be/src/format/parquet/parquet_predicate.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 <gen_cpp/parquet_types.h>

#include <cmath>
#include <cstring>
#include <vector>

#include "cctz/time_zone.h"
#include "core/data_type/data_type_decimal.h"
#include "core/data_type/primitive_type.h"
#include "exec/common/endian.h"
#include "format/format_common.h"
#include "format/parquet/parquet_block_split_bloom_filter.h"
#include "format/parquet/parquet_column_convert.h"
#include "format/parquet/parquet_common.h"
#include "format/parquet/schema_desc.h"
#include "io/fs/file_reader.h"
#include "storage/olap_scan_common.h"
#include "storage/segment/row_ranges.h"
#include "util/timezone_utils.h"

namespace doris {
class ParquetPredicate {
private:
    static inline bool _is_ascii(uint8_t byte) { return byte < 128; }

    static int _common_prefix(const std::string& encoding_min, const std::string& encoding_max) {
        size_t min_length = std::min(encoding_min.size(), encoding_max.size());
        int common_length = 0;
        while (common_length < min_length &&
               encoding_min[common_length] == encoding_max[common_length]) {
            common_length++;
        }
        return common_length;
    }

    static bool _try_read_old_utf8_stats(std::string& encoding_min, std::string& encoding_max) {
        if (encoding_min == encoding_max) {
            // If min = max, then there is a single value only
            // No need to modify, just use min
            encoding_max = encoding_min;
            return true;
        } else {
            int common_prefix_length = _common_prefix(encoding_min, encoding_max);

            // For min we can retain all-ASCII, because this produces a strictly lower value.
            int min_good_length = common_prefix_length;
            while (min_good_length < encoding_min.size() &&
                   _is_ascii(static_cast<uint8_t>(encoding_min[min_good_length]))) {
                min_good_length++;
            }

            // For max we can be sure only of the part matching the min. When they differ, we can consider only one next, and only if both are ASCII
            int max_good_length = common_prefix_length;
            if (max_good_length < encoding_max.size() && max_good_length < encoding_min.size() &&
                _is_ascii(static_cast<uint8_t>(encoding_min[max_good_length])) &&
                _is_ascii(static_cast<uint8_t>(encoding_max[max_good_length]))) {
                max_good_length++;
            }
            // Incrementing 127 would overflow. Incrementing within non-ASCII can have side-effects.
            while (max_good_length > 0 &&
                   (static_cast<uint8_t>(encoding_max[max_good_length - 1]) == 127 ||
                    !_is_ascii(static_cast<uint8_t>(encoding_max[max_good_length - 1])))) {
                max_good_length--;
            }
            if (max_good_length == 0) {
                // We can return just min bound, but code downstream likely expects both are present or both are absent.
                return false;
            }

            encoding_min.resize(min_good_length);
            encoding_max.resize(max_good_length);
            if (max_good_length > 0) {
                encoding_max[max_good_length - 1]++;
            }
            return true;
        }
    }

    static SortOrder _determine_sort_order(const tparquet::SchemaElement& parquet_schema) {
        tparquet::Type::type physical_type = parquet_schema.type;
        const tparquet::LogicalType& logical_type = parquet_schema.logicalType;

        // Assume string type is SortOrder::SIGNED, use ParquetPredicate::_try_read_old_utf8_stats() to handle it.
        if (logical_type.__isset.STRING &&
            (physical_type == tparquet::Type::BYTE_ARRAY ||
             physical_type == tparquet::Type::FIXED_LEN_BYTE_ARRAY)) {
            return SortOrder::SIGNED;
        }

