ClickHouse/dbms/src/Storages/MergeTree/MergeTreeDataSelectExecutor.cpp

#include <boost/rational.hpp>   /// For calculations related to sampling coefficients.
#include <optional>

#include <Common/FieldVisitors.h>
#include <Storages/MergeTree/MergeTreeDataSelectExecutor.h>
#include <Storages/MergeTree/MergeTreeBlockInputStream.h>
#include <Storages/MergeTree/MergeTreeReadPool.h>
#include <Storages/MergeTree/MergeTreeThreadBlockInputStream.h>
#include <Storages/MergeTree/KeyCondition.h>
#include <Parsers/ASTIdentifier.h>
#include <Parsers/ASTFunction.h>
#include <Parsers/ASTSampleRatio.h>

/// Allow to use __uint128_t as a template parameter for boost::rational.
// https://stackoverflow.com/questions/41198673/uint128-t-not-working-with-clang-and-libstdc
#if !defined(__GLIBCXX_BITSIZE_INT_N_0) && defined(__SIZEOF_INT128__)
namespace std
{
    template <>
    struct numeric_limits<__uint128_t>
    {
        static constexpr bool is_specialized = true;
        static constexpr bool is_signed = false;
        static constexpr bool is_integer = true;
        static constexpr int radix = 2;
        static constexpr int digits = 128;
        static constexpr __uint128_t min () { return 0; } // used in boost 1.65.1+
        static constexpr __uint128_t max () { return __uint128_t(0) - 1; } // used in boost 1.68.0+
    };
}
#endif

#include <DataStreams/ExpressionBlockInputStream.h>
#include <DataStreams/FilterBlockInputStream.h>
#include <DataStreams/CollapsingFinalBlockInputStream.h>
#include <DataStreams/AddingConstColumnBlockInputStream.h>
#include <DataStreams/CreatingSetsBlockInputStream.h>
#include <DataStreams/NullBlockInputStream.h>
#include <DataStreams/SummingSortedBlockInputStream.h>
#include <DataStreams/ReplacingSortedBlockInputStream.h>
#include <DataStreams/AggregatingSortedBlockInputStream.h>
#include <DataStreams/VersionedCollapsingSortedBlockInputStream.h>
#include <DataTypes/DataTypesNumber.h>
#include <DataTypes/DataTypeDate.h>
#include <DataTypes/DataTypeEnum.h>
#include <Storages/VirtualColumnUtils.h>


namespace ProfileEvents
{
    extern const Event SelectedParts;
    extern const Event SelectedRanges;
    extern const Event SelectedMarks;
}


namespace DB
{

namespace ErrorCodes
{
    extern const int INDEX_NOT_USED;
    extern const int SAMPLING_NOT_SUPPORTED;
    extern const int ILLEGAL_TYPE_OF_COLUMN_FOR_FILTER;
    extern const int ILLEGAL_COLUMN;
    extern const int ARGUMENT_OUT_OF_BOUND;
}


MergeTreeDataSelectExecutor::MergeTreeDataSelectExecutor(MergeTreeData & data_)
    : data(data_), log(&Logger::get(data.getLogName() + " (SelectExecutor)"))
{
}


/// Construct a block consisting only of possible values of virtual columns
static Block getBlockWithPartColumn(const MergeTreeData::DataPartsVector & parts)
{
    auto column = ColumnString::create();

    for (const auto & part : parts)
        column->insert(part->name);

    return Block{ColumnWithTypeAndName(std::move(column), std::make_shared<DataTypeString>(), "_part")};
}


size_t MergeTreeDataSelectExecutor::getApproximateTotalRowsToRead(
    const MergeTreeData::DataPartsVector & parts, const KeyCondition & key_condition, const Settings & settings) const
{
    size_t full_marks_count = 0;

    /// We will find out how many rows we would have read without sampling.
    LOG_DEBUG(log, "Preliminary index scan with condition: " << key_condition.toString());

    for (size_t i = 0; i < parts.size(); ++i)
    {
        const MergeTreeData::DataPartPtr & part = parts[i];
        MarkRanges ranges = markRangesFromPKRange(part->index, key_condition, settings);

        /** In order to get a lower bound on the number of rows that match the condition on PK,
          *  consider only guaranteed full marks.
          * That is, do not take into account the first and last marks, which may be incomplete.
          */
        for (size_t j = 0; j < ranges.size(); ++j)
            if (ranges[j].end - ranges[j].begin > 2)
                full_marks_count += ranges[j].end - ranges[j].begin - 2;
    }

    return full_marks_count * data.index_granularity;
}


using RelativeSize = boost::rational<ASTSampleRatio::BigNum>;

std::string toString(const RelativeSize & x)
{
    return ASTSampleRatio::toString(x.numerator()) + "/" + ASTSampleRatio::toString(x.denominator());
}

