核心逻辑：不等值关联转等值关联
使用场景：本案例适合满足以下条件的场景
- 大表A不等值关联小表B
- 表A和表B满足关联条件的结果集较小，即大表A之后很少一部分数据才满足关联条件
1.【原始语句】
INSERT INTO dmcon.dm_cbg_ci_inv_dtl_w_f_tmp_d_test (
begin_day,
cbg_area_cn_name,
cbg_area_code,
cbg_area_en_name,
cbg_country_cn_name,
cbg_country_code,
cbg_country_en_name,
cbg_office_cn_name,
cbg_office_code,
cbg_office_en_name,
cbg_region_cn_name,
cbg_region_code,
cbg_region_en_name,
cbg_repoffice_cn_name,
cbg_repoffice_code,
cbg_repoffice_en_name,
confirm_point_cn_name,
confirm_point_code,
confirm_point_en_name,
end_day,
honor_area_cn_name,
honor_area_code,
honor_area_en_name,
honor_country_cn_name,
honor_country_code,
honor_country_en_name,
honor_office_cn_name,
honor_office_code,
honor_office_en_name,
honor_region_cn_name,
honor_region_code,
honor_region_en_name,
honor_repoffice_cn_name,
honor_repoffice_code,
honor_repoffice_en_name,
inv_age_cn_name,
inv_age_en_name,
inv_age_type,
par_proj_cn_name,
par_proj_en_name,
par_proj_num,
product_manager_employee,
recognise_type_l1_cn_name,
recognise_type_l1_code,
recognise_type_l1_en_name,
recognise_type_l2_cn_name,
recognise_type_l2_code,
recognise_type_l2_en_name,
saleproj_manager_employee,
sub_proj_cn_name,
sub_proj_en_name,
sub_proj_num,
uuid
)
SELECT
s.begin_day,
s.cbg_area_cn_name,
s.cbg_area_code,
s.cbg_area_en_name,
s.cbg_country_cn_name,
s.cbg_country_code,
s.cbg_country_en_name,
s.cbg_office_cn_name,
s.cbg_office_code,
s.cbg_office_en_name,
s.cbg_region_cn_name,
s.cbg_region_code,
s.cbg_region_en_name,
s.cbg_repoffice_cn_name,
s.cbg_repoffice_code,
s.cbg_repoffice_en_name,
s.confirm_point_cn_name,
s.confirm_point_code,
s.confirm_point_en_name,
s.end_day,
s.honor_area_cn_name,
s.honor_area_code,
s.honor_area_en_name,
s.honor_country_cn_name,
s.honor_country_code,
s.honor_country_en_name,
s.honor_office_cn_name,
s.honor_office_code,
s.honor_office_en_name,
s.honor_region_cn_name,
s.honor_region_code,
s.honor_region_en_name,
s.honor_repoffice_cn_name,
s.honor_repoffice_code,
s.honor_repoffice_en_name,
s.inv_age_cn_name,
s.inv_age_en_name,
s.inv_age_type,
s.par_proj_cn_name,
s.par_proj_en_name,
s.par_proj_num,
s.product_manager_employee,
s.recognise_type_l1_cn_name,
s.recognise_type_l1_code,
s.recognise_type_l1_en_name,
s.recognise_type_l2_cn_name,
s.recognise_type_l2_code,
s.recognise_type_l2_en_name,
s.saleproj_manager_employee,
s.sub_proj_cn_name,
s.sub_proj_en_name,
s.sub_proj_num,
s.uuid
FROM (
SELECT
f.uuid,
par.par_proj_cn_name,
par.par_proj_num,
par.par_proj_en_name,
par.sub_proj_cn_name,
par.sub_proj_num,
par.sub_proj_en_name,
stk.product_manager_employee,
stk.saleproj_manager_employee,
agd.inv_age_cn_name,
agd.inv_age_en_name,
agd.begin_day,
agd.end_day,
agd.inv_age_type,
pnt.confirm_point_code,
pnt.confirm_point_cn_name,
pnt.confirm_point_en_name,
rec.recognise_type_l1_code,
rec.recognise_type_l1_cn_name,
rec.recognise_type_l1_en_name,
