We write out the cluster config and update the script using Chef.

In an emergency situation, chef-client is a wee bit slow ;)

Also we like having the script standalone so there's just one moving part.