背景
哈啰出行旗下包括哈啰单车、助力车、顺风车、打车、电动车、换电等几乎所有业务都基于PostgreSQL数据库构建,PostgreSQL的安全、稳定、高效为哈啰出行的上亿用户提供了强大的基础。随着时间的推移,表里的数据量越来越大,有些需要分库分表处理,有些可以通过简单的分区处理即可,表分区有很多好处:
1):只vacuum 最近分区子表而不是vacuum大表,降低io消耗,减少表膨胀。
2):方便维护表,如创建索引耗时更短,通过清理历史分区释放磁盘空间。
3):减少数据扫描等
01
测试目的
1)测试pg_pathman、native、inherit分区表和不分区表的QPS/TPS性能;2)对比:压测索引键+分区键(有索引)查询/插入方式的QPS/TPS、CPU利用率;02
测试说明
PG实例信息:实例ID:i-bp15xu7930bhkq3urjwm, PG10, 4C8GB。主表下有20个分区表,按月分区,总数据量均为5612.5504万,均匀分布在各分区。查询的对应时间段月分区的的数据量为280万。4个和10个并发压测,压测脚本:/usr/pgsql-10/bin/pgbench -f select/insert.sql -c 4/10 -j 4/10 -n -P 10 -r -T 300/180 -R 10000 -p 7474 -d postgres -U postgres -h locahost
03
测试CASE
04
测试结论
1,在5612.5504万数据量,20个分区情况下select的结论:
1)不分区表相比分区表性能更高,占用cpu更低,qps更高。
2)分区表之间对比,qps相差不多的情况下,pathman分区方式占用cpu更低,10并发下native和inherits占用cpu超过90%以上。
2,在5612.5504万数据量,20个分区情况下insert的结论:
1)不分区表相比分区表性能更高,占用cpu更低,tps较高。
2)分区表之间对比,tps性能相差不多的情况下,pathman分区方式占用cpu更低,native和inherits占用cpu都大于pathman方式。
05
测试环境硬件配置信息
CPU:
postgres@VECS04164:~$ cat /proc/cpuinfo | grep "model name"model name : Intel(R) Xeon(R) Platinum 8163 CPU @ 2.50GHzmodel name : Intel(R) Xeon(R) Platinum 8163 CPU @ 2.50GHzmodel name : Intel(R) Xeon(R) Platinum 8163 CPU @ 2.50GHzmodel name : Intel(R) Xeon(R) Platinum 8163 CPU @ 2.50GHz
内存:
postgres@VECS04164:~$ cat /proc/meminfo | grep -i totalMemTotal: 8193156 kBSwapTotal: 0 kBVmallocTotal: 34359738367 kBHugePages_Total: 0
操作系统版本:
postgres@VECS04164:~$ cat /etc/redhat-releaseCentOS release 6.9 (Final)
需要提前安装好pg_pathman插件:
下载地址:https://download.postgresql.org/pub/repos/yum/10/redhat/rhel-6.9-x86_64/
需要两个rpm包:
pg_pathman10-1.4.13-1.rhel6.x86_64.rpmpython-psycopg2-2.7.4-1.rhel6.x86_64.rpm
pg_pathman分区测试
1,创建pathman分区表
CREATE TABLE pathman_emp_20190710 (emp_id SERIAL,emp_level INTEGER,emp_name TEXT,create_time TIMESTAMP NOT NULL);
2,插入数据:
INSERT INTO pathman_emp_20190710 (create_time, emp_level, emp_name) SELECT g, random() * 6,md5(g::text) FROM generate_series('2016-01-01'::date, '2019-12-31'::date, '1 minute') as g;
3,创建索引
CREATE INDEX ON pathman_emp_20190710(create_time);CREATE INDEX ON pathman_emp_20190710(emp_name);
4,创建分区子表:
SELECT create_range_partitions('pathman_emp_20190710',--主表名'create_time', --分区字段'2016-01-01'::date, --分区起始日期'1 month'::interval, --分区间隔null, --不指定分区数量,根据时间与间隔会自动计算出数量false --默认tue立即迁移数据,false是不迁移数据);
5,验证父表的数据量
select count(*) from only pathman_emp_20190710;
6,将数据并行的迁移到分区子表中:(需要一段时间)
select partition_table_concurrently('pathman_emp_20190710',10000,1.0);
7,等迁移完成,验证父、子表数据量:
select count(*) from only pathman_emp_20190710;postgres=#\dt+
8,编辑测试脚本文件select_pathman_emp_20190710.sql 、insert_pathman_emp_20190710.sql
postgres@VECS04164:~$ cat select_pathman_new.sqlselect * from public.pathman_emp where emp_name ='e0cf722200f2833a04415347324a85f3' and create_time >='2016-07-01' and create_time<'2016-08-01';postgres@VECS04164:~$ cat insert_pathman_new.sqlinsert into pathman_emp_new values(emp_name,emp_level,create_time) values('测试',100,now());
9,调用pgbench进行测试:
/usr/pgsql-10/bin/pgbench -f select_pathman_new.sql -c 4/10 -j 4/10 -n -P 10 -r -T 180 -R 10000 -p 7474 -d postgres -U postgres -h localhost/usr/pgsql-10/bin/pgbench -f insert_pathman_new.sql -c 4/10 -j 4/10 -n -P 10 -r -T 180 -R 10000 -p 7474 -d postgres -U postgres -h localhost
native 分区测试
1,创建native分区表
CREATE TABLE native_emp_20190710 (emp_id SERIAL,emp_level INTEGER,emp_name TEXT,create_time TIMESTAMP NOT NULL)partition by range(create_time);
2,创建子分区:
create table native_emp_20190710_201601 partition of native_emp_20190710 for values from ('20160101') to ('20160201');create table native_emp_20190710_201602 partition of native_emp_20190710 for values from ('20160201') to ('20160301');create table native_emp_20190710_201603 partition of native_emp_20190710 for values from ('20160301') to ('20160401');.........create table native_emp_20190710_201910 partition of native_emp_20190710 for values from ('20191001') to ('20191101');create table native_emp_20190710_201911 partition of native_emp_20190710 for values from ('20191101') to ('20191201');create table native_emp_20190710_201912 partition of native_emp_20190710 for values from ('20191201') to ('20200101');
