There are multiple ways of keeping the master highly available.
One would be to demote and promote manually if you need to. The better way would be to delegate that task to a mechanism that knows the current state of all (possible) master nodes and can perform the failover procedure automatically.
Known methods, when only using open-source software, are building Pacemaker/ Corosync clusters with self-written OCF agents. Another way could be using keepalived.
