IT博客汇 | MySQL运行时的可观测性

MySQL运行时的可观测性

叶金荣发表于 2023-07-03 05:37:36

感知SQL运行时的状态

1. 说在前面的话

在MySQL里，一条SQL运行时产生多少磁盘I/O，占用多少内存，是否有创建临时表，这些指标如果都能观测到，有助于更快发现SQL瓶颈，扑灭潜在隐患。

从MySQL 5.7版本开始，performance_schema就默认启用了，并且还增加了sys schema，到了8.0版本又进一步得到增强提升，在SQL运行时就能观察到很多有用的信息，实现一定程度的可观测性。

下面举例说明如何进行观测，以及主要观测哪些指标。

2. 安装employees测试库

安装MySQL官方提供的employees测试数据库，戳此链接(https://dev.mysql.com/doc/index-other.html)下载，解压缩后开始安装：

$ mysql -f < employees.sql;

INFO
CREATING DATABASE STRUCTURE
INFO
storage engine: InnoDB
INFO
LOADING departments
INFO
LOADING employees
INFO
LOADING dept_emp
INFO
LOADING dept_manager
INFO
LOADING titles
INFO
LOADING salaries
data_load_time_diff
00:00:37

MySQL还提供了相应的使用文档：https://dev.mysql.com/doc/employee/en/。

本次测试采用GreatSQL 8.0.32-24版本，且运行在MGR环境中：

greatsql> \s
...
Server version:         8.0.32-24 GreatSQL, Release 24, Revision 3714067bc8c
...

greatsql> select MEMBER_ID, MEMBER_ROLE, MEMBER_VERSION from performance_schema.replication_group_members;
+--------------------------------------+-------------+----------------+
| MEMBER_ID                            | MEMBER_ROLE | MEMBER_VERSION |
+--------------------------------------+-------------+----------------+
| 2adec6d2-febb-11ed-baca-d08e7908bcb1 | SECONDARY   | 8.0.32         |
| 2f68fee2-febb-11ed-b51e-d08e7908bcb1 | ARBITRATOR  | 8.0.32         |
| 5e34a5e2-feb6-11ed-b288-d08e7908bcb1 | PRIMARY     | 8.0.32         |
+--------------------------------------+-------------+----------------+

3. 观测SQL运行状态

查看当前连接/会话的连接ID、内部线程ID：

greatsql> select processlist_id, thread_id from performance_schema.threads where processlist_id = connection_id();
+----------------+-----------+
| processlist_id | thread_id |
+----------------+-----------+
|            110 |       207 |
+----------------+-----------+

查询得到当前的连接ID=110，内部线程ID=207。

P.S，由于本文整理过程不是连续的，所以下面看到的 thread_id 值可能会有好几个，每次都不同。

3.1 观测SQL运行时的内存消耗

执行下面的SQL，查询所有员工的薪资总额，按员工号分组，并按薪资总额倒序，取前10条记录：

greatsql> explain select emp_no, sum(salary) as total_salary from salaries group by emp_no order by total_salary desc limit 10\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: salaries
   partitions: NULL
         type: index
possible_keys: PRIMARY
          key: PRIMARY
      key_len: 7
          ref: NULL
         rows: 2838426
     filtered: 100.00
        Extra: Using temporary; Using filesort

看到需要全索引扫描（其实也等同于全表扫描，因为是基于PRIMARY索引），并且还需要生成临时表，以及额外的filesort。

在正式运行该SQL之前，在另外的窗口中新建一个连接会话，执行下面的SQL先观察该连接/会话当前的内存分配情况：

greatsql> select * from sys.x$memory_by_thread_by_current_bytes where thread_id = 207\G
*************************** 1. row ***************************
         thread_id: 207
              user: root@localhost
current_count_used: 9
 current_allocated: 26266
 current_avg_alloc: 2918.4444
 current_max_alloc: 16464
   total_allocated: 30311

等到该SQL执行完了，再一次查询内存分配情况：

greatsql> select * from sys.x$memory_by_thread_by_current_bytes where thread_id = 207\G
*************************** 1. row ***************************
         thread_id: 207
              user: root@localhost
current_count_used: 13
 current_allocated: 24430
 current_avg_alloc: 1879.2308
 current_max_alloc: 16456
   total_allocated: 95719

我们注意到几个数据的变化情况，用下面表格来展示：

指标	运行前	运行后
total_allocated	30311	95719

也就是说，SQL运行时，需要分配的内存是：95719 – 30311 = 65408 字节。

3.2 观测SQL运行时的其他开销

通过观察 performance_schema.status_by_thread 表，可以知道相应连接/会话中SQL运行的一些状态指标。在SQL运行结束后，执行下面的SQL命令即可查看：

