记一次22亿大数据分析处理踩坑经历

前言：用最简单最少的语言，分享我的坑，理不理解需求不重要，问题都在shell代码中，看shell极度枯燥，希望能有帮助！

一. 起因

需求，分析hive表中两年内用户XX的所有数据，按照每天早，中，晚三个时间段统计，按照两年内的节假日统计，按照月份的上旬，中旬，下旬统计，按照周末，工作日统计等等。
假设现数据形式，手机号代表每一个用户，用户不同时间发送的短信数量作为统计目的！
最后，按照类似mobile , am_count , noon_count , pm_count , springday_count , nationalday_count,weekend_count,weekday_count形式统计为一张表！
说的太抽象，但是，你可以了解的有：

二. 解决方案

方案一

按照需求，将每一个字段对应一条sql的方式求出mobile , count的值，然后将这些字段统计起来（利用mysql的唯一键unique index和 duplicate on update方式）。

具体步骤：

1. hive脚本导出每一列数据

#!/usr/bin/env bash
echo '-----------开始从hive查数----------------'
HIVE_SETTING="
SET mapred.child.java.opts=-Xmx8192m;
SET mapreduce.reduce.memory.mb=8192;
SET mapreduce.reduce.java.opts='-Xmx8192M';
SET mapreduce.map.memory.mb=8192;
SET mapreduce.map.java.opts='-Xmx8192M';
SET mapred.child.map.java.opts='-Xmx8192M';
SET mapred.job.priority=HIGH;
SET mapred.map.tasks.speculative.execution=false;
SET mapred.reduce.tasks.speculative.execution=false;
set hive.exec.dynamic.partition.mode=nonstrict;
set hive.exec.dynamic.partition=true;
SET hive.exec.max.dynamic.partitions=100000;
SET hive.exec.max.dynamic.partitions.pernode=100000;
USE xxxdb;
set mapred.job.queue.name=wirelessdev;
set hive.exec.compress.output=true;
set mapred.output.compression.codec=org.apache.hadoop.io.compress.GzipCodec;
set hive.exec.parallel=true;
set mapred.job.name = ${0}_xxx;
"

#工作日数据查询
HIVE_SQL="
select mobile,count(mobile) from rdb_sms_outbox_financial where delivrd='DELIVRD' and pmod(datediff(optime, '2012-01-01'), 7) in (1,2,3,4,5) group by mobile;
"
#1. 将hive执行结果赋值给变量
DATA=$(hive -e "
${HIVE_SETTING}
${HIVE_SQL};
")

#2. 将hive结果输出到文件中
hive -e "
${HIVE_SETTING}
${HIVE_SQL};
" >/home/q/hive_data.txt
echo '-----------结束从hive查数----------------'

查询出来数据22亿, 约占45G磁盘空间.

2. mysql脚本导入数据

#!/usr/bin/env bash
echo "进程Pid: $$"
#将文本文件里面(mobile,count)字段插入到mysql中
insertIntoMysql(){
    #获取参数
    path=${1}
    col=${2}
    TIMESTAMP=$(date +%Y%m%d%H%M%S)
    echo "path:${path},column:${col},time:${TIMESTAMP}"
    #遍历文件每一行
    cat ${path} | while read line
    do
        #获取每一行中的每一列
        mobile=$(echo -e "${line}" | cut -f 1)
        count=$(echo -e "${line}" | cut -f 2)
        #写入myusql
        cmd="INSERT INTO sms.sms_financial(mobile,${col}) VALUES ('${mobile}',${count}) ON DUPLICATE KEY UPDATE ${col}=${count};"
        eval $(mysql -uroot -pxxx --default-character-set=utf8 -e "${cmd}")
        echo "mobile:${mobile},count:${count}"
    done

    TIMESTAMP=$(date +%Y%m%d%H%M%S)
    echo "end time:${TIMESTAMP}"
}

#保存每个字段(mobile,count)的文件目录
path="/home/q/part1"
eval cd ${path}

line=$(find ${path} -type f)
for s in ${line[@]}
do
    #截取文件名，即mysql table中对应的列名！
    col=$(echo ${s} |cut -d "/" -f5)
    insertIntoMysql ${s} ${col}
done
exit;

到此,似乎是完了,多开几个脚本一起往mysql中导数就行了. 但是,这只是开始!