rec.recognise_type_l2_code,
rec.recognise_type_l2_cn_name,
rec.recognise_type_l2_en_name,
rcd.cbg_area_code,
rcd.cbg_area_cn_name,
rcd.cbg_area_en_name,
rcd.cbg_region_code,
rcd.cbg_region_cn_name,
rcd.cbg_region_en_name,
rcd.cbg_repoffice_code,
rcd.cbg_repoffice_cn_name,
rcd.cbg_repoffice_en_name,
rcd.cbg_office_code,
rcd.cbg_office_cn_name,
rcd.cbg_office_en_name,
rcd.cbg_country_cn_name,
rcd.cbg_country_en_name,
rcd.country_code AS cbg_country_code,
hor.cbg_area_code AS honor_area_code,
hor.cbg_area_cn_name AS honor_area_cn_name,
hor.cbg_area_en_name AS honor_area_en_name,
hor.cbg_region_code AS honor_region_code,
hor.cbg_region_cn_name AS honor_region_cn_name,
hor.cbg_region_en_name AS honor_region_en_name,
hor.cbg_repoffice_code AS honor_repoffice_code,
hor.cbg_repoffice_cn_name AS honor_repoffice_cn_name,
hor.cbg_repoffice_en_name AS honor_repoffice_en_name,
hor.cbg_office_code AS honor_office_code,
hor.cbg_office_cn_name AS honor_office_cn_name,
hor.cbg_office_en_name AS honor_office_en_name,
hor.cbg_country_cn_name AS honor_country_cn_name,
hor.cbg_country_en_name AS honor_country_en_name,
hor.country_code AS honor_country_code
FROM
dmcon.dm_cbg_ci_inv_dtl_f_test f
INNER JOIN (
SELECT
par.par_proj_cn_name,
par.par_proj_num,
par.par_proj_en_name,
par.sub_proj_cn_name,
par.sub_proj_num,
par.sub_proj_en_name,
par.proj_key
FROM (
SELECT
DISTINCT proj_key
FROM dmcon.dm_cbg_ci_inv_dtl_f_test
WHERE period_id = 202302
) t,
dmdim.dwr_inv_dim_par_proj_d par
WHERE t.proj_key = par.proj_key
) par ON f.proj_key = par.proj_key
LEFT JOIN dmdim.dm_dim_inv_contract_stkhold_v stk ON f.contract_key = stk.contract_key
LEFT JOIN dmdim.dm_dim_inv_aging_d agd ON f.inv_age BETWEEN agd.begin_day AND agd.end_day AND agd.inv_age_type = 2
INNER JOIN dmdim.dm_dim_inv_conf_point_d pnt ON f.confirm_point_id = pnt.confirm_point_id
LEFT JOIN dmdim.dm_inv_recognise_type_d rec ON f.recognise_type_id = rec.recognise_type_id
LEFT JOIN dmdim.dm_dim_region_rc_cbg_d rcd ON f.geo_pc_key = rcd.geo_pc_key AND rcd.region_tree_code = 'HUAWEI_TREE'
LEFT JOIN dmdim.dm_dim_region_rc_cbg_d hor ON f.geo_pc_key = hor.geo_pc_key AND hor.region_tree_code = 'HONOR_TREE'
WHERE f.period_id = 202302 AND f.version_id = 'R02'
) s2.【性能分析】
发现此语句执行耗时波动较大，在200s（详细见附件 《原始语句-性能波动-耗时短的perf信息.sql》）到1000s左右（详细见附件 《原始语句-性能波动-耗时长的perf信息.sql》）的之间波动。分析发现执行耗时在200s左右和1000s左右的执行计划一样，但是在比较顶层的id=3和id=4的算子执行时间相差特别大，进一步分析发现id=4的算子下层存在id=8的Nest Loop Left Join算子，怀疑是大数据关联走NestLoop导致的Cache Miss影响语句的执行性能。
因此我们尝试消除语句的中的Nest Loop Left Join的执行算子，对比原始语句，发现Nest Loop Left Join是因为如下的不等值关联导致的
LEFT JOIN dmdim.dm_dim_inv_aging_d agd ON f.inv_age BETWEEN agd.begin_day AND agd.end_day AND agd.inv_age_type = 2在DWS中不等值关联只能走Nest Loop Left Join，因此我们尝试改写SQL，尝试把Nest Loop Left Join从主执行路径提取出来，让主执行路径上只有Hash Jion操作。
进一步分析语句如下特征
1. 别名为adf的表dmdim.dm_dim_inv_aging_d较小
2. 别名为f的表dmcon.dm_cbg_ci_inv_dtl_f_test的字段inv_age只有几千个枚举值
