These slides were used by Wei Shan from GMO GlobalSign while presenting at pgDay Asia 2016. He discussed about challenges with the maintenance of Pacemaker/Corosync HA Clusters and how he migrated over to repmgr. He also did a short demo
6. ● Open-source alternative to Red Hat Cluster Suite
● Extremely popular choice in the open-source world
● Made up of 2 different stack of software
○ Pacemaker
○ Corosync/Heartbeat
● Complicated to get the configuration correct
10. ● Lightweight connection pooler for PostgreSQL
● Open-source
● Acts as the single point of entry to the database
● Useful for managing huge number of incoming
connections to the database
● Latest version - v1.7.2
12. ● Common Address Redundancy Protocol (CARP)
● Linux’s implementation of CARP from FreeBSD
● Allows multiple hosts to share a single IP address
● Management of Virtual IP for failover purpose
● For client connectivity to Pg_bouncer
● Latest version => v1.5.2
14. ● Developed by 2ndQuadrant
● Open-source
● Manages replication and failover for your
PostgreSQL HA cluster
● Latest version - v3.1.1
15. ● Linux or Unix only
● repmgr 2.0 is for PostgreSQL 9.0 to 9.4
● repmgr 3.0 is for PostgreSQL 9.3 or higher
● Does not take care of client failover!!
16. ● Automatic failover capabilities
● Provisioning of standby servers
● 2 main tools
○ repmgr => Perform administrative tasks
○ repmgrd => Perform monitoring, automatic failover and
notification events
17. ● Requires a database to store cluster metadata
● Runs as postgres user
● Password-less SSH connectivity between all
hosts
● Recommended to run at with an odd number
cluster
18. The decision whether a server can be promoted depends whether the majority of servers are
"visible". If you have three servers - primary and standby in one location, and a second
standby in another location - and the network to the second standby goes down, the second
standby will see it's in the minority (its location represents 1/3 of the servers) and won't
promote itself.
If you have two servers in each location, you'd need an additional witness server so one
location still has a "majority" - otherwise in the event of a network disconnection you might
end up with one standby in each location promoting itself.