I'm at a big company that uses hadoop, and someone else maintains the infrastructure. But, the problem we have is that hadoop is a very leaky abstraction and it's very easy for someone that doesn't know implementation details to break the grid. It's relatively easy to exhaust the memory of the jobtracker, we constantly have problems with mapper starvation because someones seemingly innocent job is causing issues, etc, etc.

My experience is that it doesn't just work, even when someone else maintains the infrastructure.