ClickHouse/src/Interpreters/ExecuteScalarSubqueriesVisitor.cpp

358 lines
13 KiB
C++
Raw Normal View History

#include <Interpreters/ExecuteScalarSubqueriesVisitor.h>
#include <Columns/ColumnNullable.h>
#include <Columns/ColumnTuple.h>
#include <DataTypes/DataTypeNullable.h>
#include <DataTypes/DataTypeTuple.h>
#include <IO/WriteHelpers.h>
#include <Interpreters/Context.h>
#include <Interpreters/InterpreterSelectWithUnionQuery.h>
#include <Interpreters/addTypeConversionToAST.h>
#include <Interpreters/misc.h>
#include <Parsers/ASTExpressionList.h>
#include <Parsers/ASTFunction.h>
#include <Parsers/ASTLiteral.h>
#include <Parsers/ASTSelectQuery.h>
#include <Parsers/ASTSubquery.h>
#include <Parsers/ASTTablesInSelectQuery.h>
#include <Parsers/ASTWithElement.h>
2021-07-19 12:35:55 +00:00
#include <Parsers/queryToString.h>
#include <Processors/Executors/PullingAsyncPipelineExecutor.h>
#include <Common/ProfileEvents.h>
2023-05-25 22:54:54 +00:00
#include <Common/FieldVisitorToString.h>
#include <IO/WriteBufferFromString.h>
2021-01-13 18:58:00 +00:00
namespace ProfileEvents
{
extern const Event ScalarSubqueriesGlobalCacheHit;
extern const Event ScalarSubqueriesLocalCacheHit;
extern const Event ScalarSubqueriesCacheMiss;
}
2021-07-17 18:06:46 +00:00
namespace DB
{
namespace ErrorCodes
{
extern const int INCORRECT_RESULT_OF_SCALAR_SUBQUERY;
}
2019-01-09 16:16:59 +00:00
bool ExecuteScalarSubqueriesMatcher::needChildVisit(ASTPtr & node, const ASTPtr & child)
{
/// Processed
2019-03-11 13:22:51 +00:00
if (node->as<ASTSubquery>() || node->as<ASTFunction>())
return false;
/// Don't descend into subqueries in FROM section
2019-03-11 13:22:51 +00:00
if (node->as<ASTTableExpression>())
return false;
2020-09-12 17:00:04 +00:00
/// Do not go to subqueries defined in with statement
if (node->as<ASTWithElement>())
return false;
2019-03-11 13:22:51 +00:00
if (node->as<ASTSelectQuery>())
2019-01-09 16:16:59 +00:00
{
/// Do not go to FROM, JOIN, UNION.
2019-03-11 13:22:51 +00:00
if (child->as<ASTTableExpression>() || child->as<ASTSelectQuery>())
2019-01-09 16:16:59 +00:00
return false;
}
return true;
}
2019-02-22 13:33:56 +00:00
void ExecuteScalarSubqueriesMatcher::visit(ASTPtr & ast, Data & data)
{
2019-03-11 13:22:51 +00:00
if (const auto * t = ast->as<ASTSubquery>())
visit(*t, ast, data);
2019-03-11 13:22:51 +00:00
if (const auto * t = ast->as<ASTFunction>())
2019-02-22 13:33:56 +00:00
visit(*t, ast, data);
}
static auto getQueryInterpreter(const ASTSubquery & subquery, ExecuteScalarSubqueriesMatcher::Data & data)
{
auto subquery_context = Context::createCopy(data.getContext());
Settings subquery_settings = data.getContext()->getSettings();
subquery_settings.max_result_rows = 1;
subquery_settings.extremes = false;
subquery_context->setSettings(subquery_settings);
if (subquery_context->hasQueryContext())
{
/// When execute `INSERT INTO t WITH ... SELECT ...`, it may lead to `Unknown columns`
/// exception with this settings enabled(https://github.com/ClickHouse/ClickHouse/issues/52494).
