At a high-level, if a connection (representing a real-time web request) is all reads, it can run off any of the read-replicas. If a connection does read, read, read, write, read, the connection is switched to the master nodes on the first write and then stays on the master for the remainder of the connection. There is no replication lag concern for the application to make sure it reads the data that was just written since it's pinned to the master.