需求：一个SQL执行特别慢，无法返回结果，需要进行优化，最终返回结果即可。
一、SQL分析
二、尝试执行，观测执行计划
三、修改SQL
四、问题总结
一、SQL分析
1）SQL文本，执行时间，执行用户 用户brjljk sql执行时间，2935分钟 sql_text select c.hphm, c.ccdjrq, c.clpp1, c.clxh, c.zt, c.syr, c.wfsj, c.wfxw, c.dsr, c.xxly, c.syq, c.wfsj1, d.wfnr, e.dlmc, c.xxly1, c.dsr1 from (select /*+ index(b idx_violation_wfsj)*/ a.hphm, a.ccdjrq, a.clpp1, a.clxh, a.zt, a.syr, a.wfsj, a.wfxw, a.dsr, a.xxly, a.syq, b.wfsj wfsj1, b.wfxw wfxw1, b.wfdd wfdd1, b.xxly xxly1, b.dsr dsr1 from A a right join B b on a.hphm = b.hphm where a.wfsj <> b.wfsj and (b.wfsj < add_months(a.wfsj, 12) and b.wfsj > add_months(a.wfsj, -12)) and a.wfsj > to_date('2018-08-01', 'yyyy-mm-dd') and a.wfsj < to_date('2018-09-01', 'yyyy-mm-dd') order by a.hphm, a.wfsj, b.wfsj) c, D d, E e where c.wfxw1 = d.wfxw and c.wfdd1 = e.dldm
2）查询会话等待事件
SQL> select sid,serial#,event,sql_id,status,(sysdate-logon_time)*86400 as "s",
last_call_et,username,inst_id,MACHINE from gv$session where status='ACTIVE' and username is not null;
SID SERIAL# EVENT SQL_ID STATUS s LAST_CALL_ET USERNAME INST_ID MACHINE
---------- ---------- ------------------------------ ------------- -------- ---------- -------
1776 28345 db file sequential read 2vcdzpaknk46s ACTIVE 180100 176352 xxx 1 xx
3)查询sql文本
SQL> select sql_text from v$sqlarea where sql_id ='&a';
SQL> select sql_text from v$sqltext where sql_id ='2vcdzpaknk46s';
SQL> select sql_text from v$sqlstats where sql_id ='2vcdzpaknk46s';
4）查询执行计划
select * from table(dbms_xplan.display_cursor('&sql',null,'PEEKED_BINDS'));
SQL> select * from table(dbms_xplan.display_awr('2vcdzpaknk46s',null));
------------------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
------------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | | | 129M(100)| |
| 1 | SORT ORDER BY | | 309 | 56238 | 129M (1)|432:58:14 |
| 2 | HASH JOIN | | 309 | 56238 | 129M (1)|432:58:14 |
| 3 | HASH JOIN | | 309 | 50985 | 129M (1)|432:58:13 |
| 4 | HASH JOIN | | 309 | 37389 | 129M (1)|432:58:13 |
| 5 | TABLE ACCESS FULL | SJS20181022 | 10455 | 847K| 891 (2)| 00:00:11 |
| 6 | TABLE ACCESS BY INDEX ROWID| VIO_VIOLATION | 144M| 5226M| 129M (1)|432:57:51 |
| 7 | INDEX FULL SCAN | IDX_VIOLATION_WFSJ | 144M| | 583K (1)| 01:56:47 |
| 8 | TABLE ACCESS FULL | VIO_CODEWFDM | 1069 | 47036 | 13 (0)| 00:00:01 |
| 9 | TABLE ACCESS FULL | FRM_ROADITEM | 5212 | 88604 | 22 (0)| 00:00:01 |
------------------------------------------------------------------------------------------------------
5）查询SQL等待事件
