HTAP数据库 PostgreSQL 场景与性能测试之 26 - (OLTP) NOT IN、NOT EXISTS 查询
背景
PostgreSQL是一个历史悠久的数据库，历史可以追溯到1973年，最早由2014计算机图灵奖得主，关系数据库的鼻祖Michael_Stonebraker 操刀设计，PostgreSQL具备与Oracle类似的功能、性能、架构以及稳定性。
PostgreSQL社区的贡献者众多，来自全球各个行业，历经数年，PostgreSQL 每年发布一个大版本，以持久的生命力和稳定性著称。
2017年10月，PostgreSQL 推出10 版本，携带诸多惊天特性，目标是胜任OLAP和OLTP的HTAP混合场景的需求：
《最受开发者欢迎的HTAP数据库PostgreSQL 10特性》
1、多核并行增强
2、fdw 聚合下推
3、逻辑订阅
4、分区
5、金融级多副本
6、json、jsonb全文检索
7、还有插件化形式存在的特性，如 向量计算、JIT、SQL图计算、SQL流计算、分布式并行计算、时序处理、基因测序、化学分析、图像分析 等。
在各种应用场景中都可以看到PostgreSQL的应用：
PostgreSQL近年来的发展非常迅猛，从知名数据库评测网站dbranking的数据库评分趋势，可以看到PostgreSQL向上发展的趋势：
从每年PostgreSQL中国召开的社区会议，也能看到同样的趋势，参与的公司越来越多，分享的公司越来越多，分享的主题越来越丰富，横跨了 传统企业、互联网、医疗、金融、国企、物流、电商、社交、车联网、共享XX、云、游戏、公共交通、航空、铁路、军工、培训、咨询服务等 行业。
接下来的一系列文章，将给大家介绍PostgreSQL的各种应用场景以及对应的性能指标。
环境
环境部署方法参考：
《PostgreSQL 10 + PostGIS + Sharding(pg_pathman) + MySQL(fdw外部表) on ECS 部署指南(适合新用户)》
阿里云 ECS：56核，224G，1.5TB*2 SSD云盘。
操作系统：CentOS 7.4 x64
数据库版本：PostgreSQL 10
PS：ECS的CPU和IO性能相比物理机会打一定的折扣，可以按下降1倍性能来估算。跑物理主机可以按这里测试的性能乘以2来估算。
场景 - NOT IN、NOT EXISTS 查询 (OLTP)
1、背景
not in 查询，多用在排除多个输入值场景。
实际上PostgreSQL支持很多种排除多个输入值的语法。
1、not in (...)
2、not in (table or subquery or srf)
3、<> all (array)
4、not exists (select 1 from (values (),(),...) as t(id) where x.?=t.id)
5、<>? and <>? and <>? and .....
6、left join others b on (a.?=b.?) where b.* is null
7、select ? from a except select ? from b，适用于输出字段与条件字段相同的情形。
他们的执行计划分别如下，（1亿记录，排除多个输入值。）：
表越大、或Filter的值越多，使用 left join, not exist, except 的效果越好。
postgres=# explain select * from a where id not in (1,2,3,4,5);
QUERY PLAN
--------------------------------------------------------------
Seq Scan on a (cost=0.00..255958.10 rows=10000001 width=45)
Filter: (id <> ALL ('{1,2,3,4,5}'::integer[]))
(2 rows)
postgres=# explain select * from a where id <> all (array[1,2,3,4,5]);
QUERY PLAN
--------------------------------------------------------------
Seq Scan on a (cost=0.00..255958.10 rows=10000001 width=45)
Filter: (id <> ALL ('{1,2,3,4,5}'::integer[]))
(2 rows)
postgres=# explain select * from a where id <> all (array(select generate_series(1,10)));
QUERY PLAN
-------------------------------------------------------------
Seq Scan on a (cost=5.02..318463.15 rows=9999996 width=45)
Filter: (id <> ALL ($0))
InitPlan 1 (returns $0)
-> ProjectSet (cost=0.00..5.02 rows=1000 width=4)
-> Result (cost=0.00..0.01 rows=1 width=0)
(5 rows)
postgres=# explain select * from a where id <> all (array(select id from (values (1),(2),(3),(4),(5)) t (id)));
QUERY PLAN
---------------------------------------------------------------------
Seq Scan on a (cost=0.06..318458.20 rows=9999996 width=45)
Filter: (id <> ALL ($0))
InitPlan 1 (returns $0)
-> Values Scan on "*VALUES*" (cost=0.00..0.06 rows=5 width=4)
(4 rows)