3【优化改写】
因此我们对原始语句进行如下改写
INSERT INTO dmcon.dm_cbg_ci_inv_dtl_w_f_tmp_d_test (
begin_day,
cbg_area_cn_name,
cbg_area_code,
cbg_area_en_name,
cbg_country_cn_name,
cbg_country_code,
cbg_country_en_name,
cbg_office_cn_name,
cbg_office_code,
cbg_office_en_name,
cbg_region_cn_name,
cbg_region_code,
cbg_region_en_name,
cbg_repoffice_cn_name,
cbg_repoffice_code,
cbg_repoffice_en_name,
confirm_point_cn_name,
confirm_point_code,
confirm_point_en_name,
end_day,
honor_area_cn_name,
honor_area_code,
honor_area_en_name,
honor_country_cn_name,
honor_country_code,
honor_country_en_name,
honor_office_cn_name,
honor_office_code,
honor_office_en_name,
honor_region_cn_name,
honor_region_code,
honor_region_en_name,
honor_repoffice_cn_name,
honor_repoffice_code,
honor_repoffice_en_name,
inv_age_cn_name,
inv_age_en_name,
inv_age_type,
par_proj_cn_name,
par_proj_en_name,
par_proj_num,
product_manager_employee,
recognise_type_l1_cn_name,
recognise_type_l1_code,
recognise_type_l1_en_name,
recognise_type_l2_cn_name,
recognise_type_l2_code,
recognise_type_l2_en_name,
saleproj_manager_employee,
sub_proj_cn_name,
sub_proj_en_name,
sub_proj_num,
uuid
)
SELECT
s.begin_day,
s.cbg_area_cn_name,
s.cbg_area_code,
s.cbg_area_en_name,
s.cbg_country_cn_name,
s.cbg_country_code,
s.cbg_country_en_name,
s.cbg_office_cn_name,
s.cbg_office_code,
s.cbg_office_en_name,
s.cbg_region_cn_name,
s.cbg_region_code,
s.cbg_region_en_name,
s.cbg_repoffice_cn_name,
s.cbg_repoffice_code,
s.cbg_repoffice_en_name,
s.confirm_point_cn_name,
s.confirm_point_code,
s.confirm_point_en_name,
s.end_day,
s.honor_area_cn_name,
s.honor_area_code,
s.honor_area_en_name,
s.honor_country_cn_name,
s.honor_country_code,
s.honor_country_en_name,
s.honor_office_cn_name,
s.honor_office_code,
s.honor_office_en_name,
s.honor_region_cn_name,
s.honor_region_code,
s.honor_region_en_name,
s.honor_repoffice_cn_name,
s.honor_repoffice_code,
s.honor_repoffice_en_name,
s.inv_age_cn_name,
s.inv_age_en_name,
s.inv_age_type,
s.par_proj_cn_name,
s.par_proj_en_name,
s.par_proj_num,
s.product_manager_employee,
s.recognise_type_l1_cn_name,
s.recognise_type_l1_code,
s.recognise_type_l1_en_name,
s.recognise_type_l2_cn_name,
s.recognise_type_l2_code,
s.recognise_type_l2_en_name,
s.saleproj_manager_employee,
s.sub_proj_cn_name,
s.sub_proj_en_name,
s.sub_proj_num,
s.uuid
FROM (
SELECT
f.uuid,
par.par_proj_cn_name,
par.par_proj_num,
par.par_proj_en_name,
par.sub_proj_cn_name,
par.sub_proj_num,
par.sub_proj_en_name,
stk.product_manager_employee,
stk.saleproj_manager_employee,
agd.inv_age_cn_name,
agd.inv_age_en_name,
agd.begin_day,
agd.end_day,
agd.inv_age_type,
pnt.confirm_point_code,
pnt.confirm_point_cn_name,
pnt.confirm_point_en_name,
rec.recognise_type_l1_code,
rec.recognise_type_l1_cn_name,
rec.recognise_type_l1_en_name,
rec.recognise_type_l2_code,
rec.recognise_type_l2_cn_name,