subquery_context->getQueryContext()->setSetting("use_structure_from_insertion_table_in_table_functions", false);
if (!data.only_analyze)
{
/// Save current cached scalars in the context before analyzing the query
/// This is specially helpful when analyzing CTE scalars
auto context = subquery_context->getQueryContext();
for (const auto & it : data.scalars)
context->addScalar(it.first, it.second);
}
}
ASTPtr subquery_select = subquery.children.at(0);
auto options = SelectQueryOptions(QueryProcessingStage::Complete, data.subquery_depth + 1, true);
options.is_create_parameterized_view = data.is_create_parameterized_view;
options.analyze(data.only_analyze);
return std::make_unique<InterpreterSelectWithUnionQuery>(subquery_select, subquery_context, options);
}
void ExecuteScalarSubqueriesMatcher::visit(const ASTSubquery & subquery, ASTPtr & ast, Data & data)
{
2023-08-29 12:31:12 +00:00
/// subquery and ast can be the same object and ast will be moved.
/// Save these fields to avoid use after move.
String subquery_alias = subquery.alias;
bool prefer_alias_to_column_name = subquery.prefer_alias_to_column_name;
auto hash = subquery.getTreeHash(/*ignore_aliases=*/ true);
2023-07-06 00:35:44 +00:00
const auto scalar_query_hash_str = toString(hash);
2019-10-19 20:36:35 +00:00
std::unique_ptr<InterpreterSelectWithUnionQuery> interpreter = nullptr;
bool hit = false;
bool is_local = false;
2019-10-19 20:36:35 +00:00
Block scalar;
if (data.only_analyze)
{
/// Don't use scalar cache during query analysis
}
else if (data.local_scalars.contains(scalar_query_hash_str))
{
hit = true;
scalar = data.local_scalars[scalar_query_hash_str];
is_local = true;
ProfileEvents::increment(ProfileEvents::ScalarSubqueriesLocalCacheHit);
2021-07-24 01:25:00 +00:00
}
else if (data.scalars.contains(scalar_query_hash_str))
2021-07-24 01:25:00 +00:00
{
hit = true;
2019-10-19 20:36:35 +00:00
scalar = data.scalars[scalar_query_hash_str];
ProfileEvents::increment(ProfileEvents::ScalarSubqueriesGlobalCacheHit);
2021-07-24 01:25:00 +00:00
}
2019-10-19 20:36:35 +00:00
else
{
if (data.getContext()->hasQueryContext() && data.getContext()->getQueryContext()->hasScalar(scalar_query_hash_str))
{
if (!data.getContext()->getViewSource())
{
/// We aren't using storage views so we can safely use the context cache
scalar = data.getContext()->getQueryContext()->getScalar(scalar_query_hash_str);
ProfileEvents::increment(ProfileEvents::ScalarSubqueriesGlobalCacheHit);
hit = true;
}
else
{
/// If we are under a context that uses views that means that the cache might contain values that reference
/// the original table and not the view, so in order to be able to check the global cache we need to first
/// make sure that the query doesn't use the view
/// Note in any case the scalar will end up cached in *data* so this won't be repeated inside this context
interpreter = getQueryInterpreter(subquery, data);
if (!interpreter->usesViewSource())
{
scalar = data.getContext()->getQueryContext()->getScalar(scalar_query_hash_str);
ProfileEvents::increment(ProfileEvents::ScalarSubqueriesGlobalCacheHit);
hit = true;
}
}
}
}
2019-10-19 20:36:35 +00:00
if (!hit)
{
if (!interpreter)
interpreter = getQueryInterpreter(subquery, data);
2019-10-19 20:36:35 +00:00
ProfileEvents::increment(ProfileEvents::ScalarSubqueriesCacheMiss);
is_local = interpreter->usesViewSource();
2019-10-19 20:36:35 +00:00
Block block;
if (data.only_analyze)
{
/// If query is only analyzed, then constants are not correct.