postgres=# explain select * from a where id not in (select id from (values (1),(2),(3),(4),(5)) t (id));
QUERY PLAN
---------------------------------------------------------------------
Seq Scan on a (cost=0.07..218458.15 rows=5000003 width=45)
Filter: (NOT (hashed SubPlan 1))
SubPlan 1
-> Values Scan on "*VALUES*" (cost=0.00..0.06 rows=5 width=4)
(4 rows)
postgres=# explain select * from a where not exists (select 1 from (values (1),(2),(3),(4),(5)) t (id) where t.id=a.id);
QUERY PLAN
-----------------------------------------------------------------------------------
Merge Anti Join (cost=0.56..301364.14 rows=10000001 width=45)
Merge Cond: (a.id = "*VALUES*".column1)
-> Index Scan using a_pkey on a (cost=0.43..276363.92 rows=10000006 width=45)
-> Sort (cost=0.12..0.13 rows=5 width=4)
Sort Key: "*VALUES*".column1
-> Values Scan on "*VALUES*" (cost=0.00..0.06 rows=5 width=4)
(6 rows)
postgres=# explain select * from a where id<>1 and id<>2 and id<>3 and id<>4 and id<>5;
QUERY PLAN
-------------------------------------------------------------------------------
Seq Scan on a (cost=0.00..318458.14 rows=10000001 width=45)
Filter: ((id <> 1) AND (id <> 2) AND (id <> 3) AND (id <> 4) AND (id <> 5))
(2 rows)
postgres=# explain with t1 as (select id from (values (1),(2),(3),(4),(5)) as t(id))
select a.* from a left join t1 b on (a.id=b.id) where b.* is null;
QUERY PLAN
---------------------------------------------------------------------
Hash Left Join (cost=0.23..230958.36 rows=50000 width=45)
Hash Cond: (a.id = b.id)
Filter: (b.* IS NULL)
CTE t1
-> Values Scan on "*VALUES*" (cost=0.00..0.06 rows=5 width=4)
-> Seq Scan on a (cost=0.00..193458.06 rows=10000006 width=45)
-> Hash (cost=0.10..0.10 rows=5 width=32)
-> CTE Scan on t1 b (cost=0.00..0.10 rows=5 width=32)
(8 rows)
postgres=# explain select id from a except select id from (values (1),(2),(3),(4),(5)) as t(id);
QUERY PLAN
-----------------------------------------------------------------------------------------------
SetOp Except (cost=1538166.72..1588166.78 rows=10000006 width=8)
-> Sort (cost=1538166.72..1563166.75 rows=10000011 width=8)
Sort Key: "*SELECT* 1".id
-> Append (cost=0.00..293458.23 rows=10000011 width=8)
-> Subquery Scan on "*SELECT* 1" (cost=0.00..293458.12 rows=10000006 width=8)
-> Seq Scan on a (cost=0.00..193458.06 rows=10000006 width=4)
-> Subquery Scan on "*SELECT* 2" (cost=0.00..0.11 rows=5 width=8)
-> Values Scan on "*VALUES*" (cost=0.00..0.06 rows=5 width=4)
(8 rows)
2、设计
1亿记录，查询匹配多个输入值的性能。分别输入1,10,100,1000,10000,100000,1000000个值作为匹配条件。
1、not in (...)
2、not in (table or subquery or srf)
3、<> all (array)
4、not exists (select 1 from (values (),(),...) as t(id) where x.?=t.id)
5、<>? and <>? and <>? and .....
