thevar1able/ClickHouse
1
0
Fork 0
mirror of https://github.com/ClickHouse/ClickHouse.git synced 2024-11-21 15:12:02 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge f559037c4d into 44b4bd38b9

Browse Source
This commit is contained in:
Arthur Passos 2024-11-20 15:12:00 -08:00 committed by GitHub
commit 7fdf3efe63
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
9 changed files with 304 additions and 330 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

73
src/Processors/Formats/Impl/Parquet/ParquetBloomFilterCondition.h
View File

@ -1,73 +0,0 @@
#pragma once
#include <config.h>
#if USE_PARQUET
#include <Storages/MergeTree/KeyCondition.h>
#include <parquet/metadata.h>
#include <Processors/Formats/Impl/ArrowFieldIndexUtil.h>
namespace parquet
{
class BloomFilter;
}
namespace DB
{
class ParquetBloomFilterCondition
{
public:
struct ConditionElement
{
enum Function
{
/// Atoms of a Boolean expression.
FUNCTION_IN,
FUNCTION_NOT_IN,
/// Can take any value.
FUNCTION_UNKNOWN,
/// Operators of the logical expression.
FUNCTION_NOT,
FUNCTION_AND,
FUNCTION_OR,
/// Constants
ALWAYS_FALSE,
ALWAYS_TRUE,
};
using ColumnPtr = IColumn::Ptr;
using HashesForColumns = std::vector<std::vector<uint64_t>>;
using KeyColumns = std::vector<std::size_t>;
Function function;
// each entry represents a list of hashes per column
// suppose there are three columns with 2 rows each
// hashes_per_column.size() == 3 and hashes_per_column[0].size() == 2
HashesForColumns hashes_per_column;
KeyColumns key_columns;
};
using RPNElement = KeyCondition::RPNElement;
using ColumnIndexToBF = std::unordered_map<std::size_t, std::unique_ptr<parquet::BloomFilter>>;
explicit ParquetBloomFilterCondition(const std::vector<ConditionElement> & condition_, const Block & header_);
bool mayBeTrueOnRowGroup(const ColumnIndexToBF & column_index_to_column_bf) const;
std::unordered_set<std::size_t> getFilteringColumnKeys() const;
private:
std::vector<ParquetBloomFilterCondition::ConditionElement> condition;
Block header;
};
std::vector<ParquetBloomFilterCondition::ConditionElement> keyConditionRPNToParquetBloomFilterCondition(
const std::vector<KeyCondition::RPNElement> & rpn,
const std::vector<ArrowFieldIndexUtil::ClickHouseIndexToParquetIndex> & clickhouse_column_index_to_parquet_index,
const std::unique_ptr<parquet::RowGroupMetaData> & parquet_rg_metadata);
}
#endif

264
src/Processors/Formats/Impl/Parquet/ParquetBloomFilterCondition.cpp → src/Processors/Formats/Impl/Parquet/keyConditionRPNToParquetRPN.cpp
View File

