基于binlog来分析mysql的行记录修改情况(python脚本分析)
阅读目录(Content)
最近写完mysql flashback,突然发现还有有这种使用场景:有些情况下,可能会统计在某个时间段内,MySQL修改了多少数据量?发生了多少事务?主要是哪些表格发生变动?变动的数量是怎么样的? 但是却不需要行记录的修改内容,只需要了解 行数据的 变动情况。故也整理了下。
昨晚写的脚本,因为个人python能力有限,本来想这不发这文,后来想想,没准会有哪位园友给出优化建议。
如果转载,请注明博文来源: www.cnblogs.com/xinysu/ ,版权归 博客园 苏家小萝卜 所有。望各位支持!
1 实现内容
有些情况下,可能会统计在某个时间段内,MySQL修改了多少数据量?发生了多少事务?主要是哪些表格发生变动?变动的数量是怎么样的? 但是却不需要行记录的修改内容,只需要了解 行数据的 变动情况。
这些情况部分可以通过监控来大致了解,但是也可以基于binlog来全盘分析,binlog的格式是row模式。
在写flashback的时候,顺带把这个也写了个脚步,使用python编写,都差不多原理,只是这个简单些,介于个人python弱的不行,性能可能还有很大的提升空间,也希望园友能协助优化下。
先贴python脚步的分析结果图如下,分为4个部分:事务耗时情况、事务影响行数情况、DML行数情况以及操作最频繁表格情况。
2 脚本简单描述
脚本依赖的模块中,pymysql需要自行安装。
创建类queryanalyse,其中有5个函数定义:_get_db、create_tab、rowrecord、binlogdesc跟closeconn。
2.1 _get_db
该函数用来解析输入参数值,参数值一共有7个,都是必须填写的。分别为host,user,password,port,table name for transaction,table name for records,对应的简写如下:
ALL options need to assign:
-h : host, the database host,which database will store the results after analysis
-u : user, the db user
-p : password, the db user's password
-P : port, the db port
-f : file path, the binlog file
-tr : table name for record , the table name to store the row record
-tt : table name for transaction, the table name to store transactions
比如,执行脚本:python queryanalyse.py -h=127.0.0.1 -P=3310 -u=root -p=password -f=/tmp/stock_binlog.log -tt=flashback.tbtran -tr=flashback.tbrow,该函数负责处理各个选项的参数值情况,并存储。
2.2 create_tab
创建两个表格,分别用来存储 binlog file文件的分析结果。一个用来存储事务的执行开始时间跟结束时间,由选项 -tt来赋值表名;一个是用来存储每一行记录的修改情况,由选项 -tr来赋值表名。
事务表记录内容:事务的开始时间及事务的结束时间。
行记录表的内容:库名,表名,DML类型以及事务对应事务表的编号。
root@localhost:mysql3310.sock 14:42:29 [flashback]>show create table tbrow \G *************************** 1. row *************************** Table: tbrow Create Table: CREATE TABLE `tbrow` ( `auto_id` int(10) unsigned NOT NULL AUTO_INCREMENT, `sqltype` int(11) NOT NULL COMMENT '1 is insert,2 is update,3 is delete', `tran_num` int(11) NOT NULL COMMENT 'the transaction number', `dbname` varchar(50) NOT NULL, `tbname` varchar(50) NOT NULL, PRIMARY KEY (`auto_id`), KEY `sqltype` (`sqltype`), KEY `dbname` (`dbname`), KEY `tbname` (`tbname`) ) ENGINE=InnoDB AUTO_INCREMENT=295151 DEFAULT CHARSET=utf8 1 row in set (0.00 sec) root@localhost:mysql3310.sock 14:42:31 [flashback]>SHOW CREATE TABLE TBTRAN \G *************************** 1. row *************************** Table: TBTRAN Create Table: CREATE TABLE `tbtran` ( `auto_id` int(10) unsigned NOT NULL AUTO_INCREMENT, `begin_time` datetime NOT NULL, `end_time` datetime NOT NULL, PRIMARY KEY (`auto_id`) ) ENGINE=InnoDB AUTO_INCREMENT=6390 DEFAULT CHARSET=utf8 1 row in set (0.00 sec)
2.3 rowrecord
重点函数,分析binlog文件内容。这里有几个规律:
- 每个事务的结束点,是以 'Xid = ' 来查找
- 事务的开始时间,是事务内的第一个 'Table_map' 行里边的时间
- 事务的结束时间,是以 'Xid = '所在行的 里边的时间
- 每个行数据是属于哪个表格,是以 'Table_map'来查找
- DML的类型是按照 行记录开头的情况是否为:'### INSERT INTO' 、'### UPDATE' 、'### DELETE FROM'
- 注意,单个事务可以包含多个表格多种DML多行数据修改的情况。
2.4 binlogdesc
描述分析结果,简单4个SQL分析。
- 分析修改行数据的 事务耗时情况
- 分析修改行数据的 事务影响行数情况
- 分析DML分布情况
