C++ 列式内存布局数据存储格式 Arrow

2024-07-25
本文作者原创，首发在本博客；CSDN 上的同名文章已标注为转载并回链到本页：https://blog.csdn.net/qq_30220519/article/details/140699186
Apache Arrow 优点 :
   高性能数据处理： Arrow 使用列式内存布局，这特别适合于数据分析和查询操作，因为它允许对数据进行高效批量处理，减少CPU缓存未命中，从而提升处理速度。
   零拷贝数据共享： Arrow 允许不同系统和进程之间直接共享内存中的数据而无需复制，这对于提高数据密集型应用的效率至关重要，减少了内存使用和CPU开销。
   跨平台兼容性： Arrow 是一个跨语言开发平台，支持C++, Java, Python等多种编程语言，促进了不同软件组件间的互操作性。
   标准化数据格式：定义了一套统一的数据格式规范，使得数据可以在不同系统间无缝传递，降低了数据转换的成本和复杂性。
   优化大数据处理：特别是在与大数据框架（如Spark、Pandas）集成时，Arrow 可显著加速数据加载、处理和分析的速度，例如，与PySpark集成后数据处理速度提升高达53倍。
   集成广泛：被众多数据处理工具和库采用，如Pandas、Parquet、Drill、Spark等，形成了强大的生态系统。
Apache Arrow 缺点 :
   内存消耗：列式存储相对于行式存储可能需要更多的内存，尤其是在处理稀疏数据或宽表时，因为每一列都需要分配连续的内存空间。
   不适合所有场景：对于需要频繁随机访问记录或更新操作的场景，Arrow 的列式存储可能不如传统的行式存储高效。
   学习曲线：对于新用户来说，理解和掌握Arrow的数据结构和API可能需要一定时间，尤其是当他们习惯于使用其他数据处理模型时。
   生态成熟度：虽然Arrow的生态系统正在快速发展，但在某些特定领域或小众技术栈中，相关支持和工具可能不够丰富或成熟。
   实现复杂性：对于开发者来说，实现Arrow的高效利用可能涉及到复杂的内存管理和优化策略，这在某些情况下可能会增加开发难度。

#define ARROW_COMPUTE

#include <arrow/compute/api.h>
#include "arrow/pretty_print.h"
#include <arrow/api.h>
#include <arrow/csv/api.h>
#include <arrow/json/api.h>
#include <arrow/io/api.h>
#include <arrow/table.h>
#include <arrow/pretty_print.h>
#include <arrow/result.h>
#include <arrow/status.h>
#include <arrow/ipc/api.h>
#include <parquet/arrow/reader.h>
#include <parquet/arrow/writer.h>
#include <parquet/exception.h>
#include <memory>
#include <iostream>

template <typename T>
using numbuildT = arrow::NumericBuilder<T>;


struct ArrowUtil {
    static arrow::Status read_csv(char const* file_name, std::shared_ptr<arrow::Table>& tb);
    static arrow::Status read_ipc(char const* file_name, std::shared_ptr<arrow::Table>& tb);
    static arrow::Status read_parquet(char const* file_name, std::shared_ptr<arrow::Table>& tb);
    static arrow::Status read_json(char const* file_name, std::shared_ptr<arrow::Table>& tb);

    static arrow::Status write_ipc(arrow::Table const& tb, char const* file_name);
    static arrow::Status write_parquet(arrow::Table const& tb, char const* file_name);

    template <typename T, typename buildT, typename arrayT>
    inline static std::shared_ptr<arrow::Array> chunked_array_to_array(std::shared_ptr<arrow::ChunkedArray> const& array_a) {
        buildT int64_builder;
        int64_builder.Resize(array_a->length());
        std::vector<T> int64_values;
        int64_values.reserve(array_a->length());
        for (int i = 0; i < array_a->num_chunks(); ++i) {
            auto inner_arr = array_a->chunk(i);
            auto int_a = std::static_pointer_cast<arrayT>(inner_arr);
            for (int j = 0; j < int_a->length(); ++j) {
                int64_values.push_back(int_a->Value(j));
            }
        }

