Given the scale described, it sounds like they could have had a single machine that held the data in memory and periodically flushed to disk/DB to support failing over to a standby.
You're basically describing how we used redis for this project.
I suppose so, but then what did you gain from the extra hop to redis?
Not having to implement redis ourselves.
Did you even bother reading the first few paragraphs? They talked about their usage of Redis for this. Next time please read the article before replying.
Perhaps, rather than posting an insult, you should consider the possibility that you misinterpreted my comment?
I would be slightly more careful and just use a cluster of servers with a simple consensus algorithm (like raft). A simple C++ server with a raft library plus uWebSockets should be able to handle a lot of load.