窗口函数查询优化案例
窗口函数常用用于分组排序运算中，方便用户实现各种分组需求。由于窗口函数需要通常需要全表扫描数据，同时还需排序聚集，消耗大量的CPU资源，执行效率较低。以下介绍一例窗口函数的优化案例。
准备例子
有这样一个功能需求。系统中存在资讯信息这样一个模块，用于发布一些和业务相关的活动动态，其中每条资讯信息都有一个所属类型（如科技类的资讯、娱乐类、军事类···）和浏览量字段。官网上需要滚动展示一些热门资讯信息列表（浏览量越大代表越热门），而且每个类别的相关资讯记录至多显示3条，换句话：“按照资讯分类分组，取每组的前3条资讯信息列表”。 表结构及初始数据如下：
Create table info( id numeric not null primary key , title varchar(100) , Viewnum numeric , info_type_id numeric , Code text ); create index info_infotypeid on info (info_type_id); Create table info_type( Id numeric not null primary key, Name varchar(100) ); --插入100个新闻分类 Insert into info_type select id, 'TYPE' || lpad(id::text, 5, '0' ) from generate_series(1, 100) id; --插入1000000个新闻 Insert into info select id, 'TTL' || lpad(id::text, 20, '0' ) title, ceil(random()*1000000) Viewnum, ceil(random()*100) info_type_id , md5(id) code from generate_series(1, 1000000) id; vacuum analyse info_type,info;
方法一：使用窗口函数
explain (analyse ,buffers )
with i as ( select i.*, row_number() over (partition by i.info_type_id order by i.viewnum desc) sn from info i)
select * from info_type t left join i on i.sn <= 3 and i.info_type_id = t.id;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------------------
Hash Right Join (cost=211867.09..245279.17 rows=333333 width=97) (actual time=4223.126..6169.377 rows=300 loops=1)
Hash Cond: (i.info_type_id = t.id)
Buffers: shared hit=11582 read=1753, temp read=17860 written=17901
-> Subquery Scan on i (cost=211863.84..244363.84 rows=333333 width=82) (actual time=4223.080..6168.742 rows=300 loops=1)
Filter: (i.sn <= 3)
Rows Removed by Filter: 999700
Buffers: shared hit=11582 read=1752, temp read=17860 written=17901
-> WindowAgg (cost=211863.84..231863.84 rows=1000000 width=82) (actual time=4223.079..6080.518 rows=1000000 loops=1)
Buffers: shared hit=11582 read=1752, temp read=17860 written=17901
-> Sort (cost=211863.84..214363.84 rows=1000000 width=74) (actual time=4223.065..5224.438 rows=1000000 loops=1)
Sort Key: i_1.info_type_id, i_1.viewnum DESC
Sort Method: external merge Disk: 84128kB
Buffers: shared hit=11582 read=1752, temp read=17860 written=17901
-> Seq Scan on info i_1 (cost=0.00..23334.00 rows=1000000 width=74) (actual time=0.006..249.981 rows=1000000 loops=1)
Buffers: shared hit=11582 read=1752
-> Hash (cost=2.00..2.00 rows=100 width=15) (actual time=0.037..0.037 rows=100 loops=1)
Buckets: 1024 Batches: 1 Memory Usage: 13kB
Buffers: shared read=1
-> Seq Scan on info_type t (cost=0.00..2.00 rows=100 width=15) (actual time=0.015..0.021 rows=100 loops=1)
Buffers: shared read=1
Planning Time: 0.328 ms
Execution Time: 6182.496 ms
(22 rows)
可以看到，这里消耗资源最大的是在 sort 操作上。那么，我们能否避免sort 操作了？ 使用索引可以避免sort 操作。
方法二：只取第3名的记录
方法一，由于读取了大量数据块，耗时过多。本方法暂时先简化例子，功能要求只需返回每组1条记录。新的SQL特点，每个类型使用子查询通过info表的info_type_id列的索引，可以避免读取多余的数据。select list的子查询作为计算列，只能返回一个值，所以使用row (i.*)::info 先整合，然后使用 (inf).* 再分解，同时使用 offset2 limit 1获取第三名的一行记录。
explain (analyse ,buffers )
select id, name, (inf).*
from (select t.*,
(select row (i.*)::info
from info i