        int64_builder.AppendValues(int64_values);
        std::shared_ptr<arrow::Array> array_a_res;
        int64_builder.Finish(&array_a_res);
        return array_a_res;
    }


    template <typename T, typename arrayT>
    inline static std::vector<T> chunked_array_to_vector(std::shared_ptr<arrow::ChunkedArray> const& array_a) {
        std::vector<T> int64_values;
        int64_values.reserve(array_a->length());
        for (int i = 0; i < array_a->num_chunks(); ++i) {
            auto inner_arr = array_a->chunk(i);
            auto int_a = std::static_pointer_cast<arrayT>(inner_arr);
            for (int j = 0; j < int_a->length(); ++j) {
                int64_values.push_back(int_a->Value(j));
            }
        }
        return int64_values;
    }

    inline static std::vector<std::string> chunked_array_to_str_vector(std::shared_ptr<arrow::ChunkedArray> const& array_a) {
        std::vector<std::string> int64_values;
        int64_values.reserve(array_a->length());
        for (int i = 0; i < array_a->num_chunks(); ++i) {
            auto inner_arr = array_a->chunk(i);
            auto int_a = std::static_pointer_cast<arrow::StringArray>(inner_arr);
            for (int j = 0; j < int_a->length(); ++j) {
                int64_values.push_back(int_a->Value(j).data());
            }
        }
        return int64_values;
    }


    inline static std::shared_ptr<arrow::Array> chunked_array_to_str_array(std::shared_ptr<arrow::ChunkedArray> const& array_a) {
        arrow::StringBuilder int64_builder;
        int64_builder.Resize(array_a->length());
        std::vector<std::string> int64_values;
        int64_values.reserve(array_a->length());
        for (int i = 0; i < array_a->num_chunks(); ++i) {
            auto inner_arr = array_a->chunk(i);
            auto int_a = std::static_pointer_cast<arrow::StringArray>(inner_arr);
            for (int j = 0; j < int_a->length(); ++j) {
                int64_values.push_back(int_a->Value(j).data());
            }
        }
        int64_builder.AppendValues(int64_values);
        std::shared_ptr<arrow::Array> array_a_res;
        int64_builder.Finish(&array_a_res);
        return array_a_res;
    }
};


arrow::Status ArrowUtil::read_csv(char const* file_name, std::shared_ptr<arrow::Table>& tb) {
    ARROW_ASSIGN_OR_RAISE(auto input_file,
        arrow::io::ReadableFile::Open(file_name));
    ARROW_ASSIGN_OR_RAISE(auto csv_reader,
        arrow::csv::TableReader::Make(
            arrow::io::default_io_context(), input_file,
            arrow::csv::ReadOptions::Defaults(),
            arrow::csv::ParseOptions::Defaults(),
            arrow::csv::ConvertOptions::Defaults()));
    ARROW_ASSIGN_OR_RAISE(auto table, csv_reader->Read());
    tb = table;
    return arrow::Status::OK();
}

arrow::Status ArrowUtil::read_ipc(char const* file_name, std::shared_ptr<arrow::Table>& tb) {
    ARROW_ASSIGN_OR_RAISE(auto input_file,
        arrow::io::ReadableFile::Open(file_name));

    ARROW_ASSIGN_OR_RAISE(auto ipc_reader, arrow::ipc::RecordBatchFileReader::Open(input_file));

    std::vector<std::shared_ptr<arrow::RecordBatch>> batches;
    batches.reserve(ipc_reader->num_record_batches());

    for (int i = 0; i < ipc_reader->num_record_batches(); ++i) {
        ARROW_ASSIGN_OR_RAISE(auto a_record, ipc_reader->ReadRecordBatch(i));
        batches.emplace_back(std::move(a_record));
    }

    arrow::Table::FromRecordBatches(ipc_reader->schema(), std::move(batches)).Value(&tb);

    return arrow::Status::OK();
}

arrow::Status ArrowUtil::read_parquet(char const* file_name, std::shared_ptr<arrow::Table>& tb) {
    std::shared_ptr<arrow::io::ReadableFile> infile;
    PARQUET_ASSIGN_OR_THROW(infile,
        arrow::io::ReadableFile::Open(file_name,
            arrow::default_memory_pool()));

