be/src/storage/segment/bitshuffle_page.h

Source
// Licensed to the Apache Software Foundation (ASF) under one
// or more contributor license agreements.  See the NOTICE file
// distributed with this work for additional information
// regarding copyright ownership.  The ASF licenses this file
// to you under the Apache License, Version 2.0 (the
// "License"); you may not use this file except in compliance
// with the License.  You may obtain a copy of the License at
//
//   http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing,
// software distributed under the License is distributed on an
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, either express or implied.  See the License for the
// specific language governing permissions and limitations
// under the License.

#pragma once

#include <glog/logging.h>

#include <algorithm>
#include <cstddef>
#include <cstdint>
#include <cstring>
#include <ostream>
#include <type_traits>

#include "common/cast_set.h"
#include "common/compiler_util.h" // IWYU pragma: keep
#include "common/status.h"
#include "core/column/column.h"
#include "core/data_type/data_type.h"
#include "core/types.h"
#include "storage/olap_common.h"
#include "storage/segment/bitshuffle_wrapper.h"
#include "storage/segment/common.h"
#include "storage/segment/options.h"
#include "storage/segment/page_builder.h"
#include "storage/segment/page_decoder.h"
#include "storage/types.h"
#include "util/alignment.h"
#include "util/coding.h"
#include "util/faststring.h"
#include "util/slice.h"

namespace doris {
namespace segment_v2 {

enum { BITSHUFFLE_PAGE_HEADER_SIZE = 16 };

void warn_with_bitshuffle_error(int64_t val);

// BitshufflePageBuilder bitshuffles and compresses the bits of fixed
// size type blocks with lz4.
//
// The page format is as follows:
//
// 1. Header: (16 bytes total)
//
//    <num_elements> [32-bit]
//      The number of elements encoded in the page.
//
//    <compressed_size> [32-bit]
//      The post-compression size of the page, including this header.
//
//    <padded_num_elements> [32-bit]
//      Padding is needed to meet the requirements of the bitshuffle
//      library such that the input/output is a multiple of 8. Some
//      ignored elements are appended to the end of the page if necessary
//      to meet this requirement.
//
//      This header field is the post-padding element count.
//
//    <elem_size_bytes> [32-bit]
//      The size of the elements, in bytes, as actually encoded. In the
//      case that all of the data in a page can fit into a smaller
//      integer type, then we may choose to encode that smaller type
//      to save CPU costs.
//
//      This is currently only implemented in the UINT32 page type.
//
//   NOTE: all on-disk ints are encoded little-endian
//
// 2. Element data
//
//    The header is followed by the bitshuffle-compressed element data.
//
template <FieldType Type>
class BitshufflePageBuilder : public PageBuilderHelper<BitshufflePageBuilder<Type>> {
public:
    using Self = BitshufflePageBuilder<Type>;
    friend class PageBuilderHelper<Self>;

    Status init() override { return reset(); }

    bool is_page_full() override { return _remain_element_capacity == 0; }

    Status add(const uint8_t* vals, size_t* count) override {
        return add_internal<false>(vals, count);
    }

    Status single_add(const uint8_t* vals, size_t* count) {
        return add_internal<true>(vals, count);
    }

    template <bool single>
    inline Status add_internal(const uint8_t* vals, size_t* num_written) {
        DCHECK(!_finished);
        if (_remain_element_capacity == 0) {
            *num_written = 0;
            return Status::OK();
        }

