thevar1able/ClickHouse
1
0
Fork 0
mirror of https://github.com/ClickHouse/ClickHouse.git synced 2024-12-16 19:32:07 +00:00
Code Issues Packages Projects Releases Wiki Activity
cfe2207be3
ClickHouse/src/Storages/MergeTree/KeyCondition.cpp
Kruglov Pavel f559c34113
Merge pull request #28302 from amosbird/binaryconstantwrap 
Always monotonic for non-zero division
2021-09-03 20:02:54 +03:00

2240 lines
79 KiB
C++
Raw Blame History

#include <Storages/MergeTree/KeyCondition.h>
#include <Storages/MergeTree/BoolMask.h>
#include <DataTypes/DataTypesNumber.h>
#include <DataTypes/FieldToDataType.h>
#include <DataTypes/getLeastSupertype.h>
#include <Interpreters/TreeRewriter.h>
#include <Interpreters/ExpressionAnalyzer.h>
#include <Interpreters/ExpressionActions.h>
#include <Interpreters/castColumn.h>
#include <Interpreters/misc.h>
#include <Functions/FunctionFactory.h>
#include <Functions/FunctionsConversion.h>
#include <Functions/CastOverloadResolver.h>
#include <Functions/IFunction.h>
#include <Common/FieldVisitorsAccurateComparison.h>
#include <Common/FieldVisitorToString.h>
#include <Common/typeid_cast.h>
#include <Interpreters/convertFieldToType.h>
#include <Interpreters/Set.h>
#include <Parsers/queryToString.h>
#include <Parsers/ASTLiteral.h>
#include <Parsers/ASTSubquery.h>
#include <Parsers/ASTIdentifier.h>
#include <IO/WriteBufferFromString.h>
#include <IO/Operators.h>
#include <Storages/KeyDescription.h>
#include <cassert>
#include <stack>
#include <limits>
namespace DB
{
namespace ErrorCodes
{
extern const int LOGICAL_ERROR;
extern const int BAD_TYPE_OF_FIELD;
}
String Range::toString() const
{
WriteBufferFromOwnString str;
str << (left_included ? '[' : '(') << applyVisitor(FieldVisitorToString(), left) << ", ";
str << applyVisitor(FieldVisitorToString(), right) << (right_included ? ']' : ')');
return str.str();
}
/// Example: for `Hello\_World% ...` string it returns `Hello_World`, and for `%test%` returns an empty string.
static String extractFixedPrefixFromLikePattern(const String & like_pattern)
{
String fixed_prefix;
const char * pos = like_pattern.data();
const char * end = pos + like_pattern.size();
while (pos < end)
{
switch (*pos)
{
case '%':
[[fallthrough]];
case '_':
return fixed_prefix;
case '\\':
++pos;
if (pos == end)
break;
[[fallthrough]];
default:
fixed_prefix += *pos;
break;
}
++pos;
}
return fixed_prefix;
}
/** For a given string, get a minimum string that is strictly greater than all strings with this prefix,
* or return an empty string if there are no such strings.
*/
static String firstStringThatIsGreaterThanAllStringsWithPrefix(const String & prefix)
{
/** Increment the last byte of the prefix by one. But if it is max (255), then remove it and increase the previous one.
* Example (for convenience, suppose that the maximum value of byte is `z`)
* abcx -> abcy
* abcz -> abd
* zzz -> empty string
* z -> empty string
*/
String res = prefix;
while (!res.empty() && static_cast<UInt8>(res.back()) == std::numeric_limits<UInt8>::max())
res.pop_back();
if (res.empty())
return res;
res.back() = static_cast<char>(1 + static_cast<UInt8>(res.back()));
return res;
}
/// A dictionary containing actions to the corresponding functions to turn them into `RPNElement`
const KeyCondition::AtomMap KeyCondition::atom_map
{
{
"notEquals",
[] (RPNElement & out, const Field & value)
{
out.function = RPNElement::FUNCTION_NOT_IN_RANGE;
out.range = Range(value);
return true;
}
},
{
"equals",
[] (RPNElement & out, const Field & value)
{
out.function = RPNElement::FUNCTION_IN_RANGE;
out.range = Range(value);
return true;
}
},
{
"less",
[] (RPNElement & out, const Field & value)
{
out.function = RPNElement::FUNCTION_IN_RANGE;
out.range = Range::createRightBounded(value, false);
return true;
}
},
{
"greater",
[] (RPNElement & out, const Field & value)
{
out.function = RPNElement::FUNCTION_IN_RANGE;
out.range = Range::createLeftBounded(value, false);
return true;
}
},
{
"lessOrEquals",
[] (RPNElement & out, const Field & value)
{
out.function = RPNElement::FUNCTION_IN_RANGE;
out.range = Range::createRightBounded(value, true);
return true;
}
},
{
"greaterOrEquals",
[] (RPNElement & out, const Field & value)
{
out.function = RPNElement::FUNCTION_IN_RANGE;
out.range = Range::createLeftBounded(value, true);
return true;
}
},
{
"in",
[] (RPNElement & out, const Field &)
{
out.function = RPNElement::FUNCTION_IN_SET;
return true;
}
},
{
"notIn",
[] (RPNElement & out, const Field &)
{
out.function = RPNElement::FUNCTION_NOT_IN_SET;
return true;
}
},
{
"globalIn",
[] (RPNElement & out, const Field &)
{
out.function = RPNElement::FUNCTION_IN_SET;
return true;
}
},
{
"globalNotIn",
[] (RPNElement & out, const Field &)
{
out.function = RPNElement::FUNCTION_NOT_IN_SET;
return true;
}
},
{
"nullIn",
[] (RPNElement & out, const Field &)
{
out.function = RPNElement::FUNCTION_IN_SET;
return true;
}
},
{
"notNullIn",
[] (RPNElement & out, const Field &)
{
out.function = RPNElement::FUNCTION_NOT_IN_SET;
return true;
}
},
{
"globalNullIn",
[] (RPNElement & out, const Field &)
{
out.function = RPNElement::FUNCTION_IN_SET;
return true;
}
},
{
"globalNotNullIn",
[] (RPNElement & out, const Field &)
{
out.function = RPNElement::FUNCTION_NOT_IN_SET;
return true;
}
},
{
"empty",
[] (RPNElement & out, const Field & value)
{
if (value.getType() != Field::Types::String)
return false;
out.function = RPNElement::FUNCTION_IN_RANGE;
out.range = Range("");
return true;
}
},
{
"notEmpty",
[] (RPNElement & out, const Field & value)
{
if (value.getType() != Field::Types::String)
return false;
out.function = RPNElement::FUNCTION_NOT_IN_RANGE;
out.range = Range("");
return true;
}
},
{
"like",
[] (RPNElement & out, const Field & value)
{
if (value.getType() != Field::Types::String)
return false;
String prefix = extractFixedPrefixFromLikePattern(value.get<const String &>());
if (prefix.empty())
return false;
String right_bound = firstStringThatIsGreaterThanAllStringsWithPrefix(prefix);
out.function = RPNElement::FUNCTION_IN_RANGE;
out.range = !right_bound.empty()
? Range(prefix, true, right_bound, false)
: Range::createLeftBounded(prefix, true);
return true;
}
},
{
"notLike",
[] (RPNElement & out, const Field & value)
{
if (value.getType() != Field::Types::String)
return false;
String prefix = extractFixedPrefixFromLikePattern(value.get<const String &>());
if (prefix.empty())
return false;
String right_bound = firstStringThatIsGreaterThanAllStringsWithPrefix(prefix);
out.function = RPNElement::FUNCTION_NOT_IN_RANGE;
out.range = !right_bound.empty()
? Range(prefix, true, right_bound, false)
: Range::createLeftBounded(prefix, true);
return true;
}
},
{
"startsWith",
[] (RPNElement & out, const Field & value)
{
if (value.getType() != Field::Types::String)
return false;
String prefix = value.get<const String &>();
if (prefix.empty())
return false;
String right_bound = firstStringThatIsGreaterThanAllStringsWithPrefix(prefix);
out.function = RPNElement::FUNCTION_IN_RANGE;
out.range = !right_bound.empty()
? Range(prefix, true, right_bound, false)
: Range::createLeftBounded(prefix, true);
return true;
}
},
{
"isNotNull",
[] (RPNElement & out, const Field &)
{
out.function = RPNElement::FUNCTION_IS_NOT_NULL;
// isNotNull means (-Inf, +Inf), which is the default Range
out.range = Range();
return true;
}
},
{
"isNull",
[] (RPNElement & out, const Field &)
{
out.function = RPNElement::FUNCTION_IS_NULL;
// When using NULL_LAST, isNull means [+Inf, +Inf]
out.range = Range(Field(POSITIVE_INFINITY));
return true;
}
}
};
static const std::map<std::string, std::string> inverse_relations = {
{"equals", "notEquals"},
{"notEquals", "equals"},
{"less", "greaterOrEquals"},
{"greaterOrEquals", "less"},
{"greater", "lessOrEquals"},
{"lessOrEquals", "greater"},
{"in", "notIn"},
{"notIn", "in"},
{"globalIn", "globalNotIn"},
{"globalNotIn", "globalIn"},
{"nullIn", "notNullIn"},
{"notNullIn", "nullIn"},
{"globalNullIn", "globalNotNullIn"},
{"globalNullNotIn", "globalNullIn"},
{"isNull", "isNotNull"},
{"isNotNull", "isNull"},
{"like", "notLike"},
{"notLike", "like"},
{"empty", "notEmpty"},
{"notEmpty", "empty"},
};
bool isLogicalOperator(const String & func_name)
{
return (func_name == "and" || func_name == "or" || func_name == "not" || func_name == "indexHint");
}
/// The node can be one of:
/// - Logical operator (AND, OR, NOT and indexHint() - logical NOOP)
/// - An "atom" (relational operator, constant, expression)
/// - A logical constant expression
/// - Any other function
ASTPtr cloneASTWithInversionPushDown(const ASTPtr node, const bool need_inversion = false)
{
const ASTFunction * func = node->as<ASTFunction>();
if (func && isLogicalOperator(func->name))
{
if (func->name == "not")
{
return cloneASTWithInversionPushDown(func->arguments->children.front(), !need_inversion);
}
const auto result_node = makeASTFunction(func->name);
/// indexHint() is a special case - logical NOOP function
if (result_node->name != "indexHint" && need_inversion)
{
result_node->name = (result_node->name == "and") ? "or" : "and";
}
if (func->arguments)
{
for (const auto & child : func->arguments->children)
{
result_node->arguments->children.push_back(cloneASTWithInversionPushDown(child, need_inversion));
}
}
return result_node;
}
auto cloned_node = node->clone();
if (func && inverse_relations.find(func->name) != inverse_relations.cend())
{
if (need_inversion)
{
cloned_node->as<ASTFunction>()->name = inverse_relations.at(func->name);
}
return cloned_node;
}
return need_inversion ? makeASTFunction("not", cloned_node) : cloned_node;
}
inline bool Range::equals(const Field & lhs, const Field & rhs) { return applyVisitor(FieldVisitorAccurateEquals(), lhs, rhs); }
inline bool Range::less(const Field & lhs, const Field & rhs) { return applyVisitor(FieldVisitorAccurateLess(), lhs, rhs); }
/** Calculate expressions, that depend only on constants.
