The cluster runs on apt-based systems. It is designed for high-availability: failure of one server is not critical. However, there is no automatic failover configuration. Instead, manual recovery is possible within minutes.
- Load balancers run varnish. Its configuration file takes into account the context_breakpoint cookie that's used to implement responsive delivery. The same server also has memcached.
- Application servers run nginx and a recent php5-fpm through unix sockets. There's also drush. The filesystem is shared through glusterfs.
- MariaDb is configured in master-slave mode on database servers. They also run Apache Solr.
All servers set up exim to work as smarthost, sending mails through gmail. In practice, gmail limits outgoing emails to a few thousands per day, so it is better to replace it by a dedicated solution such as Mailchimp. There's also newrelic for server monitoring on all servers.
The playbook assumes that all servers have public IPs on eth0 and sit in the private network on eth1.
For the rest, check the code in github.