        if (logical_type.__isset.INTEGER) {
            if (logical_type.INTEGER.isSigned) {
                return SortOrder::SIGNED;
            } else {
                return SortOrder::UNSIGNED;
            }
        } else if (logical_type.__isset.DATE) {
            return SortOrder::SIGNED;
        } else if (logical_type.__isset.ENUM) {
            return SortOrder::UNSIGNED;
        } else if (logical_type.__isset.BSON) {
            return SortOrder::UNSIGNED;
        } else if (logical_type.__isset.JSON) {
            return SortOrder::UNSIGNED;
        } else if (logical_type.__isset.STRING) {
            return SortOrder::UNSIGNED;
        } else if (logical_type.__isset.DECIMAL) {
            return SortOrder::UNKNOWN;
        } else if (logical_type.__isset.MAP) {
            return SortOrder::UNKNOWN;
        } else if (logical_type.__isset.LIST) {
            return SortOrder::UNKNOWN;
        } else if (logical_type.__isset.TIME) {
            return SortOrder::SIGNED;
        } else if (logical_type.__isset.TIMESTAMP) {
            return SortOrder::SIGNED;
        } else if (logical_type.__isset.UNKNOWN) {
            return SortOrder::UNKNOWN;
        } else {
            switch (physical_type) {
            case tparquet::Type::BOOLEAN:
            case tparquet::Type::INT32:
            case tparquet::Type::INT64:
            case tparquet::Type::FLOAT:
            case tparquet::Type::DOUBLE:
                return SortOrder::SIGNED;
            case tparquet::Type::BYTE_ARRAY:
            case tparquet::Type::FIXED_LEN_BYTE_ARRAY:
                return SortOrder::UNSIGNED;
            case tparquet::Type::INT96:
                return SortOrder::UNKNOWN;
            default:
                return SortOrder::UNKNOWN;
            }
        }
    }

public:
    static constexpr int BLOOM_FILTER_MAX_HEADER_LENGTH = 64;
    struct ColumnStat {
        std::string encoded_min_value;
        std::string encoded_max_value;
        bool has_null;
        bool is_all_null;
        const FieldSchema* col_schema;
        const cctz::time_zone* ctz;
        std::unique_ptr<ParquetBlockSplitBloomFilter> bloom_filter;
        std::function<bool(ParquetPredicate::ColumnStat*, const int)>* get_stat_func = nullptr;
        std::function<bool(ParquetPredicate::ColumnStat*, const int)>* get_bloom_filter_func =
                nullptr;
    };

    static bool bloom_filter_supported(PrimitiveType type) {
        // Only support types where physical type == logical type (no conversion needed)
        // For types like DATEV2, DATETIMEV2, DECIMAL, Parquet stores them in physical format
        // (INT32, INT64, etc.) but Doris uses different internal representations.
        // Bloom filter works with physical bytes, but we only have logical type values,
        // and there's no reverse conversion (logical -> physical) available.
        // TINYINT/SMALLINT also need conversion via LittleIntPhysicalConverter.
        switch (type) {
        case TYPE_BOOLEAN:
        case TYPE_INT:
        case TYPE_BIGINT:
        case TYPE_FLOAT:
        case TYPE_DOUBLE:
        case TYPE_CHAR:
        case TYPE_VARCHAR:
        case TYPE_STRING:
            return true;
        default:
            return false;
        }
    }

    struct PageIndexStat {
        // Indicates whether the page index information in this column can be used.
        bool available = false;
        int64_t num_of_pages;
        std::vector<std::string> encoded_min_value;
        std::vector<std::string> encoded_max_value;
        std::vector<bool> has_null;
        std::vector<bool> is_all_null;
        const FieldSchema* col_schema;

        // Record the row range corresponding to each page.
        std::vector<segment_v2::RowRange> ranges;
    };

    struct CachedPageIndexStat {
        const cctz::time_zone* ctz;
        std::map<int, PageIndexStat> stats;
        std::function<bool(PageIndexStat**, int)> get_stat_func;
        RowRange row_group_range;
    };

    // The encoded Parquet min-max value is parsed into `fields`;
    // Can be used in row groups and page index statistics.
    static Status parse_min_max_value(const FieldSchema* col_schema, const std::string& encoded_min,
                                      const std::string& encoded_max, const cctz::time_zone& ctz,
                                      Field* min_field, Field* max_field) {
        auto logical_data_type = remove_nullable(col_schema->data_type);
        // UUID statistics are used only with Iceberg's binary carrier. Ordinary UUID text
        // scans reject these predicates in ParquetReader::_type_matches.
        auto converter = parquet::PhysicalToLogicalConverter::get_converter(
                col_schema, logical_data_type, logical_data_type, &ctz, false, true);
        ColumnPtr physical_column;
        switch (col_schema->parquet_schema.type) {
        case tparquet::Type::type::BOOLEAN: {
            auto physical_col = ColumnUInt8::create();
            physical_col->get_data().data();
            physical_col->resize(2);
            physical_col->get_data()[0] = *reinterpret_cast<const bool*>(encoded_min.data());
            physical_col->get_data()[1] = *reinterpret_cast<const bool*>(encoded_max.data());
            physical_column = std::move(physical_col);
            break;
        }
        case tparquet::Type::type::INT32: {
            auto physical_col = ColumnInt32::create();
            physical_col->resize(2);

