Perhaps we could change versioning and detection with different loads with respect to time. Say if we are getting too many aborts we may start using a lazy optimistic instead of eager pessimistic and vice versa. This may help with burst operations. It sounds too complicated to implement though.