@ -1,12 +1,10 @@
#include <Processors/Formats/Impl/Parquet/ParquetBloomFilterCondition.h>
#include <iostream>
#include <Processors/Formats/Impl/Parquet/keyConditionRPNToParquetRPN.h>
#if USE_PARQUET
#include <parquet/bloom_filter.h>
#include <parquet/metadata.h>
#include <parquet/xxhasher.h>
#include <Interpreters/convertFieldToType.h>
#include <Columns/ColumnConst.h>
namespace DB
{
@ -16,8 +14,43 @@ namespace ErrorCodes
extern const int LOGICAL_ERROR;
}
namespace
const parquet::ColumnDescriptor * getColumnDescriptorIfBloomFilterIsPresent(
const std::unique_ptr<parquet::RowGroupMetaData> & parquet_rg_metadata,
const std::vector<ArrowFieldIndexUtil::ClickHouseIndexToParquetIndex> & clickhouse_column_index_to_parquet_index,
std::size_t clickhouse_column_index)
{
if (clickhouse_column_index_to_parquet_index.size() <= clickhouse_column_index)
{
return nullptr;
}
const auto & parquet_indexes = clickhouse_column_index_to_parquet_index[clickhouse_column_index].parquet_indexes;
// complex types like structs, tuples and maps will have more than one index.
// we don't support those for now
if (parquet_indexes.size() > 1)
{
return nullptr;
}
if (parquet_indexes.empty())
{
throw Exception(ErrorCodes::LOGICAL_ERROR, "Something bad happened, raise an issue and try the query with `input_format_parquet_bloom_filter_push_down=false`");
}
auto parquet_column_index = parquet_indexes[0];
const auto * parquet_column_descriptor = parquet_rg_metadata->schema()->Column(parquet_column_index);
bool column_has_bloom_filter = parquet_rg_metadata->ColumnChunk(parquet_column_index)->bloom_filter_offset().has_value();
if (!column_has_bloom_filter)
{
return nullptr;
}
return parquet_column_descriptor;
}
bool isParquetStringTypeSupportedForBloomFilters(
const std::shared_ptr<const parquet::LogicalType> & logical_type,
@ -48,9 +81,9 @@ bool isParquetIntegerTypeSupportedForBloomFilters(const std::shared_ptr<const pa
}
if (parquet::ConvertedType::type::NONE != converted_type && !(converted_type == parquet::ConvertedType::INT_8 || converted_type == parquet::ConvertedType::INT_16
|| converted_type == parquet::ConvertedType::INT_32 || converted_type == parquet::ConvertedType::INT_64
|| converted_type == parquet::ConvertedType::UINT_8 || converted_type == parquet::ConvertedType::UINT_16
|| converted_type == parquet::ConvertedType::UINT_32 || converted_type == parquet::ConvertedType::UINT_64))
|| converted_type == parquet::ConvertedType::INT_32 || converted_type == parquet::ConvertedType::INT_64
|| converted_type == parquet::ConvertedType::UINT_8 || converted_type == parquet::ConvertedType::UINT_16
|| converted_type == parquet::ConvertedType::UINT_32 || converted_type == parquet::ConvertedType::UINT_64))
{
return false;
}
@ -197,194 +230,25 @@ std::optional<std::vector<uint64_t>> hash(const IColumn * data_column, const par
return hashes;
}
bool maybeTrueOnBloomFilter(const std::vector<uint64_t> & hashes, const std::unique_ptr<parquet::BloomFilter> & bloom_filter)
KeyCondition::RPN keyConditionRPNToParquetRPN(const std::vector<KeyCondition::RPNElement> & rpn,
const std::vector<ArrowFieldIndexUtil::ClickHouseIndexToParquetIndex> & clickhouse_column_index_to_parquet_index,
const std::unique_ptr<parquet::RowGroupMetaData> & parquet_rg_metadata)
{
for (const auto hash : hashes)
{
if (bloom_filter->FindHash(hash))
{
return true;
}
}
return false;
}
const parquet::ColumnDescriptor * getColumnDescriptorIfBloomFilterIsPresent(
const std::unique_ptr<parquet::RowGroupMetaData> & parquet_rg_metadata,
const std::vector<ArrowFieldIndexUtil::ClickHouseIndexToParquetIndex> & clickhouse_column_index_to_parquet_index,
std::size_t clickhouse_column_index)
{
if (clickhouse_column_index_to_parquet_index.size() <= clickhouse_column_index)
{
return nullptr;
}
const auto & parquet_indexes = clickhouse_column_index_to_parquet_index[clickhouse_column_index].parquet_indexes;
// complex types like structs, tuples and maps will have more than one index.
// we don't support those for now
