MySQL并行DDL

发布时间 2023-03-27 21:49:09作者: 吃饭端住碗
  随着MySQL版本的不断更新,对DDL操作的支持也在不断的完善和更新:比如从MySQL 5.6引入Online DDL,在MySQL 5.7对Online DDL进一步完善,到现在的8.0版本,则对DDL的实现重新进行了设计,比如DDL操作支持原子特性,在MySQL 8.0.27引入并行DDL。本篇就来探究一下MySQL 8.0.27的并行DDL对于DDL操作速度的提升。
  MySQL 8.0.14引入了innodb_parallel_read_threads变量来控制扫描聚簇索引的并行线程。MySQL 8.0.27引入了innodb_ddl_threads变量来控制用于创建二级索引时的并行线程数量,此参数一般和一并引入的innodb_ddl_buffer_size一起使用,innodb_ddl_buffer_size用于指定进行并行DDL操作时能够使用的buffer大小,buffer是在所有的DDL并行线程中平均分配的,所以一般如果调大innodb_ddl_threads变量时,也需要调大innodb_ddl_buffer_size的大小。
  innodb_ddl_threads、innodb_ddl_buffer_size和innodb_parallel_read_threads的默认大小分别为:
mysql> select @@global.innodb_ddl_threads;
+-----------------------------+
| @@global.innodb_ddl_threads |
+-----------------------------+
|                           4 |
+-----------------------------+
1 row in set (0.00 sec)

mysql> select @@global.innodb_ddl_buffer_size;
+---------------------------------+
| @@global.innodb_ddl_buffer_size |
+---------------------------------+
|                         1048576 |
+---------------------------------+
1 row in set (0.00 sec)

mysql> select @@global.innodb_parallel_read_threads;
+---------------------------------------+
| @@global.innodb_parallel_read_threads |
+---------------------------------------+
|                                     4 |
+---------------------------------------+
1 row in set (0.00 sec)
接下来测试一下调大innodb_ddl_threads、innodb_ddl_buffer_size和innodb_parallel_read_threads参数值对DDL操作的性能提升。
首先创建一张5000万的表:
-- 数据库版本为8.0.28
mysql> select @@version;
+----------+
| @@version|
+----------+
| 8.0.28   |
+----------+
1 row in set (0.00 sec)

-- buffer pool大小为24G
mysql> select @@global.innodb_buffer_pool_size;
+----------------------------------+
| @@global.innodb_buffer_pool_size |
+----------------------------------+
|                       1073741824 |
+----------------------------------+
1 row in set (0.001 sec)

mysql> create database action;
Query OK, 1 row affected (0.01 sec)

# sysbench /usr/share/sysbench/oltp_read_write.lua --mysql-socket=/data/mysql/data/3306/mysqld.sock  --mysql-user=root --mysql-password='123' --mysql-db=action --tables=1 --table-size=50000000 --report-interval=1 --threads=8 prepare

mysql> select count(*) from action.sbtest1;
+----------+
| count(*) |
+----------+
| 50000000 |
+----------+
1 row in set (21.64 sec)

-- 表空间大小为12G
# ll -h
total 12G
-rw-r-----. 1 mysql mysql 12G Jan 20 17:38 sbtest1.ibd
分别测试不同的线程数量和缓冲区大小的DDL操作时间,例如:
-- 设置并发DDL线程为1
mysql> set innodb_ddl_threads = 1;
Query OK, 0 rows affected (0.01 sec)

-- 设置buffer大小为512M
mysql> set innodb_ddl_buffer_size = 536870912;
Query OK, 0 rows affected (0.00 sec)

-- 设置并行索引扫描线程为1
mysql> set innodb_parallel_read_threads = 1;
Query OK, 0 rows affected (0.01 sec)

-- 执行DDL操作
mysql> alter table action.sbtest1 add index idx_c(c);
Query OK, 0 rows affected (6 min 54.21 sec)
Records: 0  Duplicates: 0  Warnings: 0

-- 查看DDL的内存最大占用
mysql> select event_name,CURRENT_NUMBER_OF_BYTES_USED/1024/1024 from performance_schema.memory_summary_global_by_event_name where event_name='memory/innodb/ddl';
+-------------------+----------------------------------------+
| event_name        | CURRENT_NUMBER_OF_BYTES_USED/1024/1024 |
+-------------------+----------------------------------------+
| memory/innodb/ddl |                           513.08750916 |
+-------------------+----------------------------------------+
1 row in set (0.00 sec)
通过不断调整相关参数得到以下结果
innodb_ddl_threads
innodb_ddl_buffer_size
innodb_parallel_read_threads
DDL占用最大内存
DDL时间
1
512M
1
513M
6 min 54.21 sec
2
1G
2
1230M
4 min 12.08 sec
4
2G
4
2735M
3 min 43.01 sec
8
4G
8
5791M
3 min 19.63 sec
16
8G
16
5975M
3 min 12.33 sec
32
16G
32
6084M
3 min 11.11 sec
可以看到,随着并发线程的增多和buffer的增加,DDL操作所占用的资源也越多,而DDL操作所花费的时间则越少。不过通过对比资源的消耗和DDL速度的提升比例,最合理的并行线程数量为4-8个,而buffer大小可以根据情况进行调整。