            physical_col->get_data()[0] = *reinterpret_cast<const int32_t*>(encoded_min.data());
            physical_col->get_data()[1] = *reinterpret_cast<const int32_t*>(encoded_max.data());

            physical_column = std::move(physical_col);
            break;
        }
        case tparquet::Type::type::INT64: {
            auto physical_col = ColumnInt64::create();
            physical_col->resize(2);
            physical_col->get_data()[0] = *reinterpret_cast<const int64_t*>(encoded_min.data());
            physical_col->get_data()[1] = *reinterpret_cast<const int64_t*>(encoded_max.data());
            physical_column = std::move(physical_col);
            break;
        }
        case tparquet::Type::type::FLOAT: {
            auto physical_col = ColumnFloat32::create();
            physical_col->resize(2);
            physical_col->get_data()[0] = *reinterpret_cast<const float*>(encoded_min.data());
            physical_col->get_data()[1] = *reinterpret_cast<const float*>(encoded_max.data());
            physical_column = std::move(physical_col);
            break;
        }
        case tparquet::Type::type::DOUBLE: {
            auto physical_col = ColumnFloat64 ::create();
            physical_col->resize(2);
            physical_col->get_data()[0] = *reinterpret_cast<const double*>(encoded_min.data());
            physical_col->get_data()[1] = *reinterpret_cast<const double*>(encoded_max.data());
            physical_column = std::move(physical_col);
            break;
        }
        case tparquet::Type::type::BYTE_ARRAY: {
            auto physical_col = ColumnString::create();
            physical_col->insert_data(encoded_min.data(), encoded_min.size());
            physical_col->insert_data(encoded_max.data(), encoded_max.size());
            physical_column = std::move(physical_col);
            break;
        }
        case tparquet::Type::type::FIXED_LEN_BYTE_ARRAY: {
            auto physical_col = ColumnUInt8::create();
            physical_col->resize(2 * col_schema->parquet_schema.type_length);
            DCHECK(col_schema->parquet_schema.type_length == encoded_min.length());
            DCHECK(col_schema->parquet_schema.type_length == encoded_max.length());

            auto ptr = physical_col->get_data().data();
            memcpy(ptr, encoded_min.data(), encoded_min.length());
            memcpy(ptr + encoded_min.length(), encoded_max.data(), encoded_max.length());
            physical_column = std::move(physical_col);
            break;
        }
        case tparquet::Type::type::INT96: {
            auto physical_col = ColumnInt8::create();
            physical_col->resize(2 * sizeof(ParquetInt96));
            DCHECK(sizeof(ParquetInt96) == encoded_min.length());
            DCHECK(sizeof(ParquetInt96) == encoded_max.length());

            auto ptr = physical_col->get_data().data();
            memcpy(ptr, encoded_min.data(), encoded_min.length());
            memcpy(ptr + encoded_min.length(), encoded_max.data(), encoded_max.length());
            physical_column = std::move(physical_col);
            break;
        }
        }

        ColumnPtr logical_column;
        if (converter->is_consistent()) {
            logical_column = physical_column;
        } else {
            logical_column = logical_data_type->create_column();
            RETURN_IF_ERROR(converter->physical_convert(physical_column, logical_column));
        }

        DCHECK(logical_column->size() == 2);
        *min_field = logical_column->operator[](0);
        *max_field = logical_column->operator[](1);

        auto logical_prim_type = logical_data_type->get_primitive_type();

        if (logical_prim_type == TYPE_FLOAT) {
            auto& min_value = min_field->get<TYPE_FLOAT>();
            auto& max_value = max_field->get<TYPE_FLOAT>();

            if (std::isnan(min_value) || std::isnan(max_value)) {
                return Status::DataQualityError("Can not use this parquet min/max value.");
            }
            // Updating min to -0.0 and max to +0.0 to ensure that no 0.0 values would be skipped
            if (std::signbit(min_value) == 0 && min_value == 0.0F) {
                min_value = -0.0F;
            }
            if (std::signbit(max_value) != 0 && max_value == -0.0F) {
                max_value = 0.0F;
            }
        } else if (logical_prim_type == TYPE_DOUBLE) {
            auto& min_value = min_field->get<TYPE_DOUBLE>();
            auto& max_value = max_field->get<TYPE_DOUBLE>();

            if (std::isnan(min_value) || std::isnan(max_value)) {
                return Status::DataQualityError("Can not use this parquet min/max value.");
            }
            // Updating min to -0.0 and max to +0.0 to ensure that no 0.0 values would be skipped
            if (std::signbit(min_value) == 0 && min_value == 0.0F) {
                min_value = -0.0F;
            }
            if (std::signbit(max_value) != 0 && max_value == -0.0F) {
                max_value = 0.0F;
            }
        } else if (col_schema->parquet_schema.type == tparquet::Type::type::INT96 ||
                   logical_prim_type == TYPE_DATETIMEV2) {
            auto min_value = min_field->get<TYPE_DATETIMEV2>();
            auto max_value = min_field->get<TYPE_DATETIMEV2>();