* For index to work when something like "WHERE Date = toDate(now())" is written.
*/
Block KeyCondition::getBlockWithConstants(
const ASTPtr & query, const TreeRewriterResultPtr & syntax_analyzer_result, ContextPtr context)
{
Block result
{
{ DataTypeUInt8().createColumnConstWithDefaultValue(1), std::make_shared<DataTypeUInt8>(), "_dummy" }
};
const auto expr_for_constant_folding = ExpressionAnalyzer(query, syntax_analyzer_result, context).getConstActions();
expr_for_constant_folding->execute(result);
return result;
}
static NameSet getAllSubexpressionNames(const ExpressionActions & key_expr)
{
NameSet names;
for (const auto & action : key_expr.getActions())
names.insert(action.node->result_name);
return names;
}
KeyCondition::KeyCondition(
const SelectQueryInfo & query_info,
ContextPtr context,
const Names & key_column_names,
const ExpressionActionsPtr & key_expr_,
bool single_point_,
bool strict_)
: key_expr(key_expr_)
, key_subexpr_names(getAllSubexpressionNames(*key_expr))
, prepared_sets(query_info.sets)
, single_point(single_point_)
, strict(strict_)
{
for (size_t i = 0, size = key_column_names.size(); i < size; ++i)
{
std::string name = key_column_names[i];
if (!key_columns.count(name))
key_columns[name] = i;
}
/** Evaluation of expressions that depend only on constants.
* For the index to be used, if it is written, for example `WHERE Date = toDate(now())`.
*/
Block block_with_constants = getBlockWithConstants(query_info.query, query_info.syntax_analyzer_result, context);
for (const auto & [name, _] : query_info.syntax_analyzer_result->array_join_result_to_source)
array_joined_columns.insert(name);
const ASTSelectQuery & select = query_info.query->as<ASTSelectQuery &>();
if (select.where() || select.prewhere())
{
ASTPtr filter_query;
if (select.where() && select.prewhere())
filter_query = makeASTFunction("and", select.where(), select.prewhere());
else
filter_query = select.where() ? select.where() : select.prewhere();
/** When non-strictly monotonic functions are employed in functional index (e.g. ORDER BY toStartOfHour(dateTime)),
* the use of NOT operator in predicate will result in the indexing algorithm leave out some data.
* This is caused by rewriting in KeyCondition::tryParseAtomFromAST of relational operators to less strict
* when parsing the AST into internal RPN representation.
* To overcome the problem, before parsing the AST we transform it to its semantically equivalent form where all NOT's
* are pushed down and applied (when possible) to leaf nodes.
*/
traverseAST(cloneASTWithInversionPushDown(filter_query), context, block_with_constants);
}
else
{
rpn.emplace_back(RPNElement::FUNCTION_UNKNOWN);
}
}
bool KeyCondition::addCondition(const String & column, const Range & range)
{
if (!key_columns.count(column))
return false;
rpn.emplace_back(RPNElement::FUNCTION_IN_RANGE, key_columns[column], range);
rpn.emplace_back(RPNElement::FUNCTION_AND);
return true;
}
/** Computes value of constant expression and its data type.
* Returns false, if expression isn't constant.
*/
bool KeyCondition::getConstant(const ASTPtr & expr, Block & block_with_constants, Field & out_value, DataTypePtr & out_type)
{
// Constant expr should use alias names if any
String column_name = expr->getColumnName();
if (const auto * lit = expr->as<ASTLiteral>())
{
/// By default block_with_constants has only one column named "_dummy".
/// If block contains only constants it's may not be preprocessed by
// ExpressionAnalyzer, so try to look up in the default column.
if (!block_with_constants.has(column_name))
column_name = "_dummy";
/// Simple literal
out_value = lit->value;
out_type = block_with_constants.getByName(column_name).type;
/// If constant is not Null, we can assume it's type is not Nullable as well.
if (!out_value.isNull())
out_type = removeNullable(out_type);
return true;
}
else if (block_with_constants.has(column_name) && isColumnConst(*block_with_constants.getByName(column_name).column))
{
/// An expression which is dependent on constants only
const auto & expr_info = block_with_constants.getByName(column_name);
out_value = (*expr_info.column)[0];
out_type = expr_info.type;
if (!out_value.isNull())
out_type = removeNullable(out_type);
return true;
}
else
return false;
}
static Field applyFunctionForField(
const FunctionBasePtr & func,
const DataTypePtr & arg_type,
const Field & arg_value)
{
ColumnsWithTypeAndName columns
{
{ arg_type->createColumnConst(1, arg_value), arg_type, "x" },
};
auto col = func->execute(columns, func->getResultType(), 1);
return (*col)[0];
}
/// The case when arguments may have types different than in the primary key.
static std::pair<Field, DataTypePtr> applyFunctionForFieldOfUnknownType(
const FunctionBasePtr & func,
const DataTypePtr & arg_type,
const Field & arg_value)
{
ColumnsWithTypeAndName arguments{{ arg_type->createColumnConst(1, arg_value), arg_type, "x" }};
DataTypePtr return_type = func->getResultType();
auto col = func->execute(arguments, return_type, 1);
Field result = (*col)[0];
return {std::move(result), std::move(return_type)};
}
/// Same as above but for binary operators
static std::pair<Field, DataTypePtr> applyBinaryFunctionForFieldOfUnknownType(
const FunctionOverloadResolverPtr & func,
const DataTypePtr & arg_type,
const Field & arg_value,
const DataTypePtr & arg_type2,
const Field & arg_value2)
{
ColumnsWithTypeAndName arguments{
{arg_type->createColumnConst(1, arg_value), arg_type, "x"}, {arg_type2->createColumnConst(1, arg_value2), arg_type2, "y"}};
FunctionBasePtr func_base = func->build(arguments);
DataTypePtr return_type = func_base->getResultType();
auto col = func_base->execute(arguments, return_type, 1);
Field result = (*col)[0];
return {std::move(result), std::move(return_type)};
}
static FieldRef applyFunction(const FunctionBasePtr & func, const DataTypePtr & current_type, const FieldRef & field)
{
/// Fallback for fields without block reference.
if (field.isExplicit())
return applyFunctionForField(func, current_type, field);
String result_name = "_" + func->getName() + "_" + toString(field.column_idx);
const auto & columns = field.columns;
size_t result_idx = columns->size();
for (size_t i = 0; i < result_idx; ++i)
{
if ((*columns)[i].name == result_name)
result_idx = i;
}
ColumnsWithTypeAndName args{(*columns)[field.column_idx]};
if (result_idx == columns->size())
{
field.columns->emplace_back(ColumnWithTypeAndName {nullptr, func->getResultType(), result_name});
(*columns)[result_idx].column = func->execute(args, (*columns)[result_idx].type, columns->front().column->size());
}
return {field.columns, field.row_idx, result_idx};
}
void KeyCondition::traverseAST(const ASTPtr & node, ContextPtr context, Block & block_with_constants)
{
RPNElement element;
if (const auto * func = node->as<ASTFunction>())
{
if (tryParseLogicalOperatorFromAST(func, element))
{
auto & args = func->arguments->children;
for (size_t i = 0, size = args.size(); i < size; ++i)
{
traverseAST(args[i], context, block_with_constants);
/** The first part of the condition is for the correct support of `and` and `or` functions of arbitrary arity
* - in this case `n - 1` elements are added (where `n` is the number of arguments).