block = interpreter->getSampleBlock();
for (auto & column : block)
2020-05-08 11:19:27 +00:00
{
if (column.column->empty())
2020-05-08 11:19:27 +00:00
{
auto mut_col = column.column->cloneEmpty();
mut_col->insertDefault();
column.column = std::move(mut_col);
}
}
}
else
{
auto io = interpreter->execute();
2019-10-19 20:36:35 +00:00
PullingAsyncPipelineExecutor executor(io.pipeline);
2022-05-27 20:47:35 +00:00
io.pipeline.setProgressCallback(data.getContext()->getProgressCallback());
2023-05-07 04:29:04 +00:00
while (block.rows() == 0 && executor.pull(block))
2023-05-07 04:16:30 +00:00
{
}
if (block.rows() == 0)
{
auto types = interpreter->getSampleBlock().getDataTypes();
if (types.size() != 1)
types = {std::make_shared<DataTypeTuple>(types)};
auto & type = types[0];
if (!type->isNullable())
{
if (!type->canBeInsideNullable())
throw Exception(ErrorCodes::INCORRECT_RESULT_OF_SCALAR_SUBQUERY,
"Scalar subquery returned empty result of type {} which cannot be Nullable",
type->getName());
type = makeNullable(type);
}
2021-07-19 12:35:55 +00:00
ASTPtr ast_new = std::make_shared<ASTLiteral>(Null());
ast_new = addTypeConversionToAST(std::move(ast_new), type->getName());
ast_new->setAlias(ast->tryGetAlias());
ast = std::move(ast_new);
/// Empty subquery result is equivalent to NULL
block = interpreter->getSampleBlock().cloneEmpty();
2023-08-29 12:31:12 +00:00
String column_name = block.columns() > 0 ? block.safeGetByPosition(0).name : "dummy";
block = Block({
ColumnWithTypeAndName(type->createColumnConstWithDefaultValue(1)->convertToFullColumnIfConst(), type, column_name)
});
}
if (block.rows() != 1)
throw Exception(ErrorCodes::INCORRECT_RESULT_OF_SCALAR_SUBQUERY, "Scalar subquery returned more than one row");
Block tmp_block;
2021-07-24 01:25:00 +00:00
while (tmp_block.rows() == 0 && executor.pull(tmp_block))
2023-05-07 04:16:30 +00:00
{
}
if (tmp_block.rows() != 0)
throw Exception(ErrorCodes::INCORRECT_RESULT_OF_SCALAR_SUBQUERY, "Scalar subquery returned more than one row");
}
2019-10-19 20:36:35 +00:00
block = materializeBlock(block);
size_t columns = block.columns();
if (columns == 1)
{
auto & column = block.getByPosition(0);
/// Here we wrap type to nullable if we can.
/// It is needed cause if subquery return no rows, it's result will be Null.
/// In case of many columns, do not check it cause tuple can't be nullable.
if (!column.type->isNullable() && column.type->canBeInsideNullable())
{
column.type = makeNullable(column.type);
column.column = makeNullable(column.column);
}
2019-10-19 20:36:35 +00:00
scalar = block;
}
else
2019-10-19 20:36:35 +00:00
{
2021-07-24 01:25:00 +00:00
scalar.insert({
ColumnTuple::create(block.getColumns()),
std::make_shared<DataTypeTuple>(block.getDataTypes()),
"tuple"});
2019-10-19 20:36:35 +00:00
}
}
const Settings & settings = data.getContext()->getSettingsRef();
2019-10-19 20:36:35 +00:00
// Always convert to literals when there is no query context.