select count(*),event,count(distinct session_id) from gv$active_session_history
where sql_id='2vcdzpaknk46s' group by event;
COUNT(*) EVENT COUNT(DISTINCTSESSION_ID)
---------- ------------------------------ -------------------------
27652 db file sequential read 1
22 gc cr block 2-way 1
258 gc cr disk read 1
293 1
6）查询执行计划最慢的步骤（failed）
select count(*),sql_plan_line_id
from gv$active_session_history
where sql_id='2vcdzpaknk46s'
group by sql_plan_line_id
order by 2;
--生产环境10.2.0.5，11g才有的字段
7）表碎片
表碎片会导致全表扫描更消耗资源，本次慢不是由于全表扫描的问题
8）数据量
通过dba_tables,num_rows,dba_segments,bytes查询得到信息如下
a表 hash join 驱动表，30万条记录
B表 hash join 被驱动表，1亿条记录，表100G大小
二、尝试执行，观测执行计划
1）确认优化重点四个表中，从执行计划看，重点为
| 6 | TABLE ACCESS BY INDEX ROWID| VIO_VIOLATION | 144M| 5226M| 129M (1)|432:57:51 |
| 7 | INDEX FULL SCAN | IDX_VIOLATION_WFSJ | 144M| | 583K (1)| 01:56:47 |
该SQL是a +b 的集合，转换为c最后与其它表进行关联查询
2）对a+b表的查询进行优化及测试
思路A，是否由于时间取值范围导致的问题
and (b.wfsj < add_months(a.wfsj, 12) and
b.wfsj > add_months(a.wfsj, -12))
and a.wfsj > to_date('2018-08-01', 'yyyy-mm-dd')
and a.wfsj < to_date('2018-09-01', 'yyyy-mm-dd')
explain plan for select a.hphm,a.wfsj,b.wfsj from A a right join B
b on a.hphm=b.hphm where b.wfsj > to_date('2017-01-01','yyyy-mm-dd')
and b.wfsj to_date('2018-08-01','yyyy-mm-dd')
and a.wfsjTO_DATE(' 2017-01-01 00:00:00', 'syyyy-mm-dd hh24:mi:ss') AND
"B"."WFSJ"TO_DATE(' 2018-08-01 00:00:00', 'syyyy-mm-dd hh24:mi:ss') AND
"A"."WFSJ"add_months(a.wfsj, -12)) and a.wfsj>to_date('2018-08-01','yyyy-mm-dd') and
a.wfsjADD_MONTHS(INTERNAL_FUNCTION("A"."WFSJ"),-12)))
4 - filter(("A"."WFSJ">TO_DATE(' 2018-08-01 00:00:00', 'syyyy-mm-dd hh24:mi:ss')
AND
"A"."WFSJ" explain plan for select a.hphm,a.wfsj,b.wfsj from A a
right join B b on a.hphm=b.hphm
where (b.wfsj < add_months(a.wfsj, 12) and b.wfsj >add_months(a.wfsj, -12))
and a.wfsj>to_date('2018-08-01','yyyy-mm-dd')
and a.wfsj select * from table(dbms_xplan.display());
PLAN_TABLE_OUTPUT
---------------------------------------------------------------------------
Plan hash value: 1015943026
---------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
--------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 309 | 10506 | 115K (1)| 00:23:12 |
| 1 | SORT ORDER BY | | 309 | 10506 | 115K (1)| 00:23:12 |
|* 2 | TABLE ACCESS BY INDEX ROWID| VIO_VIOLATION | 1 | 17 | 13 (0)| 00:00:01 |
| 3 | NESTED LOOPS | | 309 | 10506 | 115K (1)| 00:23:12 |
|* 4 | INDEX FAST FULL SCAN | SJS20181022_IND_HPHM | 10455 | 173K| 277 (2)| 00:00:04 |