if (parquet_indexes.size() > 1)
{
return nullptr;
}
if (parquet_indexes.empty())
{
throw Exception(ErrorCodes::LOGICAL_ERROR, "Something bad happened, raise an issue and try the query with `input_format_parquet_bloom_filter_push_down=false`");
}
auto parquet_column_index = parquet_indexes[0];
const auto * parquet_column_descriptor = parquet_rg_metadata->schema()->Column(parquet_column_index);
bool column_has_bloom_filter = parquet_rg_metadata->ColumnChunk(parquet_column_index)->bloom_filter_offset().has_value();
if (!column_has_bloom_filter)
{
return nullptr;
}
return parquet_column_descriptor;
}
}
ParquetBloomFilterCondition::ParquetBloomFilterCondition(const std::vector<ConditionElement> & condition_, const Block & header_)
: condition(condition_), header(header_)
{
}
bool ParquetBloomFilterCondition::mayBeTrueOnRowGroup(const ColumnIndexToBF & column_index_to_column_bf) const
{
using Function = ConditionElement::Function;
std::vector<BoolMask> rpn_stack;
for (const auto & element : condition)
{
if (element.function == Function::FUNCTION_IN
|| element.function == Function::FUNCTION_NOT_IN)
{
bool maybe_true = true;
for (auto column_index = 0u; column_index < element.hashes_per_column.size(); column_index++)
{
// in case bloom filter is not present for this row group
// https://github.com/ClickHouse/ClickHouse/pull/62966#discussion_r1722361237
if (!column_index_to_column_bf.contains(element.key_columns[column_index]))
{
rpn_stack.emplace_back(true, true);
continue;
}
bool column_maybe_contains = maybeTrueOnBloomFilter(
element.hashes_per_column[column_index],
column_index_to_column_bf.at(element.key_columns[column_index]));
if (!column_maybe_contains)
{
maybe_true = false;
break;
}
}
rpn_stack.emplace_back(maybe_true, true);
if (element.function == Function::FUNCTION_NOT_IN)
rpn_stack.back() = !rpn_stack.back();
}
else if (element.function == Function::FUNCTION_NOT)
{
rpn_stack.back() = !rpn_stack.back();
}
else if (element.function == Function::FUNCTION_OR)
{
auto arg1 = rpn_stack.back();
rpn_stack.pop_back();
auto arg2 = rpn_stack.back();
rpn_stack.back() = arg1 | arg2;
}
else if (element.function == Function::FUNCTION_AND)
{
auto arg1 = rpn_stack.back();
rpn_stack.pop_back();
auto arg2 = rpn_stack.back();
rpn_stack.back() = arg1 & arg2;
}
else if (element.function == Function::ALWAYS_TRUE)
{
rpn_stack.emplace_back(true, false);
}
else if (element.function == Function::ALWAYS_FALSE)
{
rpn_stack.emplace_back(false, true);
}
else
{
rpn_stack.emplace_back(true, true);
}
}
if (rpn_stack.size() != 1)
throw Exception(ErrorCodes::LOGICAL_ERROR, "Unexpected stack size in KeyCondition::mayBeTrueOnRowGroup");
return rpn_stack[0].can_be_true;
}
std::unordered_set<std::size_t> ParquetBloomFilterCondition::getFilteringColumnKeys() const
{
std::unordered_set<std::size_t> column_keys;
for (const auto & element : condition)
{
for (const auto index : element.key_columns)
{
column_keys.insert(index);
}
}
return column_keys;
}
/*
* `KeyCondition::rpn` is overly complex for bloom filters, some operations are not even supported. Not only that, but to avoid hashing each time
* we loop over a rpn element, we need to store hashes instead of where predicate values. To address this, we loop over `KeyCondition::rpn`
* and build a simplified RPN that holds hashes instead of values.
*
* `KeyCondition::RPNElement::FUNCTION_IN_RANGE` becomes:
* `FUNCTION_IN`
* `FUNCTION_UNKNOWN` when range limits are different
* `KeyCondition::RPNElement::FUNCTION_IN_SET` becomes
* `FUNCTION_IN`
*
* Complex types and structs are not supported.
* There are two sources of data types being analyzed, and they need to be compatible: DB::Field type and parquet type.
* This is determined by the `isColumnSupported` method.
*
* Some interesting examples:
* 1. file(..., 'str_column UInt64') where str_column = 50; Field.type == UInt64. Parquet type string. Not supported.
* 2. file(...) where str_column = 50; Field.type == String (conversion already taken care by `KeyCondition`). Parquet type string.