rec.recognise_type_l2_en_name,
rcd.cbg_area_code,
rcd.cbg_area_cn_name,
rcd.cbg_area_en_name,
rcd.cbg_region_code,
rcd.cbg_region_cn_name,
rcd.cbg_region_en_name,
rcd.cbg_repoffice_code,
rcd.cbg_repoffice_cn_name,
rcd.cbg_repoffice_en_name,
rcd.cbg_office_code,
rcd.cbg_office_cn_name,
rcd.cbg_office_en_name,
rcd.cbg_country_cn_name,
rcd.cbg_country_en_name,
rcd.country_code AS cbg_country_code,
hor.cbg_area_code AS honor_area_code,
hor.cbg_area_cn_name AS honor_area_cn_name,
hor.cbg_area_en_name AS honor_area_en_name,
hor.cbg_region_code AS honor_region_code,
hor.cbg_region_cn_name AS honor_region_cn_name,
hor.cbg_region_en_name AS honor_region_en_name,
hor.cbg_repoffice_code AS honor_repoffice_code,
hor.cbg_repoffice_cn_name AS honor_repoffice_cn_name,
hor.cbg_repoffice_en_name AS honor_repoffice_en_name,
hor.cbg_office_code AS honor_office_code,
hor.cbg_office_cn_name AS honor_office_cn_name,
hor.cbg_office_en_name AS honor_office_en_name,
hor.cbg_country_cn_name AS honor_country_cn_name,
hor.cbg_country_en_name AS honor_country_en_name,
hor.country_code AS honor_country_code
FROM
dmcon.dm_cbg_ci_inv_dtl_f_test f
INNER JOIN (
SELECT
par.par_proj_cn_name,
par.par_proj_num,
par.par_proj_en_name,
par.sub_proj_cn_name,
par.sub_proj_num,
par.sub_proj_en_name,
par.proj_key
FROM (
SELECT
DISTINCT proj_key
FROM dmcon.dm_cbg_ci_inv_dtl_f_test
WHERE period_id = 202302
) t,
dmdim.dwr_inv_dim_par_proj_d par
WHERE t.proj_key = par.proj_key
) par ON f.proj_key = par.proj_key
LEFT JOIN dmdim.dm_dim_inv_contract_stkhold_v stk ON f.contract_key = stk.contract_key
LEFT JOIN (
SELECT
f.inv_age,
agd.inv_age_cn_name,
agd.inv_age_en_name,
agd.begin_day,
agd.end_day,
agd.inv_age_type,
FROM (SELECT DISTINCT inv_age FROM dmcon.dm_cbg_ci_inv_dtl_f_test f WHERE f.period_id = 202302) f
LEFT JOIN dmdim.dm_dim_inv_aging_d agd ON f.inv_age BETWEEN agd.begin_day AND agd.end_day AND agd.inv_age_type = 2
) AGD ON f.inv_age = agd.inv_age
INNER JOIN dmdim.dm_dim_inv_conf_point_d pnt ON f.confirm_point_id = pnt.confirm_point_id
LEFT JOIN dmdim.dm_inv_recognise_type_d rec ON f.recognise_type_id = rec.recognise_type_id
LEFT JOIN dmdim.dm_dim_region_rc_cbg_d rcd ON f.geo_pc_key = rcd.geo_pc_key AND rcd.region_tree_code = 'HUAWEI_TREE'
LEFT JOIN dmdim.dm_dim_region_rc_cbg_d hor ON f.geo_pc_key = hor.geo_pc_key AND hor.region_tree_code = 'HONOR_TREE'
WHERE f.period_id = 202302 AND f.version_id = 'R02'
) s新旧语句差别如下
改写完之后别名为agd的子查询的结果集只有2687条
改写完执行的执行信息（详细见附件 《改写语句-主查询消除NestLoop之后的perf信息.sql》）如下，可以发现语句性能提升到60s左右
id | operation | A-time | A-rows | E-rows | E-distinct | Peak Memory | A-width | E-width | E-costs
----+------------------------------------------------------------------------------------------------------------------+------------------------+-----------+------------+------------+----------------+-----------+---------+-------------