/// Converts sample size to an approximate number of rows (ex. `SAMPLE 1000000`) to relative value (ex. `SAMPLE 0.1`).
static RelativeSize convertAbsoluteSampleSizeToRelative(const ASTPtr & node, size_t approx_total_rows)
{
    if (approx_total_rows == 0)
        return 1;

    const ASTSampleRatio & node_sample = typeid_cast<const ASTSampleRatio &>(*node);

    auto absolute_sample_size = node_sample.ratio.numerator / node_sample.ratio.denominator;
    return std::min(RelativeSize(1), RelativeSize(absolute_sample_size) / RelativeSize(approx_total_rows));
}


BlockInputStreams MergeTreeDataSelectExecutor::read(
    const Names & column_names_to_return,
    const SelectQueryInfo & query_info,
    const Context & context,
    const size_t max_block_size,
    const unsigned num_streams,
    Int64 max_block_number_to_read) const
{
    return readFromParts(
        data.getDataPartsVector(), column_names_to_return, query_info, context,
        max_block_size, num_streams, max_block_number_to_read);
}

BlockInputStreams MergeTreeDataSelectExecutor::readFromParts(
    MergeTreeData::DataPartsVector parts,
    const Names & column_names_to_return,
    const SelectQueryInfo & query_info,
    const Context & context,
    const size_t max_block_size,
    const unsigned num_streams,
    Int64 max_block_number_to_read) const
{
    size_t part_index = 0;

    /// If query contains restrictions on the virtual column `_part` or `_part_index`, select only parts suitable for it.
    /// The virtual column `_sample_factor` (which is equal to 1 / used sample rate) can be requested in the query.
    Names virt_column_names;
    Names real_column_names;

    bool part_column_queried = false;

    bool sample_factor_column_queried = false;
    Float64 used_sample_factor = 1;

    for (const String & name : column_names_to_return)
    {
        if (name == "_part")
        {
            part_column_queried = true;
            virt_column_names.push_back(name);
        }
        else if (name == "_part_index")
        {
            virt_column_names.push_back(name);
        }
        else if (name == "_sample_factor")
        {
            sample_factor_column_queried = true;
            virt_column_names.push_back(name);
        }
        else
        {
            real_column_names.push_back(name);
        }
    }

    NamesAndTypesList available_real_columns = data.getColumns().getAllPhysical();

    NamesAndTypesList available_real_and_virtual_columns = available_real_columns;
    for (const auto & name : virt_column_names)
        available_real_and_virtual_columns.emplace_back(data.getColumn(name));

    /// If there are only virtual columns in the query, you must request at least one non-virtual one.
    if (real_column_names.empty())
        real_column_names.push_back(ExpressionActions::getSmallestColumn(available_real_columns));

    /// If `_part` virtual column is requested, we try to use it as an index.
    Block virtual_columns_block = getBlockWithPartColumn(parts);
    if (part_column_queried)
        VirtualColumnUtils::filterBlockWithQuery(query_info.query, virtual_columns_block, context);

    std::multiset<String> part_values = VirtualColumnUtils::extractSingleValueFromBlock<String>(virtual_columns_block, "_part");

    data.check(real_column_names);

    const Settings & settings = context.getSettingsRef();
    Names primary_sort_columns = data.getPrimarySortColumns();

    KeyCondition key_condition(
        query_info, context, available_real_and_virtual_columns,
        primary_sort_columns, data.getPrimaryExpression());

    if (settings.force_primary_key && key_condition.alwaysUnknownOrTrue())
    {
        std::stringstream exception_message;
        exception_message << "Primary key (";
        for (size_t i = 0, size = primary_sort_columns.size(); i < size; ++i)
            exception_message << (i == 0 ? "" : ", ") << primary_sort_columns[i];
        exception_message << ") is not used and setting 'force_primary_key' is set.";

        throw Exception(exception_message.str(), ErrorCodes::INDEX_NOT_USED);
    }

    std::optional<KeyCondition> minmax_idx_condition;
    if (data.minmax_idx_expr)
    {
        minmax_idx_condition.emplace(
            query_info, context, available_real_and_virtual_columns,
            data.minmax_idx_columns, data.minmax_idx_expr);

        if (settings.force_index_by_date && minmax_idx_condition->alwaysUnknownOrTrue())
        {
            String msg = "MinMax index by columns (";
            bool first = true;
            for (const String & col : data.minmax_idx_columns)
            {
                if (first)
                    first = false;
                else
                    msg += ", ";
                msg += col;
            }
            msg += ") is not used and setting 'force_index_by_date' is set";

