Vertical scaling and scaling out via replicas is only excusable in a world where sharding isn't a solved problem.

With a bunch of read replicas you have to deal with eventual consistency when you get replication lag as well as the inability to scale writes.

If you shard out horizontally you get write scaleability, and read scalability with much less chance of replication delay. You like likely be reading from the master node most of the time. You also gain a bunch of operational flexibility, and smaller failure domains.