            // From Trino: Parquet INT96 timestamp values were compared incorrectly
            // for the purposes of producing statistics by older parquet writers,
            // so PARQUET-1065 deprecated them. The result is that any writer that produced stats
            // was producing unusable incorrect values, except the special case where min == max
            // and an incorrect ordering would not be material to the result.
            // PARQUET-1026 made binary stats available and valid in that special case.
            if (min_value != max_value) {
                return Status::DataQualityError("invalid min/max value");
            }
        }

        return Status::OK();
    }

    static Status read_column_stats(const FieldSchema* col_schema,
                                    const tparquet::ColumnMetaData& column_meta_data,
                                    std::unordered_map<tparquet::Type::type, bool>* ignored_stats,
                                    const std::string& file_created_by, ColumnStat* ans_stat) {
        auto& statistic = column_meta_data.statistics;

        if (!statistic.__isset.null_count) [[unlikely]] {
            return Status::DataQualityError("This parquet Column meta no set null_count.");
        }
        ans_stat->has_null = statistic.null_count > 0;
        ans_stat->is_all_null = statistic.null_count == column_meta_data.num_values;
        if (ans_stat->is_all_null) {
            return Status::OK();
        }
        auto prim_type = remove_nullable(col_schema->data_type)->get_primitive_type();

        // Min-max of statistic is plain-encoded value
        if (statistic.__isset.min_value && statistic.__isset.max_value) {
            ColumnOrderName column_order =
                    col_schema->physical_type == tparquet::Type::INT96 ||
                                    col_schema->parquet_schema.logicalType.__isset.UNKNOWN
                            ? ColumnOrderName::UNDEFINED
                            : ColumnOrderName::TYPE_DEFINED_ORDER;
            if ((statistic.min_value != statistic.max_value) &&
                (column_order != ColumnOrderName::TYPE_DEFINED_ORDER)) {
                return Status::DataQualityError("Can not use this parquet min/max value.");
            }
            ans_stat->encoded_min_value = statistic.min_value;
            ans_stat->encoded_max_value = statistic.max_value;

            if (prim_type == TYPE_VARCHAR || prim_type == TYPE_CHAR || prim_type == TYPE_STRING) {
                auto encoded_min_copy = ans_stat->encoded_min_value;
                auto encoded_max_copy = ans_stat->encoded_max_value;
                if (!_try_read_old_utf8_stats(encoded_min_copy, encoded_max_copy)) {
                    return Status::DataQualityError("Can not use this parquet min/max value.");
                }
                ans_stat->encoded_min_value = encoded_min_copy;
                ans_stat->encoded_max_value = encoded_max_copy;
            }

        } else if (statistic.__isset.min && statistic.__isset.max) {
            bool max_equals_min = statistic.min == statistic.max;

            SortOrder sort_order = _determine_sort_order(col_schema->parquet_schema);
            bool sort_orders_match = SortOrder::SIGNED == sort_order;
            if (!sort_orders_match && !max_equals_min) {
                return Status::NotSupported("Can not use this parquet min/max value.");
            }

            bool should_ignore_corrupted_stats = false;
            if (ignored_stats != nullptr) {
                if (ignored_stats->count(col_schema->physical_type) == 0) {
                    if (CorruptStatistics::should_ignore_statistics(file_created_by,
                                                                    col_schema->physical_type)) {
                        ignored_stats->emplace(col_schema->physical_type, true);
                        should_ignore_corrupted_stats = true;
                    } else {
                        ignored_stats->emplace(col_schema->physical_type, false);
                    }
                } else if (ignored_stats->at(col_schema->physical_type)) {
                    should_ignore_corrupted_stats = true;
                }
            } else if (CorruptStatistics::should_ignore_statistics(file_created_by,
                                                                   col_schema->physical_type)) {
                should_ignore_corrupted_stats = true;
            }

            if (should_ignore_corrupted_stats) {
                return Status::DataQualityError("Error statistics, should ignore.");
            }