问题

1. 为什么不用mysql的批量导入?
2. 一行一行的插入22亿数据,要插入多久?

答: 批量导入的原子操作整行数据 , 无法做到聚合列! 22亿数据多个脚本,24小时插入量在2000W左右!

改进1: ok单表插入太慢,我分表插入会快一些吧! 改进脚本!

#!/usr/bin/env bash

#多表插入,根据mobile确定表名
# (sms_financial,sms_financial11,sms_financial0,...,sms_financial9)
getTableName(){
    mobile=${1}
    table="sms_financial"
    if [ ${mobile} -a -n ${mobile} ]
    then
        prefix=$(echo "${mobile:0:5}")
        #861开头的手机号太多,所以又分十张表
        if [ ${prefix:0:3} == "861" ]; then
            model=`expr ${prefix} % 10`
            table=${table}"${model}"
        fi

        #11位手机号的分一张sms_financial11
        if [ ${#mobile} == 11 -a ${mobile:0:1} == "1" ]; then
            table=${table}"11"
        fi
        echo "${table}"
    else
        #国际,其他的分一张sms_financial
        echo "${table}"
    fi
}

#将文本文件里面（mobile，count）字段插入到mysql中
insertIntoMysql(){
    path=${1}
    col=${2}
    echo "path:${path},column:${col}"
    cat ${path} | while read line
    do
        mobile=$(echo -e "${line}" | cut -f 1)
        count=$(echo -e "${line}" | cut -f 2)
        table=`getTableName ${mobile}`
        cmd="INSERT INTO sms.${table}(mobile,${col}) VALUES ('${mobile}',${count}) ON DUPLICATE KEY UPDATE ${col}=${count};"
        eval $(mysql -h127.0.0.1 -P3306 -uroot -p'xxx' --default-character-set=utf8 -e "${cmd}")
        echo "table:${table},mobile:${mobile},count:${count}"
    done
}

path="/home/q/data_hive/hive1"
eval cd ${path}
line=$(find ${path} -type f)
for s in ${line[@]}
do
    col=$(echo ${s} |cut -d "/" -f6)
    insertIntoMysql ${s} ${col}
done
exit;

问题

1. 的确横向分表后插入数据的确快很多,但是会出现数据集中同时插入同一张表的情况,依旧不能容忍!

改进2: ok一条一条的插入不可以,我批量插入!

但是,上面横向分表逻辑不能使用了!因为每一个手机号对应的表不一样,sql语句拼接很困难!既然,横切表不行,为了简单,我选择纵切表(将表的列切开mobile, count1,mobile,count2的形式).

#!/usr/bin/env bash

echo "进程Pid: $$"
#将文本文件里面（mobile，count）字段插入到mysql中
insertIntoMysql(){
   path=${1}
   col=${2}
   TIMESTAMP=$(date +%Y%m%d%H%M%S)
   echo "path:${path},column:${col},time:${TIMESTAMP}"
   str1="INSERT INTO sms.sms_financial99(mobile,${col}) VALUES "
   str2=" ON DUPLICATE KEY UPDATE ${col}=VALUES(${col});"
   n=0
   cat ${path} | while read line
   do
       mobile=$(echo -e "${line}" | cut -f 1)
       count=$(echo -e "${line}" | cut -f 2)
       let n++
       if [ `expr ${n} % 5000` == 0 ];
       then
           cmd=${cmd}"('${mobile}',${count})"
           cmd=${str1}${cmd}${str2}
           #echo ${cmd}
           eval $(mysql -h127.0.0.1 -P3306 -uroot -p'xxx' --default-character-set=utf8 -e "${cmd}")
           cmd=" "
       else
           cmd=${cmd}"('${mobile}',${count}),"
       fi
       #echo "mobile:${mobile},count:${count}"
   done
   TIMESTAMP=$(date +%Y%m%d%H%M%S)
   echo "end ${col} time:${TIMESTAMP}"
}
path="/home/q/data_hive/hive2"
eval cd ${path}

line=$(find ${path} -type f)
for s in ${line[@]}
do
   col=$(echo ${s} |cut -d "/" -f6)
   insertIntoMysql ${s} ${col}
done
exit;

