postgresql无序uuid性能测试
无序uuid对数据库的影响
由于最近在做超大表的性能测试，在该过程中发现了无序uuid做主键对表插入性能有一定影响。结合实际情况发现当表的数据量越大，对表插入性能的影响也就越大。
测试环境
PostgreSQL创建插入脚本，测试各种情况的tps。
数据库版本：PostgreSQL 10.4 (ArteryBase 5.0.0, Thunisoft)
操作系统配置：CentOS Linux release 7 ，32GB内存，8 cpu
测试参数：pgbench -M prepared -r -n -j 8 -c 8 -T 60 -f /opt/thunisoft/pgbench_uuid_v4.sql -U sa pgbenchdb
空表，1000w数据，5000w数据，一亿数据的各种主键测试。
测试无序的uuid，有序的uuid，序列，有普通btree，有唯一索引和没有主键的情况
测试
1.创建表
--无序的uuid
pgbenchdb=# create table test_uuid_v4(id char(32) primary key);
CREATE TABLE
--有序的uuid
pgbenchdb=# create table test_time_nextval(id char(32) primary key);
CREATE TABLE
--递增序列
pgbenchdb=# create table test_seq_bigint(id int8 primary key);
CREATE TABLE
--创建序列
create sequence test_seq start with 1 ;
2.测试脚本
--测试无序uuid脚本
vi pgbench_uuid_v4.sql
insert into test_uuid_v4 (id) values (replace(uuid_generate_v4()::text,'-',''));
--测试有序uuid脚本
vi pgbench_time_nextval.sql
insert into test_time_nextval (id) values (replace(uuid_time_nextval()::text,'-',''));
--测试序列脚本
vi pgbench_seq_bigint.sql
insert into test_seq_bigint (id) values (nextval('test_seq'::regclass));
无序uuid,无数据情况
磁盘使用情况
avg-cpu: %user %nice %system %iowait %steal %idle
0.76 0.00 0.38 4.67 0.00 94.19
Device: rrqm/s wrqm/s r/s w/s rkB/s wkB/s avgrq-sz avgqu-sz await r_await w_await svctm %util
sdb 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
sda 0.00 0.00 0.00 96.00 0.00 2048.00 42.67 1.02 10.67 0.00 10.67 10.33 99.20
dm-0 0.00 0.00 0.00 96.00 0.00 2048.00 42.67 1.02 10.66 0.00 10.66 10.32 99.10
dm-1 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
dm-2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
tps：
[thunisoft@localhost thunisoft]$ pgbench -M prepared -r -n -j 8 -c 8 -T 60 -f /opt/thunisoft/pgbench_uuid_v4.sql -U sa pgbenchdb
transaction type: /opt/thunisoft/pgbench_uuid_v4.sql
scaling factor: 1
query mode: prepared
number of clients: 8
number of threads: 8
duration: 60 s
number of transactions actually processed: 53494
latency average = 8.974 ms
tps = 891.495404 (including connections establishing)
tps = 891.588967 (excluding connections establishing)
script statistics:
- statement latencies in milliseconds:
9.006 insert into test_uuid_v4 (id) values (replace(uuid_generate_v4()::text,'-',''));
无数据情况下，tps
类别 | 第一次 | 第二次 | 第三次 | 平均值(tps) |%util |await
---------------+---------+---------+---------+---------+-------+-------
无序uuid | 919 | 907 | 891 | 906 | 99.2% | 10.66
有序uuid | 985 | 882 | 932 | 933 | 98.7% | 4.4
序列 | 1311 | 1277 | 1280 | 1289 | 97.5% | 3.4
向表里面初始化100w数据
pgbenchdb=# insert into test_uuid_v4 (id) select replace(uuid_generate_v4()::text,'-','') from generate_series(1,1000000);
INSERT 0 1000000
Time: 43389.817 ms (00:43.390)