1 | -> Row Adapter | 66083.969 | 0 | 1 | | 8KB | | 941 | 43123074.62
2 | -> Vector Streaming (type: GATHER) | 66083.958 | 0 | 1 | | 56KB | | 941 | 43123074.62
3 | -> Vector Insert on dmcon.dm_cbg_ci_inv_dtl_w_f_tmp_d_test | [55709.424, 65744.467] | 108880389 | 3170973572 | | [276KB, 276KB] | | 941 | 43123067.96
4 | -> Vector Hash Left Join (5, 47) | [33466.773, 43244.343] | 108880389 | 3170973572 | | [19MB, 19MB] | | 941 | 13557764.90
5 | -> Vector Hash Left Join (6, 46) | [19940.411, 25422.558] | 108880389 | 2378362457 | 185 | [5MB, 5MB] | | 748 | 3621210.45
6 | -> Vector Sonic Hash Join (7,34) | [13657.117, 16496.235] | 108880389 | 1783871057 | 185 | [1MB, 1MB] | | 551 | 2324787.63
7 | -> Vector Hash Left Join (8, 32) | [11221.391, 13126.028] | 108880389 | 416236580 | 7 | [524MB, 524MB] | | 459 | 1523083.57
8 | -> Vector Hash Left Join (9, 23) | [7540.558, 8453.214] | 108880389 | 416236580 | 4851 | [2MB, 2MB] | | 422 | 420677.39
9 | -> Vector Hash Left Join (10, 22) | [5951.733, 6865.294] | 108880389 | 108880389 | 838 | [1MB, 1MB] | | 380 | 220746.49
10 | -> Vector Sonic Hash Join (11,13) | [3349.834, 3776.304] | 108880389 | 108880389 | 8 | [1MB, 1MB] | | 287 | 164336.67
11 | -> Vector Partition Iterator | [922.961, 1088.882] | 108880389 | 108880389 | 6 | [25KB, 25KB] | | 71 | 23332.51
12 | -> Partitioned CStore Scan on dmcon.dm_cbg_ci_inv_dtl_f_test f | [917.327, 1080.917] | 108880389 | 108880389 | | [4MB, 4MB] | | 71 | 23332.51
13 | -> Vector Streaming(type: BROADCAST) | [2.333, 4.494] | 24864 | 420 | 15 | [937KB, 937KB] | | 240 | 91813.37
14 | -> Vector Sonic Hash Join (15,16) | [1086.778, 1220.204] | 888 | 15 | | [925KB, 925KB] | | 240 | 91800.65
15 | -> CStore Scan on dmdim.dwr_inv_dim_par_proj_d par | [536.444, 593.697] | 57191754 | 57191754 | 2042563 | [1MB, 2MB] | | 232 | 63047.56
16 | -> Vector Streaming(type: BROADCAST) | [274.599, 374.617] | 24864 | 420 | 15 | [520KB, 520KB] | | 8 | 23340.10
17 | -> Vector Sonic Hash Aggregate | [335.765, 380.204] | 888 | 15 | | [500KB, 500KB] | [17,17] | 8 | 23333.73
18 | -> Vector Streaming(type: REDISTRIBUTE) | [335.520, 379.965] | 17535 | 168 | | [520KB, 520KB] | | 8 | 23333.16
19 | -> Vector Sonic Hash Aggregate | [197.785, 256.138] | 17535 | 168 | | [500KB, 500KB] | [17,17] | 8 | 23332.57
20 | -> Vector Partition Iterator | [109.947, 121.609] | 108880389 | 108880389 | | [25KB, 41KB] | | 8 | 13611.05
21 | -> Partitioned CStore Scan on dmcon.dm_cbg_ci_inv_dtl_f_test | [105.029, 115.601] | 108880389 | 108880389 | | [3MB, 3MB] | | 8 | 13611.05
22 | -> CStore Scan on dmdim.dm_inv_recognise_type_d rec | [0.259, 0.513] | 700 | 700 | 25 | [1MB, 1MB] | [236,236] | 102 | 25.02
23 | -> Vector Streaming(type: BROADCAST) | [5.089, 218.248] | 75236 | 102172 | 948 | [1MB, 1MB] | [177,177] | 54 | 24398.13
24 | -> Vector Nest Loop (25,30) | [520.410, 624.103] | 2687 | 3649 | | [256KB, 256KB] | | 54 | 23571.34
