When client node is asked to finish commit, it sends a single requests to primary master node. That node will then take a read-lock for all objects involved in the transaction. When storage nodes take that lock, they write received objects to a temporary location, where they will survive a node restart but are not accessible by other nodes.
Once all storage acknowledged the lock take, primary master node asks them to release all locks on transaction objects. When storage nodes release this lock, they also move previously written objects to the definitive location, where they become available as the current revision of given object to other nodes.
Also, primary master node answers committing client that commit is over, and sends cache invalidations to other client nodes. In both cases, this is used to invalidate client caches, with the only difference that committing client should store committed data in its local cache rather than just flushing.
This locking scheme achieves a "network barrier": it prevents being able to read simultaneously different versions of the same object while storage nodes are independently committing data.