ClickHouse/src/Storages/MergeTree/KeyCondition.h
Ivan Babrou d9d8d0242e Optimize PK lookup for queries that match exact PK range
Existing code that looks up marks that match the query has a pathological
case, when most of the part does in fact match the query.

The code works by recursively splitting a part into ranges and then discarding
the ranges that definitely do not match the query, based on primary key.

The problem is that it requires visiting every mark that matches the query,
making the complexity of this sort of look up O(n).

For queries that match exact range on the primary key, we can find
both left and right parts of the range with O(log 2) complexity.

This change implements exactly that.

To engage this optimization, the query must:

* Have a prefix list of the primary key.
* Have only range or single set element constraints for columns.
* Have only AND as a boolean operator.

Consider a table with `(service, timestamp)` as the primary key.

The following conditions will be optimized:

* `service = 'foo'`
* `service = 'foo' and timestamp >= now() - 3600`
* `service in ('foo')`
* `service in ('foo') and timestamp >= now() - 3600 and timestamp <= now`

The following will fall back to previous lookup algorithm:

* `timestamp >= now() - 3600`
* `service in ('foo', 'bar') and timestamp >= now() - 3600`
* `service = 'foo'`

Note that the optimization won't engage when PK has a range expression
followed by a point expression, since in that case the range is not continuous.

Trace query logging provides the following messages types of messages,
each representing a different kind of PK usage for a part:

```
Used optimized inclusion search over index for part 20200711_5710108_5710108_0 with 9 steps
Used generic exclusion search over index for part 20200711_5710118_5710228_5 with 1495 steps
Not using index on part 20200710_5710473_5710473_0
```

Number of steps translates to computational complexity.

Here's a comparison for before and after for a query over 24h of data:

```
Read 4562944 rows, 148.05 MiB in 45.19249672 sec.,   100966 rows/sec.,   3.28 MiB/sec.
Read 4183040 rows, 135.78 MiB in 0.196279627 sec., 21311636 rows/sec., 691.75 MiB/sec.
```

This is especially useful for queries that read data in order
and terminate early to return "last X things" matching a query.

