PostgreSQL 13支持增量排序(Incremental Sorting)
PostgreSQL 13支持增量排序(Incremental Sorting)
PostgreSQL 13一个重要的功能是支持增量排序，使用order by 时可以加速排序，SQL如下
select * from test order by a,b limit 10;
如果在字段a上面建立了索引，需要对字段a、b进行排序，如果一个结果已经按几个前导键排序，这就允许对附加的b进行批量排序。
enable_incremental_sort
PostgreSQL新增了配置enable_incremental_sort用于控制是否开启增量排序，此参数默认开启
测试准备
在PostgreSQL 13中创建测试表进行测试
postgres=# create table test(id int,c1 int ,c2 int,info varchar(300),crt_time timestamp);
CREATE TABLE
postgres=# insert into test select t,t,2,'test',clock_timestamp() from generate_series(1,1000000)t;
INSERT 0 1000000
postgres=# create index i_test_id on test(id);
CREATE INDEX
--查看数据如下
postgres=# select * from test order by id,c1 limit 10;
id | c1 | c2 | info | crt_time
----+----+----+------+----------------------------
1 | 1 | 2 | test | 2022-06-02 14:23:38.253289
2 | 2 | 2 | test | 2022-06-02 14:23:38.253777
3 | 3 | 2 | test | 2022-06-02 14:23:38.253785
4 | 4 | 2 | test | 2022-06-02 14:23:38.253787
5 | 5 | 2 | test | 2022-06-02 14:23:38.25379
6 | 6 | 2 | test | 2022-06-02 14:23:38.253791
7 | 7 | 2 | test | 2022-06-02 14:23:38.253793
8 | 8 | 2 | test | 2022-06-02 14:23:38.253795
9 | 9 | 2 | test | 2022-06-02 14:23:38.253809
10 | 10 | 2 | test | 2022-06-02 14:23:38.25381
(10 rows)
PostgreSQL 13 测试
- 这里我是在pg14中做的测试，pg13这个参数名叫
enable_incrementalsort
postgres=# show enable_incremental_sort;
enable_incremental_sort
-------------------------
on
(1 row)
postgres=# explain analyze select * from test order by id,c1 limit 10;
QUERY PLAN
----------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=0.46..1.16 rows=10 width=25) (actual time=0.159..0.163 rows=10 loops=1)
-> Incremental Sort (cost=0.46..70373.03 rows=1000000 width=25) (actual time=0.157..0.159 rows=10 loops=1)
Sort Key: id, c1
Presorted Key: id
Full-sort Groups: 1 Sort Method: quicksort Average Memory: 25kB Peak Memory: 25kB
-> Index Scan using i_test_id on test (cost=0.42..25373.02 rows=1000000 width=25) (actual time=0.103..0.106 rows=11 loops=1)
Planning Time: 0.427 ms
Execution Time: 0.265 ms
(8 rows)
- 可以看到
Incremental Sort和Presorted Key: id并且走了i_test_id索引，SQL耗时0.265ms
关闭enable_incremental_sort
postgres=# set enable_incremental_sort=off;
SET
postgres=# explain analyze select * from test order by id,c1 limit 10;
QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------------
Limit (cost=38962.64..38962.67 rows=10 width=25) (actual time=272.945..272.953 rows=10 loops=1)
-> Sort (cost=38962.64..41462.64 rows=1000000 width=25) (actual time=272.933..272.937 rows=10 loops=1)
Sort Key: id, c1
Sort Method: top-N heapsort Memory: 25kB
-> Seq Scan on test (cost=0.00..17353.00 rows=1000000 width=25) (actual time=0.028..118.098 rows=1000000 loops=1)
Planning Time: 0.305 ms
Execution Time: 273.023 ms
(7 rows)
- 关闭增量排序后SQL耗时
273.023 ms，性能差了几个数量级
PostgreSQL 12 测试
- Abase 7.0基于PostgreSQL 12.3
同样使用上面的建表语句，执行SQL如下
postgres=# explain analyze select * from test order by id,c1 limit 10;
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------
Limit (cost=38962.64..38962.67 rows=10 width=536) (actual time=288.847..288.851 rows=10 loops=1)
-> Sort (cost=38962.64..41462.64 rows=1000000 width=536) (actual time=288.839..288.840 rows=10 loops=1)
Sort Key: id, c1
Sort Method: top-N heapsort Memory: 25kB
-> Seq Scan on test (cost=0.00..17353.00 rows=1000000 width=536) (actual time=0.078..173.460 rows=1000000 loops=1)
Planning Time: 24.726 ms