greatsql> select * from performance_schema.status_by_thread where thread_id = 207;
...
|       207 | Created_tmp_disk_tables             | 0                        |
|       207 | Created_tmp_tables                  | 0                        |
...
|       207 | Handler_read_first                  | 1                        |
|       207 | Handler_read_key                    | 1                        |
|       207 | Handler_read_last                   | 0                        |
|       207 | Handler_read_next                   | 2844047                  |
|       207 | Handler_read_prev                   | 0                        |
|       207 | Handler_read_rnd                    | 0                        |
|       207 | Handler_read_rnd_next               | 0                        |
|       207 | Handler_rollback                    | 0                        |
|       207 | Handler_savepoint                   | 0                        |
|       207 | Handler_savepoint_rollback          | 0                        |
|       207 | Handler_update                      | 0                        |
|       207 | Handler_write                       | 0                        |
|       207 | Last_query_cost                     | 286802.914893            |
|       207 | Last_query_partial_plans            | 1                        |
...
|       207 | Select_full_join                    | 0                        |
|       207 | Select_full_range_join              | 0                        |
|       207 | Select_range                        | 0                        |
|       207 | Select_range_check                  | 0                        |
|       207 | Select_scan                         | 1                        |
|       207 | Slow_launch_threads                 | 0                        |
|       207 | Slow_queries                        | 1                        |
|       207 | Sort_merge_passes                   | 0                        |
|       207 | Sort_range                          | 0                        |
|       207 | Sort_rows                           | 1                       |
|       207 | Sort_scan                           | 1                        |
...

上面我们只罗列了部分比较重要的状态指标。从这个结果也可以佐证slow query log中的结果，确实没创建临时表。

作为参照，查看这条SQL对应的slow query log记录：

# Query_time: 0.585593  Lock_time: 0.000002 Rows_sent: 10  Rows_examined: 2844057 Thread_id: 110 Errno: 0 Killed: 0 Bytes_received: 115 Bytes_sent: 313 Read_first: 1 Read_last: 0 Read_key: 1 Read_next: 2844047 Read_prev: 0 Read_rnd: 0 Read_rnd_next: 0 Sort_merge_passes: 0 Sort_range_count: 0 Sort_rows: 10 Sort_scan_count: 1 Created_tmp_disk_tables: 0 Created_tmp_tables: 0 Start: 2023-07-06T10:06:01.438376+08:00 End: 2023-07-06T10:06:02.023969+08:00 Schema: employees Rows_affected: 0
# Tmp_tables: 0  Tmp_disk_tables: 0  Tmp_table_sizes: 0
# InnoDB_trx_id: 0
# Full_scan: Yes  Full_join: No  Tmp_table: No  Tmp_table_on_disk: No
# Filesort: Yes  Filesort_on_disk: No  Merge_passes: 0
#   InnoDB_IO_r_ops: 0  InnoDB_IO_r_bytes: 0  InnoDB_IO_r_wait: 0.000000
#   InnoDB_rec_lock_wait: 0.000000  InnoDB_queue_wait: 0.000000
#   InnoDB_pages_distinct: 4281
use employees;
SET timestamp=1688609161;
select emp_no, sum(salary) as total_salary from salaries group by emp_no order by total_salary desc limit 10;

可以看到，Created_tmp_disk_tables, Created_tmp_tables, Handler_read_next, Select_full_join, Select_scan, Sort_rows, Sort_scan, 等几个指标的数值是一样的。

还可以查看该SQL运行时的I/O latency情况，SQL运行前后两次查询对比：

greatsql> select * from sys.io_by_thread_by_latency where thread_id = 207;
+----------------+-------+---------------+-------------+-------------+-------------+-----------+----------------+
| user           | total | total_latency | min_latency | avg_latency | max_latency | thread_id | processlist_id |
+----------------+-------+---------------+-------------+-------------+-------------+-----------+----------------+
| root@localhost |     7 | 75.39 us      | 5.84 us     | 10.77 us    | 22.12 us    |       207 |            110 |
+----------------+-------+---------------+-------------+-------------+-------------+-----------+----------------+

...

greatsql> select * from sys.io_by_thread_by_latency where thread_id = 207;
+----------------+-------+---------------+-------------+-------------+-------------+-----------+----------------+
| user           | total | total_latency | min_latency | avg_latency | max_latency | thread_id | processlist_id |
+----------------+-------+---------------+-------------+-------------+-------------+-----------+----------------+
| root@localhost |     8 | 85.29 us      | 5.84 us     | 10.66 us    | 22.12 us    |       207 |            110 |
+----------------+-------+---------------+-------------+-------------+-------------+-----------+----------------+

可以看到这个SQL运行时的I/O latency是：85.29 – 75.39 = 9.9us。

3.3 观测SQL运行进度

我们知道，运行完一条SQL后，可以利用PROFLING功能查看它各个阶段的耗时，但是在运行时如果也想查看各阶段耗时该怎么办呢？

从MySQL 5.7版本开始，可以通过 performance_schema.events_stages_% 相关表查看SQL运行过程以及各阶段耗时，需要先修改相关设置：

# 确认是否对所有主机&用户都启用
greatsql> SELECT * FROM performance_schema.setup_actors;
+------+------+------+---------+---------+
| HOST | USER | ROLE | ENABLED | HISTORY |
+------+------+------+---------+---------+
| %    | %    | %    | NO      | NO      |
+------+------+------+---------+---------+