6、left join others b on (a.?=b.?) where b.* is null
7、select ? from a except select ? from b，适用于输出字段与条件字段相同的情形。
3、准备测试表
create table t_in_test (id int primary key, info text, crt_time timestamp);
4、准备测试函数(可选)
5、准备测试数据
insert into t_in_test select generate_series(1,100000000), md5(random()::text), clock_timestamp();
6、准备测试脚本
set parallel_setup_cost =0;
set parallel_tuple_cost =0;
set max_parallel_workers_per_gather =28;
alter table t_in_test set (parallel_workers =28);
1、not in (...)
1,10,100,1000,10000,100000,1000000 个输入值的测试性能
do language plpgsql $$
declare
arr text;
ts timestamp := clock_timestamp();
mx int8;
begin
for i in 0..6 loop
mx := (1*(10^i))::int8;
select string_agg((random()*100000)::int::text, ',') into arr from generate_series(1, mx);
ts := clock_timestamp();
execute 'select * from t_in_test where id not in ('||arr||')';
raise notice '%: %', mx, clock_timestamp()-ts;
end loop;
end;
$$ ;
2、not in (table or subquery or srf)
1,10,100,1000,10000,100000,1000000 个输入值的测试性能
do language plpgsql $$
declare
arr text;
ts timestamp := clock_timestamp();
mx int8;
begin
for i in 0..6 loop
mx := (1*(10^i))::int8;
ts := clock_timestamp();
perform * from t_in_test where not id in ( select (random()*100000)::int from generate_series(1, mx) );
raise notice '%: %', mx, clock_timestamp()-ts;
end loop;
end;
$$ ;
3、<> all (array)
1,10,100,1000,10000,100000,1000000 个输入值的测试性能
do language plpgsql $$
declare
arr int[];
ts timestamp := clock_timestamp();
mx int8;
begin
for i in 0..6 loop
mx := (1*(10^i))::int8;
select array_agg((random()*100000)::int) into arr from generate_series(1, mx);
ts := clock_timestamp();
perform * from t_in_test where id <> all ( arr );
raise notice '%: %', mx, clock_timestamp()-ts;
end loop;
end;
$$ ;
4、not exists (select 1 from (values (),(),...) as t(id) where x.?=t.id)
1,10,100,1000,10000,100000,1000000 个输入值的测试性能
do language plpgsql $$
declare
ts timestamp := clock_timestamp();
mx int8;
begin
for i in 0..6 loop
mx := (1*(10^i))::int8;
ts := clock_timestamp();
perform * from t_in_test where not exists ( select 1 from ( select (random()*100000)::int id from generate_series(1,mx) ) t where t_in_test.id=t.id );
raise notice '%: %', mx, clock_timestamp()-ts;
end loop;
end;
$$ ;
6、left join others b on (a.?=b.?) where b.* is null
do language plpgsql $$
declare
ts timestamp := clock_timestamp();
mx int8;
begin
for i in 0..6 loop
mx := (1*(10^i))::int8;
ts := clock_timestamp();
perform a.* from t_in_test a left join (select (random()*100000)::int id from generate_series(1,mx)) b on (a.id=b.id) where b.* is null;
raise notice '%: %', mx, clock_timestamp()-ts;
end loop;
end;
$$ ;
7、select ? from a except select ? from b，适用于输出字段与条件字段相同的情形。
do language plpgsql $$
declare
ts timestamp := clock_timestamp();
mx int8;
begin
for i in 0..6 loop
mx := (1*(10^i))::int8;
ts := clock_timestamp();
perform a.id from t_in_test a except select (random()*100000)::int id from generate_series(1,mx);
raise notice '%: %', mx, clock_timestamp()-ts;
end loop;
end;
$$ ;
7、测试
1、not in (...)
1,10,100,1000,10000,100000,1000000 个输入值的测试性能
NOTICE: 1: 00:00:20.760034
NOTICE: 10: 00:00:27.766224
NOTICE: 100: 00:01:22.95002
NOTICE: 1000: 00:10:16.690793
..........