- 分析 最多DML操作的表格 ,取前十个分析
2.5 closeconn
关闭数据库连接。
3 使用说明
首先,确保python安装了pymysql模块,把python脚本拷贝到文件 queryanalyse.py。
然后,把要分析的binlog文件先用 mysqlbinlog 指令分析存储,具体binlog的文件说明,可以查看之前的博文:关于binary log那些事——认真码了好长一篇。mysqlbinlog的指令使用方法,可以详细查看文档:https://dev.mysql.com/doc/refman/5.7/en/mysqlbinlog.html 。
比较常用通过指定开始时间跟结束时间来分析 binlog文件。
mysqlbinlog --start-datetime='2017-04-23 00:00:03' --stop-datetime='2017-04-23 00:30:00' --base64-output=decode-rows -v /data/mysql/logs/mysql-bin.007335 > /tmp/binlog_test.log
分析后,可以把这个 binlog_test.log文件拷贝到其他空闲服务器执行分析,只需要有个空闲的DB来存储分析记录即可。
假设这个时候,拷贝 binlog_test.log到测试服务器上,测试服务器上的数据库可以用来存储分析内容,则可以执行python脚本了,注意要进入到python脚本的目录中,或者指定python脚本路径。
python queryanalyse.py -h=127.0.0.1 -P=3310 -u=root -p=password -f= /tmp/binlog_test.log -tt=flashback.tbtran -tr=flashback.tbrow
没了,就等待输出吧。
性能是硬伤,在虚拟机上测试,大概500M的binlog文件需要分析2-3min,有待提高!
4 python脚本
1 import pymysql 2 from pymysql.cursors import DictCursor 3 import re 4 import os 5 import sys 6 import datetime 7 import time 8 import logging 9 import importlib 10 importlib.reload(logging) 11 logging.basicConfig(level=logging.DEBUG,format='%(asctime)s %(levelname)s %(message)s ') 12 13 14 usage=''' usage: python [script's path] [option] 15 ALL options need to assign: 16 17 -h : host, the database host,which database will store the results after analysis 18 -u : user, the db user 19 -p : password, the db user's password 20 -P : port, the db port 21 -f : file path, the binlog file 22 -tr : table name for record , the table name to store the row record 23 -tt : table name for transaction, the table name to store transactions 24 Example: python queryanalyse.py -h=127.0.0.1 -P=3310 -u=root -p=password -f=/tmp/stock_binlog.log -tt=flashback.tbtran -tr=flashback.tbrow 25 26 ''' 27 28 class queryanalyse: 29 def __init__(self): 30 #初始化 31 self.host='' 32 self.user='' 33 self.password='' 34 self.port='3306' 35 self.fpath='' 36 self.tbrow='' 37 self.tbtran='' 38 39 self._get_db() 40 logging.info('assign values to parameters is done:host={},user={},password=***,port={},fpath={},tb_for_record={},tb_for_tran={}'.format(self.host,self.user,self.port,self.fpath,self.tbrow,self.tbtran)) 41 42 self.mysqlconn = pymysql.connect(host=self.host, user=self.user, password=self.password, port=self.port,charset='utf8') 43 self.cur = self.mysqlconn.cursor(cursor=DictCursor) 44 logging.info('MySQL which userd to store binlog event connection is ok') 45 46 self.begin_time='' 47 self.end_time='' 48 self.db_name='' 49 self.tb_name='' 50 51 def _get_db(self): 52 #解析用户输入的选项参数值,这里对password的处理是明文输入,可以自行处理成是input格式, 53 #由于可以拷贝binlog文件到非线上环境分析,所以password这块,没有特殊处理 54 logging.info('begin to assign values to parameters') 55 if len(sys.argv) == 1: 56 print(usage) 57 sys.exit(1) 58 elif sys.argv[1] == '--help': 59 print(usage) 60 sys.exit() 61 elif len(sys.argv) > 2: 62 for i in sys.argv[1:]: 63 _argv = i.split('=') 64 if _argv[0] == '-h': 65 self.host = _argv[1] 66 elif _argv[0] == '-u': 67 self.user = _argv[1] 68 elif _argv[0] == '-P': 69 self.port = int(_argv[1]) 70 elif _argv[0] == '-f': 71 self.fpath = _argv[1] 72 elif _argv[0] == '-tr': 73 self.tbrow = _argv[1] 74 elif _argv[0] == '-tt': 75 self.tbtran = _argv[1] 76 elif _argv[0] == '-p': 77 self.password = _argv[1] 78 else: 79 print(usage) 80 81 def create_tab(self): 82 #创建两个表格:一个用户存储事务情况,一个用户存储每一行数据修改的情况 83 #注意,一个事务可以存储多行数据修改的情况 84 logging.info('creating table ...') 