where i.info_type_id = t.id
order by i.viewnum desc
offset 2
limit 1) inf
from info_type t
) t;
QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------------------
Seq Scan on info_type t (cost=0.00..6708942.94 rows=100 width=361) (actual time=127.552..10513.868 rows=100 loops=1)
Buffers: shared hit=3544406 read=3255
SubPlan 1
-> Limit (cost=13417.88..13417.88 rows=1 width=38) (actual time=21.744..21.745 rows=1 loops=100)
Buffers: shared hit=706280 read=3252
-> Sort (cost=13417.87..13442.87 rows=10000 width=38) (actual time=21.740..21.740 rows=3 loops=100)
Sort Key: i.viewnum DESC
Sort Method: top-N heapsort Memory: 25kB
Buffers: shared hit=706280 read=3252
-> Bitmap Heap Scan on info i (cost=185.93..13288.63 rows=10000 width=38) (actual time=3.985..18.371 rows=10000 loops=100)
Recheck Cond: (info_type_id = t.id)
Heap Blocks: exact=706728
Buffers: shared hit=706280 read=3252
-> Bitmap Index Scan on info_infotypeid (cost=0.00..183.43 rows=10000 width=0) (actual time=2.615..2.615 rows=10000 loops=100)
Index Cond: (info_type_id = t.id)
Buffers: shared hit=1272 read=1532
SubPlan 2
-> Limit (cost=13417.88..13417.88 rows=1 width=38) (actual time=20.599..20.600 rows=1 loops=100)
Buffers: shared hit=709529 read=3
-> Sort (cost=13417.87..13442.87 rows=10000 width=38) (actual time=20.595..20.595 rows=3 loops=100)
Sort Key: i_1.viewnum DESC
Sort Method: top-N heapsort Memory: 25kB
Buffers: shared hit=709529 read=3
-> Bitmap Heap Scan on info i_1 (cost=185.93..13288.63 rows=10000 width=38) (actual time=3.640..17.373 rows=10000 loops=100)
Recheck Cond: (info_type_id = t.id)
Heap Blocks: exact=706728
Buffers: shared hit=709529 read=3
-> Bitmap Index Scan on info_infotypeid (cost=0.00..183.43 rows=10000 width=0) (actual time=2.291..2.291 rows=10000 loops=100)
Index Cond: (info_type_id = t.id)
Buffers: shared hit=2801 read=3
SubPlan 3
-> Limit (cost=13417.88..13417.88 rows=1 width=38) (actual time=21.284..21.285 rows=1 loops=100)
Buffers: shared hit=709532
-> Sort (cost=13417.87..13442.87 rows=10000 width=38) (actual time=21.279..21.279 rows=3 loops=100)
Sort Key: i_2.viewnum DESC
Sort Method: top-N heapsort Memory: 25kB
Buffers: shared hit=709532
-> Bitmap Heap Scan on info i_2 (cost=185.93..13288.63 rows=10000 width=38) (actual time=3.609..17.868 rows=10000 loops=100)
Recheck Cond: (info_type_id = t.id)
Heap Blocks: exact=706728
Buffers: shared hit=709532
-> Bitmap Index Scan on info_infotypeid (cost=0.00..183.43 rows=10000 width=0) (actual time=2.267..2.267 rows=10000 loops=100)
Index Cond: (info_type_id = t.id)
Buffers: shared hit=2804
SubPlan 4
-> Limit (cost=13417.88..13417.88 rows=1 width=38) (actual time=20.763..20.763 rows=1 loops=100)
Buffers: shared hit=709532
-> Sort (cost=13417.87..13442.87 rows=10000 width=38) (actual time=20.759..20.759 rows=3 loops=100)
Sort Key: i_3.viewnum DESC
Sort Method: top-N heapsort Memory: 25kB
Buffers: shared hit=709532
-> Bitmap Heap Scan on info i_3 (cost=185.93..13288.63 rows=10000 width=38) (actual time=3.769..17.505 rows=10000 loops=100)
