_attribute);
} else {
new_attribute_list.emplace_back(attr);
project_expressions.emplace_back(attr);
}
}
DCHECK(build_attribute != nullptr);
const P::TableReferencePtr build_side_table =
P::TableReference::Create(source_table->relation(),
source_table->relation()->getName(),
new_attribute_list);
output = P::HashJoin::Create(output,
build_side_table,
{probe_attribute},
{build_attribute},
nullptr,
project_expressions,
P::HashJoin::JoinType::kInnerJoin);
}
return output;
}
bool OneToOne::generatePlan(const L::LogicalPtr &logical_input,
P::PhysicalPtr *physical_output) {
switch (logical_input->getLogicalType()) {
case L::LogicalType::kTopLevelPlan: {
const L::TopLevelPlanPtr top_level_plan = std::static_pointer_cast<const L::TopLevelPlan>(logical_input);
const P::PhysicalPtr main_physical_plan =
physical_mapper_->createOrGetPhysicalFromLogical(top_level_plan->plan());
std::vector<P::PhysicalPtr> shared_physical_subplans;
for (const L::LogicalPtr &shared_logical_subplan : top_level_plan->shared_subplans()) {
shared_physical_subplans.emplace_back(
physical_mapper_->createOrGetPhysicalFromLogical(shared_logical_subplan));
}
*physical_output = P::TopLevelPlan::Create(main_physical_plan,
shared_physical_subplans);
return true;
}
case L::LogicalType::kSharedSubplanReference: {
const L::SharedSubplanReferencePtr shared_subplan_reference =
std::static_pointer_cast<const L::SharedSubplanReference>(logical_input);
*physical_output = P::SharedSubplanReference::Create(shared_subplan_reference->subplan_id(),
shared_subplan_reference->referenced_attributes(),
shared_subplan_reference->output_attributes());
return true;
}
case L::LogicalType::kTableReference: {
const L::TableReferencePtr table_reference =
std::static_pointer_cast<const L::TableReference>(logical_input);
*physical_output = P::TableReference::Create(table_reference->catalog_relation(),
table_reference->relation_alias(),
table_reference->attribute_list());
return true;
}
case L::LogicalType::kCopyFrom: {
const L::CopyFromPtr copy_from =
std::static_pointer_cast<const L::CopyFrom>(logical_input);
*physical_output = P::CopyFrom::Create(
copy_from->catalog_relation(), copy_from->file_name(),
copy_from->column_delimiter(), copy_from->escape_strings());
return true;
}
case L::LogicalType::kCreateIndex: {
const L::CreateIndexPtr create_index =
std::static_pointer_cast<const L::CreateIndex>(logical_input);
*physical_output = P::CreateIndex::Create(physical_mapper_->createOrGetPhysicalFromLogical(
create_index->input()),
create_index->index_name(),
create_index->index_attributes(),
create_index->index_description());
return true;
}
case L::LogicalType::kCreateTable: {
const L::CreateTablePtr create_table =
std::static_pointer_cast<const L::CreateTable>(logical_input);
*physical_output = P::CreateTable::Create(create_table->relation_name(),
create_table->attributes(),
create_table->block_properties());
return true;
}
case L::LogicalType::kDeleteTuples: {
const L::DeleteTuplesPtr delete_tuples =
std::static_pointer_cast<const L::DeleteTuples>(logical_input);
*physical_output = P::DeleteTuples::Create(
physical_mapper_->createOrGetPhysicalFromLogical(delete_tuples->input()),
delete_tuples->predicate());
return true;
}
case L::LogicalType::kDropTable: {
const L::DropTablePtr drop_table =
std::static_pointer_cast<const L::DropTable>(logical_input);
*physical_output = P::DropTable::Create(drop_table->catalog_relation());
return true;
}
case L::LogicalType::kInsertSelection: {
const L::InsertSelectionPtr insert_selection =
std::static_pointer_cast<const L::InsertSelection>(logical_input);
*physical_output = P::InsertSelection::Create(
physical_mapper_->createOrGetPhysicalFromLogical(insert_selection->destination()),
physical_mapper_->createOrGetPhysicalFromLogical(insert_selection->selection()));
return true;
}
case L::LogicalType::kInsertTuple: {
const L::InsertTuplePtr insert_tuple =
std::static_pointer_cast<const L::InsertTuple>(logical_input);
*physical_output = P::InsertTuple::Create(
physical_mapper_->createOrGetPhysicalFromLogical(insert_tuple->input()),
insert_tuple->column_values());
return true;
}
case L::LogicalType::kSample: {
const L::SamplePtr sample =
std::static_pointer_cast<const L::Sample>(logical_input);
*physical_output = P::Sample::Create(
physical_mapper_->createOrGetPhysicalFromLogical(sample->input()),
sample->is_block_sample(),
sample->percentage());
return true;
}
case L::LogicalType::kSort: {
const L::Sort *sort =
static_cast<const L::Sort*>(logical_input.get());
const P::PhysicalPtr physical_input =
physical_mapper_->createOrGetPhysicalFromLogical(sort->input());
// Find non-sort attributes.
const std::vector<E::AttributeReferencePtr> input_attributes =
physical_input->getOutputAttributes();
E::UnorderedNamedExpressionSet sort_attributes_set(sort->sort_attributes().begin(),
sort->sort_attributes().end());
std::vector<E::AttributeReferencePtr> non_sort_attributes;
for (const E::AttributeReferencePtr &input_attribute : input_attributes) {
if (sort_attributes_set.find(input_attribute) == sort_attributes_set.end()) {
non_sort_attributes.emplace_back(input_attribute);
}
}
*physical_output = P::Sort::Create(
physical_mapper_->createOrGetPhysicalFromLogical(sort->input()),
sort->sort_attributes(),
non_sort_attributes,
sort->sort_ascending(),
sort->nulls_first_flags(),
sort->limit());
return true;
}
case L::LogicalType::kTableGenerator: {
const L::TableGeneratorPtr table_generator =
std::static_pointer_cast<const L::TableGenerator>(logical_input);
*physical_output = P::TableGenerator::Create(
table_generator->generator_function_handle(),
table_generator->table_alias(),
table_generator->attribute_list());
return true;
}
case L::LogicalType::kUpdateTable: {
const L::UpdateTablePtr update_table =
std::static_pointer_cast<const L::UpdateTable>(logical_input);
*physical_output = P::UpdateTable::Create(
physical_mapper_->createOrGetPhysicalFromLogical(update_table->input()),
update_table->assignees(),
update_table->assignment_expressions(),
update_table->predicate());
return true;
}
default:
return false;
}
}