MySQL8.0.17推出了一个重量级的功能:clone plugin。允许用户可以将当前实例进行本地或者远程的clone。这在某些场景尤其想快速搭建复制备份或者在group replication里加入新成员时非常有用。本文主要试玩下该功能,并试图阐述下其实现的机制是什么。
我们以本地clone为例,因为去除网络部分,理解起来会相对简单点。 也不会过度接触代码部分,仅仅做简单的原理性阐述
示例
本地clone
本地clone无需启动额外mysqld, 只要在实例上执行一条sql语句,指定下目标目录即可:
CLONE LOCAL DATA DIRECTORY [=] 'clone_dir';
root@test 03:49:43>SELECT STAGE, STATE, END_TIME FROM performance_schema.clone_progress;
+-----------+-------------+----------------------------+
| STAGE | STATE | END_TIME |
+-----------+-------------+----------------------------+
| DROP DATA | Completed | 2019-07-26 12:07:12.285611 |
| FILE COPY | Completed | 2019-07-26 12:07:18.270998 |
| PAGE COPY | Completed | 2019-07-26 12:07:18.472560 |
| REDO COPY | Completed | 2019-07-26 12:07:18.673061 |
| FILE SYNC | Completed | 2019-07-26 12:07:32.090219 |
| RESTART | Not Started | NULL |
| RECOVERY | Not Started | NULL |
+-----------+-------------+----------------------------+
7 rows in set (0.00 sec)
需要BACKUP_ADMIN权限
远程clone:
CLONE INSTANCE FROM USER@HOST:PORT
IDENTIFIED BY 'password'
[DATA DIRECTORY [=] 'clone_dir']
[REQUIRE [NO] SSL];
mysql> SET GLOBAL clone_valid_donor_list = 'example.donor.host.com:3306';
mysql> CLONE INSTANCE FROM clone_user@example.donor.host.com:3306
IDENTIFIED BY 'password';
mysql> CLONE INSTANCE FROM user_name@example.donor.host.com:3306
IDENTIFIED BY 'password'
DATA DIRECTORY = '/path/to/clone_dir';
- 需要指定绝对路径,并且路径目录必须不存在
- 在接受机器上启动mysqld,执行上述语句连接到目标机器,就能从目标机器上clone数据到本地,注意如果没有指定data directory的话,就默认配置的目录,已有的文件会被清理掉,并在clone完成后重启
- 两个实例上都需要安装clone plugin
- 必须有相同的字符集设置
官方文档列出的一些限制:
- ddl包括truncate table在clone期间不允许执行 //被block住
- An instance cannot be cloned from a different MySQL server version. The donor and recipient must have the same MySQL server version.
- the X Protocol port specified by mysqlx_port is not supported for remote cloning operations
- The clone plugin does not support cloning of MySQL server configurations
- 不支持clone binlog
- The clone plugin only clones data stored in InnoDB. Other storage engine data is not cloned
- Connecting to the donor MySQL server instance through MySQL Router is not supported.
- Local cloning operations do not support cloning of general tablespaces that were created with an absolute path. A cloned tablespace file with the same path as the source tablespace file would cause a conflict.
主要流程
主要流程包含如下几个过程:
[INIT] ---> [FILE COPY] ---> [PAGE COPY] ---> [REDO COPY] -> [Done]
INIT阶段
需要持有backup lock, 阻止ddl进行
FILE COPY
按照文件进行拷贝,同时开启page tracking功能,记录在拷贝过程中修改的page, 此时会设置buf_pool->track_page_lsn为当前lsn,track_page_lsn在flush page阶段用到:
buf_flush_page:
if (!fsp_is_system_temporary(bpage->id.space()) &&
buf_pool->track_page_lsn != LSN_MAX) {
page_t *frame;
lsn_t frame_lsn;
frame = bpage->zip.data;
if (!frame) {
frame = ((buf_block_t *)bpage)->frame;
}
frame_lsn = mach_read_from_8(frame + FIL_PAGE_LSN); //对于在track_page_lsn之后的page, 如果frame_Lsn大于track_page_lsn, 表示已经记录下page id了,无需重复记录
arch_page_sys->track_page(bpage, buf_pool->track_page_lsn, frame_lsn,
false); // 将page id记录下来,表示在track_page_lsn后修改过的page
}
会创建一个后套线程page_archiver_thread(),将内存记录的page id flush到disk上
PAGE COPY
这里有两个动作
- 开启redo archiving功能,从当前点开始存储新增的redo log,这样从当前点开始所有的增量修改都不会丢失
- 同时上一步在page track的page被发送到目标端。确保当前点之前所做的变更一定发送到目标端
关于redo archiving,实际上这是官方早就存在的功能,主要用于官方的企业级备份工具,但这里clone利用了该特性来维持增量修改产生的redo。 在开始前会做一次checkpoint, 开启一个后台线程log_archiver_thread()来做日志归档。当有新的写入时(notify_about_advanced_write_lsn
)也会通知他去archive
当arch_log_sys处于活跃状态时,他会控制日志写入以避免未归档的日志被覆盖(log_writer_wait_on_archiver
), 注意如果log_writer等待时间过长的话, archive任务会被中断掉
Redo Copy
停止Redo Archiving", 所有归档的日志被发送到目标端,这些日志包含了从page copy阶段开始到现在的所有日志,另外可能还需要记下当前的复制点,例如最后一个事务提交时的binlog位点或者gtid信息,在系统页中可以找到
Done
目标端重启实例,通过crash recovery将redo log应用上去。
本文作者 :zhaiwx_yinfeng