If the business got that big, I'd be hiring people to do it for me and there would be some benefits to using more standard approaches since getting people who think like I do would suck and it would suck for employees to have to follow their boss's crazy ideas on the right way to make a responsive database for a point of sale environment.
But there's nothing preventing an in-memory database being duplicated, load balanced, redundantized (or whatever the right word is for adding redundant mirror servers) etc. In fact the consistency transactions are easier because the parties are algorithms talking to each other, rather than algorithms talking via a storage medium. You would have to write it yourself though since the world still seems to have a hard on for SQL long after they left Cobol behind, which is bad for the same reasons. In a point of sale situation, the order of transactions is not that important except for specific things, like you can't return an item before you've purchased it. So it's easy to post-hoc order the transaction log, a bit like the bitcoin block chain, where the current head of the list is indeterminate, but it's resolved a few steps back.
My issue isn't with disk vs. memory. That's a normal tradeoff. My issue is with SQL, which is horrible in every way that an insecure, computationaly undecidable lump of middleware can be.