See #11564 for more thoughts on this.
2020-07-11 12:26:54 -07:00

419 lines
14 KiB
C++

#pragma once
#include <sstream>
#include <optional>
#include <Interpreters/Set.h>
#include <Core/SortDescription.h>
#include <Parsers/ASTExpressionList.h>
#include <Parsers/ASTSelectQuery.h>
#include <Parsers/ASTFunction.h>
#include <Storages/SelectQueryInfo.h>
namespace DB
{
class Context;
class IFunction;
using FunctionBasePtr = std::shared_ptr<IFunctionBase>;
class ExpressionActions;
using ExpressionActionsPtr = std::shared_ptr<ExpressionActions>;
/** A field, that can be stored in two representations:
* - A standalone field.
* - A field with reference to its position in a block.
* It's needed for execution of functions on ranges during
* index analysis. If function was executed once for field,
* its result would be cached for whole block for which field's reference points to.
*/
struct FieldRef : public Field
{
FieldRef() = default;
/// Create as explicit field without block.
template <typename T>
FieldRef(T && value) : Field(std::forward<T>(value)) {}
/// Create as reference to field in block.
FieldRef(Block * block_, size_t row_idx_, size_t column_idx_)
: Field((*block_->getByPosition(column_idx_).column)[row_idx_]),
block(block_), row_idx(row_idx_), column_idx(column_idx_) {}
bool isExplicit() const { return block == nullptr; }
Block * block = nullptr;
size_t row_idx = 0;
size_t column_idx = 0;
};
/** Range with open or closed ends; possibly unbounded.
*/
struct Range
{
private:
static bool equals(const Field & lhs, const Field & rhs);
static bool less(const Field & lhs, const Field & rhs);
public:
FieldRef left; /// the left border, if any
FieldRef right; /// the right border, if any
bool left_bounded = false; /// bounded at the left
bool right_bounded = false; /// bounded at the right
bool left_included = false; /// includes the left border, if any
bool right_included = false; /// includes the right border, if any
/// The whole unversum.
Range() {}
/// One point.
Range(const FieldRef & point)
: left(point), right(point), left_bounded(true), right_bounded(true), left_included(true), right_included(true) {}
/// A bounded two-sided range.
Range(const FieldRef & left_, bool left_included_, const FieldRef & right_, bool right_included_)
: left(left_), right(right_),
left_bounded(true), right_bounded(true),
left_included(left_included_), right_included(right_included_)
{
shrinkToIncludedIfPossible();
}
static Range createRightBounded(const FieldRef & right_point, bool right_included)
{
Range r;
r.right = right_point;
r.right_bounded = true;
r.right_included = right_included;
r.shrinkToIncludedIfPossible();
return r;
}
static Range createLeftBounded(const FieldRef & left_point, bool left_included)
{
Range r;
r.left = left_point;
r.left_bounded = true;
r.left_included = left_included;
r.shrinkToIncludedIfPossible();
return r;
}
/** Optimize the range. If it has an open boundary and the Field type is "loose"
* - then convert it to closed, narrowing by one.
* That is, for example, turn (0,2) into [1].
*/
void shrinkToIncludedIfPossible()
{
if (left.isExplicit() && left_bounded && !left_included)
{
if (left.getType() == Field::Types::UInt64 && left.get<UInt64>() != std::numeric_limits<UInt64>::max())
{
++left.get<UInt64 &>();
left_included = true;
}
if (left.getType() == Field::Types::Int64 && left.get<Int64>() != std::numeric_limits<Int64>::max())
{
++left.get<Int64 &>();
left_included = true;
}
}
if (right.isExplicit() && right_bounded && !right_included)
{
if (right.getType() == Field::Types::UInt64 && right.get<UInt64>() != std::numeric_limits<UInt64>::min())
{
--right.get<UInt64 &>();
right_included = true;
}
if (right.getType() == Field::Types::Int64 && right.get<Int64>() != std::numeric_limits<Int64>::min())
{
--right.get<Int64 &>();
right_included = true;
}
}
}
bool empty() const
{
return left_bounded && right_bounded
&& (less(right, left)
|| ((!left_included || !right_included) && !less(left, right)));
}
/// x contained in the range
bool contains(const FieldRef & x) const
{
return !leftThan(x) && !rightThan(x);
}
/// x is to the left
bool rightThan(const FieldRef & x) const
{
return (left_bounded
? !(less(left, x) || (left_included && equals(x, left)))
: false);
}
/// x is to the right
bool leftThan(const FieldRef & x) const
{
return (right_bounded
? !(less(x, right) || (right_included && equals(x, right)))
: false);
}
bool intersectsRange(const Range & r) const
{
/// r to the left of me.
if (r.right_bounded
&& left_bounded
&& (less(r.right, left)
|| ((!left_included || !r.right_included)
&& equals(r.right, left))))
return false;
/// r to the right of me.
if (r.left_bounded
&& right_bounded
&& (less(right, r.left) /// ...} {...
|| ((!right_included || !r.left_included) /// ...) [... or ...] (...
&& equals(r.left, right))))
return false;
return true;
}
bool containsRange(const Range & r) const
{
/// r starts to the left of me.
if (left_bounded
&& (!r.left_bounded
|| less(r.left, left)
|| (r.left_included
&& !left_included
&& equals(r.left, left))))
return false;
/// r ends right of me.
if (right_bounded
&& (!r.right_bounded
|| less(right, r.right)
|| (r.right_included
&& !right_included
&& equals(r.right, right))))
return false;
return true;
}
void swapLeftAndRight()
{
std::swap(left, right);
std::swap(left_bounded, right_bounded);
std::swap(left_included, right_included);
}
String toString() const;
};
/** Condition on the index.
*
* Consists of the conditions for the key belonging to all possible ranges or sets,
* as well as logical operators AND/OR/NOT above these conditions.
*
* Constructs a reverse polish notation from these conditions
* and can calculate (interpret) its satisfiability over key ranges.
*/
class KeyCondition
{
public:
/// Does not take into account the SAMPLE section. all_columns - the set of all columns of the table.
KeyCondition(
const SelectQueryInfo & query_info,
const Context & context,
const Names & key_column_names,
const ExpressionActionsPtr & key_expr);
