数据库内核月报－ 2016/09 - PgSQL · 最佳实践 · pg_rman源码浅析与使用 - 《数据库内核月报》

pg_rman 介绍
初始化pg_rman backup catalog
全量备份
增量备份
列出备份集
物理删除已从catalog删除的备份集
例子
pg_rman 可以优化的地方
注意

PostgreSQL增量备份是怎么做到的呢？

一个数据页的框架如下

数据页头部的数据结构


{  
        /* XXX LSN is member of *any* block, not only page-organized ones */  
        PageXLogRecPtr pd_lsn;          /* LSN: next byte after last byte of xlog  
                                         * record for last change to this page */  
        uint16          pd_checksum;    /* checksum */   
        uint16          pd_flags;               /* flag bits, see below */  
        LocationIndex pd_lower;         /* offset to start of free space */  
        LocationIndex pd_upper;         /* offset to end of free space */  
        LocationIndex pd_special;       /* offset to start of special space */  
        uint16          pd_pagesize_version;  
        TransactionId pd_prune_xid; /* oldest prunable XID, or zero if none */  
        ItemIdData      pd_linp[FLEXIBLE_ARRAY_MEMBER]; /* line pointer array */  
} PageHeaderData;

因为如果对象是持久化的，那么它的所有变更都会记录REDO，数据页头部的pd_lsn表示该数据页最后一次变化时，变化产生的REDO在xlog file中的结束位置.
即如果xlog flush的xlog地址位大于或等于此页pd_lsn，那么这个页的更改就可以认为是可靠的。

 *              pd_lsn          - identifies xlog record for last change to this page.  
 *              pd_checksum - page checksum, if set.  
 *              pd_flags        - flag bits.  
 *              pd_lower        - offset to start of free space.  
 *              pd_upper        - offset to end of free space.  
 *              pd_special      - offset to start of special space.  
 *              pd_pagesize_version - size in bytes and page layout version number.  
 *              pd_prune_xid - oldest XID among potentially prunable tuples on page.

好了，既然每次块的变化都包含了LSN的修改，那么也即是说，我们可以通过第一次备份开始时的全局LSN，以及当前需要备份的数据的page LSN来判断此页是否发生过修改。

如果修改了就备份，没修改，就不需要备份，从而实现数据库的块级增量备份。

pg_rman 介绍

pg_rman是一个，类似Oracle的RMAN。
documentation见

pg_rman使用的是pg_start_backup(), copy, pg_stop_backup()的备份模式。
pg_rman跑的不是流复制协议，而是文件拷贝，所以pg_rman必须和数据库节点跑在一起。
如果在standby节点跑pg_rman，pg_rman则需要通过网络连接到主节点执行pg_start_backup和pg_stop_backup。
pg_rman的用法非常简单，支持以下几种运行模式。

使用pg_rman的前提

开启归档
配置csvlog
建议的配置

postgres=# show log_destination ;  
 log_destination   
-----------------  
 csvlog  
(1 row)  
postgres=# SHOW log_directory ;  
 log_directory  
---------------  
 pg_log  
(1 row)  
postgres=# SHOW archive_command ;  
              archive_command  
--------------------------------------------  
 cp %p /data04/digoal/arc_log/%f  
(1 row)

初始化pg_rman backup catalog

首先需要初始化一个backup catalog，实际上就是需要一个目录，这个目录将用于存放备份的文件。
同时这个目录也会存放一些元数据，例如备份的配置文件，数据库的systemid，时间线文件历史等等。
初始化命令需要两个参数，分别为备份目标目录，以及数据库的$PGDATA

$ mkdir /data05/digoal/pgbbk  
$ /home/digoal/pgsql9.5/bin/pg_rman init -B /data05/digoal/pgbbk -D /data04/digoal/pg_root  
INFO: ARCLOG_PATH is set to '/data04/digoal/arc_log'  
INFO: SRVLOG_PATH is set to '/data04/digoal/pg_root/pg_log'

生成备份元数据如下

[digoal@iZ28tqoemgtZ ~]$ cd /data05/digoal/pgbbk/  
[digoal@iZ28tqoemgtZ pgbbk]$ ll  
total 16  
drwx------ 4 digoal digoal 4096 Aug 26 19:29 backup  
-rw-rw-r-- 1 digoal digoal   82 Aug 26 19:29 pg_rman.ini  
-rw-rw-r-- 1 digoal digoal   40 Aug 26 19:29 system_identifier  
drwx------ 2 digoal digoal 4096 Aug 26 19:29 timeline_history

生成的配置文件

$ cat pg_rman.ini   
ARCLOG_PATH='/data04/digoal/arc_log'  
SRVLOG_PATH='/data04/digoal/pg_root/pg_log'

你可以把将来要使用的配置写在这个配置文件中，或者写在pg_rman的命令行中。
我后面的测试会直接使用命令行参数。
生成的数据库system id，用于区分备份的数据库是不是一个数据库，防止被冲。

$ cat system_identifier   
SYSTEM_IDENTIFIER='6318621837015461309'

