mydumper,myloader原理及实战

mydumper 特性

(1)多线程备份（和mysqlpump的多线程不同，mysqlpump多线程备份的粒度是表，mydumper多线程备份的粒度是行，这对于备份大表特别有用）
(2)因为是多线程逻辑备份，备份后会生成多个备份文件
(3)备份时对 MyISAM 表施加 FTWRL (FLUSH TABLES WITH READ LOCK), 会阻塞 DML 语句
(4)保证备份数据的一致性
(5)支持文件压缩
(6)支持导出binlog
(7)支持多线程恢复
(8)支持以守护进程模式工作，定时快照和连续二进制日志
(9)支持将备份文件切块

下载安装包
mydumper-0.9.1.tar.gz

安装依赖包：
yum install glib2-devel mysql-devel zlib-devel pcre-devel openssl-devel cmake

tar zxvf mydumper-0.9.1.tar.gz
cd mydumper-0.9.1
cmake .
make
make install

mydumper主要流程概括：

0、连接目标数据库
1、通过show processlist来判断是否有长查询，根据参数long-query-guard和kill-long-queries决定退出或杀掉长查询
2、主线程，针对MyISAM引擎表，FLUSH TABLES WITH READ LOCK, 施加全局只读锁，以阻止DML语句写入，保证数据的一致性
3、主线程，针对InnoDB引擎表，开启事务，start transaction，读取当前时间点的二进制日志文件名和日志写入的位置并记录在metadata文件中，以供即使点恢复使用
4、创建worker子线程，N个（线程数可以指定，默认是4）dump线程 START TRANSACTION WITH CONSISTENT SNAPSHOT; 开启读一致的事务
5、确定候选表，根据类别分别插入innodb_table，non_innodb_table以及table_schemas链表(表结构)
6、将候选表通过g_async_queue_push加入任务队列(队列最后元素是thread shutdown)，由worker子线程从队列中读取表信息并执行数据导出
7、dump non-InnoDB tables, 首先导出非事务引擎的表
8、主线程 UNLOCK TABLES 非 事务引擎备份完后，释放全局只读锁
9、dump InnoDB tables, 基于 事务导出InnoDB表
10、事务结束
11、等待worker退出

mydumper使用--less-locking可以减少锁等待时间，此时mydumper的执行机制大致为
1、主线程 FLUSH TABLES WITH READ LOCK （全局锁）
2、Dump线程 START TRANSACTION WITH CONSISTENT SNAPSHOT;
3、LL Dump线程 LOCK TABLES non-InnoDB （线程内部锁）
4、主线程UNLOCK TABLES
5、LL Dump线程 dump non-InnoDB tables
6、LL DUmp线程 UNLOCK non-InnoDB
7、Dump线程 dump InnoDB tables

备份文件相关信息：

1、所有的备份文件在一个目录中，未指定时为当前目录，且自动生成备份日志时间文件夹，如export-20150703-145806
2、如果是在从库进行备份，还会记录备份时同步至主库的二进制日志文件及写入位置
3、每个表有两个备份文件：database.table-schema.sql表结构文件，database.table.sql表数据文件
4、如果对表文件分片，将生成多个备份数据文件，可以指定行数或指定大小分片

