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。