    std::unique_ptr<parquet::arrow::FileReader> reader;
    PARQUET_THROW_NOT_OK(
        parquet::arrow::OpenFile(infile, arrow::default_memory_pool(), &reader));
    std::shared_ptr<arrow::Table> table;
    PARQUET_THROW_NOT_OK(reader->ReadTable(&table));
    tb = table;
    return arrow::Status::OK();
}

arrow::Status ArrowUtil::read_json(char const* file_name, std::shared_ptr<arrow::Table>& tb) {
    std::shared_ptr<arrow::io::ReadableFile> infile;
    PARQUET_ASSIGN_OR_THROW(infile,
        arrow::io::ReadableFile::Open(file_name,
            arrow::default_memory_pool()));

    ARROW_ASSIGN_OR_RAISE(auto reader, arrow::json::TableReader::Make(arrow::default_memory_pool(), infile, arrow::json::ReadOptions::Defaults(), arrow::json::ParseOptions::Defaults()));

    ARROW_ASSIGN_OR_RAISE(auto res_tb, reader->Read());
    tb = res_tb;
    return arrow::Status::OK();
}

arrow::Status ArrowUtil::write_ipc(arrow::Table const& tb, char const* file_name) {

    ARROW_ASSIGN_OR_RAISE(auto output_file,
        arrow::io::FileOutputStream::Open(file_name));
    ARROW_ASSIGN_OR_RAISE(auto batch_writer,
        arrow::ipc::MakeFileWriter(output_file, tb.schema()));
    ARROW_RETURN_NOT_OK(batch_writer->WriteTable(tb));
    ARROW_RETURN_NOT_OK(batch_writer->Close());

    return arrow::Status::OK();
}

arrow::Status ArrowUtil::write_parquet(arrow::Table const& tb, char const* file_name) {
    std::shared_ptr<arrow::io::FileOutputStream> outfile;
    PARQUET_ASSIGN_OR_THROW(
        outfile, arrow::io::FileOutputStream::Open(file_name));
    // The last argument to the function call is the size of the RowGroup in
    // the parquet file. Normally you would choose this to be rather large but
    // for the example, we use a small value to have multiple RowGroups.
    PARQUET_THROW_NOT_OK(
        parquet::arrow::WriteTable(tb, arrow::default_memory_pool(), outfile, 3));
    return arrow::Status::OK();
}



void testReadCSV() {
    // 读取CSV文件
    char const* csv_path = "./test.csv";
    std::shared_ptr<arrow::Table> tb;
    ArrowUtil::read_csv(csv_path, tb);
    auto const& tb_ = *tb;
    arrow::PrettyPrint(tb_, {}, &std::cerr);
    assert(tb_.num_rows() == 2);
}

void testWriteIpc() {
    // 读取CSV文件并写入IPC文件
    char const* csv_path = "./test.csv";
    std::shared_ptr<arrow::Table> tb;
    ArrowUtil::read_csv(csv_path, tb);
    auto const& tb_ = *tb;

    char const* write_csv_path = "./test_dst.arrow";
    arrow::PrettyPrint(tb_, {}, &std::cerr);
    auto write_res = ArrowUtil::write_ipc(tb_, write_csv_path);
    assert(write_res == arrow::Status::OK());
}

void testReadIPC() {
    // 读取Arrow IPC 文件
    char const* ipc_path = "./test_dst.arrow";
    std::shared_ptr<arrow::Table> tb;
    ArrowUtil::read_ipc(ipc_path, tb);
    auto const& tb_ = *tb;
    arrow::PrettyPrint(tb_, {}, &std::cerr);
    assert(tb_.num_rows() == 2);
}


void testWriteParquet() {
    // 写入Parquet文件
    char const* csv_path = "./test.csv";
    std::shared_ptr<arrow::Table> tb;
    ArrowUtil::read_csv(csv_path, tb);
    auto const& tb_ = *tb;

    char const* write_parquet_path = "./test_dst.parquet";
    arrow::PrettyPrint(tb_, {}, &std::cerr);
    auto write_res = ArrowUtil::write_parquet(tb_, write_parquet_path);
    assert(write_res == arrow::Status::OK());
}