Execution Time: 289.135 ms
(7 rows)
PG 12中执行计划和PG 14关闭enable_incremental_sort参数一样，性能较低
当然这只是一个简单的查询，如果包含where，以及连表等情况是否也可以使用 Incremental Sort
带条件
加上c1 > 100000，c1没有创建索引
postgres=# explain analyze select * from test where c1 > 100000 order by id,c1 limit 10;
QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=0.47..1.23 rows=10 width=25) (actual time=49.470..49.476 rows=10 loops=1)
-> Incremental Sort (cost=0.47..68345.40 rows=899386 width=25) (actual time=49.467..49.469 rows=10 loops=1)
Sort Key: id, c1
Presorted Key: id
Full-sort Groups: 1 Sort Method: quicksort Average Memory: 25kB Peak Memory: 25kB
-> Index Scan using i_test_id on test (cost=0.42..27873.02 rows=899386 width=25) (actual time=49.383..49.387 rows=11 loops=1)
Filter: (c1 > 100000)
Rows Removed by Filter: 100000
Planning Time: 0.879 ms
Execution Time: 49.594 ms
(10 rows)
加上 id > 100000，id有索引
postgres=# explain analyze select * from test where id > 100000 order by id,c1 limit 10;
QUERY PLAN
---------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=0.46..1.19 rows=10 width=25) (actual time=0.160..0.164 rows=10 loops=1)
-> Incremental Sort (cost=0.46..65542.05 rows=899386 width=25) (actual time=0.148..0.150 rows=10 loops=1)
Sort Key: id, c1
Presorted Key: id
Full-sort Groups: 1 Sort Method: quicksort Average Memory: 25kB Peak Memory: 25kB
-> Index Scan using i_test_id on test (cost=0.42..25069.68 rows=899386 width=25) (actual time=0.115..0.119 rows=11 loops=1)
Index Cond: (id > 100000)
Planning Time: 0.408 ms
Execution Time: 0.258 ms
(9 rows)
可以看到即使where条件没有索引，排序字段有索引也可以使用增量排序功能，而且效果也还不错。做了一个过滤操作 Filter: (c1 > 100000)
PG 13 多字段排序
- 根据id，c1，c2进行排序，一样可以走增量排序
postgres=# explain analyze select * from test order by id,c1,c2 limit 10;
QUERY PLAN
----------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=0.46..1.16 rows=10 width=25) (actual time=0.175..0.179 rows=10 loops=1)
-> Incremental Sort (cost=0.46..70373.03 rows=1000000 width=25) (actual time=0.172..0.174 rows=10 loops=1)
Sort Key: id, c1, c2
Presorted Key: id
Full-sort Groups: 1 Sort Method: quicksort Average Memory: 25kB Peak Memory: 25kB
-> Index Scan using i_test_id on test (cost=0.42..25373.02 rows=1000000 width=25) (actual time=0.126..0.130 rows=11 loops=1)
Planning Time: 0.485 ms
Execution Time: 0.237 ms
(8 rows)
PG 13 join
- 复制一张test2
postgres=# create table test2 as select * from test;
SELECT 1000000
postgres=# create index i_test2_id on test2(id);
CREATE INDEX
- join连表查询，并且排序字段test.id，test.c1
postgres=# explain analyze select *from test join test2 on test.id = test2.id order by test.id,test.c1 limit 10;
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=1.93..3.04 rows=10 width=50) (actual time=0.089..0.092 rows=10 loops=1)
-> Incremental Sort (cost=1.93..110738.33 rows=1000000 width=50) (actual time=0.087..0.089 rows=10 loops=1)
Sort Key: test.id, test.c1
Presorted Key: test.id
Full-sort Groups: 1 Sort Method: quicksort Average Memory: 26kB Peak Memory: 26kB
-> Merge Join (cost=1.85..65738.33 rows=1000000 width=50) (actual time=0.044..0.068 rows=11 loops=1)
Merge Cond: (test.id = test2.id)
-> Index Scan using i_test_id on test (cost=0.42..25373.02 rows=1000000 width=25) (actual time=0.022..0.036 rows=11 loops=1)
-> Index Scan using i_test2_id on test2 (cost=0.42..25373.02 rows=1000000 width=25) (actual time=0.014..0.018 rows=11 loops=1)
Planning Time: 0.599 ms
Execution Time: 0.174 ms
