A large deployment will have a large number of nodes. If you are growing your installation from scratch, you may have started with a single Puppet master running the built-in WEBrick server and moved up to a passenger installation. At a certain point in your deployment, a single Puppet master just won't cut it—the load will become too great. In my experience, this limit was around 600 nodes. Puppet agent runs begin to fail on the nodes, and catalogs fail to compile. There are two ways to deal with this problem: divide and conquer or conquer by dividing.

That is, we can either split up our Puppet master and divide the workload among several machines or we can make each of our nodes apply our code directly using Puppet agent (this is known as a masterless configuration). We'll examine each of these solutions separately.