# 修改成对所有主机&用户都启用
greatsql> UPDATE performance_schema.setup_actors
 SET ENABLED = 'YES', HISTORY = 'YES'
 WHERE HOST = '%' AND USER = '%';

# 修改 setup_instruments & setup_consumers 设置
greatsql> UPDATE performance_schema.setup_consumers
 SET ENABLED = 'YES'
 WHERE NAME LIKE '%events_statements_%';

greatsql> UPDATE performance_schema.setup_consumers
 SET ENABLED = 'YES'
 WHERE NAME LIKE '%events_stages_%';

这就实时可以观测SQL运行过程中的状态了。

在SQL运行过程中，从另外的窗口查看该SQL对应的 EVENT_ID：

greatsql> SELECT EVENT_ID, TRUNCATE(TIMER_WAIT/1000000000000,6) as Duration, SQL_TEXT        FROM performance_schema.events_statements_history WHERE thread_id = 85 order by event_id desc limit 5;
+----------+----------+-------------------------------------------------------------------------------------------------------------------------------+
| EVENT_ID | Duration | SQL_TEXT                                                                                                                      |
+----------+----------+-------------------------------------------------------------------------------------------------------------------------------+
|   149845 |   0.6420 | select emp_no, sum(salary) as total_salary, sleep(0.000001) from salaries group by emp_no order by total_salary desc limit 10 |
|   149803 |   0.6316 | select emp_no, sum(salary) as total_salary, sleep(0.000001) from salaries group by emp_no order by total_salary desc limit 10 |
|   149782 |   0.6245 | select emp_no, sum(salary) as total_salary, sleep(0.000001) from salaries group by emp_no order by total_salary desc limit 10 |
|   149761 |   0.6361 | select emp_no, sum(salary) as total_salary, sleep(0.000001) from salaries group by emp_no order by total_salary desc limit 10 |
|   149740 |   0.6245 | select emp_no, sum(salary) as total_salary, sleep(0.000001) from salaries group by emp_no order by total_salary desc limit 10 |
+----------+----------+-------------------------------------------------------------------------------------------------------------------------------+

# 再根据 EVENT_ID 值去查询 events_stages_history_long
greatsql> SELECT thread_id ,event_Id, event_name AS Stage, TRUNCATE(TIMER_WAIT/1000000000000,6) AS Duration  FROM performance_schema.events_stages_history_long WHERE NESTING_EVENT_ID = 149845 order by event_id;
+-----------+----------+------------------------------------------------+----------+
| thread_id | event_Id | Stage                                          | Duration |
+-----------+----------+------------------------------------------------+----------+
|        85 |   149846 | stage/sql/starting                             |   0.0000 |
|        85 |   149847 | stage/sql/Executing hook on transaction begin. |   0.0000 |
|        85 |   149848 | stage/sql/starting                             |   0.0000 |
|        85 |   149849 | stage/sql/checking permissions                 |   0.0000 |
|        85 |   149850 | stage/sql/Opening tables                       |   0.0000 |
|        85 |   149851 | stage/sql/init                                 |   0.0000 |
|        85 |   149852 | stage/sql/System lock                          |   0.0000 |
|        85 |   149854 | stage/sql/optimizing                           |   0.0000 |
|        85 |   149855 | stage/sql/statistics                           |   0.0000 |
|        85 |   149856 | stage/sql/preparing                            |   0.0000 |
|        85 |   149857 | stage/sql/Creating tmp table                   |   0.0000 |
|        85 |   149858 | stage/sql/executing                            |   0.6257 |
|        85 |   149859 | stage/sql/end                                  |   0.0000 |
|        85 |   149860 | stage/sql/query end                            |   0.0000 |
|        85 |   149861 | stage/sql/waiting for handler commit           |   0.0000 |
|        85 |   149862 | stage/sql/closing tables                       |   0.0000 |
|        85 |   149863 | stage/sql/freeing items                        |   0.0000 |
|        85 |   149864 | stage/sql/logging slow query                   |   0.0000 |
|        85 |   149865 | stage/sql/cleaning up                          |   0.0000 |
+-----------+----------+------------------------------------------------+----------+

上面就是这条SQL的运行进度展示，以及各个阶段的耗时，和PROFILING的输出一样，当我们了解一条SQL运行所需要经历的各个阶段时，从上面的输出结果中也就能估算出该SQL大概还要多久能跑完，决定是否要提前kill它。

如果想要观察DDL SQL的运行进度，可以参考这篇文章：不用MariaDB/Percona也能查看DDL的进度。

更多的观测指标、维度还有待继续挖掘，以后有机会再写。

延伸阅读
– Query Profiling Using Performance Schema, https://dev.mysql.com/doc/refman/8.0/en/performance-schema-query-profiling.html
– 不用MariaDB/Percona也能查看DDL的进度
– 事件记录 | performance_schema全方位介绍
– 内存分配统计视图 | 全方位认识 sys 系统库

全文完：）