# mydumper --help
Usage:
  mydumper [OPTION...] multi-threaded MySQL dumping

Help Options:
  -?, --help                  Show help options

Application Options:
  -B, --database              Database to dump
  -T, --tables-list           Comma delimited table list to dump (does not exclude regex option)
  -o, --outputdir             Directory to output files to
  -s, --statement-size        Attempted size of INSERT statement in bytes, default 1000000
  -r, --rows                  Try to split tables into chunks of this many rows. This option turns off --chunk-filesize
  -F, --chunk-filesize        Split tables into chunks of this output file size. This value is in MB
  -c, --compress              Compress output files
  -e, --build-empty-files     Build dump files even if no data available from table
  -x, --regex                 Regular expression for 'db.table' matching
  -i, --ignore-engines        Comma delimited list of storage engines to ignore
  -m, --no-schemas            Do not dump table schemas with the data
  -d, --no-data               Do not dump table data
  -G, --triggers              Dump triggers
  -E, --events                Dump events
  -R, --routines              Dump stored procedures and functions
  -k, --no-locks              Do not execute the temporary shared read lock.  WARNING: This will cause inconsistent backups
  --less-locking              Minimize locking time on InnoDB tables.
  -l, --long-query-guard      Set long query timer in seconds, default 60
  -K, --kill-long-queries     Kill long running queries (instead of aborting)
  -D, --daemon                Enable daemon mode
  -I, --snapshot-interval     Interval between each dump snapshot (in minutes), requires --daemon, default 60
  -L, --logfile               Log file name to use, by default stdout is used
  --tz-utc                    SET TIME_ZONE='+00:00' at top of dump to allow dumping of TIMESTAMP data when a server has data in different time zones or data is being moved between servers with different time zones, defaults to on use --skip-tz-utc to disable.
  --skip-tz-utc              
  --use-savepoints            Use savepoints to reduce metadata locking issues, needs SUPER privilege
  --success-on-1146           Not increment error count and Warning instead of Critical in case of table doesn't exist
  --lock-all-tables           Use LOCK TABLE for all, instead of FTWRL
  -U, --updated-since         Use Update_time to dump only tables updated in the last U days
  --trx-consistency-only      Transactional consistency only
  -h, --host                  The host to connect to
  -u, --user                  Username with privileges to run the dump
  -p, --password              User password
  -P, --port                  TCP/IP port to connect to
  -S, --socket                UNIX domain socket file to use for connection
  -t, --threads               Number of threads to use, default 4
  -C, --compress-protocol     Use compression on the MySQL connection
  -V, --version               Show the program version and exit
  -v, --verbose               Verbosity of output, 0 = silent, 1 = errors, 2 = warnings, 3 = info, default 2

 

# myloader --help
Usage:
  myloader [OPTION...] multi-threaded MySQL loader

Help Options:
  -?, --help                        Show help options

Application Options:
  -d, --directory                   Directory of the dump to import
  -q, --queries-per-transaction     Number of queries per transaction, default 1000
  -o, --overwrite-tables            Drop tables if they already exist
  -B, --database                    An alternative database to restore into
  -s, --source-db                   Database to restore
  -e, --enable-binlog               Enable binary logging of the restore data
  -h, --host                        The host to connect to
  -u, --user                        Username with privileges to run the dump
  -p, --password                    User password
  -P, --port                        TCP/IP port to connect to
  -S, --socket                      UNIX domain socket file to use for connection
  -t, --threads                     Number of threads to use, default 4
  -C, --compress-protocol           Use compression on the MySQL connection
  -V, --version                     Show the program version and exit
  -v, --verbose                     Verbosity of output, 0 = silent, 1 = errors, 2 = warnings, 3 = info, default 2

 

备份 test 库到 /tmp/backup/01 文件夹中，并压缩备份文件 
mydumper -u dba_user -p msds007 -h 192.168.1.101 -P 3306 -B mytest -c -o /tmp/backup/01 

备份所有数据库，备份至 /tmp/backup/02 文件夹 
mydumper -u root -p msds007 -h 192.168.1.101 -P 3306 -o /tmp/backup/02 
 
备份 test.wx_edu_homework 表，且不备份表结构，备份至 /tmp/backup/03 文件夹 
mydumper -u root -p msds007 -h 127.0.0.1 -P 3306 -B test -T wx_edu_homework -m -o /tmp/backup/03

恢复test库，如果恢复的表存在则先删除 
myloader -u dba_user -p msds007 -h 192.168.1.101 -P 3306 -e -o -d /tmp/backup/01
-e参数会把恢复数据时的SQL写进binlog

 

打开general log执行

mydumper -u root -p msds007 -h 192.168.1.101 -P 3306 -o /tmp/backup/02 

myloader -u dba_user -p msds007 -h 192.168.1.101 -P 3306 -e -o -d /tmp/backup/01

看general log的详细信息 

 

原理图

mydumper和mysqldump原理类似，最大的区别在于多线程复制。并发的粒度可以到行级别，还是通过线程分别调用FTWRL并获取一致性点位。