MySQL Performance-Schema(三) 实践篇

     前一篇文章我们分析了Performance-Schema中每个表的用途，以及主要字段的含义，比较侧重于理论的介绍。这篇文章我主要从DBA的角度出发，详细介绍如何通过Performance-Schema得到DBA关心的数据，比如哪个SQL执行次数最多，哪个表访问最频繁，哪个锁最热等信息。通过充分利用Performance-Schema表的数据，让DBA更了解DB的运行状态，也更有助于排查定位问题。本文主要从两方面展开讨论，第一方面是统计信息维度，包括SQL维度，对象维度和等待事件维度三小点；第二方面是定位分析问题维度，结合实际应用场景，利用Performance-Schema搜集的数据进行分析。

统计信息(SQL维度)
     关于SQL维度的统计信息主要集中在events_statements_summary_by_digest表中，通过将SQL语句抽象出digest，可以统计某类SQL语句在各个维度的统计信息(比如：执行次数，排序次数，使用临时表等)
(1).哪类SQL执行最多？

SELECT
DIGEST_TEXT,
COUNT_STAR,
FIRST_SEEN,
LAST_SEEN
FROM events_statements_summary_by_digest
ORDER BY COUNT_STAR DESC limit 1\G
*************************** 1. row ***************************
DIGEST_TEXT: SELECT `pay_order_id` , `total_fee` , `pay_seller_id` , `pay_buyer_id` , `buyer_id` , `seller_id` FROM `pay_order` WHERE `pay_order_id` = ? 
COUNT_STAR: 5282893
FIRST_SEEN: 2015-12-15 18:25:38
LAST_SEEN: 2015-12-18 16:20:59

可以看到执行次数最多的SQL是“SELECT `pay_order_id` , `total_fee` , `pay_seller_id` , pay_buyer_id , buyer_id , seller_id FROM tc_pay_order WHERE pay_order_id = ?”，FIRST_SEEN和LAST_SEEN分别显示了语句第一次执行和最后一次执行的时间点。
(2).哪类SQL的平均响应时间最多？
AVG_TIMER_WAIT
(3).哪类SQL排序记录数最多？
SUM_SORT_ROWS
(4).哪类SQL扫描记录数最多？
SUM_ROWS_EXAMINED
(5).哪类SQL使用临时表最多？
SUM_CREATED_TMP_TABLES,SUM_CREATED_TMP_DISK_TABLES
(6).哪类SQL返回结果集最多？
SUM_ROWS_SENT
     通过上述指标我们可以间接获得某类SQL的逻辑IO(SUM_ROWS_EXAMINED)，CPU消耗(SUM_SORT_ROWS)，网络带宽(SUM_ROWS_SENT)的对比，但还无法得到某类SQL的物理IO消耗，以及某类SQL访问数据的buffer命中率。

统计信息(对象维度)
(1).哪个表物理IO最多？

SELECT
file_name,
event_name,
SUM_NUMBER_OF_BYTES_READ,
SUM_NUMBER_OF_BYTES_WRITE
FROM file_summary_by_instance
ORDER BY SUM_NUMBER_OF_BYTES_READ + SUM_NUMBER_OF_BYTES_WRITE DESC LIMIT 1\G
*************************** 1. row ***************************
file_name: /u01/my3306/data/chuck/test_01.ibd
event_name: wait/io/file/innodb/innodb_data_file
SUM_NUMBER_OF_BYTES_READ: 2168078336
SUM_NUMBER_OF_BYTES_WRITE: 1390761934848

通过file_summary_by_instance表，可以获得系统运行到现在，哪个文件(表)物理IO最多，这可能意味着这个表经常需要访问磁盘IO，从结果来看chuck库里面的test_01的数据文件访问最多。

(2).哪个表逻辑IO最多?