* 3. file(...) where uint32_column = toIPv4(5). Field.type == IPv4. Incompatible column types, resolved by `KeyCondition` itself.
* 4. file(...) where toIPv4(uint32_column) = toIPv4(5). Field.type == IPv4. We know it is safe to hash it using an int32 API.
* */
std::vector<ParquetBloomFilterCondition::ConditionElement> keyConditionRPNToParquetBloomFilterCondition(
const std::vector<KeyCondition::RPNElement> & rpn,
const std::vector<ArrowFieldIndexUtil::ClickHouseIndexToParquetIndex> & clickhouse_column_index_to_parquet_index,
const std::unique_ptr<parquet::RowGroupMetaData> & parquet_rg_metadata)
{
std::vector<ParquetBloomFilterCondition::ConditionElement> condition_elements;
std::vector<KeyCondition::RPNElement> condition_elements;
using RPNElement = KeyCondition::RPNElement;
using Function = ParquetBloomFilterCondition::ConditionElement::Function;
for (const auto & rpn_element : rpn)
{
condition_elements.emplace_back(rpn_element);
// this would be a problem for `where negate(x) = -58`.
// It would perform a bf search on `-58`, and possibly miss row groups containing this data.
if (!rpn_element.monotonic_functions_chain.empty())
{
condition_elements.emplace_back(Function::FUNCTION_UNKNOWN);
continue;
}
ParquetBloomFilterCondition::ConditionElement::HashesForColumns hashes;
KeyCondition::BloomFilterData::HashesForColumns hashes;
if (rpn_element.function == RPNElement::FUNCTION_IN_RANGE
|| rpn_element.function == RPNElement::FUNCTION_NOT_IN_RANGE)
@ -392,7 +256,6 @@ std::vector<ParquetBloomFilterCondition::ConditionElement> keyConditionRPNToParq
// Only FUNCTION_EQUALS is supported and for that extremes need to be the same
if (rpn_element.range.left != rpn_element.range.right)
{
condition_elements.emplace_back(Function::FUNCTION_UNKNOWN);
continue;
}
@ -401,7 +264,6 @@ std::vector<ParquetBloomFilterCondition::ConditionElement> keyConditionRPNToParq
if (!parquet_column_descriptor)
{
condition_elements.emplace_back(Function::FUNCTION_UNKNOWN);
continue;
}
@ -409,7 +271,6 @@ std::vector<ParquetBloomFilterCondition::ConditionElement> keyConditionRPNToParq
if (!hashed_value)
{
condition_elements.emplace_back(Function::FUNCTION_UNKNOWN);
continue;
}
@ -418,14 +279,10 @@ std::vector<ParquetBloomFilterCondition::ConditionElement> keyConditionRPNToParq
hashes.emplace_back(std::move(hashes_for_column));
auto function = rpn_element.function == RPNElement::FUNCTION_IN_RANGE
? ParquetBloomFilterCondition::ConditionElement::Function::FUNCTION_IN
: ParquetBloomFilterCondition::ConditionElement::Function::FUNCTION_NOT_IN;
std::vector<std::size_t> key_columns;
key_columns.emplace_back(rpn_element.key_column);
condition_elements.emplace_back(function, std::move(hashes), std::move(key_columns));
condition_elements.back().bloom_filter_data = KeyCondition::BloomFilterData {std::move(hashes), std::move(key_columns)};
}
else if (rpn_element.function == RPNElement::FUNCTION_IN_SET
|| rpn_element.function == RPNElement::FUNCTION_NOT_IN_SET)
@ -485,35 +342,16 @@ std::vector<ParquetBloomFilterCondition::ConditionElement> keyConditionRPNToParq
if (found_empty_column)
{
condition_elements.emplace_back(Function::ALWAYS_FALSE);
// todo arthur
continue;
}
if (hashes.empty())
{
condition_elements.emplace_back(Function::FUNCTION_UNKNOWN);
continue;
}
auto function = RPNElement::FUNCTION_IN_SET == rpn_element.function ? Function::FUNCTION_IN : Function::FUNCTION_NOT_IN;
condition_elements.emplace_back(function, hashes, key_columns);
}
else if (rpn_element.function == RPNElement::FUNCTION_NOT)
{
condition_elements.emplace_back(Function::FUNCTION_NOT);
}
else if (rpn_element.function == RPNElement::FUNCTION_OR)
{
condition_elements.emplace_back(Function::FUNCTION_OR);
}
else if (rpn_element.function == RPNElement::FUNCTION_AND)
{
condition_elements.emplace_back(Function::FUNCTION_AND);
}
else
{
condition_elements.emplace_back(Function::ALWAYS_TRUE);
condition_elements.back().bloom_filter_data = {std::move(hashes), std::move(key_columns)};
}
}

