复杂SQL治理实践 | 京东物流技术团队

一、前言

软件在持续的开发和维护过程中，会不断添加新功能和修复旧的缺陷，这往往伴随着代码的快速增长和复杂性的提升。若代码库没有得到良好的管理和重构，就可能积累大量的技术债务，包括不一致的设计、冗余代码、过时的库和框架以及不再使用的功能。这些因素都会导致软件结构的脆弱，增加系统出错的可能性，我们俗称为“代码腐化”，持续性的重构是一种好的解决方案。SQL也是我们常用的代码语言，虽然SQL本身作为一种标准化的查询语言不会"腐化"，但是使用SQL编写的数据库应用程序、查询和架构确实可能会因时间推移而面临类似于代码腐化的问题。

平台技术部一直坚持做稳定性建设，其中慢SQL就作为一个核心指标在治理。在治理进入深水区时，就会啃到因“SQL腐化”引入的复杂SQL治理这种硬骨头。本文以一个案例为依托来看看怎样像重构Java等高级编程语言一样来重构SQL。

二、JDL路由系统复杂SQL治理案例

路由规划是为保障客户体验，依据产品需求及时效目标，设计物流网络中每个节点的操作时长，然后通过节点互相串联保障全程链通且综合最优，同步输出规划方案并指导运营现场操作，双向校验优化，实现路由规划与实际运营的不断趋合。

简言之，路由系统支持的路由规划就是在做基于物流网络运营的运筹优化，网络是基础。而网络的基础又是线路，必然对线路的操作会“千奇百怪”。

1.问题SQL

select count(*) total_count 
from (
select * 
from (
select 
a.line_store_goods_id as line_resource_id, a.group_num as group_num, 
a.approval_erp as approval_erp, a.approval_person as approval_person, 
a.approval_status as approval_status, a.approval_time as approval_time, 
a.approval_remark as approval_remark, a.master_slave as master_slave, 
a.parent_line_code as parent_line_code, a.remarks as remarks, a.operator_time, 
a.same_stowage as same_stowage, b.start_org_id, b.start_org_name, 
b.start_province_id, b.start_province_name, b.start_city_id, b.start_city_name, 
b.start_node_code, b.start_node_name, b.end_org_id, b.end_org_name, 
b.end_province_id, b.end_province_name, b.end_city_id, b.end_city_name, 
b.end_node_code, b.end_node_name, b.depart_time, b.arrive_time, b.depart_wave_code, 
b.arrive_wave_code, b.enable_time, b.disable_time, a.store_enable_time, 
a.store_disable_time, a.update_name operator_name, b.line_code, 
b.line_type, b.transport_type, IF(a.store_enable_time > b.enable_time, 
IF(a.store_enable_time > c.enable_time, a.store_enable_time, c.enable_time), 
IF(b.enable_time > c.enable_time, b.enable_time, c.enable_time)) 
as insect_start_time, IF(a.store_disable_time < b.disable_time, 
IF(a.store_disable_time < c.disable_time, a.store_disable_time, c.disable_time), 
IF(b.disable_time < c.disable_time, b.disable_time, c.disable_time)) 
as insect_end_time 
FROM (
select * FROM line_store_goods WHERE yn = 1 and master_slave = 1) a 
join (
select start_org_id, start_org_name, start_province_id, start_province_name, 
start_city_id, start_city_name, start_node_code, start_node_name, end_org_id, 
end_org_name, end_province_id, end_province_name, end_city_id, end_city_name, 
end_node_code, end_node_name, depart_time, arrive_time, depart_wave_code, 
arrive_wave_code, line_code, line_type, transport_type, min(enable_time) 
as enable_time, max(disable_time) as disable_time 
from line_resource where line_code in (
select line_code from line_store_goods WHERE yn = 1 ) 
and yn=1 group by line_code) b 
ON a.line_code = b.line_code and a.start_node_code = b.start_node_code 
join (
select line_code,start_node_code, min(enable_time) as enable_time, 
max(disable_time) as disable_time from line_resource 
WHERE yn = 1 group by line_code) c 
ON a.parent_line_code = c.line_code and a.start_node_code = c.start_node_code) temp 
WHERE start_node_code = '311F001' and disable_time > '2023-11-15 00:00:00' 
and enable_time < disable_time) t_total;

这是一段运行在生产上的复杂SQL案例，通过慢SQL指标统计识别出来。一眼看过去毫无头绪（说明不仅性能差，而且可读性差，那么必然可维护性差），非功能性指标总是存在很强的关联性。