Recheck Cond: (info_type_id = t.id)
Heap Blocks: exact=706728
Buffers: shared hit=709532
-> Bitmap Index Scan on info_infotypeid (cost=0.00..183.43 rows=10000 width=0) (actual time=2.390..2.390 rows=10000 loops=100)
Index Cond: (info_type_id = t.id)
Buffers: shared hit=2804
SubPlan 5
-> Limit (cost=13417.88..13417.88 rows=1 width=38) (actual time=20.713..20.713 rows=1 loops=100)
Buffers: shared hit=709532
-> Sort (cost=13417.87..13442.87 rows=10000 width=38) (actual time=20.709..20.709 rows=3 loops=100)
Sort Key: i_4.viewnum DESC
Sort Method: top-N heapsort Memory: 25kB
Buffers: shared hit=709532
-> Bitmap Heap Scan on info i_4 (cost=185.93..13288.63 rows=10000 width=38) (actual time=3.689..17.432 rows=10000 loops=100)
Recheck Cond: (info_type_id = t.id)
Heap Blocks: exact=706728
Buffers: shared hit=709532
-> Bitmap Index Scan on info_infotypeid (cost=0.00..183.43 rows=10000 width=0) (actual time=2.288..2.288 rows=10000 loops=100)
Index Cond: (info_type_id = t.id)
Buffers: shared hit=2804
Planning Time: 0.729 ms
Execution Time: 10514.326 ms
(74 rows)
方法二针对 info_type 的每一行，info 表都要根据 info_type_id 索引访问 info 表 5 次 （5个列）。 总时间消耗： 100 （行）*5（列）* 20 （每次大概20ms），大约 10000ms。
执行计划分析：根据 info_type_id 索引，需要访问的行数太多，而且还是需要排序。基于这些考虑，我们可以创建个 info_type_id + viewnum 复合索引，减少每访问的时间消耗，避免排序。
方法三：优化索引
create index info_typeview on info(info_type_id,viewnum);
QUERY PLAN
----------------------------------------------------------------------------------------------------------------------------------------------------------
Seq Scan on info_type t (cost=0.00..4627.72 rows=100 width=361) (actual time=0.255..13.391 rows=100 loops=1)
Buffers: shared hit=2881 read=120
SubPlan 1
-> Limit (cost=6.31..9.25 rows=1 width=38) (actual time=0.041..0.041 rows=1 loops=100)
Buffers: shared hit=480 read=120
-> Index Scan Backward using info_typeview on info i (cost=0.42..29421.91 rows=10000 width=38) (actual time=0.034..0.040 rows=3 loops=100)
Index Cond: (info_type_id = t.id)
Buffers: shared hit=480 read=120
SubPlan 2
-> Limit (cost=6.31..9.25 rows=1 width=38) (actual time=0.022..0.022 rows=1 loops=100)
Buffers: shared hit=600
-> Index Scan Backward using info_typeview on info i_1 (cost=0.42..29421.91 rows=10000 width=38) (actual time=0.018..0.021 rows=3 loops=100)
Index Cond: (info_type_id = t.id)
Buffers: shared hit=600
SubPlan 3
-> Limit (cost=6.31..9.25 rows=1 width=38) (actual time=0.021..0.021 rows=1 loops=100)
Buffers: shared hit=600
-> Index Scan Backward using info_typeview on info i_2 (cost=0.42..29421.91 rows=10000 width=38) (actual time=0.018..0.020 rows=3 loops=100)
Index Cond: (info_type_id = t.id)
Buffers: shared hit=600
SubPlan 4
-> Limit (cost=6.31..9.25 rows=1 width=38) (actual time=0.021..0.021 rows=1 loops=100)
Buffers: shared hit=600
-> Index Scan Backward using info_typeview on info i_3 (cost=0.42..29421.91 rows=10000 width=38) (actual time=0.018..0.020 rows=3 loops=100)
Index Cond: (info_type_id = t.id)
Buffers: shared hit=600