SELECT
object_name,
COUNT_READ,
COUNT_WRITE,
COUNT_FETCH,
SUM_TIMER_WAIT
FROM table_io_waits_summary_by_table
ORDER BY sum_timer_wait DESC limit 1\G
*************************** 1. row ***************************
object_name: test_slow
COUNT_READ: 1986130
COUNT_WRITE: 393222
COUNT_FETCH: 1986130
SUM_TIMER_WAIT: 925309746633072

通过table_io_waits_summary_by_table表，可以获得系统运行到现在，哪个表逻辑IO最多，亦即最“热”的表，从结果来看chuck库里面的test_slow表访问次数最多。

(3).哪个索引访问最多？

SELECT
OBJECT_NAME,
INDEX_NAME,
COUNT_FETCH,
COUNT_INSERT,
COUNT_UPDATE,
COUNT_DELETE
FROM table_io_waits_summary_by_index_usage
ORDER BY SUM_TIMER_WAIT DESC limit 1\G
*************************** 1. row ***************************
OBJECT_NAME: test_slow
INDEX_NAME: PRIMARY
COUNT_FETCH: 393247
COUNT_INSERT: 0
COUNT_UPDATE: 393222
COUNT_DELETE: 0

通过table_io_waits_summary_by_index_usage表，可以获得系统运行到现在，哪个表的具体哪个索引(包括主键索引，二级索引)使用最多，从结果来看，我们知道test_slow表访问最多，并且都是通过主键访问。

(4).哪个索引从来没有使用过？

SELECT
OBJECT_SCHEMA,
OBJECT_NAME,
INDEX_NAME
FROM table_io_waits_summary_by_index_usage
WHERE INDEX_NAME IS NOT NULL
AND COUNT_STAR = 0
AND OBJECT_SCHEMA <> 'mysql'
ORDER BY OBJECT_SCHEMA,OBJECT_NAME;
*************************** 1. row ***************************
OBJECT_SCHEMA: chuck
OBJECT_NAME: test_icp
INDEX_NAME: idx_y_z

通过table_io_waits_summary_by_index_usage表，我们还可以获得系统运行到现在，哪些索引从来没有被用过。由于索引也会占用大量的空间，我们可以利用这个统计信息，结合一定的时间策略将无用的索引删除。上面的结果显示，chuck库test_icp表的idx_y_z从来没有被使用过。

统计信息(等待事件维度)
(1).哪个等待事件消耗的时间最多？

SELECT
EVENT_NAME,
COUNT_STAR,
SUM_TIMER_WAIT,
AVG_TIMER_WAIT
FROM events_waits_summary_global_by_event_name
WHERE event_name != 'idle'
ORDER BY SUM_TIMER_WAIT DESC LIMIT 1;
*************************** 1. row ***************************
EVENT_NAME: wait/synch/cond/threadpool/worker_cond
COUNT_STAR: 26369820
SUM_TIMER_WAIT: 1604134685750586132
AVG_TIMER_WAIT: 60832219664

通过events_waits_summary_global_by_event_name表，可以获取到系统运行到现在，消耗时间最多的事件，当然还可以根据其它维度排序，比如平均等待时间，从结果来看wait/synch/cond/threadpool/worker_cond这个事件消耗的累计时间最长。

具体场景分析
      前面我们讨论的基本都是汇总信息，有点类似与ORACLE-AWR(Automatic Workload Repository)的味道，那么我们如何利用peformance schema来定位问题呢？或者对于的发生的问题，比如抖动，我们是否有办法知道当时发生了什么？
(1).剖析某条SQL的执行情况，包括statement信息，stage信息和wait信息。
     有时候我们需要分析具体某条SQL，该SQL在执行各个阶段的时间消耗，通过events_statements_xxx表和events_stages_xxx表，就可以达到目的。两个表通过event_id与nesting_event_id关联，stages表的nesting_event_id为对应statements表的event_id。比如分析包含count(*)的某条SQL语句，具体如下：

SELECT
EVENT_ID,
sql_text
FROM events_statements_history
WHERE sql_text LIKE '%count(*)%';
+----------+--------------------------------------+
| EVENT_ID | sql_text |
+----------+--------------------------------------+
| 1690 | select count(*) from chuck.test_slow |
+----------+--------------------------------------+

首先得到了语句的event_id为1690，通过查找events_stages_xxx中nesting_event_id为1690的记录，可以达到目的。
a.查看每个阶段的时间消耗

