The problem I have with locking mechanisms is that they prevent any other activity from happening while the lock is active, which could be extremely inconvenient for users. In the train-and-hotel example, a single user would be able to stop all other users from booking trains or hotels for whatever the length of the lock is. Sure, you could lock individual train tickets, but that’s already a fairly application specific solution.
I think the best solutions to these problems are those solutions that recognize that shards are virtual galaxies, and not physical sets of computers, and so the nodes executing two shards can, eg. with stateless client witness protocols, temporarily be merged for one block.
As I mention here, you can even cycle through “leader shards” in each round, where transactions from the leader shard are executed first and can read and write across shards, and executors on both the leader shard and followe shards use stateless client techniques to execute the part of the operations that are happening on shards that they are not native to.