void testReadParquet() {
    // 读取 Parquet
    char const* parquet_path = "./test_dst.parquet";
    std::shared_ptr<arrow::Table> tb;
    ArrowUtil::read_parquet(parquet_path, tb);
    auto const& tb_ = *tb;
    arrow::PrettyPrint(tb_, {}, &std::cerr);
    assert(tb_.num_rows() == 2);
}

void testReadJson() {
    // 读取Json文件
    char const* json_path = "./test.json";
    std::shared_ptr<arrow::Table> tb;
    ArrowUtil::read_json(json_path, tb);
    auto const& tb_ = *tb;
    arrow::PrettyPrint(tb_, {}, &std::cerr);
    assert(tb_.num_rows() == 2);
}

void testComputeGreater() {
    // 比较两列 int 值中 int1 > int2的值， greater函数
    char const* json_path = "./comp_gt.csv";
    std::shared_ptr<arrow::Table> tb;
    ArrowUtil::read_csv(json_path, tb);
    auto const& tb_ = *tb;
    arrow::PrettyPrint(tb_, {}, &std::cerr);
    auto array_a = tb_.GetColumnByName("int1");
    auto array_b = tb_.GetColumnByName("int2");

    auto array_a_res = ArrowUtil::chunked_array_to_array<int64_t, numbuildT<arrow::Int64Type>, arrow::Int64Array>(array_a);
    auto array_b_res = ArrowUtil::chunked_array_to_array<int64_t, numbuildT<arrow::Int64Type>, arrow::Int64Array>(array_b);

    auto compared_datum = arrow::compute::CallFunction("greater", { array_a_res, array_b_res });
    auto array_a_gt_b_compute = compared_datum->make_array();

    arrow::PrettyPrint(*array_a_gt_b_compute, {}, &std::cerr);

    auto schema =
        arrow::schema({ arrow::field("int1", arrow::int64()), arrow::field("int2", arrow::int64()),
                       arrow::field("a>b? (arrow)", arrow::boolean()) });

    std::shared_ptr<arrow::Table> my_table = arrow::Table::Make(
        schema, { array_a_res, array_b_res, array_a_gt_b_compute }, tb_.num_rows());

    arrow::PrettyPrint(*my_table, {}, &std::cerr);
}

void testComputeMinMax() {
    // 计算int1列的最大值和最小值
    char const* json_path = "./comp_gt.csv";
    std::shared_ptr<arrow::Table> tb;
    ArrowUtil::read_csv(json_path, tb);
    auto const& tb_ = *tb;
    arrow::PrettyPrint(tb_, {}, &std::cerr);
    auto array_a = tb_.GetColumnByName("int1");
    auto array_a_res = ArrowUtil::chunked_array_to_array<int64_t, numbuildT<arrow::Int64Type>, arrow::Int64Array>(array_a);

    arrow::compute::ScalarAggregateOptions scalar_aggregate_options;
    scalar_aggregate_options.skip_nulls = false;

    auto min_max = arrow::compute::CallFunction("min_max", { array_a_res }, &scalar_aggregate_options);

    // Unpack struct scalar result (a two-field {"min", "max"} scalar)
    auto min_value = min_max->scalar_as<arrow::StructScalar>().value[0];
    auto max_value = min_max->scalar_as<arrow::StructScalar>().value[1];

    assert(min_value->ToString() == "1");
    assert(max_value->ToString() == "8");
}

#define GTEST_TEST(a, b) void a##_##b()
#define ASSERT_EQ(a, b) assert(a == b)

GTEST_TEST(RWTests, ComputeMean) {
    // 计算int1列的平均值
    char const* json_path = "../data/comp_gt.csv";
    std::shared_ptr<arrow::Table> tb;
    ArrowUtil::read_csv(json_path, tb);
    auto const& tb_ = *tb;
    arrow::PrettyPrint(tb_, {}, &std::cerr);
    auto array_a = tb_.GetColumnByName("int1");
    auto array_a_res = ArrowUtil::chunked_array_to_array<int64_t, numbuildT<arrow::Int64Type>, arrow::Int64Array>(array_a);

    arrow::compute::ScalarAggregateOptions scalar_aggregate_options;
    scalar_aggregate_options.skip_nulls = false;

    auto mean = arrow::compute::CallFunction("mean", { array_a_res }, &scalar_aggregate_options);

    auto const& mean_value = mean->scalar_as<arrow::Scalar>();