SELECT
event_id,
EVENT_NAME,
SOURCE,
TIMER_END - TIMER_START
FROM events_stages_history_long
WHERE NESTING_EVENT_ID = 1690;
+----------+--------------------------------+----------------------+-----------------------+
| event_id | EVENT_NAME | SOURCE | TIMER_END-TIMER_START |
+----------+--------------------------------+----------------------+-----------------------+
| 1691 | stage/sql/init | mysqld.cc:990 | 316945000 |
| 1693 | stage/sql/checking permissions | sql_parse.cc:5776 | 26774000 |
| 1695 | stage/sql/Opening tables | sql_base.cc:4970 | 41436934000 |
| 2638 | stage/sql/init | sql_select.cc:1050 | 85757000 |
| 2639 | stage/sql/System lock | lock.cc:303 | 40017000 |
| 2643 | stage/sql/optimizing | sql_optimizer.cc:138 | 38562000 |
| 2644 | stage/sql/statistics | sql_optimizer.cc:362 | 52845000 |
| 2645 | stage/sql/preparing | sql_optimizer.cc:485 | 53196000 |
| 2646 | stage/sql/executing | sql_executor.cc:112 | 3153000 |
| 2647 | stage/sql/Sending data | sql_executor.cc:192 | 7369072089000 |
| 4304138 | stage/sql/end | sql_select.cc:1105 | 19920000 |
| 4304139 | stage/sql/query end | sql_parse.cc:5463 | 44721000 |
| 4304145 | stage/sql/closing tables | sql_parse.cc:5524 | 61723000 |
| 4304152 | stage/sql/freeing items | sql_parse.cc:6838 | 455678000 |
| 4304155 | stage/sql/logging slow query | sql_parse.cc:2258 | 83348000 |
| 4304159 | stage/sql/cleaning up | sql_parse.cc:2163 | 4433000 |
+----------+--------------------------------+----------------------+-----------------------+

通过间接关联，我们能分析得到SQL语句在每个阶段的时间消耗，时间单位以皮秒表示。这里展示的结果很类似profiling功能，有了performance schema，就不再需要profiling这个功能了。另外需要注意的是，由于默认情况下events_stages_history表中只为每个连接记录了最近10条记录，为了确保获取所有记录，需要访问events_stages_history_long表。

b.查看某个阶段的锁等待情况
      针对每个stage可能出现的锁等待，一个stage会对应一个或多个wait，events_waits_history_long这个表容易爆满[默认阀值10000]。由于select count(*)需要IO(逻辑IO或者物理IO)，所以在stage/sql/Sending data阶段会有io等待的统计。通过stage_xxx表的event_id字段与waits_xxx表的nesting_event_id进行关联。

SELECT
event_id,
event_name,
source,
timer_wait,
object_name,
index_name,
operation,
nesting_event_id
FROM events_waits_history_long
WHERE nesting_event_id = 2647;
+----------+---------------------------+-----------------+------------+-------------+------------+-----------+------------------+
| event_id | event_name | source | timer_wait | object_name | index_name | operation | nesting_event_id |
+----------+---------------------------+-----------------+------------+-------------+------------+-----------+------------------+
| 190607 | wait/io/table/sql/handler | handler.cc:2842 | 1845888 | test_slow | idx_c1 | fetch | 2647 |
| 190608 | wait/io/table/sql/handler | handler.cc:2842 | 1955328 | test_slow | idx_c1 | fetch | 2647 |
| 190609 | wait/io/table/sql/handler | handler.cc:2842 | 1929792 | test_slow | idx_c1 | fetch | 2647 | 
| 190610 | wait/io/table/sql/handler | handler.cc:2842 | 1869600 | test_slow | idx_c1 | fetch | 2647 |
| 190611 | wait/io/table/sql/handler | handler.cc:2842 | 1922496 | test_slow | idx_c1 | fetch | 2647 |
+----------+---------------------------+-----------------+------------+-------------+------------+-----------+------------------+

通过上面的实验，我们知道了statement,stage,wait的三级结构，通过nesting_event_id进行关联，它表示某个事件的父event_id。

(2).模拟innodb行锁等待的例子
      会话A执行语句update test_icp set y=y+1 where x=1(x为primary key)，不commit；会话B执行同样的语句update test_icp set y=y+1 where x=1，会话B堵塞，并最终报错。通过连接连接查询events_statements_history_long和events_stages_history_long，可以看到在updating阶段花了大约60s的时间。这主要因为实例上的innodb_lock_wait_timeout设置为60，等待60s后超时报错了。