24
src/Processors/Formats/Impl/Parquet/keyConditionRPNToParquetRPN.h Normal file
View File

@ -0,0 +1,24 @@
#pragma once
#include <config.h>
#if USE_PARQUET
#include <Storages/MergeTree/KeyCondition.h>
#include <Processors/Formats/Impl/ArrowFieldIndexUtil.h>
namespace parquet
{
class RowGroupMetadata;
}
namespace DB
{
KeyCondition::RPN keyConditionRPNToParquetRPN(const std::vector<KeyCondition::RPNElement> & rpn,
const std::vector<ArrowFieldIndexUtil::ClickHouseIndexToParquetIndex> & clickhouse_column_index_to_parquet_index,
const std::unique_ptr<parquet::RowGroupMetaData> & parquet_rg_metadata);
}
#endif

107
src/Processors/Formats/Impl/ParquetBlockInputFormat.cpp
View File

@ -27,7 +27,7 @@
#include <DataTypes/DataTypeNullable.h>
#include <Common/FieldVisitorsAccurateComparison.h>
#include <Processors/Formats/Impl/Parquet/ParquetRecordReader.h>
#include <Processors/Formats/Impl/Parquet/ParquetBloomFilterCondition.h>
#include <Processors/Formats/Impl/Parquet/keyConditionRPNToParquetRPN.h>
#include <Interpreters/convertFieldToType.h>
namespace ProfileEvents
@ -275,7 +275,21 @@ static Field decodePlainParquetValueSlow(const std::string & data, parquet::Type
return field;
}
static ParquetBloomFilterCondition::ColumnIndexToBF buildColumnIndexToBF(
struct ParquetBloomFilter : public KeyCondition::BloomFilter
{
explicit ParquetBloomFilter(std::unique_ptr<parquet::BloomFilter> && parquet_bf_)
: parquet_bf(std::move(parquet_bf_)) {}
bool findHash(uint64_t hash) override
{
return parquet_bf->FindHash(hash);
}
private:
std::unique_ptr<parquet::BloomFilter> parquet_bf;
};
static KeyCondition::ColumnIndexToBloomFilter buildColumnIndexToBF(
parquet::BloomFilterReader & bf_reader,
int row_group,
const std::vector<ArrowFieldIndexUtil::ClickHouseIndexToParquetIndex> & clickhouse_column_index_to_parquet_index,
@ -289,7 +303,7 @@ static ParquetBloomFilterCondition::ColumnIndexToBF buildColumnIndexToBF(
return {};
}
ParquetBloomFilterCondition::ColumnIndexToBF index_to_column_bf;
KeyCondition::ColumnIndexToBloomFilter index_to_column_bf;
for (const auto & [clickhouse_index, parquet_indexes] : clickhouse_column_index_to_parquet_index)
{
@ -306,14 +320,14 @@ static ParquetBloomFilterCondition::ColumnIndexToBF buildColumnIndexToBF(
auto parquet_index = parquet_indexes[0];
auto bf = rg_bf->GetColumnBloomFilter(parquet_index);
auto parquet_bf = rg_bf->GetColumnBloomFilter(parquet_index);
if (!bf)
if (!parquet_bf)
{
continue;
}
index_to_column_bf[clickhouse_index] = std::move(bf);
index_to_column_bf[clickhouse_index] = std::make_unique<ParquetBloomFilter>(std::move(parquet_bf));
}
return index_to_column_bf;
@ -484,6 +498,24 @@ static std::vector<Range> getHyperrectangleForRowGroup(const parquet::FileMetaDa
return hyperrectangle;
}
std::unordered_set<std::size_t> getBloomFilterFilteringColumnKeys(const KeyCondition::RPN & rpn)
{
std::unordered_set<std::size_t> column_keys;
for (const auto & element : rpn)
{
if (auto bf_data = element.bloom_filter_data)
{
for (const auto index : bf_data->key_columns)
{
column_keys.insert(index);
}
}
}
return column_keys;
}
ParquetBlockInputFormat::ParquetBlockInputFormat(
ReadBuffer & buf,
