Python多进程导入CSV至数据库

前段时间帮同事处理了一个把 CSV 数据导入到 MySQL 的需求。两个很大的 CSV 文件， 分别有 3GB、2100 万条记录和 7GB、3500 万条记录。对于这个量级的数据，用简单的单进程／单线程导入 会耗时很久，最终用了多进程的方式来实现。具体过程不赘述，记录一下几个要点：

1. 批量插入而不是逐条插入

2. 为了加快插入速度，先不要建索引

3. 生产者和消费者模型，主进程读文件，多个 worker 进程执行插入

4. 注意控制 worker 的数量，避免对 MySQL 造成太大的压力

5. 注意处理脏数据导致的异常

6. 原始数据是 GBK 编码，所以还要注意转换成 UTF-8

7. 用 click 封装命令行工具
   

具体的代码实现如下：

#!/usr/bin/env python
# -*- coding: utf-8 -*-

import codecs
import csv
import logging
import multiprocessing
import os
import warnings

import click
import MySQLdb
import sqlalchemy

warnings.filterwarnings(&＃39;ignore&＃39;, category=MySQLdb.Warning)

# 批量插入的记录数量
BATCH = 5000

DB_URI = &＃39;mysql://root@localhost:3306/example?charset=utf8&＃39;

engine = sqlalchemy.create_engine(DB_URI)


def get_table_cols(table):
  sql = &＃39;SELECT * FROM `{table}` LIMIT 0&＃39;.format(table=table)
  res = engine.execute(sql)
  return res.keys()


def insert_many(table, cols, rows, cursor):
  sql = &＃39;INSERT INTO `{table}` ({cols}) VALUES ({marks})&＃39;.format(
      table=table,
      cols=&＃39;, &＃39;.join(cols),
      marks=&＃39;, &＃39;.join([&＃39;%s&＃39;] * len(cols)))
  cursor.execute(sql, *rows)
  logging.info(&＃39;process %s inserted %s rows into table %s&＃39;, os.getpid(), len(rows), table)


def insert_worker(table, cols, queue):
  rows = []
  # 每个子进程创建自己的 engine 对象
  cursor = sqlalchemy.create_engine(DB_URI)
  while True:
    row = queue.get()
    if row is None:
      if rows:
        insert_many(table, cols, rows, cursor)
      break

    rows.append(row)
    if len(rows) == BATCH:
      insert_many(table, cols, rows, cursor)
      rows = []


def insert_parallel(table, reader, w=10):
  cols = get_table_cols(table)

  # 数据队列，主进程读文件并往里写数据，worker 进程从队列读数据
  # 注意一下控制队列的大小，避免消费太慢导致堆积太多数据，占用过多内存
  queue = multiprocessing.Queue(maxsize=w*BATCH*2)
  workers = []
  for i in range(w):
    p = multiprocessing.Process(target=insert_worker, args=(table, cols, queue))
    p.start()
    workers.append(p)
    logging.info(&＃39;starting # %s worker process, pid: %s...&＃39;, i + 1, p.pid)

  dirty_data_file = &＃39;./{}_dirty_rows.csv&＃39;.format(table)
  xf = open(dirty_data_file, &＃39;w&＃39;)
  writer = csv.writer(xf, delimiter=reader.dialect.delimiter)

  for line in reader:
    # 记录并跳过脏数据: 键值数量不一致
    if len(line) != len(cols):
      writer.writerow(line)
      continue

    # 把 None 值替换为 &＃39;NULL&＃39;
    clean_line = [None if x == &＃39;NULL&＃39; else x for x in line]

    # 往队列里写数据
    queue.put(tuple(clean_line))
    if reader.line_num % 500000 == 0:
      logging.info(&＃39;put %s tasks into queue.&＃39;, reader.line_num)

  xf.close()

  # 给每个 worker 发送任务结束的信号
  logging.info(&＃39;send close signal to worker processes&＃39;)
  for i in range(w):
    queue.put(None)

  for p in workers:
    p.join()


def convert_file_to_utf8(f, rv_file=None):
  if not rv_file:
    name, ext = os.path.splitext(f)
    if isinstance(name, unicode):
      name = name.encode(&＃39;utf8&＃39;)
    rv_file = &＃39;{}_utf8{}&＃39;.format(name, ext)
  logging.info(&＃39;start to process file %s&＃39;, f)
  with open(f) as infd:
    with open(rv_file, &＃39;w&＃39;) as outfd:
      lines = []
      loop = 0
      chunck = 200000
      first_line = infd.readline().strip(codecs.BOM_UTF8).strip() + &＃39;\n&＃39;
      lines.append(first_line)
      for line in infd:
        clean_line = line.decode(&＃39;gb18030&＃39;).encode(&＃39;utf8&＃39;)
        clean_line = clean_line.rstrip() + &＃39;\n&＃39;
        lines.append(clean_line)
        if len(lines) == chunck:
          outfd.writelines(lines)
          lines = []
          loop += 1
          logging.info(&＃39;processed %s lines.&＃39;, loop * chunck)

      outfd.writelines(lines)
      logging.info(&＃39;processed %s lines.&＃39;, loop * chunck + len(lines))


@click.group()
def cli():
  logging.basicConfig(level=logging.INFO,
            format=&＃39;%(asctime)s - %(levelname)s - %(name)s - %(message)s&＃39;)


@cli.command(&＃39;gbk_to_utf8&＃39;)
@click.argument(&＃39;f&＃39;)
def convert_gbk_to_utf8(f):
  convert_file_to_utf8(f)


@cli.command(&＃39;load&＃39;)
@click.option(&＃39;-t&＃39;, &＃39;--table&＃39;, required=True, help=&＃39;表名&＃39;)
@click.option(&＃39;-i&＃39;, &＃39;--filename&＃39;, required=True, help=&＃39;输入文件&＃39;)
@click.option(&＃39;-w&＃39;, &＃39;--workers&＃39;, default=10, help=&＃39;worker 数量，默认 10&＃39;)
def load_fac_day_pro_nos_sal_table(table, filename, workers):
  with open(filename) as fd:
    fd.readline()  # skip header
    reader = csv.reader(fd)
    insert_parallel(table, reader, w=workers)


if name == &＃39;main&＃39;:
  cli()