|* 5 | INDEX RANGE SCAN | IDX_VIOLATION_HPHM | 12 | | 3 (0)| 00:00:01 |
---------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - filter("B"."WFSJ"ADD_MONTHS(INTERNAL_FUNCTION("A"."WFSJ"),-12))
4 - filter("A"."WFSJ">TO_DATE(' 2018-08-01 00:00:00', 'syyyy-mm-dd hh24:mi:ss') AND
"A"."WFSJ"
三、修改SQL
删除Hint，让SQL走nest loop 方式，10s内返回结果
SQL> select c.hphm,
c.ccdjrq,
c.clpp1,
c.clxh,
c.zt,
c.syr,
c.wfsj,
c.wfxw,
c.dsr,
c.xxly,
c.syq,
c.wfsj1,
d.wfnr,
e.dlmc,
c.xxly1,
c.dsr1
from (select
a.hphm,
a.ccdjrq,
a.clpp1,
a.clxh,
a.zt,
a.syr,
a.wfsj,
a.wfxw,
a.dsr,
a.xxly,
a.syq,
b.wfsj wfsj1,
b.wfxw wfxw1,
b.wfdd wfdd1,
b.xxly xxly1,
b.dsr dsr1
from A a
right join B b
on a.hphm = b.hphm
where a.wfsj <> b.wfsj
and (b.wfsj < add_months(a.wfsj, 12) and
b.wfsj > add_months(a.wfsj, -12))
and a.wfsj > to_date('2018-08-01', 'yyyy-mm-dd')
and a.wfsj < to_date('2018-09-01', 'yyyy-mm-dd')
order by a.hphm, a.wfsj, b.wfsj) c,
D d,
E e
where c.wfxw1 = d.wfxw
and c.wfdd1 = e.dldm;
52519 rows selected.
Elapsed: 00:00:05.08
Execution Plan
----------------------------------------------------------
Plan hash value: 2181500870
-----------------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
-----------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 309 | 56238 | 136K (1)| 00:27:24 |
| 1 | SORT ORDER BY | | 309 | 56238 | 136K (1)| 00:27:24 |
|* 2 | HASH JOIN | | 309 | 56238 | 136K (1)| 00:27:24 |
|* 3 | HASH JOIN | | 309 | 50985 | 136K (1)| 00:27:23 |
|* 4 | TABLE ACCESS BY INDEX ROWID| VIO_VIOLATION | 1 | 38 | 13 (0)| 00:00:01 |
| 5 | NESTED LOOPS | | 309 | 37389 | 136K (1)| 00:27:23 |
|* 6 | TABLE ACCESS FULL | SJS20181022 | 10455 | 847K| 891 (2)| 00:00:11 |
|* 7 | INDEX RANGE SCAN | IDX_VIOLATION_HPHM | 12 | | 3 (0)| 00:00:01 |
| 8 | TABLE ACCESS FULL | VIO_CODEWFDM | 1069 | 47036 | 13 (0)| 00:00:01 |
| 9 | TABLE ACCESS FULL | FRM_ROADITEM | 5212 | 88604 | 22 (0)| 00:00:01 |
-----------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
2 - access("B"."WFDD"="E"."DLDM")
3 - access("B"."WFXW"="D"."WFXW")
4 - filter("A"."WFSJ"<>"B"."WFSJ" AND "B"."WFSJ"ADD_MONTHS(INTERNAL_FUNCTION("A"."WFSJ"),-12))
6 - filter("A"."WFSJ">TO_DATE(' 2018-08-01 00:00:00', 'syyyy-mm-dd hh24:mi:ss') AND
"A"."WFSJ"
四、问题总结
1）使用Nest loop方式，被驱动表及时循环查询30万次，比想象中的快很多很多
2）本次sql未优化前走hash join方式的原因是，hint 索引是时间列，
而Nest loop方式需要驱动表的查询结果输出身份证，被驱动表拿着身份证，
去被驱动表中索取记录；驱动表在本次执行计划无变化，被驱动表从时间字段索引，转换走
IDX_VIOLATION_HPHM， 也就是说，由于索引的选择度的问题，Oracle认为 hash join的连接方式
优于 date索引（hint）找到对应的rowid,然后找到hphm字段值
3）今后，在使用hint前，通过测试，选择合适的hint