const Block & header_,
@ -577,45 +609,68 @@ void ParquetBlockInputFormat::initializeIfNeeded()
return std::min(std::max(preferred_num_rows, MIN_ROW_NUM), static_cast<size_t>(format_settings.parquet.max_block_size));
};
std::unique_ptr<ParquetBloomFilterCondition> parquet_bloom_filter_condition;
std::unordered_set<std::size_t> filtering_columns;
if (format_settings.parquet.bloom_filter_push_down && key_condition)
{
bf_reader = parquet::BloomFilterReader::Make(arrow_file, metadata, bf_reader_properties, nullptr);
const auto parquet_conditions = keyConditionRPNToParquetBloomFilterCondition(
const auto parquet_conditions = keyConditionRPNToParquetRPN(
key_condition->getRPN(),
index_mapping,
metadata->RowGroup(0));
parquet_bloom_filter_condition = std::make_unique<ParquetBloomFilterCondition>(parquet_conditions, getPort().getHeader());
filtering_columns = parquet_bloom_filter_condition->getFilteringColumnKeys();
filtering_columns = getBloomFilterFilteringColumnKeys(parquet_conditions);
}
auto skip_row_group_based_on_filters = [&](int row_group)
{
if (!format_settings.parquet.filter_push_down && !format_settings.parquet.bloom_filter_push_down)
{
return false;
}
KeyCondition::RPN possibly_modified_rpn = key_condition->getRPN();
KeyCondition::ColumnIndexToBloomFilter column_index_to_bloom_filter;
const auto & header = getPort().getHeader();
std::vector<Range> hyperrectangle(header.columns(), Range::createWholeUniverse());
if (format_settings.parquet.filter_push_down)
{
hyperrectangle = getHyperrectangleForRowGroup(*metadata, row_group, header, format_settings);getHyperrectangleForRowGroup(*metadata, row_group, getPort().getHeader(), format_settings);
}
if (format_settings.parquet.bloom_filter_push_down)
{
possibly_modified_rpn = keyConditionRPNToParquetRPN(key_condition->getRPN(),
index_mapping,
metadata->RowGroup(row_group));
column_index_to_bloom_filter = buildColumnIndexToBF(*bf_reader, row_group, index_mapping, filtering_columns);
}
bool maybe_exists = KeyCondition::checkRPNAgainstHyperrectangle(
possibly_modified_rpn,
hyperrectangle,
key_condition->key_space_filling_curves,
getPort().getHeader().getDataTypes(),
key_condition->isSinglePoint(),
column_index_to_bloom_filter).can_be_true;
return !maybe_exists;
};
for (int row_group = 0; row_group < num_row_groups; ++row_group)
{
if (skip_row_groups.contains(row_group))
continue;
if (parquet_bloom_filter_condition)
if (key_condition && skip_row_group_based_on_filters(row_group))
{
const auto column_index_to_bf = buildColumnIndexToBF(*bf_reader, row_group, index_mapping, filtering_columns);
if (!parquet_bloom_filter_condition->mayBeTrueOnRowGroup(column_index_to_bf))
{
continue;
}
}
if (format_settings.parquet.filter_push_down && key_condition
&& !key_condition
->checkInHyperrectangle(
getHyperrectangleForRowGroup(*metadata, row_group, getPort().getHeader(), format_settings),
getPort().getHeader().getDataTypes())
.can_be_true)
continue;
}
// When single-threaded parsing, can prefetch row groups, so need to put all row groups in the same row_group_batch
if (row_group_batches.empty() || (!prefetch_group && row_group_batches.back().total_bytes_compressed >= min_bytes_for_seek))

