exist和left join 性能对比
今天遇到一个性能问题，再调优过程中发现耗时最久的计划是exist 部分涉及的三个表。
然后计划用left join 来替换exist，然后查询了很多资料，大部分都说exist和left join 性能差不多。 为了验证这一结论进行了如下实验
步骤如下
1、创建测试表
drop table app_family;
CREATE TABLE app_family (
"family_id" character varying(32 char) NOT NULL,
"application_id" character varying(32 char) NULL,
"family_number" character varying(50 char) ,
"household_register_number" character varying(50 char),
"poverty_reason" character varying(32 char),
CONSTRAINT "pk_app_family_idpk" PRIMARY KEY (family_id));
insert into app_family select generate_series(1,1000000),generate_series(1,1000000),'aaaa','aaa','bbb' from dual ;
create table app_family2 as select * from app_family;
create table app_memeber as select * from app_family;
2、验证两张表join和exist 性能对比
语句1、两张表exist
explain analyze select a1.application_id,a1.family_id from app_family a1 where
a1.family_id >1000 and
EXISTS(
SELECT
1
FROM
app_family2 a2
WHERE
a2.application_id=a1.application_id
and a2.family_id > 500000
)
总计用时646.203 ms
----------------------------------------------------------------------------------------------------------------------------------------------------
Gather (cost=16927.11..44466.84 rows=111111 width=12) (actual time=354.314..621.714 rows=500000 loops=1)
Workers Planned: 2
Workers Launched: 2
-> Parallel Hash Semi Join (cost=15927.11..32355.74 rows=46296 width=12) (actual time=355.657..512.049 rows=166667 loops=3)
Hash Cond: ((a1.application_id)::text = (a2.application_id)::text)
-> Parallel Seq Scan on app_family a1 (cost=0.00..13648.00 rows=138889 width=12) (actual time=0.222..111.618 rows=333000 loops=3)
Filter: ((family_id)::integer > 1000)
Rows Removed by Filter: 333
-> Parallel Hash (cost=13648.00..13648.00 rows=138889 width=6) (actual time=149.203..149.204 rows=166667 loops=3)
Buckets: 131072 Batches: 8 Memory Usage: 3520kB
-> Parallel Seq Scan on app_family2 a2 (cost=0.00..13648.00 rows=138889 width=6) (actual time=48.576..109.251 rows=166667 loops=3)
Filter: ((family_id)::integer > 500000)
Rows Removed by Filter: 166667
Planning Time: 0.145 ms
Execution Time: 645.095 ms
(15 rows)
Time: 646.203 ms
kingbase=#
语句2 两张表join
explain analyze select a1.application_id,a1.family_id from app_family a1 LEFT JOIN app_family2 a2 ON a2.application_id=a1.application_id
WHERE a1.family_id >1000 AND a2.family_id > 500000
总计执行时间624.211 ms
---------------------------------------------------------------------------------------------------------------------------------------------------
Gather (cost=16927.11..44300.95 rows=111111 width=12) (actual time=349.752..601.304 rows=500000 loops=1)
Workers Planned: 2
Workers Launched: 2
-> Parallel Hash Join (cost=15927.11..32189.85 rows=46296 width=12) (actual time=337.548..508.139 rows=166667 loops=3)
Hash Cond: ((a1.application_id)::text = (a2.application_id)::text)
