FineGrainedAtomicMap doesn't respect the REPEATABLE_READ isolation level: a transaction holding a FGAM reference will see subkeys added by another concurrent transaction.

A test with PESSIMISTIC locking is here: https://github.com/danberindei/infinispan/commit/526ae4c414009b2a95b39a65b45bf0467cb78d73

We need to test OPTIMISTIC locking and the coarse-grained AtomicMap as well.