求 DBA 带飞， sql 优化问题，左链接字句太多，加上分析排序，效率很慢

[poupoo]

select u.final_ng_cnt,
v.color_cnt,
v.color_total,
y.weight_total,
y.weight_pass_cnt,
z.residual_total,
z.residual_total - z.last_ng_cnt as residual_pass_cnt,
p.package_total,
p.package_pass_cnt
from (select 'aaa' as batch_no,
sum(case
when t.result = 0 then
1
else
0
end) final_ng_cnt,
count(distinct t.code) as final_total
from (select a.item,
a.code,
a.result,
a.datetime,
a. recordid,
dense_rank() over(partition by a.code, a.fruit_name, a.item order by nvl(a. recordid, '0000000') desc) as final_flag
from fruit_table_a a
inner join fruit_table_b b on a. recordid =
b.productid
where instr(a.code, '000000000000') = 0
and instr(a.code, '9999999') = 0
and instr(nvl(b.batch_no, '@@@'), '@@@') = 0
and b.batch_no = 'aaa') t
where t.final_flag = 1) u
left join (select 'aaa' as batch_no,
sum(case
when c.first_rs = 1 then
1
else
0
end) color_cnt,
count(c.code) as color_total

from (select b.batch_no,
a.item,
a.code,
a.result as first_rs,
a.datetime,
dense_rank() over(partition by a.code, a.fruit_name, a.item order by nvl(a. recordid, '00000000')) as first_flag
from fruit_table_a a
inner join fruit_table_b b on a. recordid =
b.productid


where instr(a.code, '000000000000') = 0
and instr(a.code, '9999999') = 0
and instr(a.item, '0019') > 0
and b.batch_no = 'aaa') c
where c.first_flag = 1
) v on u.batch_no = v.batch_no
left join (select 'aaa' as batch_no,
sum(case
when instr(g.result, '0') = 0 then
1
else
0
end) weight_pass_cnt,
count(g.code) as weight_total
from (select k.code,
wm_concat(k.item) rank,
wm_concat(k.result) as result
from (select distinct c. recordid,
c.batch_no,
c.line_num,
c.station_id,
c.thread_id,
c.item,
c.code,
c.result,
c.datetime,
c.first_flag
from (select a. recordid,
b.batch_no,
a.line_num,
a.station_id,
a.thread_id,
a.item,
a.code,
a.result,
a.datetime,
dense_rank() over(partition by a.code, a.fruit_name, a.item order by nvl(a. recordid, '0000000')) as first_flag
from fruit_table_a a
inner join fruit_table_b b on a. recordid =
b.productid

where instr(a.code, '000000000000') = 0
and instr(a.code, '9999999') = 0
and instr(a.item, '0823') > 0
and b.batch_no = 'aaa'
union all
select a. recordid,
b.batch_no,
a.line_num,
a.station_id,
a.thread_id,
a.item,
a.code,
a.result,
a.datetime,
dense_rank() over(partition by a.code, a.fruit_name, a.item order by nvl(a. recordid, '0000000')) as first_flag
from fruit_table_a a
inner join fruit_table_b b on a. recordid =
b.productid

where instr(a.code, '000000000000') = 0
and instr(a.code, '9999999') = 0
and instr(a.item, '0923') > 0
and b.batch_no = 'aaa'
union all
select a. recordid,
b.batch_no,
a.line_num,
a.station_id,
a.thread_id,
a.item,
a.code,
a.result,
a.datetime,
dense_rank() over(partition by a.code, a.fruit_name, a.item order by nvl(a. recordid, '0000000')) as first_flag
from fruit_table_a a
inner join fruit_table_b b on a. recordid =
b.productid

where instr(a.code, '000000000000') = 0
and instr(a.code, '9999999') = 0
and instr(a.item, '1023') > 0
and b.batch_no = 'aaa') c) k
group by k.code) g
where instr(g.rank, '0823') > 0
and instr(g.rank, '0923') > 0
and instr(g.rank, '1023') > 0) y on u.batch_no =
y.batch_no
left join (select 'aaa' as batch_no,
sum(count(c.code)) as residual_total,
sum(sum(case
when c.final_flag = 1 and c.result = 0 then
1
else
0
end)) last_ng_cnt
from (select b.batch_no,
a.line_num,
a.station_id,
a.thread_id,
a.item,
a.code,
a.result,
a.datetime,
dense_rank() over(partition by a.code, a.fruit_name, a.item order by nvl(a. recordid, '0000000') desc) as final_flag
from fruit_table_a a
inner join fruit_table_b b on a. recordid =
b.productid


where instr(a.code, '000000000000') = 0
and instr(a.code, '9999999') = 0
and instr(a.item, '0823') = 0
and instr(a.item, '0923') = 0
and instr(a.item, '1023') = 0
and instr(a.item, '23') > 0
and b.batch_no = 'aaa') c
group by c.code) z on u.batch_no = z.batch_no
left join (select 'aaa' as batch_no,
sum(case
when g.first_rs = 1 and g.first_flag = 1 then
1
else
0
end) package_pass_cnt,
count(distinct g.code) as package_total

from (select 'aaa' as batch_no,
a.item,
a.code,
a.result as first_rs,
a.datetime,
dense_rank() over(partition by a.code, a.fruit_name, a.item order by nvl(a. recordid, '00000000')) as first_flag
from fruit_table_a a
inner join fruit_table_b b on a. recordid =
b.productid

where instr(a.code, '000000000000') = 0
and instr(a.code, '9999999') = 0
and b.batch_no = 'aaa'
and instr(a.item, '0020') > 0) g) p on u.batch_no =
p.batch_no

[l00t]

说 SQL 优化呢，你得先说数据库是什么，然后贴个目前的执行计划，这个是最基本的了吧…… 不然从何说起呢……

从语法看，你这个多半是 Oracle 吧。我建议你把每个子查询单独拉出来执行下，看看哪个特别慢，找到性能瓶颈。

没有执行计划，说不了太多。单从语句看，y 那段问题比较大。没有必要写那么复杂的，你可以精简出更有效率的语句来实现同样的效果。y 里面的那个 dense_rank 更是完全看不出作用。

你这个语句里所有的 inner join 里面的 where 貌似都可以改成 and，或者直接改成不需要 inner join 的模式。我看你的 inner join 貌似无非只是确认下存在而已，对于 b 表的字段一个都没用上。

left join 对你这个语句的影响应该等于 0。你的 u,v,y,z,p 全都只有一行记录，这种 left join 能有啥关系？

[poupoo]

@MeteorCat @l00t @wdlth 确实是 oracle,我考虑拆开每一项用物化视图，你们有什么更好的建议吗。

[MeteorCat]

@poupoo 只能这样来处理，我以前那家游戏的数据就是视图+触发器+存储过程，说到底，前人留的坑只能这样处理，重构是不可能重构的，你看你表那堆关联就排除了重构；什么索引优化也别想了，那堆关联 join 就压根没法命中索引，所以只能加机器性能和创建冗余的视图等