-> Parallel Seq Scan on app_family a1 (cost=0.00..13648.00 rows=138889 width=12) (actual time=0.087..111.949 rows=333000 loops=3)
Filter: ((family_id)::integer > 1000)
Rows Removed by Filter: 333
-> Parallel Hash (cost=13648.00..13648.00 rows=138889 width=6) (actual time=131.718..131.719 rows=166667 loops=3)
Buckets: 131072 Batches: 8 Memory Usage: 3488kB
-> Parallel Seq Scan on app_family2 a2 (cost=0.00..13648.00 rows=138889 width=6) (actual time=31.730..90.917 rows=166667 loops=3)
Filter: ((family_id)::integer > 500000)
Rows Removed by Filter: 166667
Planning Time: 0.093 ms
Execution Time: 623.465 ms
(15 rows)
Time: 624.211 ms
两张表场景总结
针对两张表的对比可以发现join还相对满了10几ms但是总的来说两边 差异不大。所以再两张表的关联情况下 join和exist 性能相近。
3、验证3张表join和exist 性能对比
语句1 三张表exist
本场景最开始执行时 exit 用户6 s多，原因时用到了内存排序，后来调整了work_mem 排除了内存排序的影响，最终执行时间
2911.146 ms
explain analyze select a1.application_id,a1.family_id from app_family a1 ,app_family2 a2 where
a1.family_id >1000 and a2.family_id < 900000 and
EXISTS(
SELECT
1
FROM
app_memeber m
WHERE
m.application_id=a1.application_id
and m.family_id=a2.family_id
)
------------------------------------------------------------------------------------------------------------------------------------------------------
--
Gather (cost=61282.11..88664.67 rows=111111 width=12) (actual time=2112.079..2847.233 rows=898999 loops=1)
Workers Planned: 2
Workers Launched: 2
-> Parallel Hash Join (cost=60282.11..76553.57 rows=46296 width=12) (actual time=2119.345..2705.935 rows=299666 loops=3)
Hash Cond: ((m.family_id)::text = (a2.family_id)::text)
-> Hash Join (cost=44898.00..60455.72 rows=138889 width=18) (actual time=1885.923..2264.850 rows=333000 loops=3)
Hash Cond: ((a1.application_id)::text = (m.application_id)::text)
-> Parallel Seq Scan on app_family a1 (cost=0.00..13648.00 rows=138889 width=12) (actual time=0.091..109.196 rows=333000 loops=3)
Filter: ((family_id)::integer > 1000)
Rows Removed by Filter: 333
-> Hash (cost=32398.00..32398.00 rows=1000000 width=12) (actual time=1880.027..1880.028 rows=1000000 loops=3)
Buckets: 1048576 Batches: 1 Memory Usage: 52897kB
-> HashAggregate (cost=22398.00..32398.00 rows=1000000 width=12) (actual time=957.973..1382.683 rows=1000000 loops=3)
Group Key: (m.application_id)::text, (m.family_id)::text
-> Seq Scan on app_memeber m (cost=0.00..17398.00 rows=1000000 width=12) (actual time=0.047..247.902 rows=1000000 loops=3
)
-> Parallel Hash (cost=13648.00..13648.00 rows=138889 width=6) (actual time=231.705..231.706 rows=300000 loops=3)
Buckets: 1048576 (originally 524288) Batches: 1 (originally 1) Memory Usage: 47552kB
-> Parallel Seq Scan on app_family2 a2 (cost=0.00..13648.00 rows=138889 width=6) (actual time=0.039..100.756 rows=300000 loops=3)
Filter: ((family_id)::integer < 900000)
Rows Removed by Filter: 33334
Planning Time: 0.359 ms
Execution Time: 2911.146 ms
(22 rows)
语句2 三张表join
为了保证语句的一致性，三张表的join顺序保持和语句1的执行计划中的顺序一致，join总计用时1476.651 ms