    ASSERT_EQ(mean_value.ToString(), "4.5");
}

GTEST_TEST(RWTests, ComputeAdd) {
    // 将第一列的值加3
    char const* json_path = "../data/comp_gt.csv";
    std::shared_ptr<arrow::Table> tb;
    ArrowUtil::read_csv(json_path, tb);
    auto const& tb_ = *tb;
    arrow::PrettyPrint(tb_, {}, &std::cerr);
    auto array_a = tb_.GetColumnByName("int1");
    auto array_a_res = ArrowUtil::chunked_array_to_array<int64_t, numbuildT<arrow::Int64Type>, arrow::Int64Array>(array_a);

    arrow::compute::ScalarAggregateOptions scalar_aggregate_options;
    scalar_aggregate_options.skip_nulls = false;

    std::shared_ptr<arrow::Scalar> increment = std::make_shared<arrow::Int64Scalar>(3);

    auto add = arrow::compute::CallFunction("add", { array_a_res, increment }, &scalar_aggregate_options);
    std::shared_ptr<arrow::Array> incremented_array = add->array_as<arrow::Array>();
    arrow::PrettyPrint(*incremented_array, {}, &std::cerr);
}


GTEST_TEST(RWTests, ComputeAddArray) {
    // int1和int2两列相加
    char const* json_path = "../data/comp_gt.csv";
    std::shared_ptr<arrow::Table> tb;
    ArrowUtil::read_csv(json_path, tb);
    auto const& tb_ = *tb;
    arrow::PrettyPrint(tb_, {}, &std::cerr);
    auto array_a = tb_.GetColumnByName("int1");
    auto array_a_res = ArrowUtil::chunked_array_to_array<int64_t, numbuildT<arrow::Int64Type>, arrow::Int64Array>(array_a);

    auto array_b = tb_.GetColumnByName("int2");
    auto array_b_res = ArrowUtil::chunked_array_to_array<int64_t, numbuildT<arrow::Int64Type>, arrow::Int64Array>(array_b);

    arrow::compute::ScalarAggregateOptions scalar_aggregate_options;
    scalar_aggregate_options.skip_nulls = false;

    auto add = arrow::compute::CallFunction("add", { array_a_res, array_b_res }, &scalar_aggregate_options);
    std::shared_ptr<arrow::Array> incremented_array = add->array_as<arrow::Array>();
    arrow::PrettyPrint(*incremented_array, {}, &std::cerr);
}

GTEST_TEST(RWTests, ComputeStringEqual) {
    // 比较s1和s2两列是否相等
    char const* json_path = "../data/comp_s_eq.csv";
    std::shared_ptr<arrow::Table> tb;
    ArrowUtil::read_csv(json_path, tb);
    auto const& tb_ = *tb;
    arrow::PrettyPrint(tb_, {}, &std::cerr);

    auto array_a = tb_.GetColumnByName("s1");
    auto array_a_res = ArrowUtil::chunked_array_to_str_array(array_a);

    auto array_b = tb_.GetColumnByName("s2");
    auto array_b_res = ArrowUtil::chunked_array_to_str_array(array_b);

    arrow::compute::ScalarAggregateOptions scalar_aggregate_options;
    scalar_aggregate_options.skip_nulls = false;

    auto eq_ = arrow::compute::CallFunction("equal", { array_a_res, array_b_res }, &scalar_aggregate_options);
    std::shared_ptr<arrow::Array> eq_array = eq_->array_as<arrow::Array>();
    arrow::PrettyPrint(*eq_array, {}, &std::cerr);
}