SELECT
statement.EVENT_ID,
stages.event_id,
statement.sql_text,
stages.event_name,
stages.timer_wait
FROM events_statements_history_long statement 
join events_stages_history_long stages 
on statement.event_id=stages.nesting_event_id 
WHERE statement.sql_text = 'update test_icp set y=y+1 where x=1';
+----------+----------+-------------------------------------+--------------------------------+----------------+
| EVENT_ID | event_id | sql_text | event_name | timer_wait |
+----------+----------+-------------------------------------+--------------------------------+----------------+
| 5816 | 5817 | update test_icp set y=y+1 where x=1 | stage/sql/init | 195543000 |
| 5816 | 5819 | update test_icp set y=y+1 where x=1 | stage/sql/checking permissions | 22730000 |
| 5816 | 5821 | update test_icp set y=y+1 where x=1 | stage/sql/Opening tables | 66079000 |
| 5816 | 5827 | update test_icp set y=y+1 where x=1 | stage/sql/init | 89116000 |
| 5816 | 5828 | update test_icp set y=y+1 where x=1 | stage/sql/System lock | 218744000 |
| 5816 | 5832 | update test_icp set y=y+1 where x=1 | stage/sql/updating | 6001362045000 |
| 5816 | 5968 | update test_icp set y=y+1 where x=1 | stage/sql/end | 10435000 |
| 5816 | 5969 | update test_icp set y=y+1 where x=1 | stage/sql/query end | 85979000 |
| 5816 | 5983 | update test_icp set y=y+1 where x=1 | stage/sql/closing tables | 56562000 |
| 5816 | 5990 | update test_icp set y=y+1 where x=1 | stage/sql/freeing items | 83563000 |
| 5816 | 5992 | update test_icp set y=y+1 where x=1 | stage/sql/cleaning up | 4589000 |
+----------+----------+-------------------------------------+--------------------------------+----------------+

查看wait事件

SELECT
event_id,
event_name,
source,
timer_wait,
object_name,
index_name,
operation,
nesting_event_id
FROM events_waits_history_long
WHERE nesting_event_id = 5832;
*************************** 1. row ***************************
event_id: 5832
event_name: wait/io/table/sql/handler
source: handler.cc:2782
timer_wait: 6005946156624
object_name: test_icp
index_name: PRIMARY
operation: fetch

从结果来看，waits表中记录了一个fetch等待事件，但并没有更细的innodb行锁等待事件统计。

(3).模拟MDL锁等待的例子
      会话A执行一个大查询select count(*) from test_slow，会话B执行表结构变更alter table test_slow modify c2 varchar(152);通过如下语句可以得到alter语句的执行过程，重点关注“stage/sql/Waiting for table metadata lock”阶段。