*/
if (i != 0 || element.function == RPNElement::FUNCTION_NOT)
rpn.emplace_back(element);
}
return;
}
}
if (!tryParseAtomFromAST(node, context, block_with_constants, element))
{
element.function = RPNElement::FUNCTION_UNKNOWN;
}
rpn.emplace_back(std::move(element));
}
/** The key functional expression constraint may be inferred from a plain column in the expression.
* For example, if the key contains `toStartOfHour(Timestamp)` and query contains `WHERE Timestamp >= now()`,
* it can be assumed that if `toStartOfHour()` is monotonic on [now(), inf), the `toStartOfHour(Timestamp) >= toStartOfHour(now())`
* condition also holds, so the index may be used to select only parts satisfying this condition.
*
* To check the assumption, we'd need to assert that the inverse function to this transformation is also monotonic, however the
* inversion isn't exported (or even viable for not strictly monotonic functions such as `toStartOfHour()`).
* Instead, we can qualify only functions that do not transform the range (for example rounding),
* which while not strictly monotonic, are monotonic everywhere on the input range.
*/
bool KeyCondition::transformConstantWithValidFunctions(
const String & expr_name,
size_t & out_key_column_num,
DataTypePtr & out_key_column_type,
Field & out_value,
DataTypePtr & out_type,
std::function<bool(IFunctionBase &, const IDataType &)> always_monotonic) const
{
const auto & sample_block = key_expr->getSampleBlock();
for (const auto & node : key_expr->getNodes())
{
auto it = key_columns.find(node.result_name);
if (it != key_columns.end())
{
std::stack<const ActionsDAG::Node *> chain;
const auto * cur_node = &node;
bool is_valid_chain = true;
while (is_valid_chain)
{
if (cur_node->result_name == expr_name)
break;
chain.push(cur_node);
if (cur_node->type == ActionsDAG::ActionType::FUNCTION && cur_node->children.size() <= 2)
{
is_valid_chain = always_monotonic(*cur_node->function_base, *cur_node->result_type);
const ActionsDAG::Node * next_node = nullptr;
for (const auto * arg : cur_node->children)
{
if (arg->column && isColumnConst(*arg->column))
continue;
if (next_node)
is_valid_chain = false;
next_node = arg;
}
if (!next_node)
is_valid_chain = false;
cur_node = next_node;
}
else if (cur_node->type == ActionsDAG::ActionType::ALIAS)
cur_node = cur_node->children.front();
else
is_valid_chain = false;
}
if (is_valid_chain)
{
/// Here we cast constant to the input type.
/// It is not clear, why this works in general.
/// I can imagine the case when expression like `column < const` is legal,
/// but `type(column)` and `type(const)` are of different types,
/// and const cannot be casted to column type.
/// (There could be `superType(type(column), type(const))` which is used for comparison).
///
/// However, looks like this case newer happenes (I could not find such).
/// Let's assume that any two comparable types are castable to each other.
auto const_type = cur_node->result_type;
auto const_column = out_type->createColumnConst(1, out_value);
auto const_value = (*castColumn({const_column, out_type, ""}, const_type))[0];
while (!chain.empty())
{
const auto * func = chain.top();
chain.pop();
if (func->type != ActionsDAG::ActionType::FUNCTION)
continue;
if (func->children.size() == 1)
{
std::tie(const_value, const_type)
= applyFunctionForFieldOfUnknownType(func->function_base, const_type, const_value);
}
else if (func->children.size() == 2)
{
const auto * left = func->children[0];
const auto * right = func->children[1];
if (left->column && isColumnConst(*left->column))
{
auto left_arg_type = left->result_type;
auto left_arg_value = (*left->column)[0];
std::tie(const_value, const_type) = applyBinaryFunctionForFieldOfUnknownType(
func->function_builder, left_arg_type, left_arg_value, const_type, const_value);
}
else
{
auto right_arg_type = right->result_type;
auto right_arg_value = (*right->column)[0];
std::tie(const_value, const_type) = applyBinaryFunctionForFieldOfUnknownType(
func->function_builder, const_type, const_value, right_arg_type, right_arg_value);
}
}
}
out_key_column_num = it->second;
out_key_column_type = sample_block.getByName(it->first).type;
out_value = const_value;
out_type = const_type;
return true;
}
}
}
return false;
}
bool KeyCondition::canConstantBeWrappedByMonotonicFunctions(
const ASTPtr & node,
size_t & out_key_column_num,
DataTypePtr & out_key_column_type,
Field & out_value,
DataTypePtr & out_type)
{
String expr_name = node->getColumnNameWithoutAlias();
if (array_joined_columns.count(expr_name))
return false;
if (key_subexpr_names.count(expr_name) == 0)
return false;
if (out_value.isNull())
return false;
return transformConstantWithValidFunctions(
expr_name, out_key_column_num, out_key_column_type, out_value, out_type, [](IFunctionBase & func, const IDataType & type)
{
if (!func.hasInformationAboutMonotonicity())
return false;
else
{
/// Range is irrelevant in this case.
auto monotonicity = func.getMonotonicityForRange(type, Field(), Field());
if (!monotonicity.is_always_monotonic)
return false;
}
return true;
});
}
/// Looking for possible transformation of `column = constant` into `partition_expr = function(constant)`
bool KeyCondition::canConstantBeWrappedByFunctions(
const ASTPtr & ast, size_t & out_key_column_num, DataTypePtr & out_key_column_type, Field & out_value, DataTypePtr & out_type)
{
String expr_name = ast->getColumnNameWithoutAlias();
if (array_joined_columns.count(expr_name))
return false;
if (key_subexpr_names.count(expr_name) == 0)
{
/// Let's check another one case.
/// If our storage was created with moduloLegacy in partition key,
/// We can assume that `modulo(...) = const` is the same as `moduloLegacy(...) = const`.
/// Replace modulo to moduloLegacy in AST and check if we also have such a column.
///
/// We do not check this in canConstantBeWrappedByMonotonicFunctions.
/// The case `f(modulo(...))` for totally monotonic `f ` is consedered to be rare.
///
/// Note: for negative values, we can filter more partitions then needed.
auto adjusted_ast = ast->clone();
KeyDescription::moduloToModuloLegacyRecursive(adjusted_ast);
expr_name = adjusted_ast->getColumnName();
if (key_subexpr_names.count(expr_name) == 0)
return false;
}
if (out_value.isNull())
return false;
return transformConstantWithValidFunctions(
expr_name, out_key_column_num, out_key_column_type, out_value, out_type, [](IFunctionBase & func, const IDataType &)
{
return func.isDeterministic();
});
}
bool KeyCondition::tryPrepareSetIndex(
const ASTs & args,
ContextPtr context,
RPNElement & out,
size_t & out_key_column_num)
{
const ASTPtr & left_arg = args[0];
out_key_column_num = 0;
std::vector<MergeTreeSetIndex::KeyTuplePositionMapping> indexes_mapping;
DataTypes data_types;
auto get_key_tuple_position_mapping = [&](const ASTPtr & node, size_t tuple_index)
{
MergeTreeSetIndex::KeyTuplePositionMapping index_mapping;
index_mapping.tuple_index = tuple_index;
DataTypePtr data_type;
if (isKeyPossiblyWrappedByMonotonicFunctions(
node, context, index_mapping.key_index, data_type, index_mapping.functions))
{
indexes_mapping.push_back(index_mapping);
data_types.push_back(data_type);
if (out_key_column_num < index_mapping.key_index)
out_key_column_num = index_mapping.key_index;
}
};
size_t left_args_count = 1;
const auto * left_arg_tuple = left_arg->as<ASTFunction>();
if (left_arg_tuple && left_arg_tuple->name == "tuple")
{
const auto & tuple_elements = left_arg_tuple->arguments->children;
left_args_count = tuple_elements.size();
for (size_t i = 0; i < left_args_count; ++i)
get_key_tuple_position_mapping(tuple_elements[i], i);
}
else
get_key_tuple_position_mapping(left_arg, 0);
if (indexes_mapping.empty())
return false;
const ASTPtr & right_arg = args[1];
SetPtr prepared_set;
if (right_arg->as<ASTSubquery>() || right_arg->as<ASTTableIdentifier>())
{
auto set_it = prepared_sets.find(PreparedSetKey::forSubquery(*right_arg));
if (set_it == prepared_sets.end())
return false;
prepared_set = set_it->second;
}
else
{
/// We have `PreparedSetKey::forLiteral` but it is useless here as we don't have enough information
/// about types in left argument of the IN operator. Instead, we manually iterate through all the sets
/// and find the one for the right arg based on the AST structure (getTreeHash), after that we check
/// that the types it was prepared with are compatible with the types of the primary key.
