postgresql搭建主备_pg数据库主备配置

1. 要求

1.使用流复制，配置复制槽
2.只有一个备库时不配置同步，多个备库时ANY
3.测试切换

2. 主库设置

postgresql.conf

# WAL参数
# wal_level=replica              # 默认replica
# synchronous_commit=on     # 默认为on。只要不配置synchronous_standby_names，这里不配置off结果都一样。
# full_page_writes=on            # 默认为on，必须为on
wal_compression=on    #默认为off，开启wal压缩
# wal_writer_delay   wal_writer_flush_after 这两参数就用默认值，如果写磁盘过于频繁增加该值

# checkpoint参数
checkpoint_timeout # 默认5min，增加能提高性能但要崩溃后恢复的时间更长。checkpoint对数据库的影响还是挺大的。
checkpoint_completion_target=0.8 # 用上面0.8倍的时间进行checkpoint，平滑I/O

# 归档参数
archive_mode=always  # 默认为off。主备环境要开启，万一WAL没了还能用归档拿。设置为always切换后也不用再设置归档。
archive_command='test ! -f /home/postgres/pg_arch/%f && cp %p /home/postgres/pg_arch/%f'   # 改成实际路径
# archive_timeout    默认不切换，流复制不需要发送wal文件不需要切换
# archive_cleanup_command 主库不会自动清除，备库会自动清除
# restore_command # 在备库上配置时是通过log方式复制，流复制不需要。主库上配置时仅在recovery时有效果。
archive_cleanup_command ='pg_archivecleanup /home/postgres/pg_arch %r' 

# send 配置
# max_wal_senders        # 默认10，够用
# max_replication_slots  # 默认10，够用
wal_keep_segments=64     # 保留1G应该够了吧
track_commit_timestamp=on # 开启保留时间戳功能

# 主库设置
# synchronous_standby_names 不配置，不设置同步，除非有明确需求
# vacuum_defer_cleanup_age 默认为0，立即删除。
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28

pg_hba.conf

host    replication     standby     192.168.0.38  md5
1

3. 主库创建复制槽、复制用户

create database test;
\c test;
create user replicat_user replication password 'bfd@dev24';
SELECT * FROM pg_create_physical_replication_slot('slot_01');
SELECT slot_name, slot_type, active FROM pg_replication_slots;
create table t1(id int);
insert into t1 values(1),(2),(3);
1
2
3
4
5
6
7

4. 导出主库

在备库执行，需要有superuser或者replication的权限才行.
用pg_basebackup导过来之后，整个数据目录都导过来了，包括数据文件，因此只需要修改配置文件就行了

pg_basebackup -h abc07 -U replicat_user -D pg_data -X s
1

5. 备库设置

postgres.conf

Primary_conninfo = 'host=abc07 port=5432 user=replicat_user password=bfd@dev24 '
primary_slot_name = 'slot_01'
1
2

pg_hba.conf

host    replication     standby      192.168.0.37  md5
1

6. 备库创建standby.signal

数据目录下：

touch standby.signal
1

一定要用启动数据库的用户创建

7. 启动备库

略

8. 备库创建复制槽、查询数据

同步后replication_user用户也被同步到备库，不需要再创建。

SELECT * FROM pg_create_physical_replication_slot('slot_01');
SELECT slot_name, slot_type, active FROM pg_replication_slots;
select * from t1;
1
2
3

9. 查看日志

2023-04-28 09:24:02.312 CST [26908] LOG:  started streaming WAL from primary at 0/7000000 on timeline 1
1

关闭主库后

2023-04-28 09:55:11.791 CST [26908] LOG:  replication terminated by primary server
2023-04-28 09:55:11.791 CST [26908] DETAIL:  End of WAL reached on timeline 1 at 0/80000A0.
2023-04-28 09:55:11.791 CST [26908] FATAL:  could not send end-of-streaming message to primary: no COPY in progress
2023-04-28 09:55:11.791 CST [26903] LOG:  invalid record length at 0/80000A0: wanted 24, got 0
2023-04-28 09:55:11.807 CST [32659] FATAL:  could not connect to the primary server: server closed the connection un
expectedly
                This probably means the server terminated abnormally
                before or while processing the request.
2023-04-28 09:55:16.798 CST [32669] FATAL:  could not connect to the primary server: could not connect to server: Co
nnection refused
                Is the server running on host "abc07" (192.168.0.37) and accepting
                TCP/IP connections on port 5432?
1
2
3
4
5
6
7
8
9
10
11
12

10. 测试切换

https://blog.csdn.net/pg_hgdb/article/details/125264019
不设置promote_trigger_file，使用pg_ctl promote或者pg_promote来提升备库为主库。任务有执行该函数权限的用户都可以执行。

pg_promote(wait boolean DEFAULT true, wait_seconds integer DEFAULT 60)。 当wait为true时直到提升成功或者达到wait_seconds的时间才返回true，当wait为false时，直接返回true。
1

可以使用以下两种方式：

$ pg_ctl promote    [-D DATADIR] [-W] [-t SECS] [-s]
1

或者：

postgres=# select pg_promote(true,60);
 pg_promote 
------------
 t
(1 row)
1
2
3
4
5