pgbenchdb=# insert into test_time_nextval (id) select replace(uuid_time_nextval()::text,'-','') from generate_series(1,1000000);
INSERT 0 1000000
Time: 30585.134 ms (00:30.585)
pgbenchdb=# insert into test_seq_bigint select generate_series (1,1000000);
INSERT 0 1000000
Time: 9818.639 ms (00:09.819)
无序uuid插入100w需要43s，有序需要30s，序列需要10s。
插入一百万数据后的tps
类别 | 第一次 | 第二次 | 第三次 | 平均值(tps) |%util |await
---------------+---------+---------+---------+---------+-------+-------
无序uuid | 355 | 440 | 302 | 365 | 98.8% | 13
有序uuid | 948 | 964 | 870 | 927 | 97.2% | 4.0
序列 | 1159 | 1234 | 1115 | 1169 | 96.6% | 3.5
插入一千万数据后的tps
类别 | 第一次 | 第二次 | 第三次 | 平均值(tps) |%util |await
---------------+---------+---------+---------+---------+-------+-------
无序uuid | 260 | 292 | 227 | 260 | 99.2% | 16.8
有序uuid | 817 | 960 | 883 | 870 | 97.7% | 3.9
序列 | 1305 | 1261 | 1270 | 1278 | 96.8% | 3.0
插入五千万数据后
向表中插入5kw数据，并且添加主键
pgbenchdb=# insert into test_time_nextval (id) select replace(uuid_time_nextval()::text,'-','') from generate_series(1,50000000);
INSERT 0 50000000
Time: 453985.318 ms (07:33.985)
pgbenchdb=# insert into test_seq_bigint select generate_series (1,50000000);
INSERT 0 50000000
Time: 352206.160 ms (05:52.206)
pgbenchdb=# insert into test_uuid_v4 (id) select replace(uuid_generate_v4()::text,'-','') from generate_series(1,50000000);
INSERT 0 50000000
Time: 1159689.338 ms (00:19:19.689)
在无主键情况下，插入五千万数据，有序uuid耗时7分钟，序列耗时6分钟，而无序uuid耗时接近20分钟。
pgbenchdb=# alter table test_uuid_v4 add primary key ("id");
ALTER TABLE
Time: 845199.296 ms (14:05.199)
pgbenchdb=# alter table test_time_nextval add primary key ("id");
ALTER TABLE
Time: 932151.103 ms (15:32.151)
pgbenchdb=# alter table test_seq_bigint add primary key ("id");
ALTER TABLE
Time: 148138.871 ms (02:28.139)
pgbenchdb=# select pg_size_pretty(pg_total_relation_size('test_uuid_v4'));
pg_size_pretty
----------------
6072 MB
(1 row)
Time: 0.861 ms
pgbenchdb=# select pg_size_pretty(pg_total_relation_size('test_time_nextval'));
pg_size_pretty
----------------
6072 MB
(1 row)
Time: 0.942 ms
pgbenchdb=# select pg_size_pretty(pg_total_relation_size('test_seq_bigint'));
pg_size_pretty
----------------
2800 MB
(1 row)
Time: 0.699 ms
插入5kw后
类别 | 第一次 | 第二次 | 第三次 | 平均值(tps) |%util |await
---------------+---------+---------+---------+---------+-------+-------
无序uuid | 162 | 163 | 163 | 163 | 99.6% | 18.4
有序uuid | 738 | 933 | 979 | 883 | 97.7% | 3.9
序列 | 1132 | 1264 | 1265 | 1220 | 96.8% | 3.5
插入1亿条数据后
类别 | 第一次 | 第二次 | 第三次 | 平均值(tps) |%util |await
---------------+---------+---------+---------+---------+-------+-------
无序uuid | 121 | 131 | 143 | 131 | 99.6% | 28.2
有序uuid | 819 | 795 | 888 | 834 | 99.2% | 28.7
序列 | 1193 | 1115 | 1109 | 1139 | 96.8% | 11.3
普通btree索引
上面测了无序uuid，1kw情况下，有主键的tps是260，无主键的tps是1234。尝试测试普通的索引，和唯一索引tps
--创建普通索引
