Using Barman for PostgreSQL Disaster Recovery

Nidhi Bansal

There must be many powerful tools available as a backup and restore option for PostgreSQL in general; Barman, PgBackRest, BART are to name a few in this context. What caught our attention was that Barman is a tool which is catching up fast with production deployment and market trends. 

Be it a docker based deployment, need to store backup in a different cloud storage or highly customisable disaster recovery architecture needs - Barman is a very strong contender in all such cases.

This blog explores Barman with few assumptions about deployment, however in no case this should be considered only possible feature-set. Barman is way beyond what we can capture in this blog and must be explored further if ‘backup and restore of PostgreSQL instance’ is considered.

DR Ready Deployment Assumption 

RPO=0 generally comes at a cost - synchronous standby server deployment would often meet that, but then it impacts the TPS of the primary server quite often.

Like PostgreSQL, Barman gives numerous deployment options to meet your needs when it comes to RPO vs performance. Think of deployment simplicity, RPO=0 or near zero performance impact; Barman fits in all.

We considered the following deployment to establish a disaster recovery solution for our backup and restore architecture.

Figure 1: PostgreSQL Deployment with Barman
Figure 1: PostgreSQL Deployment with Barman

There are two sites (as in general for disaster-recovery sites) - Site-X and Site-Y. 

In Site-X there is:

  • One server ‘pgServer’ hosting a PostgreSQL server instance pgServer, and one OS user ‘postgres’ 
    • PostgreSQL instance also to host a superuser role ‘bmuser’
  • One server ‘bServer’ hosting the Barman binaries and an OS user ‘bmuser’

In Site-Y there is:

  • One server ‘geobServer’ hosting the Barman binaries and an OS user ‘bmuser’

There are multiple types of connection involved in this setup.

  • Between ‘bServer’ and ‘pgServer’:
    • Management-plane connectivity from Barman to the PostgreSQL Instance
    • rsync connectivity to do actual base backup from Barman to the PostgreSQL Instance
    • WAL archiving using barman-wal-archive from the PostgreSQL Instance to Barman
    • WAL streaming using pg_receivexlog at Barman
  • Between ‘bServer’ and ‘geobserver’:
    • Synchronisation between Barman servers to provide geo-replication

Connectivity First 

The primary connectivity needs between the servers is via ssh. In order to make it password-less ssh-keys are used. Let’s establish the ssh keys and exchange them.

On pgServer:

[email protected]$ ssh-keygen -q -t rsa -N '' -f ~/.ssh/id_rsa <<<y 2>&1 >/dev/null

[email protected]$ ssh-copy-id -i ~/.ssh/id_rsa.pub [email protected]

[email protected]$ ssh [email protected] "chmod 600 ~/.ssh/authorized_keys"

On bServer:

[email protected]$ ssh-keygen -q -t rsa -N '' -f ~/.ssh/id_rsa <<<y 2>&1 >/dev/null

[email protected]$ ssh-copy-id -i ~/.ssh/id_rsa.pub [email protected]

[email protected]$ ssh [email protected] "chmod 600 ~/.ssh/authorized_keys"

On geobServer:

[email protected]$ ssh-keygen -q -t rsa -N '' -f ~/.ssh/id_rsa <<<y 2>&1 >/dev/null

[email protected]$ ssh-copy-id -i ~/.ssh/id_rsa.pub [email protected]

[email protected]$ ssh [email protected] "chmod 600 ~/.ssh/authorized_keys"

PostgreSQL Instance Configuration 

There are two main things that we need to re-constitute a postgres instance - The base directory and the WAL / Transactions logs generated thereafter. Barman server intelligently keeps track of them. What we need is to ensure that proper feeds are generated for Barman to collect these artefacts.

Add following lines to postgresql.conf:

listen_addresses = '172.25.130.180'     #as per above deployment assumption

wal_level = replica                     #or higher 

archive_mode = on

archive_command = 'barman-wal-archive -U bmuser bserver pgserver %p'

Archive command is ensuring that when WAL is to be archived by postgres instance, barman-wal-archive utility submits it to the Barman Server. It should be noted that the barman-cli package hence should be made available on ‘pgServer’. There is another option of using rsync if we don’t want to use barman-wal-archive utility.

Add following to pg_hba.conf:

host     all                    all     172.25.130.186/32     md5

host     replication            all     172.25.130.186/32     md5

It is basically allowing a replication and a normal connection from ‘bmserver’ to this postgres instance.

Now just restart the instance and create a super user role called bmuser:

[email protected]$ pg_ctl restart

[email protected]$ createuser -s -P bmuser 

If required, we can avoid using bmuser as a super-user too; that would need privileges assigned to this user. For the above example, we used bmuser as password too.  But that’s pretty much all, as far as a PostgreSQL instance configuration is required.

