DM数据库 回表优化案例

京华开发一哥们找我优化条SQL，反馈在DM数据库执行时间很慢需要 40s 才能出结果，安排。

原SQL：

SELECT A.IND_CODE,
       A.IND_NAME                                                                                AS "specialName",
       COUNT(C.ORDER_ID)                                                                         AS "orderCount",
       COUNT(CASE WHEN D.MEDIATE_RESULT IN ('达成调解协议', '双方自行和解') THEN C.ORDER_ID END) AS "successCount"
FROM WCWCWC.AAAAAAAAA A
         LEFT JOIN WCWCWC.BBBBBBBBB B
             ON A.IND_CODE = B.INDUSTRY_CD
         LEFT JOIN
     WCWCWC.ccccccccccccccccc C
         ON B.UNISCID = C.UNISCID
         LEFT JOIN WCWCWC.DDDDDDDDDDDDDD D
             ON C.ORDER_ID = D.ORDER_ID
WHERE A.IND_CODE = 'C'
GROUP BY A.IND_CODE,
         A.IND_NAME;

执行计划：

1   #NSET2: [491, 1, 96]
2     #PRJT2: [491, 1, 96]; exp_num(4), is_atom(FALSE)
3       #HAGR2: [491, 1, 96]; grp_num(2), sfun_num(2), distinct_flag[0,0]; slave_empty(0) keys(DMTEMPVIEW_889674440.TMPCOL0, DMTEMPVIEW_889674440.TMPCOL1)
4         #PRJT2: [466, 324467, 96]; exp_num(4), is_atom(FALSE)
5           #HASH RIGHT JOIN2: [466, 324467, 96]; key_num(1), ret_null(0), KEY(D.ORDER_ID=C.ORDER_ID)
6             #CSCN2: [1, 15287, 56]; INDEX33557062(DDDDDDDDDDDDDD as D)
7             #HASH RIGHT JOIN2: [437, 324467, 96]; key_num(1), ret_null(0), KEY(C.UNISCID=B.UNISCID)
8               #CSCN2: [11, 97535, 56]; INDEX33557061(ccccccccccccccccc as C)
9               #HASH LEFT JOIN2: [386, 324467, 96]; key_num(1), partition_keys_num(0), ret_null(0), mix(0) KEY(A.IND_CODE=B.INDUSTRY_CD)
10                #INDEX JOIN LEFT JOIN2: [386, 324467, 96]  ret_null(0)
11                  #ACTRL: [386, 324467, 96];
12                    #BLKUP2: [1, 1, 96]; INDEX33557119(A)
13                      #SSEK2: [1, 1, 96]; scan_type(ASC), INDEX33557119(AAAAAAAAA as A), scan_range['C','C']
14                  #BLKUP2: [351, 324467, 48]; IND_INDUSTRY_CD(B)
15                    #SSEK2: [351, 324467, 48]; scan_type(ASC), IND_INDUSTRY_CD(BBBBBBBBB as B), scan_range[A.IND_CODE,A.IND_CODE]
16                #CSCN2: [727, 5840415, 96]; INDEX33558123(BBBBBBBBB as B)

A、B、C、D 表 所有关联列都有索引，整体SQL返回一行数据，但是要 39s 左右，非常不合理。

数据量如下：

select count(1) ,'A' from WCWCWC.AAAAAAAAA
union all
select count(1) ,'B' from WCWCWC.BBBBBBBBB
union all
select count(1) ,'C' from WCWCWC.ccccccccccccccccc
union all
select count(1) ,'D' from WCWCWC.DDDDDDDDDDDDDD;

 监控下缓慢的节点：

发现慢的节点是 B 表产生BLKUP2，500多W行数据回表 ，IND_INDUSTRY_CD 索引无法找到所有数据。

B表创建联合索引：

create index idx_BBBBBBBBB_1_2 on WCWCWC.BBBBBBBBB(INDUSTRY_CD,UNISCID);

 新的执行计划：

1   #NSET2: [158, 1, 96]
2     #PRJT2: [158, 1, 96]; exp_num(4), is_atom(FALSE)
3       #HAGR2: [158, 1, 96]; grp_num(2), sfun_num(2), distinct_flag[0,0]; slave_empty(0) keys(DMTEMPVIEW_889674491.TMPCOL0, DMTEMPVIEW_889674491.TMPCOL1)
4         #PRJT2: [133, 324467, 96]; exp_num(4), is_atom(FALSE)
5           #HASH RIGHT JOIN2: [133, 324467, 96]; key_num(1), ret_null(0), KEY(D.ORDER_ID=C.ORDER_ID)
6             #CSCN2: [1, 15287, 56]; INDEX33557062(DDDDDDDDDDDDDD as D)
7             #HASH RIGHT JOIN2: [104, 324467, 96]; key_num(1), ret_null(0), KEY(C.UNISCID=B.UNISCID)
8               #CSCN2: [11, 97535, 56]; INDEX33557061(ccccccccccccccccc as C)
9               #HASH LEFT JOIN2: [54, 324467, 96]; key_num(1), partition_keys_num(0), ret_null(0), mix(0) KEY(A.IND_CODE=B.INDUSTRY_CD)
10                #INDEX JOIN LEFT JOIN2: [54, 324467, 96]  ret_null(0)
11                  #ACTRL: [54, 324467, 96];
12                    #BLKUP2: [1, 1, 96]; INDEX33557119(A)
13                      #SSEK2: [1, 1, 96]; scan_type(ASC), INDEX33557119(AAAAAAAAA as A), scan_range['C','C']
14                  #SSEK2: [49, 324467, 96]; scan_type(ASC), IDX_BBBBBBBBB_1_2(BBBBBBBBB as B), scan_range[(A.IND_CODE,min),(A.IND_CODE,max))
15                #SSCN: [727, 5840415, 96]; IDX_BBBBBBBBB_1_2(BBBBBBBBB as B)

可以看到在创建完联合索引后，#BLKUP2: [351, 324467, 48]; 回表计划消失了，SQL也能0.3S跑出结果。

B表UNISCID字段本身就是主键，而INDUSTRY_CD字段本身有索引，开发老哥以为走了两个字段都走索引不会有什么问题，而忽略了回表计划，在大表产生回表计划是非常恐怖的事情！！！

总结：近两年感觉国产数据库市场份额越来越高了，笔者SQL优化的案例更多得也是从ORACLE、MySQL变成了DM、金仓数据库，

         建议大家还是要好好深入学习下国产数据库，免得35岁被优化了找不到工作 。