explain analyze select a1.application_id,a1.family_id from app_family a1
left join app_memeber m on a1.application_id = m.application_id LEFT JOIN app_family2 a2 ON m.family_id = a2.family_id
WHERE a1.family_id >1000 AND a2.family_id < 900000
Gather (cost=32990.22..64898.93 rows=111111 width=12) (actual time=993.681..1436.895 rows=898999 loops=1)
Workers Planned: 2
Workers Launched: 2
-> Parallel Hash Join (cost=31990.22..52787.83 rows=46296 width=12) (actual time=982.512..1241.385 rows=299666 loops=3)
Hash Cond: ((m.application_id)::text = (a1.application_id)::text)
-> Parallel Hash Join (cost=15927.11..34245.98 rows=138889 width=6) (actual time=377.411..635.945 rows=300000 loops=3)
Hash Cond: ((m.family_id)::text = (a2.family_id)::text)
-> Parallel Seq Scan on app_memeber m (cost=0.00..11564.67 rows=416667 width=12) (actual time=0.034..59.470 rows=333333 loops=3)
-> Parallel Hash (cost=13648.00..13648.00 rows=138889 width=6) (actual time=232.286..232.287 rows=300000 loops=3)
Buckets: 131072 (originally 131072) Batches: 16 (originally 8) Memory Usage: 3296kB
-> Parallel Seq Scan on app_family2 a2 (cost=0.00..13648.00 rows=138889 width=6) (actual time=0.030..104.370 rows=300000 loops=
3)
Filter: ((family_id)::integer < 900000)
Rows Removed by Filter: 33334
-> Parallel Hash (cost=13648.00..13648.00 rows=138889 width=12) (actual time=271.185..271.185 rows=333000 loops=3)
Buckets: 131072 (originally 131072) Batches: 16 (originally 8) Memory Usage: 4032kB
-> Parallel Seq Scan on app_family a1 (cost=0.00..13648.00 rows=138889 width=12) (actual time=0.091..129.188 rows=333000 loops=3)
Filter: ((family_id)::integer > 1000)
Rows Removed by Filter: 333
Planning Time: 0.140 ms
Execution Time: 1475.305 ms
(20 rows)
Time: 1476.651 ms (00:01.477)
总结三张表场景
在三张表的场景下exist用时2911.146 ms ，join用时1476.651 ms 可见 join的顺序明显优于exist。
在三张表的场景下可以看到，针对中间表appmember扫描时， exist语句用到HashAggregate 并做了 Group Key，所以导致exist 执行时间增加。如果work_mem 配置不合适时间会更长。
exist和left join 性能对比的更多相关文章
- Go 字符串连接+=与strings.Join性能对比
Go字符串连接 对于字符串的连接大致有两种方式: 1.通过+号连接 func StrPlus1(a []string) string { var s, sep string for i := 0; i ...
- 自己写的轻量级PHP框架trig与laravel5.1,yii2性能对比
看了下当前最热门的php开发框架,想对比一下自己写的框架与这些框架的性能对比.先看下当前流行框架的投票情况. 看结果对比,每个测试脚本做了一个数据库的联表查询并进行print_r输出,查询的sql语句 ...
- SQL点滴10—使用with语句来写一个稍微复杂sql语句，附加和子查询的性能对比
原文:SQL点滴10-使用with语句来写一个稍微复杂sql语句,附加和子查询的性能对比 今天偶尔看到sql中也有with关键字,好歹也写了几年的sql语句,居然第一次接触,无知啊.看了一位博主的文章 ...
- python3下multiprocessing、threading和gevent性能对比----暨进程池、线程池和协程池性能对比
python3下multiprocessing.threading和gevent性能对比----暨进程池.线程池和协程池性能对比 标签: python3 / 线程池 / multiprocessi ...
- SQL Server-聚焦IN VS EXISTS VS JOIN性能分析（十九）
前言 本节我们开始讲讲这一系列性能比较的终极篇IN VS EXISTS VS JOIN的性能分析,前面系列有人一直在说场景不够,这里我们结合查询索引列.非索引列.查询小表.查询大表来综合分析,简短的内 ...
- [原] KVM 环境下MySQL性能对比
KVM 环境下MySQL性能对比 标签(空格分隔): Cloud2.0 [TOC] 测试目的 对比MySQL在物理机和KVM环境下性能情况 压测标准 压测遵循单一变量原则,所有的对比都是只改变一个变量 ...