提升后日志显示：

2023-04-28 10:05:35.761 CST [26903] LOG:  received promote request
2023-04-28 10:05:35.761 CST [26903] LOG:  redo done at 0/8000028
2023-04-28 10:05:35.762 CST [26903] LOG:  selected new timeline ID: 2
2023-04-28 10:05:35.882 CST [26903] LOG:  archive recovery complete
2023-04-28 10:05:35.891 CST [26899] LOG:  database system is ready to accept connections
1
2
3
4
5

此时数据目录中standby.signal没有了。

11. 同步测试

修改备库配置文件，或者在命令行修改。

Primary_conninfo = 'host=abc07 port=5432 user=replicat_user password=bfd@dev24 application_name=s38'
1

修改主库配置文件，或者在命令行修改。

synchronous_commit=on
synchronous_standby_names='1 (s38)'
1
2

测试关闭备库，然后看主库提交等待情况：

一直卡住不提交。
从这点看，如果没有监控备库的状态就用同步，会对主库造成致命问题。
所以，要用ANY的策略，ANY能保证至少X个备库同步完成就行，并且尝试同步所有备库。
当仅有一个备库时，不做同步复制，太危险了

12. 监控主备状态

主库上的pg_stat_replication记录了复制信息：
https://www.postgresql.org/docs/12/monitoring-stats.html

postgres=# \x
Expanded display is on.
postgres=# select * from pg_stat_replication;
-[ RECORD 1 ]----+------------------------------
pid              | 19189
usesysid         | 16386
usename          | replicat_user
application_name | s38
client_addr      | 192.168.0.38
client_hostname  | 
client_port      | 12214
backend_start    | 2023-04-28 14:42:16.180019+08
backend_xmin     | 
state            | streaming
sent_lsn         | 0/B0004B8
write_lsn        | 0/B0004B8
flush_lsn        | 0/B0004B8
replay_lsn       | 0/B0004B8
write_lag        | 
flush_lag        | 
replay_lag       | 
sync_priority    | 1
sync_state       | sync
reply_time       | 2023-04-28 15:43:23.142803+08
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

sent_lsn：发送的WAL日志位置
write_lsn：standby写入到磁盘的WAL的位置。(写到缓冲)
flush_lsn：standby flush到磁盘的WAL的位置。(写到磁盘上)
replay_lsn：已经Apply的WAL的位置
write_lag：主库WAL本地化与standby romote_write(写到缓冲)之间的lag。(同步复制时才有)
flush_lag：主库WAL本地化与standby 写到磁盘之间的lag。(同步复制时才有)
replay_lag: 主库WAL本地化与standby apply之间的lag。(同步复制时才有)

sync_state：
备库的状态，如下：
async：该standy异步复制。不在synchronous_standby_names中的都是异步。
potential：该standy现在是异步复制，但是在同步的standby down的时候有可能成为同步的
sync：该standy同步复制
quorum：备选的standbys。意义不明

reply_time：上次发送的apply的时间

主备的pg_current_wal_lsn()记录了当前的WAL位置。

postgres=# select pg_current_wal_lsn();
 pg_current_wal_lsn 
--------------------
 0/B0004B8
(1 row)
1
2
3
4
5

备库的pg_last_wal_receive_lsn()记录了接收位置 ：

postgres=# select pg_last_wal_receive_lsn();
 pg_last_wal_receive_lsn 
-------------------------
 0/B0004B8
(1 row)
1
2
3
4
5

备库的pg_stat_wal_receiver显示复制信息：

postgres=# \x
Expanded display is on.
postgres=# select * from pg_stat_wal_receiver;
-[ RECORD 1 ]---------+----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
pid                   | 21791
status                | streaming
receive_start_lsn     | 0/B000000
receive_start_tli     | 1
received_lsn          | 0/B0004B8
received_tli          | 1
last_msg_send_time    | 2023-04-28 17:23:04.078542+08
last_msg_receipt_time | 2023-04-28 17:23:04.111514+08
latest_end_lsn        | 0/B0004B8
latest_end_time       | 2023-04-28 14:42:16.182656+08
slot_name             | slot_01
sender_host           | abc07
sender_port           | 5432
conninfo              | user=replicat_user password=******** dbname=replication host=abc07 port=5432 application_name=s38 fallback_application_name=walreceiver sslmode=disable sslcompression=0 gssencmode=disable krbsrvname=postgres target_session_attrs=any
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18

通过主备的视图能比较准确的判断复制的延迟：
https://www.postgresql.org/docs/12/runtime-config-wal.html#GUC-FULL-PAGE-WRITES

select pid,usename,client_addr,state,
pg_wal_lsn_diff(pg_current_wal_lsn(),write_lsn) write_delay,
pg_wal_lsn_diff(pg_current_wal_lsn(),flush_lsn) flush_delay,
pg_wal_lsn_diff(pg_current_wal_lsn(),replay_lsn) replay_dely
from pg_stat_replication;
1
2
3
4
5