        // When increasing the size of the memtabl flush threshold to a very large value, for example 15GB.
        // the row count of a single men tbl could be very large.
        // a real log:
        /*
        I20250823 19:01:16.153575 2982018 memtable_flush_executor.cpp:185] begin to flush memtable for tablet: 1755915952737, memsize: 15.11 GB, rows: 3751968
        */
        // This is not a very wide table, actually it just has two columns, int and array<float>
        // The write process of column array has two steps: write nested column(column float here), and write offsets column.
        // The row count of column array is 3751968, which is not that big, but each row of column array has 768 float numbers (this is a common case in vector search scenario).
        // so the row num of nested column float will be 3751968 * 768 = 2,881,511,424, which is bigger than INT32_MAX.
        uint32_t to_add = cast_set<UInt32>(
                std::min(cast_set<size_t>(_remain_element_capacity), *num_written));
        // Max value of to_add_size is less than STORAGE_PAGE_SIZE_DEFAULT_VALUE
        int to_add_size = to_add * SIZE_OF_TYPE;
        size_t orig_size = _data.size();
        // This may need a large memory, should return error if could not allocated
        // successfully, to avoid BE OOM.
        RETURN_IF_CATCH_EXCEPTION(_data.resize(orig_size + to_add_size));
        _count += to_add;
        _remain_element_capacity -= to_add;
        _raw_data_size += to_add_size;
        // return added number through count
        *num_written = to_add;
        if constexpr (single) {
            if constexpr (SIZE_OF_TYPE == 1) {
                *reinterpret_cast<uint8_t*>(&_data[orig_size]) = *vals;
                return Status::OK();
            } else if constexpr (SIZE_OF_TYPE == 2) {
                *reinterpret_cast<uint16_t*>(&_data[orig_size]) =
                        *reinterpret_cast<const uint16_t*>(vals);
                return Status::OK();
            } else if constexpr (SIZE_OF_TYPE == 4) {
                *reinterpret_cast<uint32_t*>(&_data[orig_size]) =
                        *reinterpret_cast<const uint32_t*>(vals);
                return Status::OK();
            } else if constexpr (SIZE_OF_TYPE == 8) {
                *reinterpret_cast<uint64_t*>(&_data[orig_size]) =
                        *reinterpret_cast<const uint64_t*>(vals);
                return Status::OK();
            }
        }
        // when single is true and SIZE_OF_TYPE > 8 or single is false
        memcpy(&_data[orig_size], vals, to_add_size);
        return Status::OK();
    }

    Status finish(OwnedSlice* slice) override {
        RETURN_IF_CATCH_EXCEPTION({ *slice = _finish(SIZE_OF_TYPE); });
        return Status::OK();
    }

    Status reset() override {
        RETURN_IF_CATCH_EXCEPTION({
            size_t block_size = _options.data_page_size;
            _count = 0;
            _raw_data_size = 0;
            _data.clear();
            _data.reserve(block_size);
            DCHECK_EQ(reinterpret_cast<uintptr_t>(_data.data()) & (alignof(CppType) - 1), 0)
                    << "buffer must be naturally-aligned";
            _buffer.clear();
            _buffer.resize(BITSHUFFLE_PAGE_HEADER_SIZE);
            _finished = false;
            _remain_element_capacity = cast_set<UInt32>(block_size / SIZE_OF_TYPE);
        });
        return Status::OK();
    }

    size_t count() const override { return _count; }

    uint64_t size() const override { return _buffer.size(); }

    uint64_t get_raw_data_size() const override { return _raw_data_size; }

private:
    BitshufflePageBuilder(const PageBuilderOptions& options)
            : _options(options), _count(0), _remain_element_capacity(0), _finished(false) {}

    OwnedSlice _finish(int final_size_of_type) {
        _data.resize(final_size_of_type * _count);

        // Do padding so that the input num of element is multiple of 8.
        int num_elems_after_padding = ALIGN_UP(_count, 8);
        int padding_elems = num_elems_after_padding - _count;
        int padding_bytes = padding_elems * final_size_of_type;
        for (int i = 0; i < padding_bytes; i++) {
            _data.push_back(0);
        }

        // reserve enough place for compression
        _buffer.resize(
                BITSHUFFLE_PAGE_HEADER_SIZE +
                bitshuffle::compress_lz4_bound(num_elems_after_padding, final_size_of_type, 0));

        int64_t bytes =
                bitshuffle::compress_lz4(_data.data(), &_buffer[BITSHUFFLE_PAGE_HEADER_SIZE],
                                         num_elems_after_padding, final_size_of_type, 0);
        if (bytes < 0) [[unlikely]] {
            // This means the bitshuffle function fails.
            // Ideally, this should not happen.
            warn_with_bitshuffle_error(bytes);
            // It does not matter what will be returned here,
            // since we have logged fatal in warn_with_bitshuffle_error().
            return OwnedSlice();
        }
        // update header
        encode_fixed32_le(&_buffer[0], _count);
        encode_fixed32_le(&_buffer[4], cast_set<uint32_t>(BITSHUFFLE_PAGE_HEADER_SIZE + bytes));
        encode_fixed32_le(&_buffer[8], num_elems_after_padding);
        encode_fixed32_le(&_buffer[12], final_size_of_type);
        _finished = true;
        // before build(), update buffer length to the actual compressed size
        _buffer.resize(BITSHUFFLE_PAGE_HEADER_SIZE + bytes);
        return _buffer.build();
    }

    using CppType = typename TypeTraits<Type>::CppType;