85 create_tb_sql ='''CREATE TABLE IF NOT EXISTS {} ( 86 `auto_id` int(10) unsigned NOT NULL AUTO_INCREMENT, 87 `begin_time` datetime NOT NULL, 88 `end_time` datetime NOT NULL, 89 PRIMARY KEY (`auto_id`) 90 ) ENGINE=InnoDB DEFAULT CHARSET=utf8; 91 CREATE TABLE IF NOT EXISTS {} ( 92 `auto_id` int(10) unsigned NOT NULL AUTO_INCREMENT, 93 `sqltype` int(11) NOT NULL COMMENT '1 is insert,2 is update,3 is delete', 94 `tran_num` int(11) NOT NULL COMMENT 'the transaction number', 95 `dbname` varchar(50) NOT NULL, 96 `tbname` varchar(50) NOT NULL, 97 PRIMARY KEY (`auto_id`), 98 KEY `sqltype` (`sqltype`), 99 KEY `dbname` (`dbname`), 100 KEY `tbname` (`tbname`) 101 ) ENGINE=InnoDB DEFAULT CHARSET=utf8; 102 truncate table {}; 103 truncate table {}; 104 '''.format(self.tbtran,self.tbrow,self.tbtran,self.tbrow) 105 106 self.cur.execute(create_tb_sql) 107 logging.info('created table {} and {}'.format(self.tbrow,self.tbtran)) 108 109 def rowrecord(self): 110 #处理每一行binlog 111 #事务的结束采用 'Xid =' 来划分 112 #分析结果,按照一个事务为单位存储提交一次到db 113 try: 114 tran_num=1 #事务数 115 record_sql='' #行记录的insert sql 116 tran_sql='' #事务的insert sql 117 118 self.create_tab() 119 120 with open(self.fpath,'r') as binlog_file: 121 logging.info('begining to analyze the binlog file ,this may be take a long time !!!') 122 logging.info('analyzing...') 123 124 for bline in binlog_file: 125 126 if bline.find('Table_map:') != -1: 127 l = bline.index('server') 128 n = bline.index('Table_map') 129 begin_time = bline[:l:].rstrip(' ').replace('#', '20') 130 131 if record_sql=='': 132 self.begin_time = begin_time[0:4] + '-' + begin_time[4:6] + '-' + begin_time[6:] 133 134 self.db_name = bline[n::].split(' ')[1].replace('`', '').split('.')[0] 135 self.tb_name = bline[n::].split(' ')[1].replace('`', '').split('.')[1] 136 bline='' 137 138 elif bline.startswith('### INSERT INTO'): 139 record_sql=record_sql+"insert into {}(sqltype,tran_num,dbname,tbname) VALUES (1,{},'{}','{}');".format(self.tbrow,tran_num,self.db_name,self.tb_name) 140 141 elif bline.startswith('### UPDATE'): 142 record_sql=record_sql+"insert into {}(sqltype,tran_num,dbname,tbname) VALUES (2,{},'{}','{}');".format(self.tbrow,tran_num,self.db_name,self.tb_name) 143 144 elif bline.startswith('### DELETE FROM'): 145 record_sql=record_sql+"insert into {}(sqltype,tran_num,dbname,tbname) VALUES (3,{},'{}','{}');".format(self.tbrow,tran_num,self.db_name,self.tb_name) 146 147 elif bline.find('Xid =') != -1: 148 149 l = bline.index('server') 150 end_time = bline[:l:].rstrip(' ').replace('#', '20') 151 self.end_time = end_time[0:4] + '-' + end_time[4:6] + '-' + end_time[6:] 152 tran_sql=record_sql+"insert into {}(begin_time,end_time) VALUES ('{}','{}')".format(self.tbtran,self.begin_time,self.end_time) 153 154 self.cur.execute(tran_sql) 155 self.mysqlconn.commit() 156 record_sql = '' 157 tran_num += 1 158 159 except Exception: 160 return 'funtion rowrecord error' 161 162 def binlogdesc(self): 163 sql='' 164 t_num=0 165 r_num=0 166 logging.info('Analysed result printing...\n') 167 #分析总的事务数跟行修改数量 168 sql="select 'tbtran' name,count(*) nums from {} union all select 'tbrow' name,count(*) nums from {};".format(self.tbtran,self.tbrow) 169 self.cur.execute(sql) 170 rows=self.cur.fetchall() 171 for row in rows: 172 if row['name']=='tbtran': 173 t_num = row['nums'] 174 else: 175 r_num = row['nums'] 176 print('This binlog file has {} transactions, {} rows are changed '.format(t_num,r_num)) 177 178 # 计算 最耗时 的单个事务 179 # 分析每个事务的耗时情况,分为5个时间段来描述 180 # 这里正常应该是 以毫秒来分析的,但是binlog中,只精确时间到second 181 sql='''select 182 count(case when cost_sec between 0 and 1 then 1 end ) cos_1, 183 count(case when cost_sec between 1.1 and 5 then 1 end ) cos_5, 184 count(case when cost_sec between 5.1 and 10 then 1 end ) cos_10, 185 count(case when cost_sec between 10.1 and 30 then 1 end ) cos_30, 186 count(case when cost_sec >30.1 then 1 end ) cos_more, 187 max(cost_sec) cos_max 188 from 189 ( 190 select 191 auto_id,timestampdiff(second,begin_time,end_time) cost_sec 192 from {} 193 ) a;'''.format(self.tbtran) 194 self.cur.execute(sql) 195 rows=self.cur.fetchall() 196 197 for row in rows: 198 print('The most cost time : {} '.format(row['cos_max'])) 199 print('The distribution map of each transaction costed time: ') 200 print('Cost time between 0 and 1 second : {} , {}%'.format(row['cos_1'],int(row['cos_1']*100/t_num))) 201 print('Cost time between 1.1 and 5 second : {} , {}%'.format(row['cos_5'], int(row['cos_5'] * 100 / t_num))) 202 print('Cost time between 5.1 and 10 second : {} , {}%'.format(row['cos_10'], int(row['cos_10'] * 100 / t_num))) 203 print('Cost time between 10.1 and 30 second : {} , {}%'.format(row['cos_30'], int(row['cos_30'] * 100 / t_num))) 204 print('Cost time > 30.1 : {} , {}%\n'.format(row['cos_more'], int(row['cos_more'] * 100 / t_num))) 205 206 # 计算 单个事务影响行数最多 的行数量 207 # 分析每个事务 影响行数 情况,分为5个梯度来描述 208 sql='''select 209 count(case when nums between 0 and 10 then 1 end ) row_1, 210 count(case when nums between 11 and 100 then 1 end ) row_2, 211 count(case when nums between 101 and 1000 then 1 end ) row_3, 212 count(case when nums between 1001 and 10000 then 1 end ) row_4, 213 count(case when nums >10001 then 1 end ) row_5, 214 max(nums) row_max 215 from 216 ( 217 select 218 count(*) nums 219 from {} group by tran_num 220 ) a;'''.format(self.tbrow) 221 self.cur.execute(sql) 222 rows=self.cur.fetchall() 223 224 for row in rows: 225 print('The most changed rows for each row: {} '.format(row['row_max'])) 226 print('The distribution map of each transaction