64
src/Storages/MergeTree/KeyCondition.cpp
View File

@ -3004,6 +3004,57 @@ bool KeyCondition::extractPlainRanges(Ranges & ranges) const
BoolMask KeyCondition::checkInHyperrectangle(
const Hyperrectangle & hyperrectangle,
const DataTypes & data_types) const
{
return checkRPNAgainstHyperrectangle(rpn, hyperrectangle, key_space_filling_curves, data_types, single_point);
}
bool mayExistOnBloomFilter(const std::vector<uint64_t> & hashes, const std::unique_ptr<KeyCondition::BloomFilter> & bloom_filter)
{
for (const auto hash : hashes)
{
if (bloom_filter->findHash(hash))
{
return true;
}
}
return false;
}
bool mayExistOnBloomFilter(const KeyCondition::BloomFilterData & condition_bloom_filter_data,
const KeyCondition::ColumnIndexToBloomFilter & column_index_to_column_bf)
{
bool maybe_true = true;
for (auto column_index = 0u; column_index < condition_bloom_filter_data.hashes_per_column.size(); column_index++)
{
// in case bloom filter is not present for this row group
// https://github.com/ClickHouse/ClickHouse/pull/62966#discussion_r1722361237
if (!column_index_to_column_bf.contains(condition_bloom_filter_data.key_columns[column_index]))
{
continue;
}
bool column_maybe_contains = mayExistOnBloomFilter(
condition_bloom_filter_data.hashes_per_column[column_index],
column_index_to_column_bf.at(condition_bloom_filter_data.key_columns[column_index]));
if (!column_maybe_contains)
{
maybe_true = false;
break;
}
}
return maybe_true;
}
BoolMask KeyCondition::checkRPNAgainstHyperrectangle(
const RPN & rpn,
const Hyperrectangle & hyperrectangle,
const KeyCondition::SpaceFillingCurveDescriptions & key_space_filling_curves,
const DataTypes & data_types,
bool single_point,
const ColumnIndexToBloomFilter & column_index_to_column_bf)
{
std::vector<BoolMask> rpn_stack;
@ -3019,6 +3070,7 @@ BoolMask KeyCondition::checkInHyperrectangle(
{
if (element.argument_num_of_space_filling_curve.has_value())
{
// todo arthur, not sure what to do here yet
/// If a condition on argument of a space filling curve wasn't collapsed into FUNCTION_ARGS_IN_HYPERRECTANGLE,
/// we cannot process it.
rpn_stack.emplace_back(true, true);
@ -3063,6 +3115,12 @@ BoolMask KeyCondition::checkInHyperrectangle(
bool contains = element.range.containsRange(*key_range);
rpn_stack.emplace_back(intersects, !contains);
if (rpn_stack.back().can_be_true && element.bloom_filter_data)
{
rpn_stack.back().can_be_true = mayExistOnBloomFilter(*element.bloom_filter_data, column_index_to_column_bf);
}
if (element.function == RPNElement::FUNCTION_NOT_IN_RANGE)
rpn_stack.back() = !rpn_stack.back();
}
@ -3236,6 +3294,12 @@ BoolMask KeyCondition::checkInHyperrectangle(
throw Exception(ErrorCodes::LOGICAL_ERROR, "Set for IN is not created yet");
rpn_stack.emplace_back(element.set_index->checkInRange(hyperrectangle, data_types, single_point));
if (rpn_stack.back().can_be_true && element.bloom_filter_data)
{
rpn_stack.back().can_be_true = mayExistOnBloomFilter(*element.bloom_filter_data, column_index_to_column_bf);
}
if (element.function == RPNElement::FUNCTION_NOT_IN_SET)
rpn_stack.back() = !rpn_stack.back();
}