与控制文件中存储的system id一致。

注意
pg_rman只从postgresql.conf取log_directory和archive_command参数的值。
如果你的PostgreSQL的配置文件是include的或者配置在postgresql.auto.conf中，这两个值将不准确。
所以建议你仅仅把参数配置在postgresql.conf中，而不要使用其他配置文件。

pg_rman 命令行用法

pg_rman manage backup/recovery of PostgreSQL database.  
Usage:  
  pg_rman OPTION init  
  pg_rman OPTION backup  
  pg_rman OPTION restore  
  pg_rman OPTION show [DATE]  
  pg_rman OPTION show detail [DATE]  
  pg_rman OPTION validate [DATE]  
  pg_rman OPTION delete DATE  
  pg_rman OPTION purge  
Common Options:  
  -D, --pgdata=PATH         location of the database storage area  
  -A, --arclog-path=PATH    location of archive WAL storage area  
  -S, --srvlog-path=PATH    location of server log storage area  
  -B, --backup-path=PATH    location of the backup storage area  
  -c, --check               show what would have been done  
  -v, --verbose             show what detail messages  
  -P, --progress            show progress of processed files  
Backup options:  
  -b, --backup-mode=MODE    full, incremental, or archive  
  -s, --with-serverlog      also backup server log files  
  -Z, --compress-data       compress data backup with zlib  
  -C, --smooth-checkpoint   do smooth checkpoint before backup  
  -F, --full-backup-on-error   switch to full backup mode  
                               if pg_rman cannot find validate full backup  
                               on current timeline  
      NOTE: this option is only used in --backup-mode=incremental or archive.  
  --keep-data-generations=NUM keep NUM generations of full data backup  
  --keep-data-days=NUM        keep enough data backup to recover to N days ago  
  --keep-arclog-files=NUM   keep NUM of archived WAL  
  --keep-arclog-days=DAY    keep archived WAL modified in DAY days  
  --keep-srvlog-files=NUM   keep NUM of serverlogs  
  --keep-srvlog-days=DAY    keep serverlog modified in DAY days  
  --standby-host=HOSTNAME   standby host when taking backup from standby  
  --standby-port=PORT       standby port when taking backup from standby  
Restore options:  
  --recovery-target-time    time stamp up to which recovery will proceed  
  --recovery-target-xid     transaction ID up to which recovery will proceed  
  --recovery-target-inclusive whether we stop just after the recovery target  
  --recovery-target-timeline  recovering into a particular timeline  
  --hard-copy                 copying archivelog not symbolic link  
Catalog options:  
  -a, --show-all            show deleted backup too  
Delete options:  
  -f, --force               forcibly delete backup older than given DATE  
Connection options:  
  -d, --dbname=DBNAME       database to connect  
  -h, --host=HOSTNAME       database server host or socket directory  
  -p, --port=PORT           database server port  
  -U, --username=USERNAME   user name to connect as  
  -w, --no-password         never prompt for password  
  -W, --password            force password prompt  
Generic options:  
  -q, --quiet               don't show any INFO or DEBUG messages  
  --debug                   show DEBUG messages  
  --help                    show this help, then exit  
  --version                 output version information, then exit  
Read the website for details. <http://github.com/ossc-db/pg_rman>  
Report bugs to <http://github.com/ossc-db/pg_rman/issues>.

输入必要的参数或option

$ export PGPASSWORD=postgres  
$ /home/digoal/pgsql9.5/bin/pg_rman backup \  
-B /data05/digoal/pgbbk \  
-D /data04/digoal/pg_root \  
-b full \  
-s \  
-Z \  
-C \  
--keep-data-days=10 \  
--keep-arclog-files=15 \  
--keep-arclog-days=10 \  
--keep-srvlog-files=10 \  
--keep-srvlog-days=15 \  
-h 127.0.0.1 -p 1921 -U postgres -d postgres

结果

INFO: copying database files  
NOTICE:  pg_stop_backup complete, all required WAL segments have been archived  
INFO: copying archived WAL files  
INFO: copying server log files  
INFO: backup complete  
HINT: Please execute 'pg_rman validate' to verify the files are correctly copied.  
INFO: start deleting old archived WAL files from ARCLOG_PATH (keep files = 15, keep days = 10)  
INFO: the threshold timestamp calculated by keep days is "2016-08-16 00:00:00"  
INFO: start deleting old server files from SRVLOG_PATH (keep files = 10, keep days = 15)  
INFO: the threshold timestamp calculated by keep days is "2016-08-11 00:00:00"  
INFO: start deleting old backup (keep after = 2016-08-16 00:00:00)  
INFO: does not include the backup just taken  
WARNING: backup "2016-08-26 19:39:32" is not taken int account  
DETAIL: This is not valid backup.

校验备份集

备份时pg_rman会记录每个备份文件的crc，以便validate进行校验。
例如某个备份集

$ less /data05/digoal/pgbbk/20160826/195809/file_database.txt  
.s.PGSQL.1921 ? 0 0 0777 2016-08-26 19:27:05  
.s.PGSQL.1921.lock f 55 590164837 0600 2016-08-26 19:27:05  
PG_VERSION f 12 3872055064 0600 2016-07-28 10:03:42  
backup_label f 167 2985542389 0600 2016-08-26 19:58:42  
backup_label.old f 155 4273989468 0600 2016-08-23 19:43:32  
base d 0 0 0700 2016-08-23 10:28:32  
base/1 d 0 0 0700 2016-08-24 16:17:02  
base/1/112 f 57 1147028285 0600 2016-07-28 10:03:42  
base/1/113 f 57 1147028285 0600 2016-07-28 10:03:42  
base/1/1247 F 8178 1875285513 0600 2016-07-29 13:51:29  
base/1/1247_fsm f 139 3668812536 0600 2016-07-28 10:03:43