            throw Exception(msg, ErrorCodes::INDEX_NOT_USED);
        }
    }

    /// Select the parts in which there can be data that satisfy `minmax_idx_condition` and that match the condition on `_part`,
    ///  as well as `max_block_number_to_read`.
    {
        auto prev_parts = parts;
        parts.clear();

        for (const auto & part : prev_parts)
        {
            if (part_values.find(part->name) == part_values.end())
                continue;

            if (part->isEmpty())
                continue;

            if (minmax_idx_condition && !minmax_idx_condition->mayBeTrueInParallelogram(
                    part->minmax_idx.parallelogram, data.minmax_idx_column_types))
                continue;

            if (max_block_number_to_read && part->info.max_block > max_block_number_to_read)
                continue;

            parts.push_back(part);
        }
    }

    /// Sampling.
    Names column_names_to_read = real_column_names;
    std::shared_ptr<ASTFunction> filter_function;
    ExpressionActionsPtr filter_expression;

    RelativeSize relative_sample_size = 0;
    RelativeSize relative_sample_offset = 0;

    ASTSelectQuery & select = typeid_cast<ASTSelectQuery &>(*query_info.query);

    auto select_sample_size = select.sample_size();
    auto select_sample_offset = select.sample_offset();

    if (select_sample_size)
    {
        relative_sample_size.assign(
            typeid_cast<const ASTSampleRatio &>(*select_sample_size).ratio.numerator,
            typeid_cast<const ASTSampleRatio &>(*select_sample_size).ratio.denominator);

        if (relative_sample_size < 0)
            throw Exception("Negative sample size", ErrorCodes::ARGUMENT_OUT_OF_BOUND);

        relative_sample_offset = 0;
        if (select_sample_offset)
            relative_sample_offset.assign(
                typeid_cast<const ASTSampleRatio &>(*select_sample_offset).ratio.numerator,
                typeid_cast<const ASTSampleRatio &>(*select_sample_offset).ratio.denominator);

        if (relative_sample_offset < 0)
            throw Exception("Negative sample offset", ErrorCodes::ARGUMENT_OUT_OF_BOUND);

        /// Convert absolute value of the sampling (in form `SAMPLE 1000000` - how many rows to read) into the relative `SAMPLE 0.1` (how much data to read).
        size_t approx_total_rows = 0;
        if (relative_sample_size > 1 || relative_sample_offset > 1)
            approx_total_rows = getApproximateTotalRowsToRead(parts, key_condition, settings);

        if (relative_sample_size > 1)
        {
            relative_sample_size = convertAbsoluteSampleSizeToRelative(select_sample_size, approx_total_rows);
            LOG_DEBUG(log, "Selected relative sample size: " << toString(relative_sample_size));
        }

        /// SAMPLE 1 is the same as the absence of SAMPLE.
        if (relative_sample_size == RelativeSize(1))
            relative_sample_size = 0;

        if (relative_sample_offset > 0 && RelativeSize(0) == relative_sample_size)
            throw Exception("Sampling offset is incorrect because no sampling", ErrorCodes::ARGUMENT_OUT_OF_BOUND);

        if (relative_sample_offset > 1)
        {
            relative_sample_offset = convertAbsoluteSampleSizeToRelative(select_sample_offset, approx_total_rows);
            LOG_DEBUG(log, "Selected relative sample offset: " << toString(relative_sample_offset));
        }
    }

    /** Which range of sampling key values do I need to read?
      * First, in the whole range ("universe") we select the interval
      *  of relative `relative_sample_size` size, offset from the beginning by `relative_sample_offset`.
      *
      * Example: SAMPLE 0.4 OFFSET 0.3
      *
      * [------********------]
      *        ^ - offset
      *        <------> - size
      *
      * If the interval passes through the end of the universe, then cut its right side.
      *
      * Example: SAMPLE 0.4 OFFSET 0.8
      *
      * [----------------****]
      *                  ^ - offset
      *                  <------> - size
      *
      * Next, if the `parallel_replicas_count`, `parallel_replica_offset` settings are set,
      *  then it is necessary to break the received interval into pieces of the number `parallel_replicas_count`,
      *  and select a piece with the number `parallel_replica_offset` (from zero).
      *
      * Example: SAMPLE 0.4 OFFSET 0.3, parallel_replicas_count = 2, parallel_replica_offset = 1
      *
      * [----------****------]
      *        ^ - offset
      *        <------> - size
      *        <--><--> - pieces for different `parallel_replica_offset`, select the second one.
      *
      * It is very important that the intervals for different `parallel_replica_offset` cover the entire range without gaps and overlaps.
      * It is also important that the entire universe can be covered using SAMPLE 0.1 OFFSET 0, ... OFFSET 0.9 and similar decimals.
      */

    bool use_sampling = relative_sample_size > 0 || settings.parallel_replicas_count > 1;
    bool no_data = false;   /// There is nothing left after sampling.