SELECT
statement.EVENT_ID,
stages.event_id,
statement.sql_text,
stages.event_name as stage_name,
stages.timer_wait as stage_time
FROM events_statements_history_long statement 
left join events_stages_history_long stages 
on statement.event_id=stages.nesting_event_id
WHERE statement.sql_text = 'alter table test_slow modify c2 varchar(152)';
+-----------+-----------+----------------------------------------------+----------------------------------------------------+---------------+
| EVENT_ID | event_id | sql_text | stage_name | stage_time |
+-----------+-----------+----------------------------------------------+----------------------------------------------------+---------------+
| 326526744 | 326526745 | alter table test_slow modify c2 varchar(152) | stage/sql/init | 216662000 |
| 326526744 | 326526747 | alter table test_slow modify c2 varchar(152) | stage/sql/checking permissions | 18183000 |
| 326526744 | 326526748 | alter table test_slow modify c2 varchar(152) | stage/sql/checking permissions | 10294000 |
| 326526744 | 326526750 | alter table test_slow modify c2 varchar(152) | stage/sql/init | 4783000 |
| 326526744 | 326526751 | alter table test_slow modify c2 varchar(152) | stage/sql/Opening tables | 140172000 |
| 326526744 | 326526760 | alter table test_slow modify c2 varchar(152) | stage/sql/setup | 157643000 |
| 326526744 | 326526769 | alter table test_slow modify c2 varchar(152) | stage/sql/creating table | 8723217000 |
| 326526744 | 326526803 | alter table test_slow modify c2 varchar(152) | stage/sql/After create | 257332000 |
| 326526744 | 326526832 | alter table test_slow modify c2 varchar(152) | stage/sql/Waiting for table metadata lock | 1000181831000 |
| 326526744 | 326526835 | alter table test_slow modify c2 varchar(152) | stage/sql/After create | 33483000 |
| 326526744 | 326526838 | alter table test_slow modify c2 varchar(152) | stage/sql/Waiting for table metadata lock | 1000091810000 |
| 326526744 | 326526841 | alter table test_slow modify c2 varchar(152) | stage/sql/After create | 17187000 |
| 326526744 | 326526844 | alter table test_slow modify c2 varchar(152) | stage/sql/Waiting for table metadata lock | 1000126464000 |
| 326526744 | 326526847 | alter table test_slow modify c2 varchar(152) | stage/sql/After create | 27472000 |
| 326526744 | 326526850 | alter table test_slow modify c2 varchar(152) | stage/sql/Waiting for table metadata lock | 561996133000 |
| 326526744 | 326526853 | alter table test_slow modify c2 varchar(152) | stage/sql/After create | 124876000 |
| 326526744 | 326526877 | alter table test_slow modify c2 varchar(152) | stage/sql/System lock | 30659000 |
| 326526744 | 326526881 | alter table test_slow modify c2 varchar(152) | stage/sql/preparing for alter table | 40246000 |
| 326526744 | 326526889 | alter table test_slow modify c2 varchar(152) | stage/sql/altering table | 36628000 |
| 326526744 | 326526912 | alter table test_slow modify c2 varchar(152) | stage/sql/committing alter table to storage engine | 11846511000 |
| 326526744 | 326528280 | alter table test_slow modify c2 varchar(152) | stage/sql/end | 43824000 |
| 326526744 | 326528281 | alter table test_slow modify c2 varchar(152) | stage/sql/query end | 112557000 |
| 326526744 | 326528299 | alter table test_slow modify c2 varchar(152) | stage/sql/closing tables | 27707000 |
| 326526744 | 326528305 | alter table test_slow modify c2 varchar(152) | stage/sql/freeing items | 201614000 |
| 326526744 | 326528308 | alter table test_slow modify c2 varchar(152) | stage/sql/cleaning up | 3584000 |
+-----------+-----------+----------------------------------------------+----------------------------------------------------+---------------+

从结果可以看到，出现了多次stage/sql/Waiting for table metadata lock阶段，并且间隔1s，说明每隔1s钟会重试判断。找一个该阶段的event_id,通过nesting_event_id关联，确定到底在等待哪个wait事件。

SELECT
event_id,
event_name,
source,
timer_wait,
object_name,
index_name,
operation,
nesting_event_id
FROM events_waits_history_long
WHERE nesting_event_id = 326526850;
+-----------+---------------------------------------------------+------------------+--------------+-------------+------------+------------+------------------+
| event_id | event_name | source | timer_wait | object_name | index_name | operation | nesting_event_id |
+-----------+---------------------------------------------------+------------------+--------------+-------------+------------+------------+------------------+
| 326526851 | wait/synch/cond/sql/MDL_context::COND_wait_status | mdl.cc:1327 | 562417991328 | NULL | NULL | timed_wait | 326526850 |
| 326526852 | wait/synch/mutex/mysys/my_thread_var::mutex | sql_class.h:3481 | 733248 | NULL | NULL | lock | 326526850 |
+-----------+---------------------------------------------------+------------------+--------------+-------------+------------+------------+------------------+

通过结果可以知道，产生阻塞的是条件变量MDL_context::COND_wait_status，并且显示了代码的位置。

小结
     本文简单举例说明了如何通过Performance Schema得到数据库运行的统计信息，以及如何利用这些统计信息分析定位问题。通过Performance Schema，DBA可以能深入的理解系统，也能进一步定位问题的源头和本质。