解释: 路径，文件类型，大小，CRC校验值，权限，时间，第四列即crc校验值
每次备份完，必须要做一次校验，否则备份集不可用用来恢复，增量备份时也不会用它来做增量比较。

$ /home/digoal/pgsql9.5/bin/pg_rman validate -B /data05/digoal/pgbbk  
INFO: validate: "2016-08-26 19:39:50" backup, archive log files and server log files by CRC  
INFO: backup "2016-08-26 19:39:50" is valid

每个备份集都包含了一个备份状态文件，如下

cat /data05/digoal/pgbbk/20160826/201955/backup.ini
# configuration
BACKUP_MODE=INCREMENTAL
FULL_BACKUP_ON_ERROR=false
WITH_SERVERLOG=true
COMPRESS_DATA=true
# result
TIMELINEID=1
START_LSN=46/df000108
STOP_LSN=46/df000210
START_TIME='2016-08-26 20:19:55'
END_TIME='2016-08-26 20:20:48'
RECOVERY_XID=3896508593
RECOVERY_TIME='2016-08-26 20:20:47'
TOTAL_DATA_BYTES=6196524307
READ_DATA_BYTES=3199287520
READ_ARCLOG_BYTES=33554754
READ_SRVLOG_BYTES=0
WRITE_BYTES=125955
BLOCK_SIZE=8192
XLOG_BLOCK_SIZE=8192
STATUS=OK

这个文件中包含了很重要的信息，比如LSN，后面LSN将用于比对增量备份时对比数据块的LSN是否发生了变化，是否需要备份。

增量备份

$ export PGPASSWORD=postgres  
$ /home/digoal/pgsql9.5/bin/pg_rman backup \  
-B /data05/digoal/pgbbk \  
-D /data04/digoal/pg_root \  
-b incremental \  
-s \  
-Z \  
-C \  
--keep-data-days=10 \  
--keep-arclog-files=15 \  
--keep-arclog-days=10 \  
--keep-srvlog-files=10 \  
--keep-srvlog-days=15 \  
-h 127.0.0.1 -p 1921 -U postgres -d postgres

增量备份输出

校验备份集

$ /home/digoal/pgsql9.5/bin/pg_rman validate -B /data05/digoal/pgbbk  
INFO: validate: "2016-08-26 19:43:20" backup, archive log files and server log files by CRC  
INFO: backup "2016-08-26 19:43:20" is valid

列出备份集

$ /home/digoal/pgsql9.5/bin/pg_rman show -B /data05/digoal/pgbbk  
==========================================================  
 StartTime           Mode  Duration    Size   TLI  Status   
==========================================================  
2016-08-26 19:43:20  INCR        0m    54kB     1  OK  
2016-08-26 19:39:50  FULL        1m   245MB     1  OK

可以看到增量非常小，因为很少变化的块。
接下来更新一张大表的某一条记录，再看看。

                      List of relations  
 Schema |   Name   | Type  |  Owner   |  Size   | Description   
--------+----------+-------+----------+---------+-------------  
 public | hll_test | table | postgres | 208 kB  |   
 public | t        | table | postgres | 3050 MB |   
 public | tbl2     | table | postgres | 63 MB   |   
 public | test     | table | postgres | 120 MB  |   
(5 rows)  
postgres=# with t1 as (select id from t where id between 1 and 1000 limit 10) update t set info='new' where id in (select * from t1);  
UPDATE 10

更新后做一个增量备份

$ /home/digoal/pgsql9.5/bin/pg_rman backup -B /data05/digoal/pgbbk -D /data04/digoal/pg_root -b incremental -s -Z -C --keep-data-days=10 --keep-arclog-files=15 --keep-arclog-days=10 --keep-srvlog-files=10 --keep-srvlog-days=15 -h 127.0.0.1 -p 1921 -U postgres -d postgres  
INFO: copying database files  
NOTICE:  pg_stop_backup complete, all required WAL segments have been archived  
INFO: copying archived WAL files  
INFO: copying server log files  
INFO: backup complete  
HINT: Please execute 'pg_rman validate' to verify the files are correctly copied.  
INFO: start deleting old archived WAL files from ARCLOG_PATH (keep files = 15, keep days = 10)  
INFO: the threshold timestamp calculated by keep days is "2016-08-16 00:00:00"  
INFO: start deleting old server files from SRVLOG_PATH (keep files = 10, keep days = 15)  
INFO: the threshold timestamp calculated by keep days is "2016-08-11 00:00:00"  
INFO: start deleting old backup (keep after = 2016-08-16 00:00:00)  
INFO: does not include the backup just taken  
INFO: backup "2016-08-26 19:58:09" should be kept  
DETAIL: This is taken after "2016-08-16 00:00:00".  
WARNING: backup "2016-08-26 19:56:54" is not taken int account  
DETAIL: This is not valid backup.  
INFO: backup "2016-08-26 19:43:20" should be kept  
DETAIL: This is taken after "2016-08-16 00:00:00".  
INFO: backup "2016-08-26 19:39:50" should be kept  
DETAIL: This is taken after "2016-08-16 00:00:00".  
WARNING: backup "2016-08-26 19:39:32" is not taken int account  
DETAIL: This is not valid backup.