- 浅谈C++之冒泡排序、希尔排序、快速排序、插入排序、堆排序、基数排序性能对比分析之后续补充说明（有图有真相）
如果你觉得我的有些话有点唐突,你不理解可以想看看前一篇<C++之冒泡排序.希尔排序.快速排序.插入排序.堆排序.基数排序性能对比分析>. 这几天闲着没事就写了一篇<C++之冒泡排序. ...
- Java--Stream，NIO ByteBuffer，NIO MappedByteBuffer性能对比
目前Java中最IO有多种文件读取的方法,本文章对比Stream,NIO ByteBuffer,NIO MappedByteBuffer的性能,让我们知道到底怎么能写出性能高的文件读取代码. pack ...
- C正则库做DNS域名验证时的性能对比
C正则库做DNS域名验证时的性能对比 本文对C的正则库regex和pcre在做域名验证的场景下做评测. 验证DNS域名的正则表达式为: "^[0-9a-zA-Z_-]+(\\.[0-9a ...
- 开发语言性能对比，C++、Java、Python、LUA、TCC
一直想做开发语言性能对比,刚好有时间都做了给大家参考一下, 编译类:C++和Java表现还不错 脚本类:TCC脚本动态运行C语言,性能比其他脚本快好多... 想玩TCC的同学下载测试包,TCC目录下修 ...
随机推荐
- 初探富文本之文档diff算法
初探富文本之文档diff算法 当我们实现在线文档的系统时,通常需要考虑到文档的版本控制与审核能力,并且这是这是整个文档管理流程中的重要环节,那么在这个环节中通常就需要文档的diff能力,这样我们就可以 ...
- 发送HTML模板邮件
概述 为了增强邮件内容展示的样式,可以将普通的文本邮件转换为HTML内容格式. 在Java中,可以通过页面模板技术来实现.具体来说,可以使用Thymeleaf模板. 具体实现 首先,在项目中引入Thy ...
- Django之第三方平台QQ授权登录的实现
接入指南:https://wiki.connect.qq.com/成为开发者 准备工作 成为开发者 首先要有一个开发者账号,https://connect.qq.com/ 登录后点击用户头像,修改个人 ...
- 无所不谈,百无禁忌,Win11本地部署无内容审查中文大语言模型CausalLM-14B
目前流行的开源大语言模型大抵都会有内容审查机制,这并非是新鲜事,因为之前chat-gpt就曾经被"玩"坏过,如果没有内容审查,恶意用户可能通过精心设计的输入(prompt)来操纵L ...
- 使用 MyBatis 操作 Nebula Graph 的实践
本文首发于 Nebula Graph Community 公众号 我最近注意到很多同学对于 ORM 框架的需求比较迫切,而且有热心的同学已经捐赠了自己开发的项目,Nebula 社区也在 working ...
- PostgreSql一个月学习计划
1.背景 国内使用数据库最多的莫过于mysql,大部分程序员第一次接触数据库就是mysql.(毕竟免费的 = =!)但近年来,有一些黑马出现(如下图),其中表现最突出的莫过于PostgreSQL.特规 ...
- 解决windows11远程连接阿里云Centos7
本地连接CentOs7时报错 Permission denied (publickey,gssapi-keyex,gssapi-with-mic). 网上大部分说的是去修改 vim /etc/ss ...
- 英语字母z解析.drawio
英语字母z解析.drawio
- 4时4态 加被动 例句：I will have been being done - will have be be do - 频率副词位置
4时4态 频率副词的用法和位置:放在实义动词之前.放在be 动词之后.放在情态动词之后. 频率副词的位置一般是放在实义动词之前.放在be 动词之后.放在情态动词之后.放在be动词之后:She is s ...
- pod的拉取和重启策略
在Kubernetes中,Pod的拉取策略和重启策略可以通过YAML配置文件来定义. Pod的拉取策略 Pod的拉取策略指的是Kubernetes在创建或重启Pod时,如何获取Pod所需的容器镜像.这 ...