/// Whether the condition and its negation are feasible in the direct product of single column ranges specified by `hyperrectangle`.
BoolMask checkInHyperrectangle(
const std::vector<Range> & hyperrectangle,
const DataTypes & data_types) const;
/// Whether the condition and its negation are (independently) feasible in the key range.
/// left_key and right_key must contain all fields in the sort_descr in the appropriate order.
/// data_types - the types of the key columns.
/// Argument initial_mask is used for early exiting the implementation when we do not care about
/// one of the resulting mask components (see BoolMask::consider_only_can_be_XXX).
BoolMask checkInRange(
size_t used_key_size,
const FieldRef * left_key,
const FieldRef* right_key,
const DataTypes & data_types,
BoolMask initial_mask = BoolMask(false, false)) const;
/// Are the condition and its negation valid in a semi-infinite (not limited to the right) key range.
/// left_key must contain all the fields in the sort_descr in the appropriate order.
BoolMask checkAfter(
size_t used_key_size,
const FieldRef * left_key,
const DataTypes & data_types,
BoolMask initial_mask = BoolMask(false, false)) const;
/// Same as checkInRange, but calculate only may_be_true component of a result.
/// This is more efficient than checkInRange(...).can_be_true.
bool mayBeTrueInRange(
size_t used_key_size,
const FieldRef * left_key,
const FieldRef * right_key,
const DataTypes & data_types) const;
/// Same as checkAfter, but calculate only may_be_true component of a result.
/// This is more efficient than checkAfter(...).can_be_true.
bool mayBeTrueAfter(
size_t used_key_size,
const FieldRef * left_key,
const DataTypes & data_types) const;
/// Checks that the index can not be used.
bool alwaysUnknownOrTrue() const;
/// Get the maximum number of the key element used in the condition.
size_t getMaxKeyColumn() const;
bool hasMonotonicFunctionsChain() const;
/// Impose an additional condition: the value in the column `column` must be in the range `range`.
/// Returns whether there is such a column in the key.
bool addCondition(const String & column, const Range & range);
String toString() const;
/** A chain of possibly monotone functions.
* If the key column is wrapped in functions that can be monotonous in some value ranges
* (for example: -toFloat64(toDayOfWeek(date))), then here the functions will be located: toDayOfWeek, toFloat64, negate.
*/
using MonotonicFunctionsChain = std::vector<FunctionBasePtr>;
/** Computes value of constant expression and its data type.
* Returns false, if expression isn't constant.
*/
static bool getConstant(
const ASTPtr & expr, Block & block_with_constants, Field & out_value, DataTypePtr & out_type);
static Block getBlockWithConstants(
const ASTPtr & query, const SyntaxAnalyzerResultPtr & syntax_analyzer_result, const Context & context);
static std::optional<Range> applyMonotonicFunctionsChainToRange(
Range key_range,
MonotonicFunctionsChain & functions,
DataTypePtr current_type);
bool matchesExactContinuousRange() const;
private:
/// The expression is stored as Reverse Polish Notation.
struct RPNElement
{
enum Function
{
/// Atoms of a Boolean expression.
FUNCTION_IN_RANGE,
FUNCTION_NOT_IN_RANGE,
FUNCTION_IN_SET,
FUNCTION_NOT_IN_SET,
FUNCTION_UNKNOWN, /// Can take any value.
/// Operators of the logical expression.
FUNCTION_NOT,
FUNCTION_AND,
FUNCTION_OR,
/// Constants
ALWAYS_FALSE,
ALWAYS_TRUE,
};
RPNElement() {}
RPNElement(Function function_) : function(function_) {}
RPNElement(Function function_, size_t key_column_) : function(function_), key_column(key_column_) {}
RPNElement(Function function_, size_t key_column_, const Range & range_)
: function(function_), range(range_), key_column(key_column_) {}
String toString() const;
Function function = FUNCTION_UNKNOWN;
/// For FUNCTION_IN_RANGE and FUNCTION_NOT_IN_RANGE.
Range range;
size_t key_column = 0;
/// For FUNCTION_IN_SET, FUNCTION_NOT_IN_SET
using MergeTreeSetIndexPtr = std::shared_ptr<MergeTreeSetIndex>;
MergeTreeSetIndexPtr set_index;
mutable MonotonicFunctionsChain monotonic_functions_chain; /// The function execution does not violate the constancy.
};
using RPN = std::vector<RPNElement>;
using ColumnIndices = std::map<String, size_t>;
using AtomMap = std::unordered_map<std::string, bool(*)(RPNElement & out, const Field & value)>;
public:
static const AtomMap atom_map;
private:
BoolMask checkInRange(
size_t used_key_size,
const FieldRef * left_key,
const FieldRef * right_key,
const DataTypes & data_types,
bool right_bounded,
BoolMask initial_mask) const;
void traverseAST(const ASTPtr & node, const Context & context, Block & block_with_constants);
bool tryParseAtomFromAST(const ASTPtr & node, const Context & context, Block & block_with_constants, RPNElement & out);
static bool tryParseLogicalOperatorFromAST(const ASTFunction * func, RPNElement & out);
/** Is node the key column
* or expression in which column of key is wrapped by chain of functions,
* that can be monotonic on certain ranges?
* If these conditions are true, then returns number of column in key, type of resulting expression
* and fills chain of possibly-monotonic functions.
*/
bool isKeyPossiblyWrappedByMonotonicFunctions(
const ASTPtr & node,
const Context & context,
size_t & out_key_column_num,
DataTypePtr & out_key_res_column_type,
MonotonicFunctionsChain & out_functions_chain);
bool isKeyPossiblyWrappedByMonotonicFunctionsImpl(
const ASTPtr & node,
size_t & out_key_column_num,
DataTypePtr & out_key_column_type,
std::vector<const ASTFunction *> & out_functions_chain);
bool canConstantBeWrappedByMonotonicFunctions(
const ASTPtr & node,
size_t & out_key_column_num,
DataTypePtr & out_key_column_type,
Field & out_value,
DataTypePtr & out_type);
/// If it's possible to make an RPNElement
/// that will filter values (possibly tuples) by the content of 'prepared_set',
/// do it and return true.
bool tryPrepareSetIndex(
const ASTs & args,
const Context & context,
RPNElement & out,
size_t & out_key_column_num);
RPN rpn;
ColumnIndices key_columns;
ExpressionActionsPtr key_expr;
PreparedSets prepared_sets;
};
}