校验备份集

[digoal@iZ28tqoemgtZ pg_rman]$ /home/digoal/pgsql9.5/bin/pg_rman validate -B /data05/digoal/pgbbk  
INFO: validate: "2016-08-26 20:19:55" backup, archive log files and server log files by CRC  
INFO: backup "2016-08-26 20:19:55" is valid

输出当前备份

[digoal@iZ28tqoemgtZ pg_rman]$ /home/digoal/pgsql9.5/bin/pg_rman show -B /data05/digoal/pgbbk  
==========================================================  
 StartTime           Mode  Duration    Size   TLI  Status   
==========================================================  
2016-08-26 20:19:55  INCR        0m   125kB     1  OK  
2016-08-26 19:58:09  FULL       11m  3094MB     1  OK  
2016-08-26 19:56:54  FULL        1m      0B     0  ERROR  
2016-08-26 19:43:20  INCR        0m    54kB     1  OK  
2016-08-26 19:39:50  FULL        1m   245MB     1  OK  
2016-08-26 19:39:32  FULL        0m      0B     0  ERROR

按指定时间从catalog删除备份集

例如我只需要我的备份集能恢复到2016-08-26 19:59:00，在这个时间点以前，不需要用来恢复到这个时间点的备份全删掉。

$ /home/digoal/pgsql9.5/bin/pg_rman delete "2016-08-26 19:59:00" -B /data05/digoal/pgbbk  
WARNING: cannot delete backup with start time "2016-08-26 19:58:09"  
DETAIL: This is the latest full backup necessary for successful recovery.  
INFO: delete the backup with start time: "2016-08-26 19:56:54"  
INFO: delete the backup with start time: "2016-08-26 19:43:20"  
INFO: delete the backup with start time: "2016-08-26 19:39:50"  
INFO: delete the backup with start time: "2016-08-26 19:39:32"

保留的备份集合可以将数据库恢复到2016-08-26 19:59:00

$ /home/digoal/pgsql9.5/bin/pg_rman show -B /data05/digoal/pgbbk  
==========================================================  
 StartTime           Mode  Duration    Size   TLI  Status   
==========================================================  
2016-08-26 20:19:55  INCR        0m   125kB     1  OK  
2016-08-26 19:58:09  FULL       11m  3094MB     1  OK

$ /home/digoal/pgsql9.5/bin/pg_rman purge -B /data05/digoal/pgbbk  
INFO: DELETED backup "2016-08-26 19:56:54" is purged  
INFO: DELETED backup "2016-08-26 19:43:20" is purged  
INFO: DELETED backup "2016-08-26 19:39:50" is purged  
INFO: DELETED backup "2016-08-26 19:39:32" is purged

恢复

pg_rman数据恢复时的两个必要要素

新的$PGDATA
备份目录
命令的选项也很简单，甚至可以不指定任何option

Restore options:  
The parameters which are started with –recovery are same as parameters in recovery.conf. See also “Recovery Configuration” for details.

接下来的几个配置，与recovery.conf的意思对齐。

如果不指定时间线，则使用$PGDATA/global/pg_control，如果没有$PGDATA/global/pg_control，则使用最新的全量备份集的时间线。

--recovery-target-timeline TIMELINE  
Specifies recovering into a particular timeline. If not specified, the current timeline from ($PGDATA/global/pg_control) is used.

如果不指定，则恢复到最新时间

--recovery-target-time TIMESTAMP  
This parameter specifies the time stamp up to which recovery will proceed. If not specified, continue recovery to the latest time.

如果不指定，则恢复到最新xid

--recovery-target-xid XID  
This parameter specifies the transaction ID up to which recovery will proceed. If not specified, continue recovery to the latest xid.

如果不指定，则默认使用true，即恢复到包含恢复目标XID的commit record为止，或者第一笔commit record ts>指定ts的 commit redo record为止；
如果是false则不apply恢复目标XID的commit record，或者不apply第一笔commit record ts>=指定ts的 commit redo record。

--recovery-target-inclusive  
Specifies whether we stop just after the specified recovery target (true), or just before the recovery target (false). Default is true.

是否使用硬链接复制archive log，而不需要拷贝文件

例子

恢复时，需要注意，你可以选择原地恢复（覆盖式），或者使用新的$PGDATA作为恢复目标。

但是请注意，不管是哪种恢复方式，如果在本机恢复的话，pg_rman可能会覆盖原有的数据文件，arch, pg_xlog目录中的文件，所以，如果你要保留原数据，建议先将原数据目录重命名。

原地恢复
使用新的$PGDATA恢复

在本机恢复的例子
1. 停库

pg_ctl stop -m fast -D /data04/digoal/pg_root_1922

2. 重命名原数据相关目录

  $PGDATA  
mv /data04/digoal/pg_root_1922 /data04/digoal/old_pg_root_1922  
  PG_XLOG  
mv /data05/digoal/pg_xlog_1922 /data05/digoal/old_pg_xlog_1922  
  表空间  
mv /data02/digoal/tbs1_1922 /data02/digoal/old_tbs1_1922  
  归档目录，除了要重命名，还需要新建一个原目录  
mv /data04/digoal/arc_log1922 /data04/digoal/old_arc_log1922  
mkdir /data04/digoal/arc_log1922  
...