(11 rows)
postgres=# set enable_incremental_sort=off ;
SET
postgres=# explain analyze select *from test join test2 on test.id = test2.id order by test.id,test.c1 limit 10;
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=87347.97..87347.99 rows=10 width=50) (actual time=1964.394..1964.407 rows=10 loops=1)
-> Sort (cost=87347.97..89847.97 rows=1000000 width=50) (actual time=1964.391..1964.402 rows=10 loops=1)
Sort Key: test.id, test.c1
Sort Method: top-N heapsort Memory: 26kB
-> Merge Join (cost=1.85..65738.33 rows=1000000 width=50) (actual time=0.070..1690.949 rows=1000000 loops=1)
Merge Cond: (test.id = test2.id)
-> Index Scan using i_test_id on test (cost=0.42..25373.02 rows=1000000 width=25) (actual time=0.042..571.732 rows=1000000 loops=1)
-> Index Scan using i_test2_id on test2 (cost=0.42..25373.02 rows=1000000 width=25) (actual time=0.017..585.722 rows=1000000 loops=1)
Planning Time: 1.292 ms
Execution Time: 1964.517 ms
(10 rows)
join后排序也可以走增量排序，使用增量排序耗时：0.174 ms，而关闭增量后耗时1964.517 ms
- 如果join后排序的字段来自不同的表test.id，test2.c1
postgres=# explain analyze select *from test join test2 on test.id = test2.id order by test.id,test2.c1 limit 10;
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=1.93..3.04 rows=10 width=50) (actual time=0.151..0.155 rows=10 loops=1)
-> Incremental Sort (cost=1.93..110738.33 rows=1000000 width=50) (actual time=0.149..0.151 rows=10 loops=1)
Sort Key: test.id, test2.c1
Presorted Key: test.id
Full-sort Groups: 1 Sort Method: quicksort Average Memory: 26kB Peak Memory: 26kB
-> Merge Join (cost=1.85..65738.33 rows=1000000 width=50) (actual time=0.075..0.088 rows=11 loops=1)
Merge Cond: (test.id = test2.id)
-> Index Scan using i_test_id on test (cost=0.42..25373.02 rows=1000000 width=25) (actual time=0.040..0.044 rows=11 loops=1)
-> Index Scan using i_test2_id on test2 (cost=0.42..25373.02 rows=1000000 width=25) (actual time=0.025..0.028 rows=11 loops=1)
Planning Time: 0.778 ms
Execution Time: 0.230 ms
(11 rows)
postgres=# set enable_incremental_sort=off ;
SET
postgres=# explain analyze select *from test join test2 on test.id = test2.id order by test.id,test2.c1 limit 10;
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=87347.97..87347.99 rows=10 width=50) (actual time=1493.513..1493.519 rows=10 loops=1)
-> Sort (cost=87347.97..89847.97 rows=1000000 width=50) (actual time=1493.510..1493.513 rows=10 loops=1)
Sort Key: test.id, test2.c1
Sort Method: top-N heapsort Memory: 26kB
-> Merge Join (cost=1.85..65738.33 rows=1000000 width=50) (actual time=0.065..1228.403 rows=1000000 loops=1)
Merge Cond: (test.id = test2.id)
-> Index Scan using i_test_id on test (cost=0.42..25373.02 rows=1000000 width=25) (actual time=0.027..318.044 rows=1000000 loops=1)
-> Index Scan using i_test2_id on test2 (cost=0.42..25373.02 rows=1000000 width=25) (actual time=0.027..390.231 rows=1000000 loops=1)
Planning Time: 0.761 ms
Execution Time: 1493.685 ms
(10 rows)
join后排序的字段来自不同的表test.id，test2.c1，也可以走增量排序，开启增量耗时：0.230，关闭后耗时：1493.685 ms
来看看一个比较慢的SQL：
- 这个SQL两表关联，而且使用了c2=2这一列全部为2，并且使用offset 100000
postgres=# explain analyze select *from test join test2 on test.id = test2.id where test.c2 = 2 order by test.id,test2.c1 limit 10 offset 100000;
QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=11325.58..11326.72 rows=10 width=50) (actual time=198.125..198.131 rows=10 loops=1)