GTEST_TEST(RWTests, ComputeCustom) {
    // 自己写算法逐个比较相等 
    char const* json_path = "../data/comp_s_eq.csv";
    std::shared_ptr<arrow::Table> tb;
    ArrowUtil::read_csv(json_path, tb);
    auto const& tb_ = *tb;
    arrow::PrettyPrint(tb_, {}, &std::cerr);
    auto arr1 = tb_.GetColumnByName("s1");
    auto arr2 = tb_.GetColumnByName("s2");
    auto v1 = ArrowUtil::chunked_array_to_str_vector(arr1);
    auto v2 = ArrowUtil::chunked_array_to_str_vector(arr2);
    for (std::size_t i = 0; i < v1.size(); ++i) {
        if (v1[i] != v2[i]) {
            std::cerr << v1[i] << "!=" << v2[i] << "\n";
        }
    }
}

GTEST_TEST(RWTests, ComputeCustomDbl) {
    // 自己写算法比较double值
    char const* json_path = "../data/custom_dbl.csv";
    std::shared_ptr<arrow::Table> tb;
    ArrowUtil::read_csv(json_path, tb);
    auto const& tb_ = *tb;
    arrow::PrettyPrint(tb_, {}, &std::cerr);
    auto arr1 = tb_.GetColumnByName("dbl1");
    auto arr2 = tb_.GetColumnByName("dbl2");
    auto v1 = ArrowUtil::chunked_array_to_vector<double, arrow::DoubleArray>(arr1);
    auto v2 = ArrowUtil::chunked_array_to_vector<double, arrow::DoubleArray>(arr2);
    for (std::size_t i = 0; i < v1.size(); ++i) {
        if (v1[i] != v2[i]) {
            std::cerr << v1[i] << "!=" << v2[i] << "\n";
        }
    }
}

GTEST_TEST(RWTests, ComputeEqualDbl) {
    // 使用equal函数比较double值
    char const* json_path = "../data/custom_dbl.csv";
    std::shared_ptr<arrow::Table> tb;
    ArrowUtil::read_csv(json_path, tb);
    auto const& tb_ = *tb;
    arrow::PrettyPrint(tb_, {}, &std::cerr);
    auto arr1 = tb_.GetColumnByName("dbl1");
    auto arr2 = tb_.GetColumnByName("dbl2");

    auto dbl_arr1 = ArrowUtil::chunked_array_to_array<double, numbuildT<arrow::DoubleType>, arrow::DoubleArray>(arr1);
    auto dbl_arr2 = ArrowUtil::chunked_array_to_array<double, numbuildT<arrow::DoubleType>, arrow::DoubleArray>(arr2);

    arrow::compute::ScalarAggregateOptions scalar_aggregate_options;
    scalar_aggregate_options.skip_nulls = false;

    auto eq_ = arrow::compute::CallFunction("equal", { dbl_arr1, dbl_arr2 }, &scalar_aggregate_options);
    std::shared_ptr<arrow::Array> eq_array = eq_->array_as<arrow::Array>();
    arrow::PrettyPrint(*eq_array, {}, &std::cerr);
}

GTEST_TEST(RWTests, StrStartsWith) {
    // 计算s1列以是否以 Zha开头的值
    char const* json_path = "../data/comp_s_eq.csv";
    std::shared_ptr<arrow::Table> tb;
    ArrowUtil::read_csv(json_path, tb);
    auto const& tb_ = *tb;
    arrow::PrettyPrint(tb_, {}, &std::cerr);

    auto array_a = tb_.GetColumnByName("s1");
    auto array_a_res = ArrowUtil::chunked_array_to_str_array(array_a);

    arrow::compute::MatchSubstringOptions options("Zha");

    auto eq_ = arrow::compute::CallFunction("starts_with", { array_a_res }, &options);
    std::shared_ptr<arrow::Array> eq_array = eq_->array_as<arrow::Array>();
    arrow::PrettyPrint(*eq_array, {}, &std::cerr);
}



using arrow::Int32Builder;
using arrow::Int64Builder;
using arrow::DoubleBuilder;
using arrow::StringBuilder;

struct row_data {
    int32_t col1;
    int64_t col2;
    double col3;
    std::string col4;
};//行结构