3. pg_rman restore

pg_rman restore -B /data05/digoal/pgstdbak -D /data04/digoal/pg_restore_root   
WARNING: could not open pg_controldata file "/data04/digoal/pg_restore_root/global/pg_control": No such file or directory  
WARNING: could not open pg_controldata file "/data04/digoal/pg_restore_root/global/pg_control": No such file or directory  
INFO: the recovery target timeline ID is not given  
INFO: use timeline ID of latest full backup as recovery target: 1  
INFO: calculating timeline branches to be used to recovery target point  
INFO: searching latest full backup which can be used as restore start point  
INFO: found the full backup can be used as base in recovery: "2016-08-29 15:05:32"  
INFO: copying online WAL files and server log files  
INFO: clearing restore destination  
INFO: validate: "2016-08-29 15:05:32" backup, archive log files and server log files by SIZE  
INFO: backup "2016-08-29 15:05:32" is valid  
INFO: restoring database files from the full mode backup "2016-08-29 15:05:32"  
INFO: searching incremental backup to be restored  
INFO: validate: "2016-08-29 15:13:10" backup, archive log files and server log files by SIZE  
INFO: backup "2016-08-29 15:13:10" is valid  
INFO: restoring database files from the incremental mode backup "2016-08-29 15:13:10"  
INFO: searching backup which contained archived WAL files to be restored  
INFO: backup "2016-08-29 15:13:10" is valid  
INFO: restoring WAL files from backup "2016-08-29 15:13:10"  
INFO: restoring online WAL files and server log files  
INFO: generating recovery.conf  
INFO: restore complete  
HINT: Recovery will start automatically when the PostgreSQL server is started.

4. 修改postgresql.conf, recovery.conf
根据实际情况调整，本例不修改

cd /data04/digoal/pg_restore_root/  
digoal@iZ28tqoemgtZ-> ll  
total 124K  
-rw------- 1 digoal digoal  193 Aug 29 17:05 backup_label.old  
drwx------ 7 digoal digoal 4.0K Aug 29 17:05 base  
drwx------ 2 digoal digoal 4.0K Aug 29 17:05 global  
drwx------ 2 digoal digoal 4.0K Aug 29 17:05 pg_clog  
drwx------ 2 digoal digoal 4.0K Aug 29 17:05 pg_commit_ts  
drwx------ 2 digoal digoal 4.0K Aug 29 17:05 pg_dynshmem  
-rw------- 1 digoal digoal 4.4K Aug 29 17:05 pg_hba.conf  
-rw------- 1 digoal digoal 1.6K Aug 29 17:05 pg_ident.conf  
drwx------ 2 digoal digoal 4.0K Aug 29 17:05 pg_log  
drwx------ 4 digoal digoal 4.0K Aug 29 17:05 pg_logical  
drwx------ 4 digoal digoal 4.0K Aug 29 17:05 pg_multixact  
drwx------ 2 digoal digoal 4.0K Aug 29 17:05 pg_notify  
drwx------ 2 digoal digoal 4.0K Aug 29 17:05 pg_replslot  
drwx------ 2 digoal digoal 4.0K Aug 29 17:05 pg_serial  
drwx------ 2 digoal digoal 4.0K Aug 29 17:05 pg_snapshots  
drwx------ 2 digoal digoal 4.0K Aug 29 17:05 pg_stat  
drwx------ 2 digoal digoal 4.0K Aug 29 17:05 pg_stat_tmp  
drwx------ 2 digoal digoal 4.0K Aug 29 17:05 pg_subtrans  
drwx------ 2 digoal digoal 4.0K Aug 29 17:05 pg_tblspc  
drwx------ 2 digoal digoal 4.0K Aug 29 17:05 pg_twophase  
-rw------- 1 digoal digoal    4 Aug 29 17:05 PG_VERSION  
lrwxrwxrwx 1 digoal digoal   27 Aug 29 17:05 pg_xlog -> /data05/digoal/pg_xlog_1922  
-rw------- 1 digoal digoal   88 Aug 29 17:05 postgresql.auto.conf  
-rw------- 1 digoal digoal  22K Aug 29 17:05 postgresql.conf  
-rw------- 1 digoal digoal   44 Aug 29 17:05 postmaster.opts  
-rw-r--r-- 1 digoal digoal  130 Aug 29 17:06 recovery.conf  
digoal@iZ28tqoemgtZ-> cd pg_tblspc/  
digoal@iZ28tqoemgtZ-> ll  
total 0  
lrwxrwxrwx 1 digoal digoal 24 Aug 29 17:05 16719 -> /data02/digoal/tbs1_1922  
cd ..  
vi postgresql.conf  
archive_command = 'cp %p /data04/digoal/arc_log1922/%f'  
vi recovery.conf  
# recovery.conf generated by pg_rman 1.3.2  
restore_command = 'cp /data04/digoal/arc_log1922/%f %p'  
recovery_target_timeline = '1'

