My understanding of the "premature" part is the fact that without profiling your application in some way, you have no idea apriori where the bottlenecks are. This "optimization" is premature in that its cost in w/e metric you want (readability, etc.) is payed and you have no idea about its relative performance cost. For example, after performing this "optimization" you save a few milliseconds but the Mnesia sync between two instances is orders of magnitude more time.