我只是进行了问题分析和文档整理
一、案例描述
MGR在遇到表不存在的情况下，节点没有退出节点而是爆出一个警告，并且节点状态也正常，警告如下：
2019-10-17T21:16:11.564211+08:00 10 [Warning] Slave SQL for channel
group_replication_applier': Worker 1 failed executing transaction 'aaaaaaaa-aaaa-aaaa-aaaa-aaaaaaaaaaaa:8' at master log , end_log_pos 220;
Error executing row event: 'Table 'test.a_1' doesn't exist', Error_code: 1146
集群状态如下：
[[email protected]][test]>select * from performance_schema.replication_group_members;
+---------------------------+--------------------------------------+-------------+-------------+--------------+
| CHANNEL_NAME | MEMBER_ID | MEMBER_HOST | MEMBER_PORT | MEMBER_STATE |
+---------------------------+--------------------------------------+-------------+-------------+--------------+
| group_replication_applier | 9fd479bb-f0d8-11e9-9381-000c29105312 | mysql_1 | 3306 | ONLINE |
| group_replication_applier | a8833a96-f0d8-11e9-a9f4-000c291fd9a5 | mysql_2 | 3306 | ONLINE |
| group_replication_applier | b2968fe2-f0d8-11e9-a8ff-000c29c89e42 | mysql_3 | 3306 | ONLINE |
+---------------------------+--------------------------------------+-------------+-------------+--------------+
3 rows in set (0.00 sec)
当时觉得很奇怪，我们知道这种错误即便是在主从情况下也是报错的SQL线程退出的，MGR居然还能在线，这种情况数据已经不同步了，应该报错并且剔除节点才对。
二、问题分析
随即一些感兴趣的同学马上进行了测试，测试结果和上面不一致，测试结果是报错而不是出警告如下：
2019-10-17T09:16:34.317542Z 84 [ERROR] Slave SQL for channel
'group_replication_applier': Error executing row event:
'Table 'test.emp1' doesn't exist', Error_code: 1146
并且这种情况表不存在的节点已经被剔除掉了。下面是正常情况的节点状态：
secondary 1节点：
[[email protected]][test]>select * from performance_schema.replication_group_members;
+---------------------------+--------------------------------------+-------------+-------------+--------------+
| CHANNEL_NAME | MEMBER_ID | MEMBER_HOST | MEMBER_PORT | MEMBER_STATE |
+---------------------------+--------------------------------------+-------------+-------------+--------------+
| group_replication_applier | a8833a96-f0d8-11e9-a9f4-000c291fd9a5 | mysql_2 | 3306 | ERROR |
+---------------------------+--------------------------------------+-------------+-------------+--------------+
1 row in set (0.00 sec)
secondary 2节点：
[[email protected]][test]>select * from performance_schema.replication_group_members;
+---------------------------+--------------------------------------+-------------+-------------+--------------+
| CHANNEL_NAME | MEMBER_ID | MEMBER_HOST | MEMBER_PORT | MEMBER_STATE |
+---------------------------+--------------------------------------+-------------+-------------+--------------+
| group_replication_applier | b2968fe2-f0d8-11e9-a8ff-000c29c89e42 | mysql_3 | 3306 | ERROR |
+---------------------------+--------------------------------------+-------------+-------------+--------------+
1 row in set (0.00 sec)
那么疑问就是为什么同样是MGR一个是警告一个是错误呢，并且前者还能处于正常同步状态。不错看到题目就知道这里和slave_skip_errors参数有关。
三、测试模拟
我们知道再Master-Slave中如果遇到从库表不存在肯定是报错的，除非设置slave_skip_errors参数，当然我在线上重来没有设置过这个参数，并且通过这个案例我们发现本参数对MGR也有影响，如下测试方法：
我们在3个节点都开启slave-skip-errors= ddl_exist_errors
如下图：
image.png
然后搭建3节点single-primary模式的MGR集群。
image.png
集群搭建正常。
然后执行如下操作：
[[email protected]][(none)]>set sql_log_bin=0;
Query OK, 0 rows affected (0.00 sec)
