The following is what happens when L1 is not enabled for distribution:
Let's say you call put on 5 keys [k1-k5] on node A of a two node cluster [A,B] and the distribution should be something like: A[k1,k2,k5] and B[k3,k4].
What really happens is that A stores all keys [k1,k2,k3,k4,k5] and B stores [k3,k4]. So, if you now call remove[k1-k5] from B, all the keys in B will be deleted and only the keys that should have belonged to A would be deleted, so the result is that A will still contain [k3,k4].
With L1, this is not an issue cos whenever there's a write on that key, the L1 is invalidated, so the above test would result on both A and B not containing any keys.
I'm in the process of replicating this in a smaller unit test. I'm not sure yet what the fix for this would be.