Redis is not mentioned here at all so maybe the author is thinking mostly of other NoSQL software here, but well, in the case of Redis the whole point was not just the in-memory performance part, but the data model as well. My claim is that you can't really exploit the advantage of using memory if you perpetuate in using the memory to represent the same data model that you were using with relational databases. For instance think at Redis sorted sets in the use case of leaderboards in popular games (the same pattern is used in a number of applications that have nothing to do with games). To use SQL, even an in memory one, in such use case is not going to work. So in the case of Redis the point was to remove the interface between the user and the way the data is actually fetched from data structures, to make the user do the choices in a very direct way. Thus this article does not apply to Redis in my opinion. I've the feeling that many other NoSQL products could argument like that, but in other areas of their research and difference. For instance I've issues to see how modern SQL systems can replace CRDTs based stores.