Server-less configuration not fault-tolerant

[pron]

The problem is that during item migration (getx-putx), the old owner needs to both transfer ownership to the requesting node (putx), as well as notify its slave (inv) that it is no longer the owner (of the item). If one of them (slave or peer) do not get the message and the old owner fails, the item may be lost (both peer and slave think they're not the owner) or become conflicted (both think they're the owner).

I suspect that doing this correctly would require a consensus protocol, which might make the whole thing not worthwhile. The only hope is that we could somehow take advantage of the fact that we know that the old owner has failed, i.e. that the node-switch event is received by everyone (we have consensus over that provided by ZooKeeper/JGroups).