3,导入数据到native_emp_20190710
insert into native_emp_20190710 select * from pathman_emp_new;
4,创建子分区表索引
CREATE INDEX ON native_emp_20190710_201601(emp_name);
CREATE INDEX ON native_emp_20190710_201602(emp_name);
CREATE INDEX ON native_emp_20190710_201603(emp_name);
...
..
CREATE INDEX ON native_emp_20190710_201910(emp_name);
CREATE INDEX ON native_emp_20190710_201911(emp_name);
CREATE INDEX ON native_emp_20190710_201912(emp_name);
CREATE INDEX ON native_emp_20190710_201601(create_time);
CREATE INDEX ON native_emp_20190710_201602(create_time);
CREATE INDEX ON native_emp_20190710_201603(create_time);
...
..
CREATE INDEX ON native_emp_20190710_201910(create_time);
CREATE INDEX ON native_emp_20190710_201911(create_time);
CREATE INDEX ON native_emp_20190710_201912(create_time);
5,验证数据是否已经到分区表
postgres=# select count(*) from native_emp_20190710;count---------2102400(1 row)postgres=# select count(*) from only native_emp_20190710;count-------0(1 row)
6,编辑测试脚本文件select_native_emp_20190710.sql 、insert_native_emp_20190710.sql
postgres@VECS04164:~$ cat select_native_new.sqlselect * from public.native_emp_new where emp_id=87289589 and create_time >='2016-07-01' and create_time<'2016-08-01' limit 1;postgres@VECS04164:~$ cat insert_native_emp_20190710.sqlinsert into native_emp_20190710 (emp_name,emp_level,create_time) values('测试',100,now());
7,调用pgbench进行测试:
/usr/pgsql-10/bin/pgbench -f select_native_emp_20190710.sql -c 4/10 -j 4/10 -n -P 10 -r -T 180 -R 10000 -p 7474 -d postgres -U postgres -h localhost/usr/pgsql-10/bin/pgbench -f insert_native_emp_20190710 -c 4/10 -j 4/10 -n -P 10 -r -T 180 -R 10000 -p 7474 -d postgres -U postgres -h localhost
Inherits 分区测试
1,建父表:
CREATE TABLE inherits_emp_20190710 (emp_id SERIAL,emp_level INTEGER,emp_name TEXT,create_time TIMESTAMP NOT NULL);
2,继承父表建分区子表:
CREATE TABLE inherits_emp_20190710_201601 ( CHECK ( create_time >= DATE '2016-01-01' AND create_time < DATE '2016-02-01' )) INHERITS (inherits_emp_20190710);CREATE TABLE inherits_emp_20190710_201602 ( CHECK ( create_time >= DATE '2016-02-01' AND create_time < DATE '2016-03-01' )) INHERITS (inherits_emp_20190710);CREATE TABLE inherits_emp_20190710_201603 ( CHECK ( create_time >= DATE '2016-03-01' AND create_time < DATE '2016-04-01' )) INHERITS (inherits_emp_20190710);.....CREATE TABLE inherits_emp_20190710_201910 ( CHECK ( create_time >= DATE '2019-10-01' AND create_time < DATE '2019-11-01' )) INHERITS (inherits_emp_20190710);CREATE TABLE inherits_emp_20190710_201911 ( CHECK ( create_time >= DATE '2019-11-01' AND create_time < DATE '2019-12-01' )) INHERITS (inherits_emp_20190710);CREATE TABLE inherits_emp_20190710_201912 ( CHECK ( create_time >= DATE '2019-12-01' AND create_time < DATE '2020-01-01' )) INHERITS (inherits_emp_20190710);
3,建函数和触发器:
CREATE OR REPLACE FUNCTION insert_inherits_emp_20190710_trigger()RETURNS trigger AS$BODY$DECLAREpartition_date TEXT;partition TEXT;BEGINpartition_date := to_char(NEW.create_time,'YYYYMM');partition := TG_TABLE_NAME || '_' || partition_date;EXECUTE 'INSERT INTO ' || partition || ' SELECT(' || TG_TABLE_NAME || ' ' || quote_literal(NEW) || ').*;';RETURN NULL;END;$BODY$LANGUAGE plpgsql VOLATILECOST 100;
CREATE TRIGGER insert_inherits_emp_20190710BEFORE INSERT ON inherits_emp_20190710FOR EACH ROW EXECUTE PROCEDURE insert_inherits_emp_20190710_trigger();
4,导入数据:
insert into inherits_emp_20190710 select * from pathman_emp_new;INSERT 0 0
5,查看是否进到分区子表
postgres=#\dt+
6,建索引
CREATE INDEX ON inherits_emp_20190710_201601(create_time);
CREATE INDEX ON inherits_emp_20190710_201602(create_time);
CREATE INDEX ON inherits_emp_20190710_201603(create_time);
....