auto set_ast_hash = right_arg->getTreeHash();
auto set_it = std::find_if(
prepared_sets.begin(), prepared_sets.end(),
[&](const auto & candidate_entry)
{
if (candidate_entry.first.ast_hash != set_ast_hash)
return false;
for (size_t i = 0; i < indexes_mapping.size(); ++i)
if (!candidate_entry.second->areTypesEqual(indexes_mapping[i].tuple_index, data_types[i]))
return false;
return true;
});
if (set_it == prepared_sets.end())
return false;
prepared_set = set_it->second;
}
/// The index can be prepared if the elements of the set were saved in advance.
if (!prepared_set->hasExplicitSetElements())
return false;
prepared_set->checkColumnsNumber(left_args_count);
for (size_t i = 0; i < indexes_mapping.size(); ++i)
prepared_set->checkTypesEqual(indexes_mapping[i].tuple_index, data_types[i]);
out.set_index = std::make_shared<MergeTreeSetIndex>(prepared_set->getSetElements(), std::move(indexes_mapping));
return true;
}
/** Allow to use two argument function with constant argument to be analyzed as a single argument function.
* In other words, it performs "currying" (binding of arguments).
* This is needed, for example, to support correct analysis of `toDate(time, 'UTC')`.
*/
class FunctionWithOptionalConstArg : public IFunctionBase
{
public:
enum Kind
{
NO_CONST = 0,
LEFT_CONST,
RIGHT_CONST,
};
explicit FunctionWithOptionalConstArg(const FunctionBasePtr & func_) : func(func_) {}
FunctionWithOptionalConstArg(const FunctionBasePtr & func_, const ColumnWithTypeAndName & const_arg_, Kind kind_)
: func(func_), const_arg(const_arg_), kind(kind_)
{
}
String getName() const override { return func->getName(); }
const DataTypes & getArgumentTypes() const override { return func->getArgumentTypes(); }
const DataTypePtr & getResultType() const override { return func->getResultType(); }
ExecutableFunctionPtr prepare(const ColumnsWithTypeAndName & arguments) const override { return func->prepare(arguments); }
ColumnPtr
execute(const ColumnsWithTypeAndName & arguments, const DataTypePtr & result_type, size_t input_rows_count, bool dry_run) const override
{
if (kind == Kind::LEFT_CONST)
{
ColumnsWithTypeAndName new_arguments;
new_arguments.reserve(arguments.size() + 1);
new_arguments.push_back(const_arg);
for (const auto & arg : arguments)
new_arguments.push_back(arg);
return func->prepare(new_arguments)->execute(new_arguments, result_type, input_rows_count, dry_run);
}
else if (kind == Kind::RIGHT_CONST)
{
auto new_arguments = arguments;
new_arguments.push_back(const_arg);
return func->prepare(new_arguments)->execute(new_arguments, result_type, input_rows_count, dry_run);
}
else
return func->prepare(arguments)->execute(arguments, result_type, input_rows_count, dry_run);
}
bool isDeterministic() const override { return func->isDeterministic(); }
bool isDeterministicInScopeOfQuery() const override { return func->isDeterministicInScopeOfQuery(); }
bool hasInformationAboutMonotonicity() const override { return func->hasInformationAboutMonotonicity(); }
bool isSuitableForShortCircuitArgumentsExecution(const DataTypesWithConstInfo & arguments) const override { return func->isSuitableForShortCircuitArgumentsExecution(arguments); }
IFunctionBase::Monotonicity getMonotonicityForRange(const IDataType & type, const Field & left, const Field & right) const override
{
return func->getMonotonicityForRange(type, left, right);
}
Kind getKind() const { return kind; }
const ColumnWithTypeAndName & getConstArg() const { return const_arg; }
private:
FunctionBasePtr func;
ColumnWithTypeAndName const_arg;
Kind kind = Kind::NO_CONST;
};
bool KeyCondition::isKeyPossiblyWrappedByMonotonicFunctions(
const ASTPtr & node,
ContextPtr context,
size_t & out_key_column_num,
DataTypePtr & out_key_res_column_type,
MonotonicFunctionsChain & out_functions_chain)
{
std::vector<const ASTFunction *> chain_not_tested_for_monotonicity;
DataTypePtr key_column_type;
if (!isKeyPossiblyWrappedByMonotonicFunctionsImpl(node, out_key_column_num, key_column_type, chain_not_tested_for_monotonicity))
return false;
for (auto it = chain_not_tested_for_monotonicity.rbegin(); it != chain_not_tested_for_monotonicity.rend(); ++it)
{
const auto & args = (*it)->arguments->children;
auto func_builder = FunctionFactory::instance().tryGet((*it)->name, context);
if (!func_builder)
return false;
ColumnsWithTypeAndName arguments;
ColumnWithTypeAndName const_arg;
FunctionWithOptionalConstArg::Kind kind = FunctionWithOptionalConstArg::Kind::NO_CONST;
if (args.size() == 2)
{
if (const auto * arg_left = args[0]->as<ASTLiteral>())
{
auto left_arg_type = applyVisitor(FieldToDataType(), arg_left->value);
const_arg = { left_arg_type->createColumnConst(0, arg_left->value), left_arg_type, "" };
arguments.push_back(const_arg);
arguments.push_back({ nullptr, key_column_type, "" });
kind = FunctionWithOptionalConstArg::Kind::LEFT_CONST;
}
else if (const auto * arg_right = args[1]->as<ASTLiteral>())
{
arguments.push_back({ nullptr, key_column_type, "" });
auto right_arg_type = applyVisitor(FieldToDataType(), arg_right->value);
const_arg = { right_arg_type->createColumnConst(0, arg_right->value), right_arg_type, "" };
arguments.push_back(const_arg);
kind = FunctionWithOptionalConstArg::Kind::RIGHT_CONST;
}
}
else
arguments.push_back({ nullptr, key_column_type, "" });
auto func = func_builder->build(arguments);
/// If we know the given range only contains one value, then we treat all functions as positive monotonic.
if (!func || (!single_point && !func->hasInformationAboutMonotonicity()))
return false;
key_column_type = func->getResultType();
if (kind == FunctionWithOptionalConstArg::Kind::NO_CONST)
out_functions_chain.push_back(func);
else
out_functions_chain.push_back(std::make_shared<FunctionWithOptionalConstArg>(func, const_arg, kind));
}
out_key_res_column_type = key_column_type;
return true;
}
bool KeyCondition::isKeyPossiblyWrappedByMonotonicFunctionsImpl(
const ASTPtr & node,
size_t & out_key_column_num,
DataTypePtr & out_key_column_type,
std::vector<const ASTFunction *> & out_functions_chain)
{
/** By itself, the key column can be a functional expression. for example, `intHash32(UserID)`.
* Therefore, use the full name of the expression for search.
*/
const auto & sample_block = key_expr->getSampleBlock();
// Key columns should use canonical names for index analysis
String name = node->getColumnNameWithoutAlias();
if (array_joined_columns.count(name))
return false;
auto it = key_columns.find(name);
if (key_columns.end() != it)
{
out_key_column_num = it->second;
out_key_column_type = sample_block.getByName(it->first).type;
return true;
}
if (const auto * func = node->as<ASTFunction>())
{
if (!func->arguments)
return false;
const auto & args = func->arguments->children;
if (args.size() > 2 || args.empty())
return false;
out_functions_chain.push_back(func);
bool ret = false;
if (args.size() == 2)
{
if (args[0]->as<ASTLiteral>())
{
ret = isKeyPossiblyWrappedByMonotonicFunctionsImpl(args[1], out_key_column_num, out_key_column_type, out_functions_chain);
}
else if (args[1]->as<ASTLiteral>())
{
ret = isKeyPossiblyWrappedByMonotonicFunctionsImpl(args[0], out_key_column_num, out_key_column_type, out_functions_chain);
}
}
else
{
ret = isKeyPossiblyWrappedByMonotonicFunctionsImpl(args[0], out_key_column_num, out_key_column_type, out_functions_chain);
}
return ret;
}
return false;
}
static void castValueToType(const DataTypePtr & desired_type, Field & src_value, const DataTypePtr & src_type, const ASTPtr & node)
{
try
{
src_value = convertFieldToType(src_value, *desired_type, src_type.get());
}
catch (...)