SubPlan 5
-> Limit (cost=6.31..9.25 rows=1 width=38) (actual time=0.023..0.023 rows=1 loops=100)
Buffers: shared hit=600
-> Index Scan Backward using info_typeview on info i_4 (cost=0.42..29421.91 rows=10000 width=38) (actual time=0.020..0.022 rows=3 loops=100)
Index Cond: (info_type_id = t.id)
Buffers: shared hit=600
Planning Time: 0.730 ms
Execution Time: 13.552 ms
(34 rows)
可以看到，创建新索引后，单次的访问从 20ms 降低到 0.023ms ，将近降了 1000 倍。
存在问题：限制了返回行数，仅一行，同时info表有5个列，所以有5个subplan，其中4个是冗余的。
方法四：使用array，一次返回多行
以下再修改新的SQL，新的SQL特点，select list的子查询作为计算列，只能返回一行值，所以使用array() 先转换成数组类型，然后使用 unnest() 再分解成多行，同时使用 limit 3获取前三名的三行记录。
explain (analyse ,buffers )
select id, name, (inf).*
from (select t.id, t.name, unnest(inf) inf
from (select t.*,
array(select row (i.*)::info
from info i
where i.info_type_id = t.id
order by i.viewnum desc
limit 3) inf
from info_type t
) t) t;
QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------------------------
Subquery Scan on t (cost=0.00..942.89 rows=1000 width=361) (actual time=0.092..2.526 rows=300 loops=1)
Buffers: shared hit=601
-> ProjectSet (cost=0.00..932.89 rows=1000 width=47) (actual time=0.089..2.406 rows=300 loops=1)
Buffers: shared hit=601
-> Seq Scan on info_type t_1 (cost=0.00..2.00 rows=100 width=15) (actual time=0.008..0.020 rows=100 loops=1)
Buffers: shared hit=1
SubPlan 1
-> Limit (cost=0.42..9.25 rows=3 width=38) (actual time=0.018..0.021 rows=3 loops=100)
Buffers: shared hit=600
-> Index Scan Backward using info_typeview on info i (cost=0.42..29421.91 rows=10000 width=38) (actual time=0.017..0.020 rows=3 loops=100)
Index Cond: (info_type_id = t_1.id)
Buffers: shared hit=600
Planning Time: 0.295 ms
Execution Time: 2.639 ms
(14 rows)
方法五：使用lateral
新的SQL特点，简洁迅速，使用LATERAL子查询，允许它们引用前面的FROM项提供的列。
explain(analyze ,buffers )
select t.*, inf.*
from info_type t
left join LATERAL (select i.*
from info i
where i.info_type_id = t.id
order by i.viewnum desc offset 0
limit 3) inf on true;
QUERY PLAN
--------------------------------------------------------------------------------------------------------------------------------------------------
Nested Loop Left Join (cost=0.42..1267.72 rows=300 width=88) (actual time=0.026..1.087 rows=264 loops=1)
Buffers: shared hit=554
-> Seq Scan on info_type t (cost=0.00..2.00 rows=100 width=15) (actual time=0.005..0.011 rows=100 loops=1)
Buffers: shared hit=1
-> Limit (cost=0.42..12.60 rows=3 width=73) (actual time=0.007..0.010 rows=2 loops=100)
Buffers: shared hit=553
-> Index Scan Backward using info_typeview on info i (cost=0.42..406.17 rows=100 width=73) (actual time=0.007..0.010 rows=2 loops=100)
Index Cond: (info_type_id = t.id)
Buffers: shared hit=553
Planning Time: 0.114 ms
Execution Time: 1.114 ms
(11 行记录)
结论
- 整个优化关键点是创建了 info_type_id + viewnum 复合索引，也就是窗口查询 partition by 和 order by 两部分列的复合索引。
- array 的应用也是关键的地方，解决了需要返回多行的问题。
- 多个subplan在嵌套了 array 之后，变成 1 个。
- 适用场景：适用于排序的队列比较长的情景，比如本例：每个info_type_id 有1万条记录，这种情况下，通过索引只需要访问3行。而如果通过窗口函数，即使使用索引，也必须全部排序后再去前3行。
KINGBASE研究院