2.开始治理

step1.格式化

对工程人员而言：要重构，格式化很重要，保证一定的可读性

select count(*) total_count from (select * from (select a.line_store_goods_id as line_resource_id, a.group_num as group_num, a.approval_erp as approval_erp, a.approval_person as approval_person, a.approval_status as approval_status, a.approval_time as approval_time, a.approval_remark as approval_remark, a.master_slave as master_slave, a.parent_line_code as parent_line_code, a.remarks as remarks, a.operator_time, a.same_stowage as same_stowage, b.start_org_id, b.start_org_name, b.start_province_id, b.start_province_name, b.start_city_id, b.start_city_name, b.start_node_code, b.start_node_name, b.end_org_id, b.end_org_name, b.end_province_id, b.end_province_name, b.end_city_id, b.end_city_name, b.end_node_code, b.end_node_name, b.depart_time, b.arrive_time, b.depart_wave_code, b.arrive_wave_code, b.enable_time, b.disable_time, a.store_enable_time, a.store_disable_time, a.update_name operator_name, b.line_code, b.line_type, b.transport_type, IF(a.store_enable_time > b.enable_time, IF(a.store_enable_time > c.enable_time, a.store_enable_time, c.enable_time), IF(b.enable_time > c.enable_time, b.enable_time, c.enable_time)) as insect_start_time, IF(a.store_disable_time < b.disable_time, IF(a.store_disable_time < c.disable_time, a.store_disable_time, c.disable_time), IF(b.disable_time < c.disable_time, b.disable_time, c.disable_time)) as insect_end_time FROM (select * FROM line_store_goods WHERE yn = 1 and master_slave = 1) a join (select start_org_id, start_org_name, start_province_id, start_province_name, start_city_id, start_city_name, start_node_code, start_node_name, end_org_id, end_org_name, end_province_id, end_province_name, end_city_id, end_city_name, end_node_code, end_node_name, depart_time, arrive_time, depart_wave_code, arrive_wave_code, line_code, line_type, transport_type, min(enable_time) as enable_time, max(disable_time) as disable_time from line_resource where line_code in (select line_code from line_store_goods WHERE yn = 1 ) and yn=1 group by line_code) b ON a.line_code = b.line_code and a.start_node_code = b.start_node_code join (select line_code,start_node_code, min(enable_time) as enable_time, max(disable_time) as disable_time from line_resource WHERE yn = 1 group by line_code) c ON a.parent_line_code = c.line_code and a.start_node_code = c.start_node_code) temp WHERE start_node_code = '311F001' and disable_time > '2023-11-15 00:00:00' and enable_time < disable_time) t_total;

经过格式化之后，能简单判断出SQL的功能是检索满足某条件的线路数量统计。

注意：格式化作为一个重要的工具可以在任意阶段发生作用。

step2.分层拆解

·level0

select count(*) total_count from t_total

·level1 - t_total

select * from temp 
WHERE start_node_code = '311F001' 
and disable_time > '2023-11-15 00:00:00' 
and enable_time < disable_time

·level2 - temp

select 
a.line_store_goods_id as line_resource_id, a.group_num as group_num, a.approval_erp as approval_erp, a.approval_person as approval_person, a.approval_status as approval_status, a.approval_time as approval_time, a.approval_remark as approval_remark, a.master_slave as master_slave, a.parent_line_code as parent_line_code, a.remarks as remarks, a.operator_time, a.same_stowage as same_stowage, b.start_org_id, b.start_org_name, b.start_province_id, b.start_province_name, b.start_city_id, b.start_city_name, b.start_node_code, b.start_node_name, b.end_org_id, b.end_org_name, b.end_province_id, b.end_province_name, b.end_city_id, b.end_city_name, b.end_node_code, b.end_node_name, b.depart_time, b.arrive_time, b.depart_wave_code, b.arrive_wave_code, b.enable_time, b.disable_time, a.store_enable_time, a.store_disable_time, a.update_name operator_name, b.line_code, b.line_type, b.transport_type, IF(a.store_enable_time > b.enable_time, IF(a.store_enable_time > c.enable_time, a.store_enable_time, c.enable_time), IF(b.enable_time > c.enable_time, b.enable_time, c.enable_time)) as insect_start_time, IF(a.store_disable_time < b.disable_time, IF(a.store_disable_time < c.disable_time, a.store_disable_time, c.disable_time), IF(b.disable_time < c.disable_time, b.disable_time, c.disable_time)) as insect_end_time 
FROM join_table

·level3 - join_table

(select * FROM line_store_goods WHERE yn = 1 and master_slave = 1) a 
join (select start_org_id, start_org_name, start_province_id, start_province_name, start_city_id, start_city_name, start_node_code, start_node_name, end_org_id, end_org_name, end_province_id, end_province_name, end_city_id, end_city_name, end_node_code, end_node_name, depart_time, arrive_time, depart_wave_code, arrive_wave_code, line_code, line_type, transport_type, min(enable_time) as enable_time, max(disable_time) as disable_time from line_resource where line_code in (select line_code from line_store_goods WHERE yn = 1 ) and yn=1 group by line_code) b ON a.line_code = b.line_code and a.start_node_code = b.start_node_code 
join (select line_code,start_node_code, min(enable_time) as enable_time, max(disable_time) as disable_time from line_resource WHERE yn = 1 group by line_code) c ON a.parent_line_code = c.line_code and a.start_node_code = c.start_node_code