{
throw Exception("Key expression contains comparison between inconvertible types: " +
desired_type->getName() + " and " + src_type->getName() +
" inside " + queryToString(node),
ErrorCodes::BAD_TYPE_OF_FIELD);
}
}
bool KeyCondition::tryParseAtomFromAST(const ASTPtr & node, ContextPtr context, Block & block_with_constants, RPNElement & out)
{
/** Functions < > = != <= >= in `notIn` isNull isNotNull, where one argument is a constant, and the other is one of columns of key,
* or itself, wrapped in a chain of possibly-monotonic functions,
* or constant expression - number.
*/
Field const_value;
DataTypePtr const_type;
if (const auto * func = node->as<ASTFunction>())
{
const ASTs & args = func->arguments->children;
DataTypePtr key_expr_type; /// Type of expression containing key column
size_t key_column_num = -1; /// Number of a key column (inside key_column_names array)
MonotonicFunctionsChain chain;
std::string func_name = func->name;
if (atom_map.find(func_name) == std::end(atom_map))
return false;
if (args.size() == 1)
{
if (!(isKeyPossiblyWrappedByMonotonicFunctions(args[0], context, key_column_num, key_expr_type, chain)))
return false;
if (key_column_num == static_cast<size_t>(-1))
throw Exception("`key_column_num` wasn't initialized. It is a bug.", ErrorCodes::LOGICAL_ERROR);
}
else if (args.size() == 2)
{
size_t key_arg_pos; /// Position of argument with key column (non-const argument)
bool is_set_const = false;
bool is_constant_transformed = false;
/// We don't look for inversed key transformations when strict is true, which is required for trivial count().
/// Consider the following test case:
///
/// create table test1(p DateTime, k int) engine MergeTree partition by toDate(p) order by k;
/// insert into test1 values ('2020-09-01 00:01:02', 1), ('2020-09-01 20:01:03', 2), ('2020-09-02 00:01:03', 3);
/// select count() from test1 where p > toDateTime('2020-09-01 10:00:00');
///
/// toDate(DateTime) is always monotonic, but we cannot relax the predicates to be
/// >= toDate(toDateTime('2020-09-01 10:00:00')), which returns 3 instead of the right count: 2.
bool strict_condition = strict;
/// If we use this key condition to prune partitions by single value, we cannot relax conditions for NOT.
if (single_point
&& (func_name == "notLike" || func_name == "notIn" || func_name == "globalNotIn" || func_name == "notNullIn"
|| func_name == "globalNotNullIn" || func_name == "notEquals" || func_name == "notEmpty"))
strict_condition = true;
if (functionIsInOrGlobalInOperator(func_name))
{
if (tryPrepareSetIndex(args, context, out, key_column_num))
{
key_arg_pos = 0;
is_set_const = true;
}
else
return false;
}
else if (getConstant(args[1], block_with_constants, const_value, const_type))
{
if (isKeyPossiblyWrappedByMonotonicFunctions(args[0], context, key_column_num, key_expr_type, chain))
{
key_arg_pos = 0;
}
else if (
!strict_condition
&& canConstantBeWrappedByMonotonicFunctions(args[0], key_column_num, key_expr_type, const_value, const_type))
{
key_arg_pos = 0;
is_constant_transformed = true;
}
else if (
single_point && func_name == "equals" && !strict_condition
&& canConstantBeWrappedByFunctions(args[0], key_column_num, key_expr_type, const_value, const_type))
{
key_arg_pos = 0;
is_constant_transformed = true;
}
else
return false;
}
else if (getConstant(args[0], block_with_constants, const_value, const_type))
{
if (isKeyPossiblyWrappedByMonotonicFunctions(args[1], context, key_column_num, key_expr_type, chain))
{
key_arg_pos = 1;
}
else if (
!strict_condition
&& canConstantBeWrappedByMonotonicFunctions(args[1], key_column_num, key_expr_type, const_value, const_type))
{
key_arg_pos = 1;
is_constant_transformed = true;
}
else if (
single_point && func_name == "equals" && !strict_condition
&& canConstantBeWrappedByFunctions(args[1], key_column_num, key_expr_type, const_value, const_type))
{
key_arg_pos = 0;
is_constant_transformed = true;
}
else
return false;
}
else
return false;
if (key_column_num == static_cast<size_t>(-1))
throw Exception("`key_column_num` wasn't initialized. It is a bug.", ErrorCodes::LOGICAL_ERROR);
/// Replace <const> <sign> <data> on to <data> <-sign> <const>
if (key_arg_pos == 1)
{
if (func_name == "less")
func_name = "greater";
else if (func_name == "greater")
func_name = "less";
else if (func_name == "greaterOrEquals")
func_name = "lessOrEquals";
else if (func_name == "lessOrEquals")
func_name = "greaterOrEquals";
else if (func_name == "in" || func_name == "notIn" ||
func_name == "like" || func_name == "notLike" ||
func_name == "ilike" || func_name == "notIlike" ||
func_name == "startsWith")
{
/// "const IN data_column" doesn't make sense (unlike "data_column IN const")
return false;
}
}
bool cast_not_needed = is_set_const /// Set args are already casted inside Set::createFromAST
|| ((isNativeNumber(key_expr_type) || isDateTime(key_expr_type))
&& (isNativeNumber(const_type) || isDateTime(const_type))); /// Numbers and DateTime are accurately compared without cast.
if (!cast_not_needed && !key_expr_type->equals(*const_type))
{
if (const_value.getType() == Field::Types::String)
{
const_value = convertFieldToType(const_value, *key_expr_type);
if (const_value.isNull())
return false;
// No need to set is_constant_transformed because we're doing exact conversion
}
else
{
DataTypePtr common_type = getLeastSupertype({key_expr_type, const_type});
if (!const_type->equals(*common_type))
{
castValueToType(common_type, const_value, const_type, node);
// Need to set is_constant_transformed unless we're doing exact conversion
if (!key_expr_type->equals(*common_type))
is_constant_transformed = true;
}
if (!key_expr_type->equals(*common_type))
{
ColumnsWithTypeAndName arguments{
{nullptr, key_expr_type, ""}, {DataTypeString().createColumnConst(1, common_type->getName()), common_type, ""}};
FunctionOverloadResolverPtr func_builder_cast = CastInternalOverloadResolver<CastType::nonAccurate>::createImpl();
auto func_cast = func_builder_cast->build(arguments);
/// If we know the given range only contains one value, then we treat all functions as positive monotonic.
if (!func_cast || (!single_point && !func_cast->hasInformationAboutMonotonicity()))
return false;
chain.push_back(func_cast);
}
}
}
/// Transformed constant must weaken the condition, for example "x > 5" must weaken to "round(x) >= 5"
if (is_constant_transformed)
{
if (func_name == "less")
func_name = "lessOrEquals";
else if (func_name == "greater")
func_name = "greaterOrEquals";
}
}
else
return false;
const auto atom_it = atom_map.find(func_name);
out.key_column = key_column_num;
out.monotonic_functions_chain = std::move(chain);
return atom_it->second(out, const_value);
}
else if (getConstant(node, block_with_constants, const_value, const_type))
{
/// For cases where it says, for example, `WHERE 0 AND something`
if (const_value.getType() == Field::Types::UInt64)
{
out.function = const_value.safeGet<UInt64>() ? RPNElement::ALWAYS_TRUE : RPNElement::ALWAYS_FALSE;
return true;
}
else if (const_value.getType() == Field::Types::Int64)
{
out.function = const_value.safeGet<Int64>() ? RPNElement::ALWAYS_TRUE : RPNElement::ALWAYS_FALSE;
return true;
}
else if (const_value.getType() == Field::Types::Float64)
{
out.function = const_value.safeGet<Float64>() ? RPNElement::ALWAYS_TRUE : RPNElement::ALWAYS_FALSE;
return true;
}
}
return false;
}
bool KeyCondition::tryParseLogicalOperatorFromAST(const ASTFunction * func, RPNElement & out)
{
/// Functions AND, OR, NOT.