25 | -> Vector Sonic Hash Aggregate | [518.754, 622.409] | 2688 | 948 | | [525KB, 525KB] | [30,31] | 6 | 23463.72
26 | -> Vector Streaming(type: REDISTRIBUTE) | [518.319, 621.974] | 60130 | 23464 | | [536KB, 536KB] | | 6 | 23395.58
27 | -> Vector Sonic Hash Aggregate | [449.792, 547.058] | 60130 | 23464 | | [564KB, 564KB] | [30,30] | 6 | 23340.89
28 | -> Vector Partition Iterator | [185.524, 231.025] | 108880389 | 108880389 | | [25KB, 25KB] | | 6 | 13611.05
29 | -> Partitioned CStore Scan on dmcon.dm_cbg_ci_inv_dtl_f_test | [180.543, 225.683] | 108880389 | 108880389 | | [3MB, 3MB] | | 6 | 13611.05
30 | -> Vector Materialize | [0.734, 0.957] | 54320 | 560 | | [448KB, 448KB] | [109,109] | 48 | 95.43
31 | -> CStore Scan on dmdim.dm_dim_inv_aging_d agd | [0.363, 0.448] | 560 | 560 | | [1MB, 1MB] | | 48 | 95.33
32 | -> Vector Streaming(type: BROADCAST) | [329.279, 421.005] | 102601912 | 102601912 | 3664354 | [2MB, 2MB] | [104,104] | 53 | 833617.87
33 | -> CStore Scan on dmdim.dm_dim_inv_contract_stkhold_v stk | [24.561, 33.929] | 3664354 | 3664354 | | [792KB, 808KB] | | 53 | 1294.87
34 | -> Vector Streaming(type: BROADCAST) | [0.633, 0.763] | 6216 | 840 | 2 | [584KB, 584KB] | | 128 | 23.54
35 | -> Vector Subquery Scan on pnt | [16.990, 21.131] | 222 | 30 | | [120KB, 120KB] | | 128 | 17.18
36 | -> Vector Append(37, 38, 42) | [16.978, 21.114] | 222 | 30 | | [8KB, 8KB] | | 527 | 17.17
37 | -> CStore Scan on dmdim.dwi_md_class t | [16.856, 20.966] | 220 | 2 | | [1MB, 1MB] | | 1580 | 17.06
38 | -> Vector Subquery Scan on "*SELECT* 2" | [0.071, 0.123] | 1 | 1 | | [160KB, 160KB] | | 0 | 0.05
39 | -> Vector Subquery Scan on t | [0.024, 0.052] | 28 | 1 | | [120KB, 120KB] | | 0 | 0.02
40 | -> Vector Adapter | [0.005, 0.012] | 28 | 1 | | [40KB, 40KB] | | 0 | 0.01
41 | -> Result | [0.001, 0.003] | 28 | 1 | | [8KB, 8KB] | | 0 | 0.01
42 | -> Vector Subquery Scan on "*SELECT* 3" | [0.033, 0.091] | 1 | 1 | | [160KB, 160KB] | | 0 | 0.05
43 | -> Vector Subquery Scan on t | [0.015, 0.042] | 28 | 1 | | [120KB, 120KB] | | 0 | 0.02
44 | -> Vector Adapter | [0.004, 0.013] | 28 | 1 | | [40KB, 40KB] | | 0 | 0.01
45 | -> Result | [0.000, 0.002] | 28 | 1 | | [8KB, 8KB] | | 0 | 0.01
46 | -> CStore Scan on dmdim.dm_dim_region_rc_cbg_d rcd | [1.901, 2.273] | 167804 | 167804 | 2996 | [4MB, 5MB] | [509,509] | 203 | 7335.95
47 | -> CStore Scan on dmdim.dm_dim_region_rc_cbg_d hor | [7.747, 9.499] | 167804 | 167804 | 2996 | [5MB, 5MB] | [517,517] | 203 | 7335.95
新语句有如下执行特征
- 因为对表dmcon.dm_cbg_ci_inv_dtl_f_test的预聚合，参与NestLoop计算的结果集（60130条和560条数据做关联）变小，关联的结果集也很少（2688条）
- NestLoop关联操作是在Hash Join的内表中执行，HashJoin执行时会直接获取内表(即agd分支)的所有数据，然后构建成hashtable，这样的逻辑下NestLoop连续执行，直到NestLoop输出所有的结果，这样就把NestLoop动作隔离限制在一个局部内，避免对主查询的影响
经过多次测试，改写后的语句的耗时始终稳定在60s左右
原创文章，作者：奋斗，如若转载，请注明出处：https://blog.ytso.com/tech/bigdata/317425.html