10000开始很久也没跑出来。继续看后面其他方法的测试。
2、not in (table or subquery or srf)
1,10,100,1000,10000,100000,1000000 个输入值的测试性能
-----
3、<> all (array)
1,10,100,1000,10000,100000,1000000 个输入值的测试性能
-----
4、not exists (select 1 from (values (),(),...) as t(id) where x.?=t.id)
1,10,100,1000,10000,100000,1000000 个输入值的测试性能
NOTICE: 1: 00:00:35.253582
NOTICE: 10: 00:00:35.256638
NOTICE: 100: 00:00:35.164034
NOTICE: 1000: 00:00:35.417756
NOTICE: 10000: 00:00:35.205454
NOTICE: 100000: 00:00:35.458987
NOTICE: 1000000: 00:00:35.447743
DO
6、a left join others b on (a.?=b.?) where b.* is null
1,10,100,1000,10000,100000,1000000 个输入值的测试性能
NOTICE: 1: 00:00:36.474715
NOTICE: 10: 00:00:36.53191
NOTICE: 100: 00:00:36.60439
NOTICE: 1000: 00:00:36.534846
NOTICE: 10000: 00:00:36.574136
NOTICE: 100000: 00:00:36.519582
NOTICE: 1000000: 00:00:37.675594
DO
7、select ? from a except select ? from b，适用于输出字段与条件字段相同的情形。
1,10,100,1000,10000,100000,1000000 个输入值的测试性能
NOTICE: 1: 00:00:50.566741
NOTICE: 10: 00:00:50.051715
NOTICE: 100: 00:00:50.098839
NOTICE: 1000: 00:00:49.966196
NOTICE: 10000: 00:00:50.608288
NOTICE: 100000: 00:00:50.715218
NOTICE: 1000000: 00:00:51.794935
DO
TPS
平均响应时间
not exists为例，1亿记录1到100万个点的排他过滤耗时，响应时间均匀，约35秒。
NOTICE: 1: 00:00:35.253582
NOTICE: 10: 00:00:35.256638
NOTICE: 100: 00:00:35.164034
NOTICE: 1000: 00:00:35.417756
NOTICE: 10000: 00:00:35.205454
NOTICE: 100000: 00:00:35.458987
NOTICE: 1000000: 00:00:35.447743
DO
补充PostgreSQL 11 parallel hash join性能
依旧是1亿NOT IN 1到100万，或者not exists的写法。
1、测试数据
postgres=# create table a(id int);
CREATE TABLE
postgres=# create table b(id int);
CREATE TABLE
postgres=# insert into a select generate_series(1,100000000);
INSERT 0 100000000
postgres=# insert into b select generate_series(1,1000000);
INSERT 0 1000000
2、NOT IN 性能(work_mem足够大才会选择hashed)
postgres=# set work_mem ='16GB';
SET
postgres=# explain analyze select count(*) from a where id not in (select id from b limit 1);
QUERY PLAN
-----------------------------------------------------------------------------------------------------------------------------------------
Finalize Aggregate (cost=464968.97..464968.98 rows=1 width=8) (actual time=801.161..801.161 rows=1 loops=1)
-> Gather (cost=464962.40..464968.81 rows=64 width=8) (actual time=801.031..801.146 rows=65 loops=1)
Workers Planned: 64
Workers Launched: 64
-> Partial Aggregate (cost=463962.40..463962.41 rows=1 width=8) (actual time=763.825..763.825 rows=1 loops=65)
-> Parallel Seq Scan on a (cost=0.02..462009.27 rows=781250 width=0) (actual time=0.058..603.051 rows=1538462 loops=65)
Filter: (NOT (hashed SubPlan 1))
Rows Removed by Filter: 0
SubPlan 1
-> Limit (cost=0.00..0.01 rows=1 width=4) (actual time=0.026..0.026 rows=1 loops=65)
-> Seq Scan on b (cost=0.00..14425.00 rows=1000000 width=4) (actual time=0.025..0.025 rows=1 loops=65)
Planning time: 0.067 ms
Execution time: 845.690 ms
(13 rows)
postgres=# explain analyze select count(*) from a where id not in (select id from b);
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------
Finalize Aggregate (cost=481893.95..481893.96 rows=1 width=8) (actual time=1710.007..1710.007 rows=1 loops=1)