/// Also a special function `indexHint` - works as if instead of calling a function there are just parentheses
/// (or, the same thing - calling the function `and` from one argument).
const ASTs & args = func->arguments->children;
if (func->name == "not")
{
if (args.size() != 1)
return false;
out.function = RPNElement::FUNCTION_NOT;
}
else
{
if (func->name == "and" || func->name == "indexHint")
out.function = RPNElement::FUNCTION_AND;
else if (func->name == "or")
out.function = RPNElement::FUNCTION_OR;
else
return false;
}
return true;
}
String KeyCondition::toString() const
{
String res;
for (size_t i = 0; i < rpn.size(); ++i)
{
if (i)
res += ", ";
res += rpn[i].toString();
}
return res;
}
KeyCondition::Description KeyCondition::getDescription() const
{
/// This code may seem to be too difficult.
/// Here we want to convert RPN back to tree, and also simplify some logical expressions like `and(x, true) -> x`.
Description description;
/// That's a binary tree. Explicit.
/// Build and optimize it simultaneously.
struct Node
{
enum class Type
{
/// Leaf, which is RPNElement.
Leaf,
/// Leafs, which are logical constants.
True,
False,
/// Binary operators.
And,
Or,
};
Type type{};
/// Only for Leaf
const RPNElement * element = nullptr;
/// This means that logical NOT is applied to leaf.
bool negate = false;
std::unique_ptr<Node> left = nullptr;
std::unique_ptr<Node> right = nullptr;
};
/// The algorithm is the same as in KeyCondition::checkInHyperrectangle
/// We build a pair of trees on stack. For checking if key condition may be true, and if it may be false.
/// We need only `can_be_true` in result.
struct Frame
{
std::unique_ptr<Node> can_be_true;
std::unique_ptr<Node> can_be_false;
};
/// Combine two subtrees using logical operator.
auto combine = [](std::unique_ptr<Node> left, std::unique_ptr<Node> right, Node::Type type)
{
/// Simplify operators with for one constant condition.
if (type == Node::Type::And)
{
/// false AND right
if (left->type == Node::Type::False)
return left;
/// left AND false
if (right->type == Node::Type::False)
return right;
/// true AND right
if (left->type == Node::Type::True)
return right;
/// left AND true
if (right->type == Node::Type::True)
return left;
}
if (type == Node::Type::Or)
{
/// false OR right
if (left->type == Node::Type::False)
return right;
/// left OR false
if (right->type == Node::Type::False)
return left;
/// true OR right
if (left->type == Node::Type::True)
return left;
/// left OR true
if (right->type == Node::Type::True)
return right;
}
return std::make_unique<Node>(Node{
.type = type,
.left = std::move(left),
.right = std::move(right)
});
};
std::vector<Frame> rpn_stack;
for (const auto & element : rpn)
{
if (element.function == RPNElement::FUNCTION_UNKNOWN)
{
auto can_be_true = std::make_unique<Node>(Node{.type = Node::Type::True});
auto can_be_false = std::make_unique<Node>(Node{.type = Node::Type::True});
rpn_stack.emplace_back(Frame{.can_be_true = std::move(can_be_true), .can_be_false = std::move(can_be_false)});
}
else if (
element.function == RPNElement::FUNCTION_IN_RANGE
|| element.function == RPNElement::FUNCTION_NOT_IN_RANGE
|| element.function == RPNElement::FUNCTION_IS_NULL
|| element.function == RPNElement::FUNCTION_IS_NOT_NULL
|| element.function == RPNElement::FUNCTION_IN_SET
|| element.function == RPNElement::FUNCTION_NOT_IN_SET)
{
auto can_be_true = std::make_unique<Node>(Node{.type = Node::Type::Leaf, .element = &element, .negate = false});
auto can_be_false = std::make_unique<Node>(Node{.type = Node::Type::Leaf, .element = &element, .negate = true});
rpn_stack.emplace_back(Frame{.can_be_true = std::move(can_be_true), .can_be_false = std::move(can_be_false)});
}
else if (element.function == RPNElement::FUNCTION_NOT)
{
assert(!rpn_stack.empty());
std::swap(rpn_stack.back().can_be_true, rpn_stack.back().can_be_false);
}
else if (element.function == RPNElement::FUNCTION_AND)
{
assert(!rpn_stack.empty());
auto arg1 = std::move(rpn_stack.back());
rpn_stack.pop_back();
assert(!rpn_stack.empty());
auto arg2 = std::move(rpn_stack.back());
Frame frame;
frame.can_be_true = combine(std::move(arg1.can_be_true), std::move(arg2.can_be_true), Node::Type::And);
frame.can_be_false = combine(std::move(arg1.can_be_false), std::move(arg2.can_be_false), Node::Type::Or);
rpn_stack.back() = std::move(frame);
}
else if (element.function == RPNElement::FUNCTION_OR)
{
assert(!rpn_stack.empty());
auto arg1 = std::move(rpn_stack.back());
rpn_stack.pop_back();
assert(!rpn_stack.empty());
auto arg2 = std::move(rpn_stack.back());
Frame frame;
frame.can_be_true = combine(std::move(arg1.can_be_true), std::move(arg2.can_be_true), Node::Type::Or);
frame.can_be_false = combine(std::move(arg1.can_be_false), std::move(arg2.can_be_false), Node::Type::And);
rpn_stack.back() = std::move(frame);
}
else if (element.function == RPNElement::ALWAYS_FALSE)
{
auto can_be_true = std::make_unique<Node>(Node{.type = Node::Type::False});
auto can_be_false = std::make_unique<Node>(Node{.type = Node::Type::True});
rpn_stack.emplace_back(Frame{.can_be_true = std::move(can_be_true), .can_be_false = std::move(can_be_false)});
}
else if (element.function == RPNElement::ALWAYS_TRUE)
{
auto can_be_true = std::make_unique<Node>(Node{.type = Node::Type::True});
auto can_be_false = std::make_unique<Node>(Node{.type = Node::Type::False});
rpn_stack.emplace_back(Frame{.can_be_true = std::move(can_be_true), .can_be_false = std::move(can_be_false)});
}
else
throw Exception("Unexpected function type in KeyCondition::RPNElement", ErrorCodes::LOGICAL_ERROR);
}
if (rpn_stack.size() != 1)
throw Exception("Unexpected stack size in KeyCondition::checkInRange", ErrorCodes::LOGICAL_ERROR);
std::vector<std::string_view> key_names(key_columns.size());
std::vector<bool> is_key_used(key_columns.size(), false);
for (const auto & key : key_columns)
key_names[key.second] = key.first;
WriteBufferFromOwnString buf;
std::function<void(const Node *)> describe;
describe = [&describe, &key_names, &is_key_used, &buf](const Node * node)
{
switch (node->type)
{
case Node::Type::Leaf:
{
is_key_used[node->element->key_column] = true;
/// Note: for condition with double negation, like `not(x not in set)`,
/// we can replace it to `x in set` here.
/// But I won't do it, because `cloneASTWithInversionPushDown` already push down `not`.
/// So, this seem to be impossible for `can_be_true` tree.
if (node->negate)
buf << "not(";
buf << node->element->toString(key_names[node->element->key_column], true);
if (node->negate)
buf << ")";
break;
}
case Node::Type::True:
buf << "true";
break;
case Node::Type::False:
buf << "false";
break;
case Node::Type::And:
buf << "and(";
describe(node->left.get());
buf << ", ";
describe(node->right.get());
buf << ")";
break;
case Node::Type::Or:
buf << "or(";
describe(node->left.get());
buf << ", ";
describe(node->right.get());
buf << ")";
break;
}
};
describe(rpn_stack.front().can_be_true.get());
description.condition = std::move(buf.str());
for (size_t i = 0; i < key_names.size(); ++i)
if (is_key_used[i])
description.used_keys.emplace_back(key_names[i]);
return description;
}
/** Index is the value of key every `index_granularity` rows.
* This value is called a "mark". That is, the index consists of marks.
*
* The key is the tuple.
* The data is sorted by key in the sense of lexicographic order over tuples.
*
* A pair of marks specifies a segment with respect to the order over the tuples.
* Denote it like this: [ x1 y1 z1 .. x2 y2 z2 ],
* where x1 y1 z1 - tuple - value of key in left border of segment;
* x2 y2 z2 - tuple - value of key in right boundary of segment.
* In this section there are data between these marks.
*
* Or, the last mark specifies the range open on the right: [ a b c .. + inf )
*
* The set of all possible tuples can be considered as an n-dimensional space, where n is the size of the tuple.
* A range of tuples specifies some subset of this space.
*
* Hyperrectangles will be the subrange of an n-dimensional space that is a direct product of one-dimensional ranges.
* In this case, the one-dimensional range can be:
* a point, a segment, an open interval, a half-open interval;
* unlimited on the left, unlimited on the right ...
*
* The range of tuples can always be represented as a combination (union) of hyperrectangles.
