当前位置: 代码迷 >> MySQL >> MySQL Performance-Schema(3) 实践篇
  详细解决方案

MySQL Performance-Schema(3) 实践篇

热度:590   发布时间:2016-05-05 16:30:40.0
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?

SELECTDIGEST_TEXT,COUNT_STAR,FIRST_SEEN,LAST_SEENFROM events_statements_summary_by_digestORDER 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: 5282893FIRST_SEEN: 2015-12-15 18:25:38LAST_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?

SELECTfile_name,event_name,SUM_NUMBER_OF_BYTES_READ,SUM_NUMBER_OF_BYTES_WRITEFROM file_summary_by_instanceORDER 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.ibdevent_name: wait/io/file/innodb/innodb_data_fileSUM_NUMBER_OF_BYTES_READ: 2168078336SUM_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?

SELECTobject_name,COUNT_READ,COUNT_WRITE,COUNT_FETCH,SUM_TIMER_WAITFROM table_io_waits_summary_by_tableORDER BY sum_timer_wait DESC limit 1\G*************************** 1. row ***************************object_name: test_slowCOUNT_READ: 1986130COUNT_WRITE: 393222COUNT_FETCH: 1986130SUM_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?

SELECTOBJECT_NAME,INDEX_NAME,COUNT_FETCH,COUNT_INSERT,COUNT_UPDATE,COUNT_DELETEFROM table_io_waits_summary_by_index_usageORDER BY SUM_TIMER_WAIT DESC limit 1\G*************************** 1. row ***************************OBJECT_NAME: test_slowINDEX_NAME: PRIMARYCOUNT_FETCH: 393247COUNT_INSERT: 0COUNT_UPDATE: 393222COUNT_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?

SELECTOBJECT_SCHEMA,OBJECT_NAME,INDEX_NAMEFROM table_io_waits_summary_by_index_usageWHERE INDEX_NAME IS NOT NULLAND COUNT_STAR = 0AND OBJECT_SCHEMA <> 'mysql'ORDER BY OBJECT_SCHEMA,OBJECT_NAME;*************************** 1. row ***************************OBJECT_SCHEMA: chuckOBJECT_NAME: test_icpINDEX_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?

SELECTEVENT_NAME,COUNT_STAR,SUM_TIMER_WAIT,AVG_TIMER_WAITFROM events_waits_summary_global_by_event_nameWHERE event_name != 'idle'ORDER BY SUM_TIMER_WAIT DESC LIMIT 1;*************************** 1. row ***************************EVENT_NAME: wait/synch/cond/threadpool/worker_condCOUNT_STAR: 26369820SUM_TIMER_WAIT: 1604134685750586132AVG_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:

SELECTEVENT_ID,sql_textFROM events_statements_historyWHERE 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

SELECTevent_id,EVENT_NAME,SOURCE,TIMER_END - TIMER_STARTFROM events_stages_history_longWHERE 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。

SELECTevent_id,event_name,source,timer_wait,object_name,index_name,operation,nesting_event_idFROM events_waits_history_longWHERE 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。

SELECTstatement.EVENT_ID,stages.event_id,statement.sql_text,stages.event_name,stages.timer_waitFROM 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

SELECTevent_id,event_name,source,timer_wait,object_name,index_name,operation,nesting_event_idFROM events_waits_history_longWHERE nesting_event_id = 5832;*************************** 1. row ***************************event_id: 5832event_name: wait/io/table/sql/handlersource: handler.cc:2782timer_wait: 6005946156624object_name: test_icpindex_name: PRIMARYoperation: 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。

SELECTstatement.EVENT_ID,stages.event_id,statement.sql_text,stages.event_name as stage_name,stages.timer_wait as stage_timeFROM events_statements_history_long statement left join events_stages_history_long stages on statement.event_id=stages.nesting_event_idWHERE 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。

SELECTevent_id,event_name,source,timer_wait,object_name,index_name,operation,nesting_event_idFROM events_waits_history_longWHERE 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。

 

  相关解决方案