-> Gather (cost=481887.38..481893.79 rows=64 width=8) (actual time=1698.510..1709.989 rows=65 loops=1)
Workers Planned: 64
Workers Launched: 64
-> Partial Aggregate (cost=480887.38..480887.39 rows=1 width=8) (actual time=1641.223..1641.223 rows=1 loops=65)
-> Parallel Seq Scan on a (cost=16925.00..478934.26 rows=781250 width=0) (actual time=677.123..1492.332 rows=1523077 loops=65)
Filter: (NOT (hashed SubPlan 1))
Rows Removed by Filter: 15385
SubPlan 1
-> Seq Scan on b (cost=0.00..14425.00 rows=1000000 width=4) (actual time=0.043..158.634 rows=1000000 loops=65)
Planning time: 0.079 ms
Execution time: 1806.930 ms
(12 rows)
3、NOT EXISTS 性能
postgres=# explain analyze select count(*) from a where not exists (select 1 from (select id from b limit 1) b where a.id=b.id);
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------
Finalize Aggregate (cost=482742.42..482742.43 rows=1 width=8) (actual time=876.618..876.618 rows=1 loops=1)
-> Gather (cost=482735.85..482742.26 rows=64 width=8) (actual time=876.470..876.602 rows=65 loops=1)
Workers Planned: 64
Workers Launched: 64
-> Partial Aggregate (cost=481735.85..481735.86 rows=1 width=8) (actual time=840.644..840.644 rows=1 loops=65)
-> Hash Anti Join (cost=0.04..477829.60 rows=1562500 width=0) (actual time=0.257..686.240 rows=1538462 loops=65)
Hash Cond: (a.id = b.id)
-> Parallel Seq Scan on a (cost=0.00..458103.01 rows=1562500 width=4) (actual time=0.020..289.812 rows=1538462 loops=65)
-> Hash (cost=0.02..0.02 rows=1 width=4) (actual time=0.032..0.032 rows=1 loops=65)
Buckets: 1024 Batches: 1 Memory Usage: 9kB
-> Limit (cost=0.00..0.01 rows=1 width=4) (actual time=0.025..0.026 rows=1 loops=65)
-> Seq Scan on b (cost=0.00..14425.00 rows=1000000 width=4) (actual time=0.025..0.025 rows=1 loops=65)
Planning time: 0.147 ms
Execution time: 920.628 ms
(14 rows)
postgres=# explain analyze select count(*) from a where not exists (select 1 from b where a.id=b.id);
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------
Finalize Aggregate (cost=487341.22..487341.23 rows=1 width=8) (actual time=1171.201..1171.201 rows=1 loops=1)
-> Gather (cost=487334.65..487341.06 rows=64 width=8) (actual time=1169.676..1171.185 rows=65 loops=1)
Workers Planned: 64
Workers Launched: 64
-> Partial Aggregate (cost=486334.65..486334.66 rows=1 width=8) (actual time=1110.487..1110.487 rows=1 loops=65)
-> Parallel Hash Anti Join (cost=4776.56..482467.46 rows=1546876 width=0) (actual time=53.768..964.692 rows=1523077 loops=65)
Hash Cond: (a.id = b.id)
-> Parallel Seq Scan on a (cost=0.00..458103.01 rows=1562500 width=4) (actual time=0.023..288.519 rows=1538462 loops=65)
-> Parallel Hash (cost=4581.25..4581.25 rows=15625 width=4) (actual time=35.322..35.322 rows=15385 loops=65)
Buckets: 1048576 Batches: 1 Memory Usage: 48864kB
-> Parallel Seq Scan on b (cost=0.00..4581.25 rows=15625 width=4) (actual time=0.022..2.010 rows=15385 loops=65)
Planning time: 0.129 ms
Execution time: 1259.454 ms
(13 rows)
参考
《PostgreSQL、Greenplum 应用案例宝典《如来神掌》 - 目录》
《PostgreSQL 使用 pgbench 测试 sysbench 相关case》
https://www.postgresql.org/docs/10/static/pgbench.html
《PostgreSQL 与关系代数 (Equi-Join , Semi-Join , Anti-Join , Division)》