* For example, the range [ x1 y1 .. x2 y2 ] given x1 != x2 is equal to the union of the following three hyperrectangles:
* [x1] x [y1 .. +inf)
* (x1 .. x2) x (-inf .. +inf)
* [x2] x (-inf .. y2]
*
* Or, for example, the range [ x1 y1 .. +inf ] is equal to the union of the following two hyperrectangles:
* [x1] x [y1 .. +inf)
* (x1 .. +inf) x (-inf .. +inf)
* It's easy to see that this is a special case of the variant above.
*
* This is important because it is easy for us to check the feasibility of the condition over the hyperrectangle,
* and therefore, feasibility of condition on the range of tuples will be checked by feasibility of condition
* over at least one hyperrectangle from which this range consists.
*/
template <typename F>
static BoolMask forAnyHyperrectangle(
size_t key_size,
const FieldRef * left_keys,
const FieldRef * right_keys,
bool left_bounded,
bool right_bounded,
std::vector<Range> & hyperrectangle,
size_t prefix_size,
BoolMask initial_mask,
F && callback)
{
if (!left_bounded && !right_bounded)
return callback(hyperrectangle);
if (left_bounded && right_bounded)
{
/// Let's go through the matching elements of the key.
while (prefix_size < key_size)
{
if (left_keys[prefix_size] == right_keys[prefix_size])
{
/// Point ranges.
hyperrectangle[prefix_size] = Range(left_keys[prefix_size]);
++prefix_size;
}
else
break;
}
}
if (prefix_size == key_size)
return callback(hyperrectangle);
if (prefix_size + 1 == key_size)
{
if (left_bounded && right_bounded)
hyperrectangle[prefix_size] = Range(left_keys[prefix_size], true, right_keys[prefix_size], true);
else if (left_bounded)
hyperrectangle[prefix_size] = Range::createLeftBounded(left_keys[prefix_size], true);
else if (right_bounded)
hyperrectangle[prefix_size] = Range::createRightBounded(right_keys[prefix_size], true);
return callback(hyperrectangle);
}
/// (x1 .. x2) x (-inf .. +inf)
if (left_bounded && right_bounded)
hyperrectangle[prefix_size] = Range(left_keys[prefix_size], false, right_keys[prefix_size], false);
else if (left_bounded)
hyperrectangle[prefix_size] = Range::createLeftBounded(left_keys[prefix_size], false);
else if (right_bounded)
hyperrectangle[prefix_size] = Range::createRightBounded(right_keys[prefix_size], false);
for (size_t i = prefix_size + 1; i < key_size; ++i)
hyperrectangle[i] = Range();
BoolMask result = initial_mask;
result = result | callback(hyperrectangle);
/// There are several early-exit conditions (like the one below) hereinafter.
/// They are important; in particular, if initial_mask == BoolMask::consider_only_can_be_true
/// (which happens when this routine is called from KeyCondition::mayBeTrueXXX),
/// they provide significant speedup, which may be observed on merge_tree_huge_pk performance test.
if (result.isComplete())
return result;
/// [x1] x [y1 .. +inf)
if (left_bounded)
{
hyperrectangle[prefix_size] = Range(left_keys[prefix_size]);
result = result | forAnyHyperrectangle(key_size, left_keys, right_keys, true, false, hyperrectangle, prefix_size + 1, initial_mask, callback);
if (result.isComplete())
return result;
}
/// [x2] x (-inf .. y2]
if (right_bounded)
{
hyperrectangle[prefix_size] = Range(right_keys[prefix_size]);
result = result | forAnyHyperrectangle(key_size, left_keys, right_keys, false, true, hyperrectangle, prefix_size + 1, initial_mask, callback);
if (result.isComplete())
return result;
}
return result;
}
BoolMask KeyCondition::checkInRange(
size_t used_key_size,
const FieldRef * left_keys,
const FieldRef * right_keys,
const DataTypes & data_types,
BoolMask initial_mask) const
{
std::vector<Range> key_ranges(used_key_size, Range());
// std::cerr << "Checking for: [";
// for (size_t i = 0; i != used_key_size; ++i)
// std::cerr << (i != 0 ? ", " : "") << applyVisitor(FieldVisitorToString(), left_keys[i]);
// std::cerr << " ... ";
// for (size_t i = 0; i != used_key_size; ++i)
// std::cerr << (i != 0 ? ", " : "") << applyVisitor(FieldVisitorToString(), right_keys[i]);
// std::cerr << "]\n";
return forAnyHyperrectangle(used_key_size, left_keys, right_keys, true, true, key_ranges, 0, initial_mask,
[&] (const std::vector<Range> & key_ranges_hyperrectangle)
{
auto res = checkInHyperrectangle(key_ranges_hyperrectangle, data_types);
// std::cerr << "Hyperrectangle: ";
// for (size_t i = 0, size = key_ranges.size(); i != size; ++i)
// std::cerr << (i != 0 ? " x " : "") << key_ranges[i].toString();
// std::cerr << ": " << res.can_be_true << "\n";
return res;
});
}
std::optional<Range> KeyCondition::applyMonotonicFunctionsChainToRange(
Range key_range,
const MonotonicFunctionsChain & functions,
DataTypePtr current_type,
bool single_point)
{
for (const auto & func : functions)
{
/// We check the monotonicity of each function on a specific range.
/// If we know the given range only contains one value, then we treat all functions as positive monotonic.
IFunction::Monotonicity monotonicity = single_point
? IFunction::Monotonicity{true}
: func->getMonotonicityForRange(*current_type.get(), key_range.left, key_range.right);
if (!monotonicity.is_monotonic)
{
return {};
}
/// If we apply function to open interval, we can get empty intervals in result.
/// E.g. for ('2020-01-03', '2020-01-20') after applying 'toYYYYMM' we will get ('202001', '202001').
/// To avoid this we make range left and right included.
/// Any function that treats NULL specially is not monotonic.
/// Thus we can safely use isNull() as an -Inf/+Inf indicator here.
if (!key_range.left.isNull())
{
key_range.left = applyFunction(func, current_type, key_range.left);
key_range.left_included = true;
}
if (!key_range.right.isNull())
{
key_range.right = applyFunction(func, current_type, key_range.right);
key_range.right_included = true;
}
current_type = func->getResultType();
if (!monotonicity.is_positive)
key_range.invert();
}
return key_range;
}
// Returns whether the condition is one continuous range of the primary key,
// where every field is matched by range or a single element set.
// This allows to use a more efficient lookup with no extra reads.
bool KeyCondition::matchesExactContinuousRange() const
{
// Not implemented yet.
if (hasMonotonicFunctionsChain())
return false;
enum Constraint
{
POINT,
RANGE,
UNKNOWN,
};
std::vector<Constraint> column_constraints(key_columns.size(), Constraint::UNKNOWN);
for (const auto & element : rpn)
{
if (element.function == RPNElement::Function::FUNCTION_AND)
{
continue;
}
if (element.function == RPNElement::Function::FUNCTION_IN_SET && element.set_index && element.set_index->size() == 1)
{
column_constraints[element.key_column] = Constraint::POINT;
continue;
}
if (element.function == RPNElement::Function::FUNCTION_IN_RANGE)
{
if (element.range.left == element.range.right)
{
column_constraints[element.key_column] = Constraint::POINT;
}
if (column_constraints[element.key_column] != Constraint::POINT)
{
column_constraints[element.key_column] = Constraint::RANGE;
}
continue;
}
if (element.function == RPNElement::Function::FUNCTION_UNKNOWN)
{
continue;
}
return false;
}
auto min_constraint = column_constraints[0];
if (min_constraint > Constraint::RANGE)
{
return false;
}
for (size_t i = 1; i < key_columns.size(); ++i)
{
if (column_constraints[i] < min_constraint)
{
return false;
}
if (column_constraints[i] == Constraint::RANGE && min_constraint == Constraint::RANGE)
{
return false;
}
min_constraint = column_constraints[i];
}
return true;
}
BoolMask KeyCondition::checkInHyperrectangle(
const std::vector<Range> & hyperrectangle,
const DataTypes & data_types) const
{
std::vector<BoolMask> rpn_stack;
for (const auto & element : rpn)
{
if (element.function == RPNElement::FUNCTION_UNKNOWN)
{
rpn_stack.emplace_back(true, true);
}
else if (element.function == RPNElement::FUNCTION_IN_RANGE
|| element.function == RPNElement::FUNCTION_NOT_IN_RANGE)
{
const Range * key_range = &hyperrectangle[element.key_column];
/// The case when the column is wrapped in a chain of possibly monotonic functions.
