True Zero Downtime HAProxy Reloads: An Alternative Approach

• Segregate server instances in to a red and blue group
• Bring the daemon of the opposing color online
• Use iptable rules to direct incoming traffic to the current daemon color
• Repeat the process, switching colors every time a reload is needed.

This method is simpler than Yelp's method of routing TCP SYN requests to a plug queue but does seem to add a minor amount of latency to route *all* traffic through iptable rules at all times. Unbounce likes this method because it allows for daemon restarts that take too much time to hold a request in flight.