pgbenchdb=# create index i_test_uuid_v4_id on test_uuid_v4(id);
CREATE INDEX
Time: 316367.010 ms (05:16.367)
--创建普通索引后
[thunisoft@localhost thunisoft]$ pgbench -M prepared -r -n -j 8 -c 8 -T 60 -f /opt/thunisoft/pgbench_uuid_v4.sql -U sa pgbenchdb
transaction type: /opt/thunisoft/pgbench_uuid_v4.sql
scaling factor: 1
query mode: prepared
number of clients: 8
number of threads: 8
duration: 60 s
number of transactions actually processed: 13308
latency average = 36.080 ms
tps = 221.727391 (including connections establishing)
tps = 221.749660 (excluding connections establishing)
script statistics:
- statement latencies in milliseconds:
38.512 insert into test_uuid_v4 (id) values (replace(uuid_generate_v4()::text,'-',''));
--创建唯一索引
pgbenchdb=# drop index i_test_uuid_v4_id;
DROP INDEX
Time: 267.451 ms
pgbenchdb=# create unique index i_test_uuid_v4_id on test_uuid_v4(id);
CREATE INDEX
Time: 153372.622 ms (02:33.373)
[thunisoft@localhost thunisoft]$ pgbench -M prepared -r -n -j 8 -c 8 -T 60 -f /opt/thunisoft/pgbench_uuid_v4.sql -U sa pgbenchdb
^[[3~transaction type: /opt/thunisoft/pgbench_uuid_v4.sql
scaling factor: 1
query mode: prepared
number of clients: 8
number of threads: 8
duration: 60 s
number of transactions actually processed: 13847
latency average = 34.693 ms
tps = 230.593988 (including connections establishing)
tps = 230.620469 (excluding connections establishing)
script statistics:
- statement latencies in milliseconds:
36.410 insert into test_uuid_v4 (id) values (replace(uuid_generate_v4()::text,'-',''));
无论是普通btree索引和唯一索引，都会影响插入的效率。
删除所有的主键索引
--删除所有主键
alter table test_uuid_v4 drop constraint "test_uuid_v4_pkey";
alter table test_time_nextval drop constraint "test_time_nextval_pkey" ;
alter table test_seq_bigint drop constraint "test_seq_bigint_pkey";
1,--无序uuid：测试pgbench_uuid_v4.sql
[thunisoft@localhost thunisoft]$ pgbench -M prepared -r -n -j 8 -c 8 -T 60 -f /opt/thunisoft/pgbench_uuid_v4.sql -U sa pgbenchdb
transaction type: /opt/thunisoft/pgbench_uuid_v4.sql
scaling factor: 1
query mode: prepared
number of clients: 8
number of threads: 8
duration: 60 s
number of transactions actually processed: 74109
latency average = 6.479 ms
tps = 1234.842229 (including connections establishing)
tps = 1235.042674 (excluding connections establishing)
script statistics:
- statement latencies in milliseconds:
6.112 insert into test_uuid_v4 (id) values (replace(uuid_generate_v4()::text,'-',''));
2、--有序uuid，测试pgbench_time_nextval.sql
[thunisoft@localhost thunisoft]$ pgbench -M prepared -r -n -j 8 -c 8 -T 60 -f /opt/thunisoft/pgbench_time_nextval.sql -U sa pgbenchdb
transaction type: /opt/thunisoft/pgbench_time_nextval.sql
scaling factor: 1
query mode: prepared
number of clients: 8