    if (use_sampling)
    {
        if (!data.sampling_expression)
            throw Exception("Illegal SAMPLE: table doesn't support sampling", ErrorCodes::SAMPLING_NOT_SUPPORTED);

        if (sample_factor_column_queried && relative_sample_size != RelativeSize(0))
            used_sample_factor = 1.0 / boost::rational_cast<Float64>(relative_sample_size);

        RelativeSize size_of_universum = 0;
        DataTypePtr type = data.getPrimaryExpression()->getSampleBlock().getByName(data.sampling_expression->getColumnName()).type;

        if (typeid_cast<const DataTypeUInt64 *>(type.get()))
            size_of_universum = RelativeSize(std::numeric_limits<UInt64>::max()) + RelativeSize(1);
        else if (typeid_cast<const DataTypeUInt32 *>(type.get()))
            size_of_universum = RelativeSize(std::numeric_limits<UInt32>::max()) + RelativeSize(1);
        else if (typeid_cast<const DataTypeUInt16 *>(type.get()))
            size_of_universum = RelativeSize(std::numeric_limits<UInt16>::max()) + RelativeSize(1);
        else if (typeid_cast<const DataTypeUInt8 *>(type.get()))
            size_of_universum = RelativeSize(std::numeric_limits<UInt8>::max()) + RelativeSize(1);
        else
            throw Exception("Invalid sampling column type in storage parameters: " + type->getName() + ". Must be unsigned integer type.",
                ErrorCodes::ILLEGAL_TYPE_OF_COLUMN_FOR_FILTER);

        if (settings.parallel_replicas_count > 1)
        {
            if (relative_sample_size == RelativeSize(0))
                relative_sample_size = 1;

            relative_sample_size /= settings.parallel_replicas_count.value;
            relative_sample_offset += relative_sample_size * RelativeSize(settings.parallel_replica_offset.value);
        }

        if (relative_sample_offset >= RelativeSize(1))
            no_data = true;

        /// Calculate the half-interval of `[lower, upper)` column values.
        bool has_lower_limit = false;
        bool has_upper_limit = false;

        RelativeSize lower_limit_rational = relative_sample_offset * size_of_universum;
        RelativeSize upper_limit_rational = (relative_sample_offset + relative_sample_size) * size_of_universum;

        UInt64 lower = boost::rational_cast<ASTSampleRatio::BigNum>(lower_limit_rational);
        UInt64 upper = boost::rational_cast<ASTSampleRatio::BigNum>(upper_limit_rational);

        if (lower > 0)
            has_lower_limit = true;

        if (upper_limit_rational < size_of_universum)
            has_upper_limit = true;

        /*std::cerr << std::fixed << std::setprecision(100)
            << "relative_sample_size: " << relative_sample_size << "\n"
            << "relative_sample_offset: " << relative_sample_offset << "\n"
            << "lower_limit_float: " << lower_limit_rational << "\n"
            << "upper_limit_float: " << upper_limit_rational << "\n"
            << "lower: " << lower << "\n"
            << "upper: " << upper << "\n";*/

        if ((has_upper_limit && upper == 0)
            || (has_lower_limit && has_upper_limit && lower == upper))
            no_data = true;

        if (no_data || (!has_lower_limit && !has_upper_limit))
        {
            use_sampling = false;
        }
        else
        {
            /// Let's add the conditions to cut off something else when the index is scanned again and when the request is processed.

            std::shared_ptr<ASTFunction> lower_function;
            std::shared_ptr<ASTFunction> upper_function;

            if (has_lower_limit)
            {
                if (!key_condition.addCondition(data.sampling_expression->getColumnName(), Range::createLeftBounded(lower, true)))
                    throw Exception("Sampling column not in primary key", ErrorCodes::ILLEGAL_COLUMN);

                ASTPtr args = std::make_shared<ASTExpressionList>();
                args->children.push_back(data.sampling_expression);
                args->children.push_back(std::make_shared<ASTLiteral>(lower));

                lower_function = std::make_shared<ASTFunction>();
                lower_function->name = "greaterOrEquals";
                lower_function->arguments = args;
                lower_function->children.push_back(lower_function->arguments);

                filter_function = lower_function;
            }

            if (has_upper_limit)
            {
                if (!key_condition.addCondition(data.sampling_expression->getColumnName(), Range::createRightBounded(upper, false)))
                    throw Exception("Sampling column not in primary key", ErrorCodes::ILLEGAL_COLUMN);

                ASTPtr args = std::make_shared<ASTExpressionList>();
                args->children.push_back(data.sampling_expression);
                args->children.push_back(std::make_shared<ASTLiteral>(upper));

                upper_function = std::make_shared<ASTFunction>();
                upper_function->name = "less";
                upper_function->arguments = args;
                upper_function->children.push_back(upper_function->arguments);

                filter_function = upper_function;
            }

            if (has_lower_limit && has_upper_limit)
            {
                ASTPtr args = std::make_shared<ASTExpressionList>();
                args->children.push_back(lower_function);
                args->children.push_back(upper_function);

                filter_function = std::make_shared<ASTFunction>();
                filter_function->name = "and";
                filter_function->arguments = args;
                filter_function->children.push_back(filter_function->arguments);
            }

            filter_expression = ExpressionAnalyzer(filter_function, context, nullptr, available_real_columns).getActions(false);

            /// Add columns needed for `sampling_expression` to `column_names_to_read`.
            std::vector<String> add_columns = filter_expression->getRequiredColumns();
            column_names_to_read.insert(column_names_to_read.end(), add_columns.begin(), add_columns.end());
            std::sort(column_names_to_read.begin(), column_names_to_read.end());
            column_names_to_read.erase(std::unique(column_names_to_read.begin(), column_names_to_read.end()), column_names_to_read.end());
        }
    }

    if (no_data)
    {
        LOG_DEBUG(log, "Sampling yields no data.");
        return {};
    }

    LOG_DEBUG(log, "Key condition: " << key_condition.toString());
    if (minmax_idx_condition)
        LOG_DEBUG(log, "MinMax index condition: " << minmax_idx_condition->toString());

    /// PREWHERE
    String prewhere_column;
    if (select.prewhere_expression)
        prewhere_column = select.prewhere_expression->getColumnName();

    RangesInDataParts parts_with_ranges;

    /// Let's find what range to read from each part.
    size_t sum_marks = 0;
    size_t sum_ranges = 0;
    for (auto & part : parts)
    {
        RangesInDataPart ranges(part, part_index++);

        if (data.hasPrimaryKey())
            ranges.ranges = markRangesFromPKRange(part->index, key_condition, settings);
        else
            ranges.ranges = MarkRanges{MarkRange{0, part->marks_count}};

        if (!ranges.ranges.empty())
        {
            parts_with_ranges.push_back(ranges);

            sum_ranges += ranges.ranges.size();
            for (const auto & range : ranges.ranges)
                sum_marks += range.end - range.begin;
        }
    }

    LOG_DEBUG(log, "Selected " << parts.size() << " parts by date, " << parts_with_ranges.size() << " parts by key, "
        << sum_marks << " marks to read from " << sum_ranges << " ranges");

    if (parts_with_ranges.empty())
        return {};

    ProfileEvents::increment(ProfileEvents::SelectedParts, parts_with_ranges.size());
    ProfileEvents::increment(ProfileEvents::SelectedRanges, sum_ranges);
    ProfileEvents::increment(ProfileEvents::SelectedMarks, sum_marks);

    BlockInputStreams res;

    if (select.final())
    {
        /// Add columns needed to calculate primary key and the sign.
        std::vector<String> add_columns = data.getPrimaryExpression()->getRequiredColumns();
        column_names_to_read.insert(column_names_to_read.end(), add_columns.begin(), add_columns.end());

        if (!data.merging_params.sign_column.empty())
            column_names_to_read.push_back(data.merging_params.sign_column);
        if (!data.merging_params.version_column.empty())
            column_names_to_read.push_back(data.merging_params.version_column);

        std::sort(column_names_to_read.begin(), column_names_to_read.end());
        column_names_to_read.erase(std::unique(column_names_to_read.begin(), column_names_to_read.end()), column_names_to_read.end());

        res = spreadMarkRangesAmongStreamsFinal(
            std::move(parts_with_ranges),
            column_names_to_read,
            max_block_size,
            settings.use_uncompressed_cache,
            query_info.prewhere_info,
            virt_column_names,
            settings);
    }
    else
    {
        res = spreadMarkRangesAmongStreams(
            std::move(parts_with_ranges),
            num_streams,
            column_names_to_read,
            max_block_size,
            settings.use_uncompressed_cache,
            query_info.prewhere_info,
            virt_column_names,
            settings);
    }

    if (use_sampling)
        for (auto & stream : res)
            stream = std::make_shared<FilterBlockInputStream>(stream, filter_expression, filter_function->getColumnName());

