We saw in an earlier section that an atom provides atomic read-and-update operation. What if we need to perform an atomic read-and-update operation across two or even more number of atoms? This clearly poses a coordination problem. Some entity has to watch over the process of reading and updating, so that the values are not corrupted. This is what a ref provides—a Software Transactional Memory (STM) based system that takes care of concurrent atomic read-and-update operations across multiple refs, such that either all the updates go through, or in the case of failure, none does. Like atoms, on failure, refs retry the whole operation from scratch with the new values.

Clojure's STM implementation is coarse grained. It works at the application level objects and aggregates (that is, references to aggregates), scoped to only all the refs in a program, constituting the "Ref world". Any update to a ref can only happen synchronously, in a transaction, in a dosync...