[lkrubner]

"I could see doing what you're describing to build a prototype"

It's very much for prototypes, and especially greenfield projects. If I was, instead, doing something like building a new service, inside an Enterprise that was already using something like the unified log architecture that Jay Kreps has described, then I would certainly think hard about what the schema would be for the particular service I was building -- after all, in such situations you're never going to pull all of the data out of Kafka, so you automatically have to figure out what part of the data you want. LinkedIn currently stores 900 terabytes of data in its Kafka instance, and I'm unlikely to write a new service that actually needs all of the 900 terabytes of data. So merely by thinking about the question "What of this data do I need?" I'm already implicitly thinking about a schema.

Having said all of that, how often have you written a service where you got the schema 100% correct on your first try, and no further changes to the schema were needed. Possibly you are smarter than I am, but I personally have never done that. All of my first attempts need later adjustment.