#define EXIT_ON_FAILURE(expr)                      \
  do {                                             \
    arrow::Status status_ = (expr);                \
    if (!status_.ok()) {                           \
      std::cerr << status_.message() << std::endl; \
      return EXIT_FAILURE;                         \
    }                                              \
  } while (0);

arrow::Status CreateTable(const std::vector<struct row_data>& rows, std::shared_ptr<arrow::Table>* table) {

    //使用arrow::jemalloc::MemoryPool::default_pool()构建器更有效，因为这可以适当增加底层内存区域的大小.
    arrow::MemoryPool* pool = arrow::default_memory_pool();

    Int32Builder col1_builder(pool);
    Int64Builder col2_builder(pool);
    DoubleBuilder col3_builder(pool);
    StringBuilder col4_builder(pool);

    //现在我们可以循环我们现有的数据，并将其插入到构建器中。这里的' Append '调用可能会失败(例如，我们无法分配足够的额外内存)。因此我们需要检查它们的返回值。
    for (const row_data& row : rows) {
        ARROW_RETURN_NOT_OK(col1_builder.Append(row.col1));
        ARROW_RETURN_NOT_OK(col2_builder.Append(row.col2));
        ARROW_RETURN_NOT_OK(col3_builder.Append(row.col3));
        ARROW_RETURN_NOT_OK(col4_builder.Append(row.col4));
    }

    //添加空值,末尾值的元素为空
    ARROW_RETURN_NOT_OK(col1_builder.AppendNull());
    ARROW_RETURN_NOT_OK(col2_builder.AppendNull());
    ARROW_RETURN_NOT_OK(col3_builder.AppendNull());
    ARROW_RETURN_NOT_OK(col4_builder.AppendNull());

    std::shared_ptr<arrow::Array> col1_array;
    ARROW_RETURN_NOT_OK(col1_builder.Finish(&col1_array));
    std::shared_ptr<arrow::Array> col2_array;
    ARROW_RETURN_NOT_OK(col2_builder.Finish(&col2_array));
    std::shared_ptr<arrow::Array> col3_array;
    ARROW_RETURN_NOT_OK(col3_builder.Finish(&col3_array));
    std::shared_ptr<arrow::Array> col4_array;
    ARROW_RETURN_NOT_OK(col4_builder.Finish(&col4_array));

    std::vector<std::shared_ptr<arrow::Field>> schema_vector = {
            arrow::field("col1", arrow::int32()), arrow::field("col2", arrow::int64()), arrow::field("col3", arrow::float64()),
            arrow::field("col4", arrow::utf8()) };

    auto schema = std::make_shared<arrow::Schema>(schema_vector);

    //最终的' table '变量是我们可以传递给其他可以使用Apache Arrow内存结构的函数的变量。这个对象拥有所有引用数据的所有权，
    //因此一旦我们离开构建表及其底层数组的函数的作用域，就不必关心未定义的引用。
    *table = arrow::Table::Make(schema, { col1_array, col2_array, col3_array,col4_array });

    return arrow::Status::OK();
}

arrow::Status TableToVector(const std::shared_ptr<arrow::Table>& table,
    std::vector<struct row_data>* rows) {
    //检查表结构是否一致
    std::vector<std::shared_ptr<arrow::Field>> schema_vector = {
            arrow::field("col1", arrow::int32()), arrow::field("col2", arrow::int64()), arrow::field("col3", arrow::float64()),
            arrow::field("col4", arrow::utf8()) };
    auto expected_schema = std::make_shared<arrow::Schema>(schema_vector);

    if (!expected_schema->Equals(*table->schema())) {
        // The table doesn't have the expected schema thus we cannot directly
        // convert it to our target representation.
        return arrow::Status::Invalid("Schemas are not matching!");
    }