5. 如果备份集的时间线发生了变化，需要先手工拷贝到归档目录，再执行restore

digoal@iZ28tqoemgtZ-> pg_rman restore -B /data05/digoal/pgstdbak -D /data04/digoal/pg_restore_root   
WARNING: could not open pg_controldata file "/data04/digoal/pg_restore_root/global/pg_control": No such file or directory  
WARNING: could not open pg_controldata file "/data04/digoal/pg_restore_root/global/pg_control": No such file or directory  
INFO: the recovery target timeline ID is not given  
INFO: use timeline ID of latest full backup as recovery target: 2  
INFO: calculating timeline branches to be used to recovery target point  
ERROR: could not open destination file "/data04/digoal/arc_log1922/00000002.history": No such file or directory  
digoal@iZ28tqoemgtZ-> mkdir /data04/digoal/arc_log1922  
digoal@iZ28tqoemgtZ-> cp /data05/digoal/pgstdbak/timeline_history/00000002.history /data04/digoal/arc_log1922/  
digoal@iZ28tqoemgtZ-> pg_rman restore -B /data05/digoal/pgstdbak -D /data04/digoal/pg_restore_root   
WARNING: could not open pg_controldata file "/data04/digoal/pg_restore_root/global/pg_control": No such file or directory  
WARNING: could not open pg_controldata file "/data04/digoal/pg_restore_root/global/pg_control": No such file or directory  
INFO: the recovery target timeline ID is not given  
INFO: use timeline ID of latest full backup as recovery target: 2  
INFO: calculating timeline branches to be used to recovery target point  
INFO: searching latest full backup which can be used as restore start point  
INFO: found the full backup can be used as base in recovery: "2016-08-29 17:14:20"  
INFO: copying online WAL files and server log files  
INFO: clearing restore destination  
INFO: validate: "2016-08-29 17:14:20" backup and archive log files by SIZE  
INFO: backup "2016-08-29 17:14:20" is valid  
INFO: restoring database files from the full mode backup "2016-08-29 17:14:20"  
INFO: searching incremental backup to be restored  
INFO: searching backup which contained archived WAL files to be restored  
INFO: backup "2016-08-29 17:14:20" is valid  
INFO: restoring WAL files from backup "2016-08-29 17:14:20"  
INFO: restoring online WAL files and server log files  
INFO: generating recovery.conf  
INFO: restore complete  
HINT: Recovery will start automatically when the PostgreSQL server is started.  
digoal@iZ28tqoemgtZ-> cd /data04/digoal/arc_log1922  
digoal@iZ28tqoemgtZ-> ll  
total 16K  
lrwxrwxrwx 1 digoal digoal 71 Aug 29 17:18 00000002000000470000000C -> /data05/digoal/pgstdbak/20160829/171420/arclog/00000002000000470000000C  
lrwxrwxrwx 1 digoal digoal 71 Aug 29 17:18 00000002000000470000000D -> /data05/digoal/pgstdbak/20160829/171420/arclog/00000002000000470000000D  
lrwxrwxrwx 1 digoal digoal 87 Aug 29 17:18 00000002000000470000000D.00000028.backup -> /data05/digoal/pgstdbak/20160829/171420/arclog/00000002000000470000000D.00000028.backup  
-rw------- 1 digoal digoal 42 Aug 29 17:18 00000002.history

6. 启动恢复目标数据库

digoal@iZ28tqoemgtZ-> pg_ctl start -D /data04/digoal/pg_restore_root  
server starting  
digoal@iZ28tqoemgtZ-> LOG:  00000: redirecting log output to logging collector process  
HINT:  Future log output will appear in directory "pg_log".  
LOCATION:  SysLogger_Start, syslogger.c:622  
digoal@iZ28tqoemgtZ-> psql -h 127.0.0.1 -p 1922  
psql (9.5.3)  
Type "help" for help.  
postgres=# select pg_is_in_recovery();  
 pg_is_in_recovery   
-------------------  
 f  
(1 row)  
postgres=# \l+  
                                                                   List of databases  
        Name        |  Owner   | Encoding | Collate | Ctype |   Access privileges   |  Size   | Tablespace |                Description                   
--------------------+----------+----------+---------+-------+-----------------------+---------+------------+--------------------------------------------  
 contrib_regression | postgres | UTF8     | C       | C     |                       | 7137 kB | pg_default |   
 db1                | postgres | UTF8     | C       | C     |                       | 111 MB  | tbs1       |   
 postgres           | postgres | UTF8     | C       | C     |                       | 797 MB  | pg_default | default administrative connection database  
                    |          |          |         |       | postgres=CTc/postgres |         |            |   
 template1          | postgres | UTF8     | C       | C     | =c/postgres          +| 7137 kB | pg_default | default template for new databases  
                    |          |          |         |       | postgres=CTc/postgres |         |            |   
(5 rows)

注意事项

小心覆盖原有的部分
建议先将原有的目录重命名，或者在其他机器恢复，
用户可以指定$PGDATA，恢复到新的目标目录，但是arch_log, 表空间, pg_xlog目录无法指定新的位置，所以原地还原时，必须注意这些目录可能被覆盖，先重命名是比较好的手段。

pg_rman 可以优化的地方

检查哪些要恢复的块与目标块的CRC是否一致，如果一致，不需要拷贝，减少WRITE。

pg_rman 软件限制

pg_rman has the following restrictions.

Requires to read database cluster directory and write backup catalog directory.
For example, you need to mount the disk where backup catalog is placed with NFS from database server.
实际上不是必须的，如果没有指定原来的$PGDATA，则使用备份集的元数据。
Block sizes of pg_rman and server should be matched. BLCKSZ and XLOG_BLCKSZ also should be matched.
编译pg_rman时，最好使用启动数据集的集群软件的pg_config。确保块大小一致。
因为需要做块的校验。读取LSN等，都与块大小有关。
If there are some unreadable files/directories in database cluster directory, WAL directory or archived WAL directory, the backup or restore would be failed.