[[email protected]][(none)]>create table test.a_1(id bigint auto_increment primary key,name varchar(20));
Query OK, 0 rows affected (0.01 sec)
[[email protected]][(none)]>set sql_log_bin=1;
Query OK, 0 rows affected (0.00 sec)
此时primary节点是有a_1表的，但是因为binlog关闭的原因，两个secondary节点是不存在a_1表的。
然后我们插入数据：
[[email protected]][test]>insert into test.a_1 values(null,'tom');
Query OK, 1 row affected (0.02 sec)
此时，primary节点因为存在a_1表，所以能够插入，但是两个secondary节点不存在a_1表，所以插入是失败的。数据产生不一致。正常情况下这种数据不一致会导致2个secondary节点被提出集群才对。但是实际上3个节点都是正常的，集群并没有失效。
[[email protected]][test]>select * from test.a_1;
+----+------+
| id | name |
+----+------+
| 1 | tom |
+----+------+
1 row in set (0.00 sec)
[[email protected]][test]>select * from performance_schema.replication_group_members;
+---------------------------+--------------------------------------+-------------+-------------+--------------+
| CHANNEL_NAME | MEMBER_ID | MEMBER_HOST | MEMBER_PORT | MEMBER_STATE |
+---------------------------+--------------------------------------+-------------+-------------+--------------+
| group_replication_applier | 9fd479bb-f0d8-11e9-9381-000c29105312 | mysql_1 | 3306 | ONLINE |
| group_replication_applier | a8833a96-f0d8-11e9-a9f4-000c291fd9a5 | mysql_2 | 3306 | ONLINE |
| group_replication_applier | b2968fe2-f0d8-11e9-a8ff-000c29c89e42 | mysql_3 | 3306 | ONLINE |
+---------------------------+--------------------------------------+-------------+-------------+--------------+
3 rows in set (0.00 sec)
此时去2个secondary节点读取test.a_1表，表是不存在的。
secondary 1：
[[email protected]][test]>select * from test.a_1;
ERROR 1146 (42S02): Table 'test.a_1' doesn't exist
[[email protected]][test]>
secondary 2：
[[email protected]][test]>select * from test.a_1;
ERROR 1146 (42S02): Table 'test.a_1' doesn't exist
error log输出信息：（set global log_error_verbosity = 3;)
2019-10-17T21:16:11.564211+08:00 10 [Warning] Slave SQL for channel
'group_replication_applier': Worker 1 failed executing transaction 'aaaaaaaa-aaaa-aaaa-aaaa-aaaaaaaaaaaa:8' at master log ,
end_log_pos 220; Error executing row event: 'Table 'test.a_1' doesn't exist', Error_code: 1146
四、slave_skip_errors源码生效点
这个设置在Rows_log_event::do_apply_event 函数中生效，也就是DML Event开始应用的时候生效，这是常规的SQL线程（或者Worker线程）调用的。
#ifdef HAVE_REPLICATION
if (opt_slave_skip_errors)
add_slave_skip_errors(opt_slave_skip_errors);
#endif
if (open_and_lock_tables(thd, rli->tables_to_lock, 0))//打开表
{
uint actual_error= thd->get_stmt_da()->mysql_errno();
if (thd->is_slave_error || thd->is_fatal_error)
{
if (ignored_error_code(actual_error)) //这里受到 slave_skip_errors 参数控制 ignored_error_code会将slave_skip_errors的参数设置读取出来
{
if (log_warnings > 1)
rli->report(WARNING_LEVEL, actual_error,
"Error executing row event: '%s'",
(actual_error ? thd->get_stmt_da()->message_text() :
"unexpected success or fatal error"));
thd->get_stmt_da()->reset_condition_info(thd);
clear_all_errors(thd, const_cast(rli));
error= 0;
goto end;
}
else
{
rli->report(ERROR_LEVEL, actual_error,
"Error executing row event: '%s'",
(actual_error ? thd->get_stmt_da()->message_text() :
"unexpected success or fatal error"));
thd->is_slave_error= 1;
const_cast(rli)->slave_close_thread_tables(thd);
DBUG_RETURN(actual_error);
}
}
可以看到MGR的执行逻辑受到了该参数的影响。