-> Incremental Sort (cost=2.02..113237.64 rows=1000000 width=50) (actual time=0.127..193.661 rows=100010 loops=1)
Sort Key: test.id, test2.c1
Presorted Key: test.id
Full-sort Groups: 3126 Sort Method: quicksort Average Memory: 29kB Peak Memory: 29kB
-> Merge Join (cost=1.94..68237.64 rows=1000000 width=50) (actual time=0.052..152.908 rows=100011 loops=1)
Merge Cond: (test.id = test2.id)
-> Index Scan using i_test_id on test (cost=0.42..27873.02 rows=1000000 width=25) (actual time=0.026..46.138 rows=100011 loops=1)
Filter: (c2 = 2)
-> Index Scan using i_test2_id on test2 (cost=0.42..25373.02 rows=1000000 width=25) (actual time=0.020..51.088 rows=100011 loops=1)
Planning Time: 0.707 ms
Execution Time: 198.252 ms
(12 rows)
因为增量排序的缘故，查询还是很快
- 如果我们关闭增量排序功能
postgres=# explain analyze select *from test join test2 on test.id = test2.id where test.c2 = 2 order by test.id,test2.c1 limit 10 offset 100000;
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=156536.56..156536.59 rows=10 width=50) (actual time=2496.085..2496.093 rows=10 loops=1)
-> Sort (cost=156286.56..158786.56 rows=1000000 width=50) (actual time=2469.643..2491.429 rows=100010 loops=1)
Sort Key: test.id, test2.c1
Sort Method: external merge Disk: 72432kB
-> Merge Join (cost=1.94..68237.64 rows=1000000 width=50) (actual time=0.082..1371.433 rows=1000000 loops=1)
Merge Cond: (test.id = test2.id)
-> Index Scan using i_test_id on test (cost=0.42..27873.02 rows=1000000 width=25) (actual time=0.040..433.114 rows=1000000 loops=1)
Filter: (c2 = 2)
-> Index Scan using i_test2_id on test2 (cost=0.42..25373.02 rows=1000000 width=25) (actual time=0.033..401.784 rows=1000000 loops=1)
Planning Time: 0.807 ms
Execution Time: 2530.205 ms
(11 rows)
这个SQL耗时 2530.205 ms，和198.252 ms比增量排序提升还是很明显
但是我们观察到上面的SQL中使用id进行关联，且用id排序的时候查询效率较高，如果排序的字段换成crt_time效果如何?
postgres=# explain analyze select *from test join test2 on test.id = test2.id where test.c2 = 2 order by test.crt_time,test2.c1 limit 10 offset 100000;
QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------------------------------------
Limit (cost=156536.56..156536.59 rows=10 width=50) (actual time=2702.107..2702.133 rows=10 loops=1)
-> Sort (cost=156286.56..158786.56 rows=1000000 width=50) (actual time=2667.324..2697.033 rows=100010 loops=1)
Sort Key: test.crt_time, test2.c1
Sort Method: external merge Disk: 72432kB
-> Merge Join (cost=1.94..68237.64 rows=1000000 width=50) (actual time=0.161..1524.794 rows=1000000 loops=1)
Merge Cond: (test.id = test2.id)
-> Index Scan using i_test_id on test (cost=0.42..27873.02 rows=1000000 width=25) (actual time=0.074..488.803 rows=1000000 loops=1)
Filter: (c2 = 2)
-> Index Scan using i_test2_id on test2 (cost=0.42..25373.02 rows=1000000 width=25) (actual time=0.073..487.688 rows=1000000 loops=1)
Planning Time: 1.835 ms
Execution Time: 2746.486 ms
(11 rows)
当join关联的字段和order by的字段不一样时，虽然order by的字段有索引但也不能走，如果字段一致那么也能利用增量排序。
使用test.crt_time排序和上面关闭增量排序执行计划一样
总结
增量排序对于单表多字段排序来说效率还是提升明显
join连表查询如果关联的键和排序键一样也能走增量排序，如果不一样则不能走增量排序
参考资料：
https://postgres.fun/20200721193000.html
https://mp.weixin.qq.com/s/mBIL2uzIHB7qVByBIVRmhg
PostgreSQL 13支持增量排序(Incremental Sorting)的更多相关文章
- 关于VC++的增量链接(Incremental Linking)
增量链接(Incremental Linking)这个词语在使用Visual C++时经常会遇到(其实不只是VS系列,其它链接器也有这个特性), 就比如经常遇到的:上一个增量链接没有生成它, 正在执行 ...
- PostgreSQL 13.4的安装记录
PostgreSQL 13.4的安装记录 想着MySQL被Oracle给买了,总得做点别的准备不是,找了找别的开源的关系型数据库,貌似PostgreSQL的评价很不错,就试试先 因为是window10 ...