Getting backup from standby-site, pg_rman has the follow restrictions too.

The environment of replication should be built right, or the backup will not finish.
You can’t get backups on master and standby at the same time.
因为pg_rman使用的是exclusive bakcup，（pg_start_backup），所以同一时间，只能跑一次pg_start_backup。
pg_basebackup则使用的是shared backup，可以跑多个。
You can’t get backups on multi standbys at the same time too.
道理同上
Basically, the backup from standby-site is used for restoring on MASTER.
pg_rman doesn’t treat the backup as restoring on standby automatically.
If you want to restore the backup on STANDBY, you have to manage archive logs with your self.
因为备库不归档，所以从standby备份时，需要解决归档备份的问题。
我在前面的文档中已经提及，包括解决思路。

When using storage snapshot, pg_rman has the following restrictions too.

If your snapshot does not have any file update time, incremental backup is same with full backup.
Because pg_rman judges performing full backup or incremental backup by update time for files.
If files don’t have update time because of storage snapshot specification, pg_rman performs full backup every time.
You can’t backup for one side works storage with split mirror snapshot.
Before you execute pg_rman, you should perform storage “RESYNC”.
After pg_rman performs backup with split mirror snapshot, storeage will be “SPLITTED”(works on one side).
pg_rman perform SPLIT command for getting snapshot, but doesn’t perform RESYNC command.
You cant’t get snapshot from different vendor storages in a time.
You cant’t use some vendor storages which have different commands for getting snapshot.
The script and commands for getting storage snapshot should be executable.
It’s expected to have authority of root for getting snapshot or mounting volumes.
So a user, performs pg_rman, is granted to execute any commands in the script.
If you use LVM(Logical Volume Manager), it’s needed root authority for mount, umount, lvcreate, lvremove, lvscan commands.
You should granted to these commands with sudo command to non-password executable.

增量备份代码

上次备份以来，数据块的LSN是否发生了变化，如果自从上次备份的start_lsn以来没有发生变化，则不备份。

代码举例

else  
{  
        pgBackupGetPath(prev_backup, prev_file_txt, lengthof(prev_file_txt),  
                DATABASE_FILE_LIST);  
        prev_files = dir_read_file_list(pgdata, prev_file_txt);  
        /*  
         * Do backup only pages having larger LSN than previous backup.  
         */  
        lsn = &prev_backup->start_lsn;  
        xlogid = (uint32) (*lsn >> 32);  
        xrecoff = (uint32) *lsn;  
        elog(DEBUG, _("backup only the page updated after LSN(%X/%08X)"),  
                                        xlogid, xrecoff);  
}  
...
/* backup files from non-snapshot */
pgBackupGetPath(&current, path, lengthof(path), DATABASE_DIR);
backup_files(pgdata, path, files, prev_files, lsn, current.compress_data, NULL);

备份结果backup.ini相关代码

# configuration  
BACKUP_MODE=FULL  
FULL_BACKUP_ON_ERROR=false  
WITH_SERVERLOG=true  
COMPRESS_DATA=true  
# result  
TIMELINEID=1  
START_LSN=43/d5000028  
STOP_LSN=43/d5000168  
START_TIME='2016-08-26 15:43:39'  
END_TIME='2016-08-26 15:44:27'  
RECOVERY_XID=3896508572  
RECOVERY_TIME='2016-08-26 15:44:18'  
TOTAL_DATA_BYTES=823571731  
READ_DATA_BYTES=823571731  
READ_ARCLOG_BYTES=234881668  
READ_SRVLOG_BYTES=218248  
WRITE_BYTES=206009921  
BLOCK_SIZE=8192  
XLOG_BLOCK_SIZE=8192  
STATUS=OK

对应的数据结构

/*  
 * pg_rman takes backup into the directory $BACKUP_PATH/<date>/<time>.  
 *  
 * status == -1 indicates the pgBackup is invalid.  
 */  
typedef struct pgBackup  
{  
        /* Backup Level */  
        BackupMode      backup_mode;  
        bool            with_serverlog;  
        bool            compress_data;  
        bool            full_backup_on_error;  
        /* Status - one of BACKUP_STATUS_xxx */  
        BackupStatus    status;  
        /* Timestamp, etc. */  
        TimeLineID      tli;  
        XLogRecPtr      start_lsn;  
        XLogRecPtr      stop_lsn;  
        time_t          start_time;  
        time_t          end_time;  
        time_t          recovery_time;  
        uint32          recovery_xid;  
        /* Size (-1 means not-backup'ed) */  
        int64           total_data_bytes;  
        int64           read_data_bytes;  
        int64           read_arclog_bytes;  
        int64           read_srvlog_bytes;  
        int64           write_bytes;  
        /* data/wal block size for compatibility check */  
        uint32          block_size;  
        uint32          wal_block_size;  
        /* if backup from standby or not */  
        bool            is_from_standby;  
} pgBackup;