    /// By the way, if a distributed query or query to a Merge table is made, then the `_sample_factor` column can have different values.
    if (sample_factor_column_queried)
        for (auto & stream : res)
            stream = std::make_shared<AddingConstColumnBlockInputStream<Float64>>(
                stream, std::make_shared<DataTypeFloat64>(), used_sample_factor, "_sample_factor");

    return res;
}


BlockInputStreams MergeTreeDataSelectExecutor::spreadMarkRangesAmongStreams(
    RangesInDataParts && parts,
    size_t num_streams,
    const Names & column_names,
    size_t max_block_size,
    bool use_uncompressed_cache,
    const PrewhereInfoPtr & prewhere_info,
    const Names & virt_columns,
    const Settings & settings) const
{
    const size_t min_marks_for_concurrent_read =
        (settings.merge_tree_min_rows_for_concurrent_read + data.index_granularity - 1) / data.index_granularity;
    const size_t max_marks_to_use_cache =
        (settings.merge_tree_max_rows_to_use_cache + data.index_granularity - 1) / data.index_granularity;

    /// Count marks for each part.
    std::vector<size_t> sum_marks_in_parts(parts.size());
    size_t sum_marks = 0;
    for (size_t i = 0; i < parts.size(); ++i)
    {
        /// Let the ranges be listed from right to left so that the leftmost range can be dropped using `pop_back()`.
        std::reverse(parts[i].ranges.begin(), parts[i].ranges.end());

        for (const auto & range : parts[i].ranges)
            sum_marks_in_parts[i] += range.end - range.begin;

        sum_marks += sum_marks_in_parts[i];
    }

    if (sum_marks > max_marks_to_use_cache)
        use_uncompressed_cache = false;

    BlockInputStreams res;

    if (sum_marks > 0 && settings.merge_tree_uniform_read_distribution == 1)
    {
        /// Reduce the number of num_streams if the data is small.
        if (sum_marks < num_streams * min_marks_for_concurrent_read && parts.size() < num_streams)
            num_streams = std::max((sum_marks + min_marks_for_concurrent_read - 1) / min_marks_for_concurrent_read, parts.size());

        MergeTreeReadPoolPtr pool = std::make_shared<MergeTreeReadPool>(
            num_streams, sum_marks, min_marks_for_concurrent_read, parts, data, prewhere_info, true,
            column_names, MergeTreeReadPool::BackoffSettings(settings), settings.preferred_block_size_bytes, false);

        /// Let's estimate total number of rows for progress bar.
        const size_t total_rows = data.index_granularity * sum_marks;
        LOG_TRACE(log, "Reading approx. " << total_rows << " rows");

        for (size_t i = 0; i < num_streams; ++i)
        {
            res.emplace_back(std::make_shared<MergeTreeThreadBlockInputStream>(
                i, pool, min_marks_for_concurrent_read, max_block_size, settings.preferred_block_size_bytes,
                settings.preferred_max_column_in_block_size_bytes, data, use_uncompressed_cache,
                prewhere_info, settings, virt_columns));

            if (i == 0)
            {
                /// Set the approximate number of rows for the first source only
                static_cast<IProfilingBlockInputStream &>(*res.front()).addTotalRowsApprox(total_rows);
            }
        }
    }
    else if (sum_marks > 0)
    {
        const size_t min_marks_per_stream = (sum_marks - 1) / num_streams + 1;

        for (size_t i = 0; i < num_streams && !parts.empty(); ++i)
        {
            size_t need_marks = min_marks_per_stream;

            /// Loop over parts.
            /// We will iteratively take part or some subrange of a part from the back
            ///  and assign a stream to read from it.
            while (need_marks > 0 && !parts.empty())
            {
                RangesInDataPart part = parts.back();
                size_t & marks_in_part = sum_marks_in_parts.back();

                /// We will not take too few rows from a part.
                if (marks_in_part >= min_marks_for_concurrent_read &&
                    need_marks < min_marks_for_concurrent_read)
                    need_marks = min_marks_for_concurrent_read;

                /// Do not leave too few rows in the part.
                if (marks_in_part > need_marks &&
                    marks_in_part - need_marks < min_marks_for_concurrent_read)
                    need_marks = marks_in_part;

                MarkRanges ranges_to_get_from_part;