number of threads: 8
duration: 60 s
number of transactions actually processed: 74027
latency average = 6.486 ms
tps = 1233.364360 (including connections establishing)
tps = 1233.482292 (excluding connections establishing)
script statistics:
- statement latencies in milliseconds:
6.186 insert into test_time_nextval (id) values (replace(uuid_time_nextval()::text,'-',''));
3、--序列，测试pgbench_seq_bigint.sql
[thunisoft@localhost thunisoft]$ pgbench -M prepared -r -n -j 8 -c 8 -T 60 -f /opt/thunisoft/pgbench_seq_bigint.sql -U sa pgbenchdb
transaction type: /opt/thunisoft/pgbench_seq_bigint.sql
scaling factor: 1
query mode: prepared
number of clients: 8
number of threads: 8
duration: 60 s
number of transactions actually processed: 76312
latency average = 6.290 ms
tps = 1271.832907 (including connections establishing)
tps = 1272.124397 (excluding connections establishing)
script statistics:
- statement latencies in milliseconds:
5.916 insert into test_seq_bigint (id) values (nextval('test_seq'::regclass));
删除主键约束后，三种情况下tps非常接近，都达到了1200+。
Btree索引，插入操作的平均tps对比
类别/平均tps | 无数据 | 一千万 | 五千万 | 一亿 |
---------------+---------+---------+---------+---------+
无序uuid | 960 | 260 | 163 | 131 |
有序uuid | 933 | 870 | 883 | 834 |
序列 | 1289 | 1278 | 1220 | 1139 |
根据测试数据可以看出无序的uuid在数据到达1kw后插入数据的tps下降的非常厉害，而有序的uuid和递增序列下降的比较少。到一亿数据的tps有序uuid是无序的6倍，序列是无序uuid的9倍。
创建单独的表空间用来存储索引信息
如果有多快磁盘那么可以将索引和数据分开存储，以此来加快写入的速度。
创建单独的索引空间：
create tablespace indx_test owner sa location '/home/tablespace/index_test';
指定索引存储目录：
create index i_test_uuid_v4_id on test_uuid_v4 using btree(id) tablespace indx_test;
关于有序uuid
测试使用的sequential-uuids插件，生成的有序uuid。
有序uuid的结构为(block ID; random data)，实际上就是把数据拆成两部分，一部分自增，一部分随机。
提供了两种算法：
1.uuid_sequence_nextval(sequence regclass, block_size int default 65536, block_count int default 65536)
前缀为自增序列，如果块ID使用2字节存储，一个索引BLOCK里面可以存储256条记录（假设8K的BLOCK，一条记录包括uuid VALUE（16字节）以及ctid（6字节），所以一个索引页约存储363条记录（8000 /（16 + 6）））
2.uuid_time_nextval(interval_length int default 60, interval_count int default 65536) RETURNS uuid
默认每60秒内的数据的前缀是一样的，前缀递增1，到65535后循环。
使用uuid_time_nextval生成的有序uuid
pgbenchdb=# select id from test_time_nextval;
id
----------------------------------
a18b7dd0ca92b0b5c1844a402f9c6999
a18b540b8bbe0ddb2b6d0189b2e393c6
a18b83eb7320b0a90e625185421e065e
a18bade4ff15e05dab81ecd3f4c2dee4
a18b79e41c3bc8d2d4ba4b70447e6b29
a18bdad18d9e0d2fa1d9d675bc7129f0
a18b13723ec7be9a2f1a3aec5345a88b
a18bd9d866047aec69a064d30e9493d2
a18bd76e8c787c7464479502f381e6d7
a18ba5c0c966f81cfdbeff866618da8d
......
有序uuid前四位有序，后面的随机生成。
结语
1.关于有序的uuid，前4位是有序的，后面都是随机生成的。
2.在该环境中发现，无序uuid随着数据量的不断增大，tps下滑比较厉害。
3.由于btree索引的存在，无序的uuid会导致大量的离散io。导致磁盘使用率高。进而影响插入效率。随着表数据量的增大更加明显。
4.该测试是在普通的磁盘上面测试，并未在ssd上面测试。
5.如果要使用有序uuid，有多种实现方式，还需要考虑分布式情况下生成全局有序uuid。