    //获取对应列数据指针
    auto col1s =
        std::static_pointer_cast<arrow::Int32Array>(table->column(0)->chunk(0));
    auto col2s =
        std::static_pointer_cast<arrow::Int64Array>(table->column(1)->chunk(0));
    auto col3s =
        std::static_pointer_cast<arrow::DoubleArray>(table->column(2)->chunk(0));
    auto col4s =
        std::static_pointer_cast<arrow::StringArray>(table->column(3)->chunk(0));

    for (int64_t i = 0; i < table->num_rows(); i++) {
        if (col1s->IsNull(i)) {
            assert(i == 3);//第四行为null
        }
        else {
            int32_t col1 = col1s->Value(i);
            int64_t col2 = col2s->Value(i);
            double col3 = col3s->Value(i);
            std::string col4 = col4s->GetString(i);
            rows->push_back({ col1, col2, col3,col4 });
        }
    }

    return arrow::Status::OK();
}

// 行数组和列数组相互转换
int testTableConvertSTL() {
    //行数组
    std::vector<row_data> rows = {
            {1, 11,1.0, "John"}, {2, 22,2.0, "Tom"}, {3,33, 3.0,"Susan"} };

    std::shared_ptr<arrow::Table> table;
    EXIT_ON_FAILURE(CreateTable(rows, &table));

    std::vector<row_data> expected_rows;
    EXIT_ON_FAILURE(TableToVector(table, &expected_rows));
    std::cout << expected_rows.size() << std::endl;
    assert(rows.size() == expected_rows.size());
    return 0;
}

void test() {
    // 构建一个int8数组
    arrow::Int8Builder builder;
    arrow::Int16Builder int16builder;
    int8_t days_raw[5] = { 1, 12, 17, 23, 28 };
    int8_t months_raw[5] = { 1, 3, 5, 7, 1 };
    int16_t years_raw[5] = { 1990, 2000, 1995, 2000, 1995 };
    builder.AppendValues(days_raw, 5);
    std::shared_ptr<arrow::Array> days = builder.Finish().MoveValueUnsafe();    
    builder.AppendValues(months_raw, 5);
    std::shared_ptr<arrow::Array> months = builder.Finish().MoveValueUnsafe();    
    int16builder.AppendValues(years_raw, 5);
    std::shared_ptr<arrow::Array> years = int16builder.Finish().MoveValueUnsafe();

    // Schema 自定义table
    // Now, we want a RecordBatch, which has columns and labels for said columns.
    // This gets us to the 2d data structures we want in Arrow.
    // These are defined by schema, which have fields -- here we get both those object types
    // ready.
    std::shared_ptr<arrow::Field> field_day, field_month, field_year;
    std::shared_ptr<arrow::Schema> schema;

    // Every field needs its name and data type.
    field_day = arrow::field("Day", arrow::int8());
    field_month = arrow::field("Month", arrow::int8());
    field_year = arrow::field("Year", arrow::int16());

    // The schema can be built from a vector of fields, and we do so here.
    schema = arrow::schema({ field_day, field_month, field_year });

    // 打印
    // With the schema and Arrays full of data, we can make our RecordBatch! Here,
    // each column is internally contiguous. This is in opposition to Tables, which we'll
    // see next.
    std::shared_ptr<arrow::RecordBatch> rbatch;
    // The RecordBatch needs the schema, length for columns, which all must match,
    // and the actual data itself.
    rbatch = arrow::RecordBatch::Make(schema, days->length(), { days, months, years });
    std::cout << rbatch->ToString();
    /*
    Day:   [
        1,
        12,
        17,
        23,
        28
      ]
    Month:   [
        1,
        3,
        5,
        7,
        1
      ]
    Year:   [
        1990,
        2000,
        1995,
        2000,
        1995
      ]
    */

    // stl vector容器
    arrow::ArrayVector day_vecs{days};
    std::shared_ptr<arrow::ChunkedArray> day_chunks =
        std::make_shared<arrow::ChunkedArray>(day_vecs);
    testTableConvertSTL();

    testReadCSV();
    /*
    col1: string
    col2: string
    col3: string
    ----
    col1:
      [
        [
          "val1",
          "val1"
        ]
      ]
    col2:
      [
        [
          "val2",
          "val2"
        ]
      ]
    col3:
      [
        [
          "val3",
          "val3"
        ]
    ]
    */
    testWriteIpc();
    testReadIPC();
    //testComputeGreater();
    //testComputeMinMax();
}
Compute Functions — Apache Arrow v17.0.0
GitHub - apache/arrow: Apache Arrow is a multi-language toolbox for accelerated data interchange and in-memory processing