..
CREATE INDEX ON inherits_emp_20190710_201910(create_time);
CREATE INDEX ON inherits_emp_20190710_201911(create_time);
CREATE INDEX ON inherits_emp_20190710_201912(create_time);
CREATE INDEX ON inherits_emp_20190710_201601(emp_name);
CREATE INDEX ON inherits_emp_20190710_201602(emp_name);
CREATE INDEX ON inherits_emp_20190710_201603(emp_name);
....
..
CREATE INDEX ON inherits_emp_20190710_201910(emp_name);
CREATE INDEX ON inherits_emp_20190710_201911(emp_name);
CREATE INDEX ON inherits_emp_20190710_201912(emp_name);
7,编辑测试脚本文件select_inherits_emp_20190710.sql 、insert_inherits_emp_20190710.sql
postgres@VECS04164:~$ cat select_inherits_emp_20190710.sql select * from public.inherits_emp_20190710 where emp_name ='e0cf722200f2833a04415347324a85f3' and create_time >='2016-07-01' and create_time<'2016-08-01';postgres@VECS04164:~$ cat insert_inherits_emp_20190710.sql insert into inherits_emp_20190710 (emp_name,emp_level,create_time) values('测试',100,now());
8,调用pgbench进行测试:
/usr/pgsql-10/bin/pgbench -f select_inherits_emp_20190710 -c 4/10 -j 4/10 -n -P 10 -r -T 180 -R 10000 -p 7474 -d postgres -U postgres -h localhost/usr/pgsql-10/bin/pgbench -f insert_inherits_emp_20190710.sql -c 4/10 -j 4/10 -n -P 10 -r -T 180 -R 10000 -p 7474 -d postgres -U postgres -h localhost
非分区表测试
CREATE TABLE test_tmp_14m (emp_id SERIAL,emp_level INTEGER,emp_name TEXT,create_time TIMESTAMP NOT NULL);
导入同等数据到test_tmp_14m,然后创建索引,之后进行select,insert的10,4并发测试
调用pgbench进行测试:
/usr/pgsql-10/bin/pgbench -f select_test_tmp_14m.sql -c 4/10 -j 4/10 -n -P 10 -r -T 180 -R 10000 -p 7474 -d postgres -U postgres -h localhost/usr/pgsql-10/bin/pgbench -f insert_test_tmp_14m.sql -c 4/10 -j 4/10 -n -P 10 -r -T 180 -R 10000 -p 7474 -d postgres -U postgres -h localhost
压测曲线
01
峰值qps压测
调用pgbench 以三种分区方式和不分区方式进行压测,调整 -c / -j / -t / -T -R等参数把cpu压满时的tps/qps。
02
CPU利用率监控视图
1),10个并发select的CPU使用率对比:
pg_pathman 10个并发select的cpu:
native 10个并发select的cpu:
inherit 10个并发select的cpu:
不分区 10个并发select的cpu:
2),4个并发select的CPU使用率对比:
pg_pathman 4个并发select的cpu:
native 4个并发select的cpu:
Inherit 4个并发select的cpu:
不分区 4个并发select的cpu:
3),10个并发insert的CPU使用率对比:
pg_pathman 10个并发insert的cpu:
native 10个并发insert的cpu:
inherit 10个并发insert的cpu:
不分区 10个并发insert的cpu:
4),4个并发insert的CPU使用率对比:
pg_pathman 4个并发insert的cpu:
native 4个并发insert的cpu:
inherit 4个并发insert的cpu:
不分区4个并发insert的cpu: