At a high enough level, all computing is functional. If your language isn’t functional (state-change and side-effect free), your process is isolating it. If your process is modifying global state, then your hypervisor or individual machine is isolating it. If your machine is working with a cluster of machines to maintain and modify global state, then all other machines not in the cluster work in isolation.
The model of global state is a difficult one to maintain as scale increases. The time to synchronize data grows and the number of complications due to stale data gets worse. At the largest scales, the distinction between database and communication layer (think pub-sub) breaks down, and they effectively become one. This is the model of tools like RethinkDB where the query becomes an asynchronous request for updates to a particular subset of the data model.
The latest software modeling paradigms make a point of restricting state to a locale via microservices. Each microservice is allowed its own state, but communicates with all other parts of the system via a common message bus. Tools like Storm and Spark specialize in connecting these services together into a larger dataflow model.
It makes sense. Physical locality is guaranteed to eventually constrain software, regardless of the programming models we’ve gotten away with in the past. I think we would do well to recognize that, when stretched to the largest scales, software is relatively small units of computation and the data flowing between them (just like hardware). Aligning our heads, tools, and techniques around this model is likely to yield the best results long-term.
Pick up and learn a functional language today!