As Richard pointed out the best way at the moment is to use atomic batches where you update many deltas. If something goes wrong just replay the batch.
The other possible solution is to use ZooKeeper as a coordination and distributed locking service: http://ria101.wordpress.com/tag/zookeeper/
Another possible solution is to use counters so you don't need to do this
your current_balance = read_users_balance(u);new_balance = current_balance + delta_for_user(u);because with counters you don't need to read balance before updating. http://www.datastax.com/dev/blog/whats-new-in-cassandra-0-8-part-2-counters
However there is a problem with counters, they are not idempotent so if you don't receive acknowledge that your increment/decrement was successful you cannot replay that counter as it can lead to overcount.
New counters will solve that problem.