备份开始时记录pg_start_backup调用返回的lsn，写入backup->start_lsn

/*  
 * Notify start of backup to PostgreSQL server.  
 */  
static void  
pg_start_backup(const char *label, bool smooth, pgBackup *backup)  
{  
        PGresult           *res;  
        const char         *params[2];  
        params[0] = label;  
        elog(DEBUG, "executing pg_start_backup()");  
        reconnect();  
        /* Assumes PG version >= 8.4 */  
        /* 2nd argument is 'fast' (IOW, !smooth) */  
        params[1] = smooth ? "false" : "true";  
        res = execute("SELECT * from pg_xlogfile_name_offset(pg_start_backup($1, $2))", 2, params);  
        if (backup != NULL)  
                get_lsn(res, &backup->tli, &backup->start_lsn);  
        elog(DEBUG, "backup start point is (WAL file: %s, xrecoff: %s)",  
                        PQgetvalue(res, 0, 0), PQgetvalue(res, 0, 1));  
        PQclear(res);  
        disconnect();  
}

备份停止，调用pg_stop_backup，从返回结果中取出LSN，写入backup->stop_lsn

/*  
 * Notify end of backup to PostgreSQL server.  
 */  
static void  
pg_stop_backup(pgBackup *backup)  
{  
        elog(DEBUG, "executing pg_stop_backup()");  
        wait_for_archive(backup,  
                "SELECT * FROM pg_xlogfile_name_offset(pg_stop_backup())");  
}  
static void  
wait_for_archive(pgBackup *backup, const char *sql)  
{  
        PGresult           *res;  
        char                    ready_path[MAXPGPATH];  
        int                             try_count;  
        reconnect();  
        res = execute(sql, 0, NULL);  
        if (backup != NULL)  
        {  
                get_lsn(res, &backup->tli, &backup->stop_lsn);  
                elog(DEBUG, "backup end point is (WAL file: %s, xrecoff: %s)",  
                                PQgetvalue(res, 0, 0), PQgetvalue(res, 0, 1));  
        }  
        /* get filename from the result of pg_xlogfile_name_offset() */  
        elog(DEBUG, "waiting for %s is archived", PQgetvalue(res, 0, 0));  
        snprintf(ready_path, lengthof(ready_path),  
                "%s/pg_xlog/archive_status/%s.ready", pgdata, PQgetvalue(res, 0, 0));  
        PQclear(res);  
        res = execute(TXID_CURRENT_SQL, 0, NULL);  
        if(backup != NULL)  
        {  
                get_xid(res, &backup->recovery_xid);  
                backup->recovery_time = time(NULL);  
        }  
        disconnect();  
        /* wait until switched WAL is archived */  
        try_count = 0;  
        while (fileExists(ready_path))  
        {  
                sleep(1);  
                if (interrupted)  
                        ereport(FATAL,  
                                (errcode(ERROR_INTERRUPTED),  
                                 errmsg("interrupted during waiting for WAL archiving")));  
                try_count++;  
                if (try_count > TIMEOUT_ARCHIVE)  
                        ereport(ERROR,  
                                (errcode(ERROR_ARCHIVE_FAILED),  
                                 errmsg("switched WAL could not be archived in %d seconds",  
                                        TIMEOUT_ARCHIVE)));  
        }  
        elog(DEBUG, "WAL file contains backup end point is archived after %d seconds waiting",  
                        try_count);  
}

validate 时，改backup.ini的STATUS字段

validate.c  
        for (i = 0; i < parray_num(backup_list); i++)  
        {  
                pgBackup *backup = (pgBackup *)parray_get(backup_list, i);  
                /* clean extra backups (switch STATUS to ERROR) */  
                if(!another_pg_rman &&  
                   (backup->status == BACKUP_STATUS_RUNNING ||  
                        backup->status == BACKUP_STATUS_DELETING))  
                {  
                        backup->status = BACKUP_STATUS_ERROR;  
                        pgBackupWriteIni(backup);  
                }  
                /* Validate completed backups only. */  
                if (backup->status != BACKUP_STATUS_DONE)  
                        continue;  
                /* validate with CRC value and update status to OK */  
                pgBackupValidate(backup, false, false, (HAVE_DATABASE(backup)));  
        }  
...  
                /* update status to OK */  
                if (corrupted)  
                        backup->status = BACKUP_STATUS_CORRUPT;  
                else  
                        backup->status = BACKUP_STATUS_OK;  
                pgBackupWriteIni(backup);

注意

1.备份参数 -C 表示无缝checkpoint, 所以可能很慢，视checkpoint_completion_target和segment_size的配置。
如果你发现pg_rman开始很慢，可以把-C去掉，速度就快了，但是可能在高峰时，造成冲击。
建议高峰是不要备份。
2.BUG
unix socket 是$PGDATA时, validate会报错

pg_rman validate  
INFO: validate: "2016-08-26 16:19:25" backup, archive log files and server log files by CRC  
ERROR: invalid type '?' found in "/data05/digoal/pgbak/20160826/161925/file_database.txt"  
vi /data05/digoal/pgbak/20160826/161925/file_database.txt  
.s.PGSQL.1921 ? 0 0 0777 2016-08-26 15:35:05

修改一下dir.c的代码即可修复这个问题，修改如下

if (strncmp(path, ".s.PGSQL", 7) != 0 && type != 'f' && type != 'F' && type != 'd' && type != 'l')
        ereport(ERROR,
                 errmsg("invalid type '%c' found in \"%s\"", type, file_txt)));

阿里云的小伙伴们加油，努力 做好内核与服务，打造最贴地气的云数据库 。