                /// We take the whole part if it is small enough.
                if (marks_in_part <= need_marks)
                {
                    /// Restore the order of segments.
                    std::reverse(part.ranges.begin(), part.ranges.end());

                    ranges_to_get_from_part = part.ranges;

                    need_marks -= marks_in_part;
                    parts.pop_back();
                    sum_marks_in_parts.pop_back();
                }
                else
                {
                    /// Loop through ranges in part. Take enough ranges to cover "need_marks".
                    while (need_marks > 0)
                    {
                        if (part.ranges.empty())
                            throw Exception("Unexpected end of ranges while spreading marks among streams", ErrorCodes::LOGICAL_ERROR);

                        MarkRange & range = part.ranges.back();

                        const size_t marks_in_range = range.end - range.begin;
                        const size_t marks_to_get_from_range = std::min(marks_in_range, need_marks);

                        ranges_to_get_from_part.emplace_back(range.begin, range.begin + marks_to_get_from_range);
                        range.begin += marks_to_get_from_range;
                        marks_in_part -= marks_to_get_from_range;
                        need_marks -= marks_to_get_from_range;
                        if (range.begin == range.end)
                            part.ranges.pop_back();
                    }
                }

                BlockInputStreamPtr source_stream = std::make_shared<MergeTreeBlockInputStream>(
                    data, part.data_part, max_block_size, settings.preferred_block_size_bytes,
                    settings.preferred_max_column_in_block_size_bytes, column_names, ranges_to_get_from_part,
                    use_uncompressed_cache, prewhere_info, true, settings.min_bytes_to_use_direct_io,
                    settings.max_read_buffer_size, true, virt_columns, part.part_index_in_query);

                res.push_back(source_stream);
            }
        }

        if (!parts.empty())
            throw Exception("Couldn't spread marks among streams", ErrorCodes::LOGICAL_ERROR);
    }

    return res;
}

BlockInputStreams MergeTreeDataSelectExecutor::spreadMarkRangesAmongStreamsFinal(
    RangesInDataParts && parts,
    const Names & column_names,
    size_t max_block_size,
    bool use_uncompressed_cache,
    const PrewhereInfoPtr & prewhere_info,
    const Names & virt_columns,
    const Settings & settings) const
{
    const size_t max_marks_to_use_cache =
        (settings.merge_tree_max_rows_to_use_cache + data.index_granularity - 1) / data.index_granularity;

    size_t sum_marks = 0;
    for (size_t i = 0; i < parts.size(); ++i)
        for (size_t j = 0; j < parts[i].ranges.size(); ++j)
            sum_marks += parts[i].ranges[j].end - parts[i].ranges[j].begin;

    if (sum_marks > max_marks_to_use_cache)
        use_uncompressed_cache = false;

    BlockInputStreams to_merge;

    /// NOTE `merge_tree_uniform_read_distribution` is not used for FINAL

    for (size_t part_index = 0; part_index < parts.size(); ++part_index)
    {
        RangesInDataPart & part = parts[part_index];

        BlockInputStreamPtr source_stream = std::make_shared<MergeTreeBlockInputStream>(
            data, part.data_part, max_block_size, settings.preferred_block_size_bytes,
            settings.preferred_max_column_in_block_size_bytes, column_names, part.ranges, use_uncompressed_cache,
            prewhere_info, true, settings.min_bytes_to_use_direct_io, settings.max_read_buffer_size, true,
            virt_columns, part.part_index_in_query);

        to_merge.emplace_back(std::make_shared<ExpressionBlockInputStream>(source_stream, data.getPrimaryExpression()));
    }

    Names sort_columns = data.getSortColumns();
    SortDescription sort_description;
    size_t sort_columns_size = sort_columns.size();
    sort_description.reserve(sort_columns_size);

    Block header = to_merge.at(0)->getHeader();
    for (size_t i = 0; i < sort_columns_size; ++i)
        sort_description.emplace_back(header.getPositionByName(sort_columns[i]), 1, 1);

    BlockInputStreamPtr merged;
    switch (data.merging_params.mode)
    {
        case MergeTreeData::MergingParams::Ordinary:
            merged = std::make_shared<MergingSortedBlockInputStream>(to_merge, sort_description, max_block_size);
            break;

        case MergeTreeData::MergingParams::Collapsing:
            merged = std::make_shared<CollapsingFinalBlockInputStream>(
                    to_merge, sort_description, data.merging_params.sign_column);
            break;

        case MergeTreeData::MergingParams::Summing:
            merged = std::make_shared<SummingSortedBlockInputStream>(to_merge,
                    sort_description, data.merging_params.columns_to_sum, max_block_size);
            break;