- 自建数据源（RSO2），并支持增量
声明:原创作品,转载时请注明文章来自SAP师太技术博客( 博/客/园www.cnblogs.com):www.cnblogs.com/jiangzhengjun,并以超链接形式标明文章原始出处,否则将 ...
- 谢尔排序/缩减增量排序(C++)
谢尔排序/缩减增量排序(C++) 谢尔排序/缩减增量排序: 他通过比较相距一定间隔的元素来工作,各趟比较所用的距离随着算法的进行而减小,直到只比较相邻元素的最后一趟排序为止.(好复杂) 看了一下实现代 ...
- Python数据结构应用5——排序（Sorting）
在具体算法之前,首先来看一下排序算法衡量的标准: 比较:比较两个数的大小的次数所花费的时间. 交换:当发现某个数不在适当的位置时,将其交换到合适位置花费的时间. 冒泡排序(Bubble Sort) 这 ...
- java算法----排序----（6）希尔排序（最小增量排序）
package log; public class Test4 { /** * java算法---希尔排序(最小增量排序) * * @param args */ public static void ...
- 通过orderby关键字，LINQ可以实现升序和降序排序。LINQ还支持次要排序。
通过orderby关键字,LINQ可以实现升序和降序排序.LINQ还支持次要排序. LINQ默认的排序是升序排序,如果你想使用降序排序,就要使用descending关键字. static void M ...
- 拓扑排序 (Topological Sorting)
拓扑排序(Topological Sorting) 一.拓扑排序 含义 构造AOV网络全部顶点的拓扑有序序列的运算称为拓扑排序(Topological Sorting). 在图论中,拓扑排序(Topo ...
- MySQL为什么不支持中文排序？
前言 或许都知道,MySQL不支持中文排序,这样的说法可以说对也可以说也不对.接下来我们分析一下: 首先执行命令,查看编码集: SHOW VARIABLES LIKE 'character_set%' ...
随机推荐
- 如何在jsp界面进行判断再输出不同的值
C标签的out <td> <c:if test="${nowtime eq returntime}"> <c:out value="逾期&q ...
- uni-app 解析后台接口返回的HTML
正常使用rich-text是可以解决问题的,但是在支付宝小程序中不显示,在文档中看到" 支付宝小程序 nodes 属性只支持使用 Array 类型.如果需要支持 HTML String,则需 ...
- eBPF+Ftrace 合璧剑指：no space left on device？
本文地址:https://www.ebpf.top/post/no_space_left_on_devices 最近在生产环境中遇到了几次创建容器报错 "no space left on d ...
- Python夺命20问
1.请观看下列代码并回答问题: import collections from random import choice Card = collection.namedtuple('Card', [' ...
- JavaScript函数中的arguments对象
ECMAScript标准中,每个函数都有一个特殊的内置对象arguments.arguments对象是一个类Array对象(object),用以保存函数接收到的实参副本. 一.内置特性 说它是一个内置 ...
- Istio实践（4）- 故障注入、熔断及ServiceEntry
前言:接上一篇istio多服务应用部署及调用,本文介绍通过流量管理(故障注入.请求超时等)以及ServiceEntry外部服务部署应用 1.设置服务延迟 修改springbootapp-vs-v1.y ...
- 程序员延寿指南「GitHub 热点速览 v.22.17」
很多人对程序员的固有印象之一便是常加班.易"猝死"!近几年的许多报道似乎也进一步加深了这种印象.应该如何更好地健康地活着.敲喜欢的代码呢?HowToLiveLonger 教你如何从 ...
- JVM调优篇
点赞再看,养成习惯,微信搜索「小大白日志」关注这个搬砖人. 文章不定期同步公众号,还有各种一线大厂面试原题.我的学习系列笔记. 基础概念 一般JVM调优,重点在于调整JVM堆大小.调整垃圾回收器 jv ...
- [AcWing 53] 最小的 k 个数
堆排序 点击查看代码 class Solution { public: vector<int> getLeastNumbers_Solution(vector<int> inp ...
- umi框架应用服务端SSR，实现数据预渲染
当我们的应用使用服务端渲染的方式时,可能需要把初始化加载的数据例如推荐等不需要用户输入的内容直接渲染获取,也有利于SEO. 上一篇已经实现服务端渲染,本次实现服务端获取数据后在做渲染. 利用getIn ...