Barman Configuration 

Barman has three basic component on its configuration:

  • Global configuration 
  • Server level configuration 
  • User who will be running the barman 

 In our case, since Barman is installed using rpm, we have had our global configuration files stored at:

/etc/barman.conf

We wanted to store the server level configuration in bmuser home directory, hence our global configuration file had the following content:

[barman]

barman_user = bmuser

configuration_files_directory = /home/bmuser/barman.d

barman_home = /home/bmuser

barman_lock_directory = /home/bmuser/run

log_file = /home/bmuser/barman.log

log_level = INFO

Primary Barman Server Configuration 

In the deployment above, we decided to keep the primary Barman server in the same data-centre / site where the PostgreSQL instance is kept. The benefit of the same is that there is less lag and faster recovery in case required. Needless to say, less computing and/or network bandwidth needs are required on the PostgreSQL server as well. 

In order to let Barman manage the PostgreSQL instance on the pgServer, we need to add a configuration file (we named pgserver.conf) with the following content:

[pgserver]

description =  "Example pgserver configuration"

ssh_command = ssh [email protected]

conninfo = host=pgserver user=bmuser dbname=postgres

backup_method = rsync

reuse_backup = link

backup_options = concurrent_backup

parallel_jobs = 2

archiver = on

archiver_batch_size = 50

path_prefix = "/usr/pgsql-12/bin"



streaming_conninfo = host=pgserver user=bmuser dbname=postgres

streaming_archiver=on

create_slot = auto

And a .pgpass file containing the credentials for bmuser in the PostgreSQL instance:

echo 'pgserver:5432:*:bmuser:bmuser' > ~/.pgpass 

To understand the important configuration items a little more:

  • ssh_command : Used to establish connection over which rsync will be done 
  • conninfo : Connection string to let Barman establish connection with postgres server
  • reuse_backup : To allow incremental backup with less storage 
  • backup_method : method to take backup of base directory
  • path_prefix : location where pg_receivexlog binaries are stored 
  • streaming_conninfo : Connection string used to stream WAL 
  • create_slot : To make sure slots been created by postgres instance 

Passive Barman Server Configuration 

The configuration of a geo-replication site is pretty simple. All it needs is a ssh connection information over which this passive node site will do the replication.

What is interesting is that such a passive node can work in mix-mode; in other words - they can act as active Barman servers to do backups for PostgreSQL sites and in parallel act as a replication/cascaded site for other Barman servers.

Since, in our case this instance of Barman (on Site-Y) need to be just a passive node, all we need is create the file /home/bmuser/barman.d/pgserver.conf with the following configuration:

[pgserver]

description =  "Geo-replication or sync for pgserver"

primary_ssh_command = ssh [email protected]

With an assumption that the keys have been exchanged and the global configuration on this node is done like previously mentioned - we are pretty much done with the configuration.

And Here is Our First Backup and Restore 

On the bserver ensure that the background process to receive WAL has been triggered; and then check the configuration of the server:

[email protected]$ barman cron

[email protected]$ barman check pgserver

The check should be OK for all substeps. If not, refer to /home/bmuser/barman.log.

Issue backup command on Barman to ensure there is a base DATA on which WAL can be applied:

[email protected]$ barman backup pgserver

On the ‘geobmserver’ ensure that the replication is done by executing following commands:

[email protected]$ barman cron 

[email protected]$ barman list-backup pgserver

The cron should be inserted into the crontab file (if not present). For the sake of simplicity, I have not shown it here. The last command will show that the backup folder has been created on the geobmserver as well. 

Now on the Postgres instance, let's create some dummy data:

[email protected]$ psql -U postgres -c "CREATE TABLE dummy_data( i INTEGER);"

[email protected]$ psql -U postgres -c "insert into dummy_data values ( generate_series (1, 1000000 ));"

The replication of the WAL from the PostgreSQL instance can be seen using the below command:

[email protected]$ psql -U postgres -c "SELECT * from pg_stat_replication ;”

In order to re-create an instance on Site-Y, first ensure that WAL records are switch-over. or this example, to create a clean recovery:

[email protected]$ barman switch-xlog --force --archive pgserver

On the Site-X, let’s bring up a standalone PostgreSQL instance to check if backup is sane:

[email protected]$ barman cron 

barman recover --get-wal pgserver latest /tmp/data

Now, edit the postgresql.conf and the postgresql.auto.conf files as per the needs. Following explain the changes done for this example:

  • postgresql.conf : listen_addresses commented so as to default to localhost
  • postgresql.auto.conf : removed sudo bmuser from restore_command 

Bring up this DATA in /tmp/data and check the existence of your records.

Conclusion

This was just the tip of an iceberg. Barman is way deeper than this because of the functionality it provides - e.g. acting as a synchronised standby, hook scripts and so forth. Needless to say, the documentation in totality should be explored to configure it as per the needs of your production environment.

ClusterControl
The only management system you’ll ever need to take control of your open source database infrastructure.