·level4 - a,b,c

select * FROM 
line_store_goods 
WHERE yn = 1 
and master_slave = 1

select start_org_id, start_org_name, start_province_id, start_province_name, start_city_id, start_city_name, start_node_code, start_node_name, end_org_id, end_org_name, end_province_id, end_province_name, end_city_id, end_city_name, end_node_code, end_node_name, depart_time, arrive_time, depart_wave_code, arrive_wave_code, line_code, line_type, transport_type, min(enable_time) as enable_time, max(disable_time) as disable_time 
from line_resource 
where line_code in (select line_code from line_store_goods WHERE yn = 1 ) 
and yn=1 
group by line_code

select line_code,start_node_code, min(enable_time) as enable_time, max(disable_time) as disable_time 
from line_resource 
WHERE yn = 1 
group by line_code

step3.重构

对于Java程序员而言，《重构 - 改善既有代码的设计》一书应该不陌生。重构的核心在设计原则（“道”&“法”）；但是工具包（“术”）同样重要，指导具体落地。

工具包准备：

•层级合并 减少临时表个数

•条件下推 减少检索行数&临时表大小

•join优化 减少检索行数&临时表大小

•子查询删除 减少临时表个数

•子查询与join的相互转换 减少检索行数

重构1 - 层级合并

level0 & level1

如下两个SQL执行效果一致，但是性能表现会有很大差异。

select count(*) total_count from (select * from temp where a = "1")
select count(*) from temp where a = "1"

第二种方式的性能表现会更好一些。原因如下：

1.减少查询计算开销：在第二种方式中，直接对表进行 count(*) 统计，不需要额外的子查询和临时表操作，可以减少计算的开销。

2.减少内存占用：第一种方式需要在内存中创建一个临时表来存储子查询的结果，而第二种方式直接对原表进行统计，不需要额外的内存占用。

3.减少磁盘 IO：第二种方式可以直接利用表的索引进行 count(*) 统计，而第一种方式可能需要额外的磁盘 IO 来处理子查询和临时表的操作。

因此，一般情况下，推荐使用第二种方式来进行 count()统计，以获得更好的性能表现。当然，在实际情况中，也需要根据具体的业务场景和数据量来综合考虑，有时候使用子查询的方式也是必要的，但总体来说，直接对原表进行 count() 统计会更高效。

重构2 - 条件下推

start_node_code = ‘311F001’ 直接下推至level4

SQL的执行是流程化的，从执行层视角看，涉及时空资源消耗最关键的有两类：1-时间（行记录扫描）、2-空间（临时表）。

简化来看，问题SQL的执行过程是子查询形成临时表，而后基于临时表做各种形式的计算（过滤、联合）。

通过条件下推，可以将过滤动作尽可能前置，减少后续过程临时表的大小。

重构3 - join优化

按个人喜好进行格式化

条件下推

剥离冗余字段，冗余字段在SQL优化过程中是一个影响易读性的干扰信息，剥离冗余字段给工程人员一个干净的画板来尽情施为

删除无效条件。join的on条件中start_node_code条件因为条件下推已经不再是有效条件。注意，此处为了行文方便做了一定的简化，理论上之前的剥离冗余字段理论上需要包含start_node_code字段查询，在此步骤之后变为冗余字段后被剥离

删除无效子查询。此时从上往下看，表a和表b存在一个奇怪的现象 - 使用了两个类似功能（子查询和join），两者的功能完全一致。题外话：此案例作为反面教材真心不错。涉及两者的优劣决策，个人做取舍的两个点是性能和可读性。在此案例中功能实现场景特别简单，join的可读性明显更好，在条件限定后扫描行数基本一致，但子查询多一个临时表；综合考量会删除子查询。

合并冗余join。继续从上往下看，表b和表c看起来一模一样。再次重复题外话：此案例作为反面教材真心不错。

等价条件替换，再次删除冗余字段

经过优化后的join语句，可读性发生了很大的变化 - 简单的双表关联查询。

step4.结果的理论验证

select count(*) from ((select line_code FROM line_store_goods WHERE yn = 1 and parent_line_code = line_code and master_slave = 1 and start_node_code = '311F001') ajoin(select line_code, min(enable_time) as enable_time, max(disable_time) as disable_time from line_resource where yn=1 and start_node_code = '311F001' group by line_code) bON a.line_code = b.line_code
) where disable_time > '2023-11-15 00:00:00' and enable_time < disable_time

重构后的SQL具备良好的可读性，基于此很容易反推出SQL的业务功能。基于此与其理论应用场景做是否匹配的理论判断很重要。有的时候生产上的SQL不一定是正确的，因为部分场景下可用性并不完全等价于正确性。