    CppType cell(int idx) const {
        DCHECK_GE(idx, 0);
        CppType ret;
        memcpy(&ret, &_data[idx * SIZE_OF_TYPE], SIZE_OF_TYPE);
        return ret;
    }

    enum { SIZE_OF_TYPE = TypeTraits<Type>::size };
    PageBuilderOptions _options;
    uint32_t _count;
    uint32_t _remain_element_capacity;
    bool _finished;
    faststring _data;
    faststring _buffer;
    uint64_t _raw_data_size = 0;
};

inline Status parse_bit_shuffle_header(const Slice& data, size_t& num_elements,
                                       size_t& compressed_size, size_t& num_element_after_padding,
                                       int& size_of_element) {
    if (data.size < BITSHUFFLE_PAGE_HEADER_SIZE) {
        return Status::InternalError("file corruption: invalid data size:{}, header size:{}",
                                     data.size, BITSHUFFLE_PAGE_HEADER_SIZE);
    }

    num_elements = decode_fixed32_le((const uint8_t*)&data[0]);
    compressed_size = decode_fixed32_le((const uint8_t*)&data[4]);
    num_element_after_padding = decode_fixed32_le((const uint8_t*)&data[8]);
    size_of_element = decode_fixed32_le((const uint8_t*)&data[12]);
    if (num_element_after_padding != ALIGN_UP(num_elements, 8)) {
        return Status::InternalError(
                "num of element information corrupted,"
                " _num_element_after_padding:{}, _num_elements:{}, expected_padding:{},"
                " compressed_size:{}, size_of_element:{}, data_size:{}",
                num_element_after_padding, num_elements, ALIGN_UP(num_elements, 8), compressed_size,
                size_of_element, data.size);
    }
    switch (size_of_element) {
    case 1:
    case 2:
    case 3:
    case 4:
    case 8:
    case 12:
    case 16:
    case 32:
        break;
    default:
        return Status::InternalError("invalid size_of_elem:{}", size_of_element);
    }
    return Status::OK();
}

template <FieldType Type>
class BitShufflePageDecoder : public PageDecoder {
public:
    BitShufflePageDecoder(Slice data, const PageDecoderOptions& options)
            : _data(data),
              _options(options),
              _parsed(false),
              _num_elements(0),
              _num_element_after_padding(0),
              _size_of_element(0),
              _cur_index(0) {}

    Status init() override {
        CHECK(!_parsed);
        size_t unused;
        RETURN_IF_ERROR(parse_bit_shuffle_header(_data, _num_elements, unused,
                                                 _num_element_after_padding, _size_of_element));

        if (_data.size !=
            _num_element_after_padding * _size_of_element + BITSHUFFLE_PAGE_HEADER_SIZE) {
            std::stringstream ss;
            ss << "Size information unmatched, _data.size:" << _data.size
               << ", _num_elements:" << _num_elements << ", expected size is "
               << _num_element_after_padding * _size_of_element + BITSHUFFLE_PAGE_HEADER_SIZE;
            return Status::InternalError(ss.str());
        }

        // Currently, only the UINT32 block encoder supports expanding size:
        if (UNLIKELY(Type != FieldType::OLAP_FIELD_TYPE_UNSIGNED_INT &&
                     _size_of_element != SIZE_OF_TYPE)) {
            return Status::InternalError(
                    "invalid size info. size of element:{}, SIZE_OF_TYPE:{}, type:{}",
                    _size_of_element, SIZE_OF_TYPE, Type);
        }
        if (UNLIKELY(_size_of_element > SIZE_OF_TYPE)) {
            return Status::InternalError("invalid size info. size of element:{}, SIZE_OF_TYPE:{}",
                                         _size_of_element, SIZE_OF_TYPE);
        }
        _parsed = true;
        return Status::OK();
    }

