如何向后读取文件以有效地查找子字符串

Question

我有一个这种结构的巨大日志文件:

"timestamp":{"标识符":值}

"1463403600":{"AA":74.42},
"1463403601":{"AA":29.55},
"1463403603":{"AA":24.78},
"1463403604":{"AA":8.46},
"1463403605":{"AA":44.84},
"1463403607":{"AA":87.05},
"1463403608":{"AA":54.81},
"1463403609":{"AA":93.1},
"1463403611":{"AA":77.64},
"1463403612":{"AA":33.39},
"1463403613":{"AA":69.2},

我想在(!)给定的时间戳后提取内容,如:

std::ifstream * myfunc( uint32_t timestamp) 

例:

myfunc(1463403611);
/* returns
"1463403611":{"AA":77.64},
"1463403612":{"AA":33.39},
"1463403613":{"AA":69.2},
*/

日志文件很长 - 太长而无法将其保留在内存中.代码将在资源有限的嵌入式设备(80Mhz,~10kB可用内存)上运行,因此我正在寻找一些有效解决方案的想法.

日志文件可能有500k +条目,并且在99%的时间内,时间戳将在最后100行中,因此从文件的beginnig开始并检查每一行的正确时间戳将是非常低效的.

所以我想我正在寻找一个解决方案来逐行读取文件.我没有真正的解决方案如何在不将大块加载到内存中的情况下提高效率.

我尝试从EOF开始阅读200字节的块,但是面对这个问题,在很多情况下,块会将时间戳减少一半.我试图检测到并在需要时重新选择一些字节,但感觉必须有一个聪明的解决方案.

Answer 1

好吧,我发现这种有趣的东西,所以我为二元搜索的想法敲了一个概念验证.

这是一个测试不当,可能有点小马车但似乎到目前为止工作,并展示了分而治之的想法.您检查文件的中间位置,并根据您是高还是低,将数据分成两部分并搜索相关的一半.你递归地做到这一点,直到你足够接近.

#include <ctime>
#include <cmath>
#include <cstdlib>
#include <string>
#include <fstream>
#include <iostream>

// Don't use this, its just to show how many reads
// are being done to find the record.
int global_counter;

std::streampos find_stamp(std::istream& is, long stamp, std::streampos pos, std::streampos end)
{
    ++global_counter;

    if(pos == 0) // can't divide zero
        return 0;

    std::string s;
    long found_stamp;

    // extract nearest timestamp after pos
    is.seekg(pos);
    if(!(std::getline(std::getline(is, s, ','), s, '"') >> found_stamp))
        return end;

    // if its too big check first half of this region
    if(found_stamp > stamp)
        return find_stamp(is, stamp, pos / 2, pos);

    // if its not within 10 timestamp seconds check end half of this region
    if(stamp - found_stamp > 10)
        return find_stamp(is, stamp, (pos + end) / 2, end);

    // read record by record (prolly more efficient than skipping)
    pos = is.tellg();
    while(std::getline(std::getline(is, s, ','), s, '"') >> found_stamp)
    {
        if(found_stamp > stamp)
            return pos;
        pos = is.tellg();
    }
    return end;
}

void print_after(const std::string& filename, long stamp)
{
    // open at end of file (to get length)
    std::ifstream ifs(filename, std::ios::ate);

    std::streampos end = ifs.tellg();
    auto pos = end / 2; // start checking in middle

    // find position before required record
    // (may be in the middle of a record)
    if((pos = find_stamp(ifs, stamp, pos, end)) != end)
    {
        ifs.seekg(pos);

        std::string line;
        std::getline(ifs, line, ','); // skip to next whole record

        // print out all following recors
        while(std::getline(ifs, line, ','))
            std::cout << line;
    }
}

inline
std::string leading_zeros(int n, int zeros = 2)
{
    std::string s;
    for(int z = std::pow(10, zeros - 1); z; z /= 10)
        s += (n < z ? "0":"");
    return s + std::to_string(n);
}

int main()
{
    std::srand(std::time(0));

    // generate some test data
    std::ofstream ofs("test.txt");

    for(int i = 0; i < 1000; ++i)
    {
        ofs << '"' << leading_zeros(i, 10) << '"';
        ofs << ":{\"AA\":" << (std::rand() % 100);
        ofs << '.' << (std::rand() % 100) << "},\n";
    }

    ofs.close();

    global_counter = 0;
    print_after("test.txt", 993);

    std::cout << "find checked " << global_counter << " places in the file\n";
}

输出:

"0000000994":{"AA":80.6}
"0000000995":{"AA":11.90}
"0000000996":{"AA":16.43}
"0000000997":{"AA":53.11}
"0000000998":{"AA":68.43}
"0000000999":{"AA":79.77}
find checked 6 places in the file

Answer 2

由于您使用的mmap()是可能无法使用的嵌入式设备,我想唯一可行的方法是使用填充文件一部分的缓冲区,以便能够一次检查一个块的内容.请注意,您需要重叠缓冲区窗口,以避免错过缓冲区开头处减半的行.您需要在块的开头寻找第一个换行符,然后在开始检查时间戳的块之前将其丢弃.当您将前一个块加载到缓冲区时,丢弃缓冲区开头的部分行也有助于将同一行的末尾与缓冲区的末尾对齐.

在缓冲区开头处理不完整的行使得这是一种非常难看且容易出错的方法.这就是为什么我建议使用mmap()if if all available,它会让你忽略这些问题.