2023-05-25 22:54:54 +00:00
if (data.only_analyze
|| !settings.enable_scalar_subquery_optimization
|| worthConvertingScalarToLiteral(scalar, data.max_literal_size)
|| !data.getContext()->hasQueryContext())
{
2019-10-19 20:36:35 +00:00
auto lit = std::make_unique<ASTLiteral>((*scalar.safeGetByPosition(0).column)[0]);
2023-08-29 12:31:12 +00:00
lit->alias = subquery_alias;
2021-07-16 14:17:22 +00:00
lit->prefer_alias_to_column_name = prefer_alias_to_column_name;
2019-10-19 20:36:35 +00:00
ast = addTypeConversionToAST(std::move(lit), scalar.safeGetByPosition(0).type->getName());
/// If only analyze was requested the expression is not suitable for constant folding, disable it.
if (data.only_analyze)
{
ast->as<ASTFunction>()->alias.clear();
auto func = makeASTFunction("__scalarSubqueryResult", std::move(ast));
2023-08-29 12:31:12 +00:00
func->alias = subquery_alias;
2021-07-16 14:17:22 +00:00
func->prefer_alias_to_column_name = prefer_alias_to_column_name;
ast = std::move(func);
}
}
2023-05-25 22:54:54 +00:00
else if (!data.replace_only_to_literals)
{
2019-10-19 20:36:35 +00:00
auto func = makeASTFunction("__getScalar", std::make_shared<ASTLiteral>(scalar_query_hash_str));
2023-08-29 12:31:12 +00:00
func->alias = subquery_alias;
func->prefer_alias_to_column_name = prefer_alias_to_column_name;
2019-10-19 20:36:35 +00:00
ast = std::move(func);
}
2019-10-19 20:36:35 +00:00
if (is_local)
data.local_scalars[scalar_query_hash_str] = std::move(scalar);
else
data.scalars[scalar_query_hash_str] = std::move(scalar);
}
2019-02-22 13:33:56 +00:00
void ExecuteScalarSubqueriesMatcher::visit(const ASTFunction & func, ASTPtr & ast, Data & data)
{
/// Don't descend into subqueries in arguments of IN operator.
2021-03-04 05:59:57 +00:00
/// But if an argument is not subquery, then deeper may be scalar subqueries and we need to descend in them.
std::vector<ASTPtr *> out;
if (checkFunctionIsInOrGlobalInOperator(func))
{
for (auto & child : ast->children)
{
if (child != func.arguments)
out.push_back(&child);
else
for (size_t i = 0, size = func.arguments->children.size(); i < size; ++i)
2019-03-11 13:22:51 +00:00
if (i != 1 || !func.arguments->children[i]->as<ASTSubquery>())
out.push_back(&func.arguments->children[i]);
}
}
else
for (auto & child : ast->children)
out.push_back(&child);
2019-02-22 13:33:56 +00:00
for (ASTPtr * add_node : out)
2019-02-22 15:45:47 +00:00
Visitor(data).visit(*add_node);
}
2023-05-25 22:54:54 +00:00
static size_t getSizeOfSerializedLiteral(const Field & field)
{
auto field_str = applyVisitor(FieldVisitorToString(), field);
return field_str.size();
}
bool worthConvertingScalarToLiteral(const Block & scalar, std::optional<size_t> max_literal_size)
{
/// Converting to literal values might take a fair amount of overhead when the value is large, (e.g.
/// Array, BitMap, etc.), This conversion is required for constant folding, index lookup, branch
/// elimination. However, these optimizations should never be related to large values, thus we blacklist them here.
const auto * scalar_type_name = scalar.safeGetByPosition(0).type->getFamilyName();
static const std::set<std::string_view> maybe_large_literal_types = {"Array", "Tuple", "AggregateFunction", "Function", "Set", "LowCardinality"};
if (!maybe_large_literal_types.contains(scalar_type_name))
return true;
if (!max_literal_size)
return false;
/// Size of serialized literal cannot be less than size in bytes.
if (scalar.bytes() > *max_literal_size)
return false;
return getSizeOfSerializedLiteral((*scalar.safeGetByPosition(0).column)[0]) <= *max_literal_size;
}
}