            ans_stat->encoded_min_value = statistic.min;
            ans_stat->encoded_max_value = statistic.max;
        } else {
            return Status::DataQualityError("This parquet file not set min/max value");
        }

        return Status::OK();
    }

    static Status read_bloom_filter(const tparquet::ColumnMetaData& column_meta_data,
                                    io::FileReaderSPtr file_reader, io::IOContext* io_ctx,
                                    ColumnStat* ans_stat) {
        if (!column_meta_data.__isset.bloom_filter_offset) {
            return Status::NotSupported("Can not use this parquet bloom filter.");
        }
        if (column_meta_data.bloom_filter_offset < 0 ||
            (column_meta_data.__isset.bloom_filter_length &&
             column_meta_data.bloom_filter_length <= 0)) {
            return Status::Corruption("Invalid Parquet bloom filter offset or declared length");
        }

        const uint64_t bloom_offset = static_cast<uint64_t>(column_meta_data.bloom_filter_offset);
        if (bloom_offset >= file_reader->size()) {
            return Status::Corruption("Parquet bloom filter offset exceeds file size");
        }
        const size_t available = file_reader->size() - bloom_offset;
        const size_t declared_available =
                column_meta_data.__isset.bloom_filter_length
                        ? std::min<size_t>(column_meta_data.bloom_filter_length, available)
                        : available;
        const size_t header_read_size =
                std::min<size_t>(declared_available, BLOOM_FILTER_MAX_HEADER_LENGTH);
        size_t bytes_read = 0;
        std::vector<uint8_t> header_buffer(header_read_size);
        RETURN_IF_ERROR(file_reader->read_at(column_meta_data.bloom_filter_offset,
                                             Slice(header_buffer.data(), header_buffer.size()),
                                             &bytes_read, io_ctx));

        tparquet::BloomFilterHeader t_bloom_filter_header;
        uint32_t t_bloom_filter_header_size = static_cast<uint32_t>(bytes_read);
        if (!deserialize_thrift_msg(header_buffer.data(), &t_bloom_filter_header_size, true,
                                    &t_bloom_filter_header)
                     .ok()) {
            return Status::Corruption("Malformed Parquet bloom filter header");
        }

        // TODO the bloom filter could be encrypted, too, so need to double check that this is NOT the case
        if (!t_bloom_filter_header.algorithm.__isset.BLOCK ||
            !t_bloom_filter_header.compression.__isset.UNCOMPRESSED ||
            !t_bloom_filter_header.hash.__isset.XXHASH || t_bloom_filter_header.numBytes <= 0) {
            return Status::NotSupported("Can not use this parquet bloom filter.");
        }

        const int64_t payload_size = t_bloom_filter_header.numBytes;
        if (payload_size < segment_v2::BloomFilter::MINIMUM_BYTES ||
            payload_size > segment_v2::BloomFilter::MAXIMUM_BYTES || payload_size % 32 != 0) {
            return Status::Corruption("Invalid Parquet bloom filter payload size {}", payload_size);
        }
        const uint64_t total_size =
                static_cast<uint64_t>(t_bloom_filter_header_size) + payload_size;
        if (total_size > available) {
            return Status::Corruption("Parquet bloom filter range exceeds file size");
        }
        const auto expected_declared_length = static_cast<int64_t>(total_size);
        if (column_meta_data.__isset.bloom_filter_length &&
            column_meta_data.bloom_filter_length != expected_declared_length) {
            return Status::Corruption("Invalid Parquet bloom filter declared length");
        }

        auto bloom_filter = std::make_unique<ParquetBlockSplitBloomFilter>();
        // Read directly into one tracked allocation so a valid maximum-size filter is admitted
        // against the task budget without doubling its peak memory during initialization.
        RETURN_IF_ERROR(bloom_filter->init_for_read(static_cast<size_t>(payload_size),
                                                    segment_v2::HashStrategyPB::XX_HASH_64));
        RETURN_IF_ERROR(file_reader->read_at(
                static_cast<size_t>(bloom_offset) + t_bloom_filter_header_size,
                Slice(bloom_filter->mutable_data(), bloom_filter->size()), &bytes_read, io_ctx));
        if (bytes_read != bloom_filter->size()) {
            return Status::Corruption("Truncated Parquet bloom filter payload");
        }

        ans_stat->bloom_filter = std::move(bloom_filter);

        return Status::OK();
    }
};

} // namespace doris

Coverage Report

Created: 2026-09-17 22:12