其中, 一次批量插入5000条, 考虑到shell中会限制参数的长度(报错: /usr/bin/mysql: Argument list too long)!
还有mysql提交sql长度默认为4M,我们可以通过show VARIABLES like '%max_allowed_packet%'; set global max_allowed_packet=33554432;查看和修改!

方案二

上面纵切,批量插入虽然基本满足需求,但是会存在两个问题,1. 如果mysql开启了bin-log很可能会导致磁盘报警! 2. 批量插入可能会出现死锁(期间出现过一次,调整批插文件顺序(减少在同一列上操作的机会))!
其实,整个问题一个hive-sql可以搞定将多列进行聚合:

#!/usr/bin/env bash
echo '-----------开始从hive查数----------------'
TIMESTAMP=$(date +%Y%m%d%H%M%S)
echo "PID: $$,start time:${TIMESTAMP}"
HIVE_SETTING="
xxx
"
HIVE_SQL="
select a.mobile,
  if(b1.midnight_msg_no_receive_count>0, b1.midnight_msg_no_receive_count,0) as midnight_msg_no_receive_count,
  if(b2.am_msg_no_receive_count>0, b2.am_msg_no_receive_count,0) as am_msg_no_receive_count,
  if(b3.noon_msg_no_receive_count>0, b3.noon_msg_no_receive_count,0) as noon_msg_no_receive_count,
   ...
  if(e8.last_one_year_normal_msg_no_receive_count>0,e8.last_one_year_normal_msg_no_receive_count,0) as last_one_year_normal_msg_no_receive_count
from
  (select mobile from rdb_sms_outbox_financial where delivrd='UNDELIVRD' group by mobile) a left outer join
  (select mobile,count(mobile) as midnight_msg_no_receive_count from rdb_sms_outbox_financial where delivrd='UNDELIVRD' and hour(optime) in (0,1,2,3,4,5,23) group by mobile) b1 on a.mobile=b1.mobile left outer join
  (select mobile,count(mobile) as am_msg_no_receive_count from rdb_sms_outbox_financial where delivrd='UNDELIVRD' and hour(optime) in (6,7,8,9,10) group by mobile) b2 on a.mobile=b2.mobile left outer join
   ...
  (select mobile,count(mobile) as last_one_year_normal_msg_no_receive_count from rdb_sms_outbox_financial where delivrd='UNDELIVRD' and ivr=0 and to_date(optime)>='2016-04-01' and to_date(optime)<='2017-03-31' group by mobile) e8 on a.mobile=e8.mobile;
"

hive -e "
${HIVE_SETTING}
${HIVE_SQL}
" >/home/q/data_to_hive/data_hive/data_hive_undelivrd

TIMESTAMP=$(date +%Y%m%d%H%M%S)
echo "end time:${TIMESTAMP}"
echo '-----------结束从hive查数----------------'
exit;

三. 总结

这里,我认为价值在于我走的弯路上！为了解决mysql插入性能问题,实施的一系列探索上, 同时积累了用脚本对mysql这些操作的熟练性.
过程中遇到的问题都轻描淡写了(有google!)，从本文你将可以了解以下知识:

1. hive脚本相关操作
2. mysql数据插入,批量插入脚本的使用,及其中我遇到的一些坑.
3. 脚本处理数据的一些操作(遍历目录下的每一个文件, 遍历文件的每一行,获取每一行中的每一列,记录shell线程,执行时间,函数传参和返回值)
4. 理解做事情的思路是多么的重要.
5. 这是一次xxx的经历.