as someone who deployed puppet apply infra to separated and geographically distributed systems for Apple in every store worldwide, there was (at the time) a serious crash problem with teh agent… I think it was a memory leak. However, today, I would count using the server as indespensible, tbh. The amount of data about my fleet coupled with the APIs I can leverage is really hard to beat…. just the intelligence about individual nodes all in a central, queryable location…