65
src/Storages/MergeTree/KeyCondition.h
View File

@ -148,6 +148,13 @@ public:
/// TODO handle the cases when generate RPN.
bool extractPlainRanges(Ranges & ranges) const;
struct BloomFilterData
{
using HashesForColumns = std::vector<std::vector<uint64_t>>;
HashesForColumns hashes_per_column;
std::vector<std::size_t> key_columns;
};
/// The expression is stored as Reverse Polish Notation.
struct RPNElement
{
@ -224,6 +231,8 @@ public:
Polygon polygon;
MonotonicFunctionsChain monotonic_functions_chain;
std::optional<BloomFilterData> bloom_filter_data;
};
using RPN = std::vector<RPNElement>;
@ -237,6 +246,44 @@ public:
bool isRelaxed() const { return relaxed; }
/// Space-filling curves in the key
enum class SpaceFillingCurveType
{
Unknown = 0,
Morton,
Hilbert
};
static const std::unordered_map<String, SpaceFillingCurveType> space_filling_curve_name_to_type;
struct SpaceFillingCurveDescription
{
size_t key_column_pos;
String function_name;
std::vector<String> arguments;
SpaceFillingCurveType type;
};
using SpaceFillingCurveDescriptions = std::vector<SpaceFillingCurveDescription>;
SpaceFillingCurveDescriptions key_space_filling_curves;
struct BloomFilter
{
virtual ~BloomFilter() = default;
virtual bool findHash(uint64_t hash) = 0;
};
using ColumnIndexToBloomFilter = std::unordered_map<std::size_t, std::unique_ptr<BloomFilter>>;
static BoolMask checkRPNAgainstHyperrectangle(
const RPN & rpn,
const Hyperrectangle & hyperrectangle,
const KeyCondition::SpaceFillingCurveDescriptions & key_space_filling_curves,
const DataTypes & data_types,
bool single_point,
const ColumnIndexToBloomFilter & column_index_to_column_bf = {});
bool isSinglePoint() const { return single_point; }
private:
BoolMask checkInRange(
size_t used_key_size,
@ -358,24 +405,6 @@ private:
/// All intermediate columns are used to calculate key_expr.
const NameSet key_subexpr_names;
/// Space-filling curves in the key
enum class SpaceFillingCurveType
{
Unknown = 0,
Morton,
Hilbert
};
static const std::unordered_map<String, SpaceFillingCurveType> space_filling_curve_name_to_type;
struct SpaceFillingCurveDescription
{
size_t key_column_pos;
String function_name;
std::vector<String> arguments;
SpaceFillingCurveType type;
};
using SpaceFillingCurveDescriptions = std::vector<SpaceFillingCurveDescription>;
SpaceFillingCurveDescriptions key_space_filling_curves;
void getAllSpaceFillingCurves();
/// Array joined column names

8
tests/queries/0_stateless/03261_test_merge_parquet_bloom_filter_minmax_stats.reference Normal file
View File

@ -0,0 +1,8 @@
{
"data": [],
"rows": 0,
"statistics": {
"rows_read": 0,
"bytes_read": 0
}
}

29
tests/queries/0_stateless/03261_test_merge_parquet_bloom_filter_minmax_stats.sh Executable file
View File

@ -0,0 +1,29 @@
#!/usr/bin/env bash
# Tags: no-ubsan, no-fasttest
CUR_DIR=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
# shellcheck source=../shell_config.sh
. "$CUR_DIR"/../shell_config.sh
USER_FILES_PATH=$($CLICKHOUSE_CLIENT_BINARY --query "select _path,_file from file('nonexist.txt', 'CSV', 'val1 char')" 2>&1 | grep Exception | awk '{gsub("/nonexist.txt","",$9); print $9}')
WORKING_DIR="${USER_FILES_PATH}/${CLICKHOUSE_TEST_UNIQUE_NAME}"
mkdir -p "${WORKING_DIR}"
DATA_FILE="${CUR_DIR}/data_parquet/integers_1_5_no_3_bf_minmax.parquet"
DATA_FILE_USER_PATH="${WORKING_DIR}/integers_1to5_no_3_bf_minmax.parquet"
cp ${DATA_FILE} ${DATA_FILE_USER_PATH}
# Prior to this PR, bloom filter and minmax were evaluated separately.
# This was sub-optimal for conditions like `x = 3 or x > 5` where data is [1, 2, 4, 5].
# Bloom filter is not able to handle greater than operations. Therefore, it can't evaluate x > 5. Even though it can tell
# `3` is not in the set by evaluating `x = 3`, it can't discard the row group because of the `or` condition.
# On the other hand, min max can handle both. It'll evaluate x = 3 to true (because it is within the range) and the latter to false
# Therefore, bloom filter would determine `false or true` and minmax would determine `true or false`. Resulting in true.
# Since both structures are now evaluated together, the row group should be skipped
${CLICKHOUSE_CLIENT} --query="select * from file('${DATA_FILE_USER_PATH}', Parquet) WHERE int8 = 3 or int8 > 5 FORMAT Json SETTINGS input_format_parquet_filter_push_down=true, input_format_parquet_bloom_filter_push_down=true;" | jq 'del(.meta,.statistics.elapsed)'

BIN
tests/queries/0_stateless/data_parquet/integers_1_5_no_3_bf_minmax.parquet Normal file
View File

Binary file not shown.