changed rows : ') 227 print('Changed rows between 1 and 10 second : {} , {}%'.format(row['row_1'],int(row['row_1']*100/t_num))) 228 print('Changed rows between 11 and 100 second : {} , {}%'.format(row['row_2'], int(row['row_2'] * 100 / t_num))) 229 print('Changed rows between 101 and 1000 second : {} , {}%'.format(row['row_3'], int(row['row_3'] * 100 / t_num))) 230 print('Changed rows between 1001 and 10000 second : {} , {}%'.format(row['row_4'], int(row['row_4'] * 100 / t_num))) 231 print('Changed rows > 10001 : {} , {}%\n'.format(row['row_5'], int(row['row_5'] * 100 / t_num))) 232 233 # 分析 各个行数 DML的类型情况 234 # 描述 delete,insert,update的分布情况 235 sql='select sqltype ,count(*) nums from {} group by sqltype ;'.format(self.tbrow) 236 self.cur.execute(sql) 237 rows=self.cur.fetchall() 238 239 print('The distribution map of the {} changed rows : '.format(r_num)) 240 for row in rows: 241 242 if row['sqltype']==1: 243 print('INSERT rows :{} , {}% '.format(row['nums'],int(row['nums']*100/r_num))) 244 if row['sqltype']==2: 245 print('UPDATE rows :{} , {}% '.format(row['nums'],int(row['nums']*100/r_num))) 246 if row['sqltype']==3: 247 print('DELETE rows :{} , {}%\n '.format(row['nums'],int(row['nums']*100/r_num))) 248 249 # 描述 影响行数 最多的表格 250 # 可以分析是哪些表格频繁操作,这里显示前10个table name 251 sql = '''select 252 dbname,tbname , 253 count(*) ALL_rows, 254 count(*)*100/{} per, 255 count(case when sqltype=1 then 1 end) INSERT_rows, 256 count(case when sqltype=2 then 1 end) UPDATE_rows, 257 count(case when sqltype=3 then 1 end) DELETE_rows 258 from {} 259 group by dbname,tbname 260 order by ALL_rows desc 261 limit 10;'''.format(r_num,self.tbrow) 262 self.cur.execute(sql) 263 rows = self.cur.fetchall() 264 265 print('The distribution map of the {} changed rows : '.format(r_num)) 266 print('tablename'.ljust(50), 267 '|','changed_rows'.center(15), 268 '|','percent'.center(10), 269 '|','insert_rows'.center(18), 270 '|','update_rows'.center(18), 271 '|','delete_rows'.center(18) 272 ) 273 print('-------------------------------------------------------------------------------------------------------------------------------------------------') 274 for row in rows: 275 print((row['dbname']+'.'+row['tbname']).ljust(50), 276 '|',str(row['ALL_rows']).rjust(15), 277 '|',(str(int(row['per']))+'%').rjust(10), 278 '|',str(row['INSERT_rows']).rjust(10)+' , '+(str(int(row['INSERT_rows']*100/row['ALL_rows']))+'%').ljust(5), 279 '|',str(row['UPDATE_rows']).rjust(10)+' , '+(str(int(row['UPDATE_rows']*100/row['ALL_rows']))+'%').ljust(5), 280 '|',str(row['DELETE_rows']).rjust(10)+' , '+(str(int(row['DELETE_rows']*100/row['ALL_rows']))+'%').ljust(5), 281 ) 282 print('\n') 283 284 logging.info('Finished to analyse the binlog file !!!') 285 286 def closeconn(self): 287 self.cur.close() 288 logging.info('release db connections\n') 289 290 def main(): 291 p = queryanalyse() 292 p.rowrecord() 293 p.binlogdesc() 294 p.closeconn() 295 296 if __name__ == "__main__": 297 main()