Range transformed_range;
if (!element.monotonic_functions_chain.empty())
{
std::optional<Range> new_range = applyMonotonicFunctionsChainToRange(
*key_range,
element.monotonic_functions_chain,
data_types[element.key_column],
single_point
);
if (!new_range)
{
rpn_stack.emplace_back(true, true);
continue;
}
transformed_range = *new_range;
key_range = &transformed_range;
}
bool intersects = element.range.intersectsRange(*key_range);
bool contains = element.range.containsRange(*key_range);
rpn_stack.emplace_back(intersects, !contains);
if (element.function == RPNElement::FUNCTION_NOT_IN_RANGE)
rpn_stack.back() = !rpn_stack.back();
}
else if (
element.function == RPNElement::FUNCTION_IS_NULL
|| element.function == RPNElement::FUNCTION_IS_NOT_NULL)
{
const Range * key_range = &hyperrectangle[element.key_column];
/// No need to apply monotonic functions as nulls are kept.
bool intersects = element.range.intersectsRange(*key_range);
bool contains = element.range.containsRange(*key_range);
rpn_stack.emplace_back(intersects, !contains);
}
else if (
element.function == RPNElement::FUNCTION_IN_SET
|| element.function == RPNElement::FUNCTION_NOT_IN_SET)
{
if (!element.set_index)
throw Exception("Set for IN is not created yet", ErrorCodes::LOGICAL_ERROR);
rpn_stack.emplace_back(element.set_index->checkInRange(hyperrectangle, data_types));
if (element.function == RPNElement::FUNCTION_NOT_IN_SET)
rpn_stack.back() = !rpn_stack.back();
}
else if (element.function == RPNElement::FUNCTION_NOT)
{
assert(!rpn_stack.empty());
rpn_stack.back() = !rpn_stack.back();
}
else if (element.function == RPNElement::FUNCTION_AND)
{
assert(!rpn_stack.empty());
auto arg1 = rpn_stack.back();
rpn_stack.pop_back();
auto arg2 = rpn_stack.back();
rpn_stack.back() = arg1 & arg2;
}
else if (element.function == RPNElement::FUNCTION_OR)
{
assert(!rpn_stack.empty());
auto arg1 = rpn_stack.back();
rpn_stack.pop_back();
auto arg2 = rpn_stack.back();
rpn_stack.back() = arg1 | arg2;
}
else if (element.function == RPNElement::ALWAYS_FALSE)
{
rpn_stack.emplace_back(false, true);
}
else if (element.function == RPNElement::ALWAYS_TRUE)
{
rpn_stack.emplace_back(true, false);
}
else
throw Exception("Unexpected function type in KeyCondition::RPNElement", ErrorCodes::LOGICAL_ERROR);
}
if (rpn_stack.size() != 1)
throw Exception("Unexpected stack size in KeyCondition::checkInRange", ErrorCodes::LOGICAL_ERROR);
return rpn_stack[0];
}
bool KeyCondition::mayBeTrueInRange(
size_t used_key_size,
const FieldRef * left_keys,
const FieldRef * right_keys,
const DataTypes & data_types) const
{
return checkInRange(used_key_size, left_keys, right_keys, data_types, BoolMask::consider_only_can_be_true).can_be_true;
}
String KeyCondition::RPNElement::toString() const { return toString("column " + std::to_string(key_column), false); }
String KeyCondition::RPNElement::toString(const std::string_view & column_name, bool print_constants) const
{
auto print_wrapped_column = [this, &column_name, print_constants](WriteBuffer & buf)
{
for (auto it = monotonic_functions_chain.rbegin(); it != monotonic_functions_chain.rend(); ++it)
{
buf << (*it)->getName() << "(";
if (print_constants)
{
if (const auto * func = typeid_cast<const FunctionWithOptionalConstArg *>(it->get()))
{
if (func->getKind() == FunctionWithOptionalConstArg::Kind::LEFT_CONST)
buf << applyVisitor(FieldVisitorToString(), (*func->getConstArg().column)[0]) << ", ";
}
}
}
buf << column_name;
for (auto it = monotonic_functions_chain.rbegin(); it != monotonic_functions_chain.rend(); ++it)
{
if (print_constants)
{
if (const auto * func = typeid_cast<const FunctionWithOptionalConstArg *>(it->get()))
{
if (func->getKind() == FunctionWithOptionalConstArg::Kind::RIGHT_CONST)
buf << ", " << applyVisitor(FieldVisitorToString(), (*func->getConstArg().column)[0]);
}
}
buf << ")";
}
};
WriteBufferFromOwnString buf;
switch (function)
{
case FUNCTION_AND:
return "and";
case FUNCTION_OR:
return "or";
case FUNCTION_NOT:
return "not";
case FUNCTION_UNKNOWN:
return "unknown";
case FUNCTION_NOT_IN_SET:
case FUNCTION_IN_SET:
{
buf << "(";
print_wrapped_column(buf);
buf << (function == FUNCTION_IN_SET ? " in " : " notIn ");
if (!set_index)
buf << "unknown size set";
else
buf << set_index->size() << "-element set";
buf << ")";
return buf.str();
}
case FUNCTION_IN_RANGE:
case FUNCTION_NOT_IN_RANGE:
{
buf << "(";
print_wrapped_column(buf);
buf << (function == FUNCTION_NOT_IN_RANGE ? " not" : "") << " in " << range.toString();
buf << ")";
return buf.str();
}
case FUNCTION_IS_NULL:
case FUNCTION_IS_NOT_NULL:
{
buf << "(";
print_wrapped_column(buf);
buf << (function == FUNCTION_IS_NULL ? " isNull" : " isNotNull");
buf << ")";
return buf.str();
}
case ALWAYS_FALSE:
return "false";
case ALWAYS_TRUE:
return "true";
}
__builtin_unreachable();
}
bool KeyCondition::alwaysUnknownOrTrue() const
{
return unknownOrAlwaysTrue(false);
}
bool KeyCondition::anyUnknownOrAlwaysTrue() const
{
return unknownOrAlwaysTrue(true);
}
bool KeyCondition::unknownOrAlwaysTrue(bool unknown_any) const
{
std::vector<UInt8> rpn_stack;
for (const auto & element : rpn)
{
if (element.function == RPNElement::FUNCTION_UNKNOWN)
{
/// If unknown_any is true, return instantly,
/// to avoid processing it with FUNCTION_AND, and change the outcome.
if (unknown_any)
return true;
/// Otherwise, it may be AND'ed via FUNCTION_AND
rpn_stack.push_back(true);
}
else if (element.function == RPNElement::ALWAYS_TRUE)
{
rpn_stack.push_back(true);
}
else if (element.function == RPNElement::FUNCTION_NOT_IN_RANGE
|| element.function == RPNElement::FUNCTION_IN_RANGE
|| element.function == RPNElement::FUNCTION_IN_SET
|| element.function == RPNElement::FUNCTION_NOT_IN_SET
|| element.function == RPNElement::FUNCTION_IS_NULL
|| element.function == RPNElement::FUNCTION_IS_NOT_NULL
|| element.function == RPNElement::ALWAYS_FALSE)
{
rpn_stack.push_back(false);
}
else if (element.function == RPNElement::FUNCTION_NOT)
{
}
else if (element.function == RPNElement::FUNCTION_AND)
{
assert(!rpn_stack.empty());
auto arg1 = rpn_stack.back();
rpn_stack.pop_back();
auto arg2 = rpn_stack.back();
rpn_stack.back() = arg1 & arg2;
}
else if (element.function == RPNElement::FUNCTION_OR)
{
assert(!rpn_stack.empty());
auto arg1 = rpn_stack.back();
rpn_stack.pop_back();
auto arg2 = rpn_stack.back();
rpn_stack.back() = arg1 | arg2;
}
else
throw Exception("Unexpected function type in KeyCondition::RPNElement", ErrorCodes::LOGICAL_ERROR);
}
if (rpn_stack.size() != 1)
throw Exception("Unexpected stack size in KeyCondition::unknownOrAlwaysTrue", ErrorCodes::LOGICAL_ERROR);
return rpn_stack[0];
}
size_t KeyCondition::getMaxKeyColumn() const
{
size_t res = 0;
for (const auto & element : rpn)
{
if (element.function == RPNElement::FUNCTION_NOT_IN_RANGE
|| element.function == RPNElement::FUNCTION_IN_RANGE
|| element.function == RPNElement::FUNCTION_IS_NULL
|| element.function == RPNElement::FUNCTION_IS_NOT_NULL
|| element.function == RPNElement::FUNCTION_IN_SET
|| element.function == RPNElement::FUNCTION_NOT_IN_SET)
{
if (element.key_column > res)
res = element.key_column;
}
}
return res;
}
bool KeyCondition::hasMonotonicFunctionsChain() const
{
for (const auto & element : rpn)
if (!element.monotonic_functions_chain.empty()
|| (element.set_index && element.set_index->hasMonotonicFunctionsChain()))
return true;
return false;
}
}
Reference in New Issue View Git Blame Copy Permalink