        case MergeTreeData::MergingParams::Aggregating:
            merged = std::make_shared<AggregatingSortedBlockInputStream>(to_merge, sort_description, max_block_size);
            break;

        case MergeTreeData::MergingParams::Replacing:    /// TODO Make ReplacingFinalBlockInputStream
            merged = std::make_shared<ReplacingSortedBlockInputStream>(to_merge,
                    sort_description, data.merging_params.version_column, max_block_size);
            break;

        case MergeTreeData::MergingParams::VersionedCollapsing: /// TODO Make VersionedCollapsingFinalBlockInputStream
            merged = std::make_shared<VersionedCollapsingSortedBlockInputStream>(
                    to_merge, sort_description, data.merging_params.sign_column, max_block_size);
            break;

        case MergeTreeData::MergingParams::Graphite:
            throw Exception("GraphiteMergeTree doesn't support FINAL", ErrorCodes::LOGICAL_ERROR);
    }

    return {merged};
}


void MergeTreeDataSelectExecutor::createPositiveSignCondition(
    ExpressionActionsPtr & out_expression, String & out_column, const Context & context) const
{
    auto function = std::make_shared<ASTFunction>();
    auto arguments = std::make_shared<ASTExpressionList>();
    auto sign = std::make_shared<ASTIdentifier>(data.merging_params.sign_column);
    auto one = std::make_shared<ASTLiteral>(Field(static_cast<Int64>(1)));

    function->name = "equals";
    function->arguments = arguments;
    function->children.push_back(arguments);

    arguments->children.push_back(sign);
    arguments->children.push_back(one);

    out_expression = ExpressionAnalyzer(function, context, {}, data.getColumns().getAllPhysical()).getActions(false);
    out_column = function->getColumnName();
}


/// Calculates a set of mark ranges, that could possibly contain keys, required by condition.
/// In other words, it removes subranges from whole range, that definitely could not contain required keys.
MarkRanges MergeTreeDataSelectExecutor::markRangesFromPKRange(
    const MergeTreeData::DataPart::Index & index, const KeyCondition & key_condition, const Settings & settings) const
{
    MarkRanges res;

    size_t marks_count = index.at(0)->size();
    if (marks_count == 0)
        return res;

    /// If index is not used.
    if (key_condition.alwaysUnknownOrTrue())
    {
        res.push_back(MarkRange(0, marks_count));
    }
    else
    {
        size_t used_key_size = key_condition.getMaxKeyColumn() + 1;
        size_t min_marks_for_seek = (settings.merge_tree_min_rows_for_seek + data.index_granularity - 1) / data.index_granularity;

        /** There will always be disjoint suspicious segments on the stack, the leftmost one at the top (back).
            * At each step, take the left segment and check if it fits.
            * If fits, split it into smaller ones and put them on the stack. If not, discard it.
            * If the segment is already of one mark length, add it to response and discard it.
            */
        std::vector<MarkRange> ranges_stack{ {0, marks_count} };

        /// NOTE Creating temporary Field objects to pass to KeyCondition.
        Row index_left(used_key_size);
        Row index_right(used_key_size);

        while (!ranges_stack.empty())
        {
            MarkRange range = ranges_stack.back();
            ranges_stack.pop_back();

            bool may_be_true;
            if (range.end == marks_count)
            {
                for (size_t i = 0; i < used_key_size; ++i)
                {
                    index[i]->get(range.begin, index_left[i]);
                }

                may_be_true = key_condition.mayBeTrueAfter(
                    used_key_size, &index_left[0], data.primary_key_data_types);
            }
            else
            {
                for (size_t i = 0; i < used_key_size; ++i)
                {
                    index[i]->get(range.begin, index_left[i]);
                    index[i]->get(range.end, index_right[i]);
                }

                may_be_true = key_condition.mayBeTrueInRange(
                    used_key_size, &index_left[0], &index_right[0], data.primary_key_data_types);
            }

            if (!may_be_true)
                continue;

            if (range.end == range.begin + 1)
            {
                /// We saw a useful gap between neighboring marks. Either add it to the last range, or start a new range.
                if (res.empty() || range.begin - res.back().end > min_marks_for_seek)
                    res.push_back(range);
                else
                    res.back().end = range.end;
            }
            else
            {
                /// Break the segment and put the result on the stack from right to left.
                size_t step = (range.end - range.begin - 1) / settings.merge_tree_coarse_index_granularity + 1;
                size_t end;

                for (end = range.end; end > range.begin + step; end -= step)
                    ranges_stack.push_back(MarkRange(end - step, end));

                ranges_stack.push_back(MarkRange(range.begin, end));
            }
        }
    }

    return res;
}

}