Others have mentioned that the article calls out ignoring traditional in-memory cache daemons because of the additional network time, but with a targeted p50 response time (of their HTTP service fronting all of this), and caches like Redis and memcached being able to respond in hundreds of microseconds... it does feel like they didn't actually run the numbers.
The other natural alternative would simply be to run Redis/memcached and colocate this HTTP service on the same box. Now your "network latency" component is almost entirely negligible, and you've deferred the work of managing memory to applications _designed_ around doing so.
Others have mentioned that the article calls out ignoring traditional in-memory cache daemons because of the additional network time, but with a targeted p50 response time (of their HTTP service fronting all of this), and caches like Redis and memcached being able to respond in hundreds of microseconds... it does feel like they didn't actually run the numbers.
The other natural alternative would simply be to run Redis/memcached and colocate this HTTP service on the same box. Now your "network latency" component is almost entirely negligible, and you've deferred the work of managing memory to applications _designed_ around doing so.