    // If the page only contains null, then _num_elements == 0, and the nullmap has
    // some value. But in _seek_columns --> seek_to_ordinal --> _seek_to_pos_in_page
    // in this call stack it will pass pos == 0 to this method. Although we can add some
    // code such as check if the count == 0, then skip seek, but there are other method such
    // as init will also call seek with pos == 0. And the seek is useless when _num_elements
    // == 0, because next batch will return empty in this method.
    Status seek_to_position_in_page(size_t pos) override {
        DCHECK(_parsed) << "Must call init()";
        if (_num_elements == 0) [[unlikely]] {
            if (pos != 0) {
                return Status::Error<ErrorCode::INTERNAL_ERROR, false>(
                        "seek pos {} is larger than total elements  {}", pos, _num_elements);
            }
        }

        DCHECK_LE(pos, _num_elements);
        _cur_index = pos;
        return Status::OK();
    }

    Status seek_at_or_after_value(const void* value, bool* exact_match) override {
        DCHECK(_parsed) << "Must call init() firstly";

        if (_num_elements == 0) {
            return Status::Error<ErrorCode::ENTRY_NOT_FOUND>("page is empty");
        }

        size_t left = 0;
        size_t right = _num_elements;

        void* mid_value = nullptr;

        // find the first value >= target. after loop,
        // - left == index of first value >= target when found
        // - left == _num_elements when not found (all values < target)
        while (left < right) {
            size_t mid = left + (right - left) / 2;
            mid_value = get_data(mid);
            if (TypeTraits<Type>::cmp(mid_value, value) < 0) {
                left = mid + 1;
            } else {
                right = mid;
            }
        }
        if (left >= _num_elements) {
            return Status::Error<ErrorCode::ENTRY_NOT_FOUND>("all value small than the value");
        }
        void* find_value = get_data(left);
        if (TypeTraits<Type>::cmp(find_value, value) == 0) {
            *exact_match = true;
        } else {
            *exact_match = false;
        }

        _cur_index = left;
        return Status::OK();
    }

    template <bool forward_index = true>
    Status next_batch(size_t* n, MutableColumnPtr& dst) {
        DCHECK(_parsed);
        if (*n == 0 || _cur_index >= _num_elements) [[unlikely]] {
            *n = 0;
            return Status::OK();
        }

        size_t max_fetch = std::min(*n, static_cast<size_t>(_num_elements - _cur_index));

        dst->insert_many_fix_len_data(get_data(_cur_index), max_fetch);
        *n = max_fetch;
        if constexpr (forward_index) {
            _cur_index += max_fetch;
        }

        return Status::OK();
    }

    Status next_batch(size_t* n, MutableColumnPtr& dst) override { return next_batch<>(n, dst); }

    Status read_by_rowids(const rowid_t* rowids, ordinal_t page_first_ordinal, size_t* n,
                          MutableColumnPtr& dst) override {
        DCHECK(_parsed);
        if (*n == 0) [[unlikely]] {
            *n = 0;
            return Status::OK();
        }

        auto total = *n;
        auto read_count = 0;
        _buffer.resize(total);
        for (size_t i = 0; i < total; ++i) {
            ordinal_t ord = rowids[i] - page_first_ordinal;
            if (UNLIKELY(ord >= _num_elements)) {
                break;
            }

            _buffer[read_count++] = *reinterpret_cast<CppType*>(get_data(ord));
        }

        if (LIKELY(read_count > 0)) {
            dst->insert_many_fix_len_data((char*)_buffer.data(), read_count);
        }

        *n = read_count;
        return Status::OK();
    }

    Status peek_next_batch(size_t* n, MutableColumnPtr& dst) override {
        return next_batch<false>(n, dst);
    }

    size_t count() const override { return _num_elements; }

    size_t current_index() const override { return _cur_index; }

    char* get_data(size_t index) const {
        return &_data.data[BITSHUFFLE_PAGE_HEADER_SIZE + index * SIZE_OF_TYPE];
    }

private:
    void _copy_next_values(size_t n, void* data) {
        memcpy(data, get_data(_cur_index), n * SIZE_OF_TYPE);
    }

    using CppType = typename TypeTraits<Type>::CppType;

    enum { SIZE_OF_TYPE = TypeTraits<Type>::size };

    Slice _data;
    PageDecoderOptions _options;
    bool _parsed;
    size_t _num_elements;
    size_t _num_element_after_padding;

    int _size_of_element;
    size_t _cur_index;

    std::vector<std::conditional_t<std::is_same_v<CppType, bool>, uint8_t, CppType>> _buffer;

    friend class BinaryDictPageDecoder;
};

} // namespace segment_v2
} // namespace doris

Coverage Report

Created: 2026-06-07 22:11