I have noticed that in most use cases of redis/memcached/etc, one will abstract it into some domain specific data-structure. At scale that abstraction becomes a

I have noticed that in most use cases of redis/memcached/etc, one will abstract it into some domain specific data-structure. At scale that abstraction becomes a service of it's own (job servers, caches, pub sub, etc.) at which point it doesn't make sense to keep the data separated from the service, because all it does is it adds an extra layer of indirection. I presume at google it's really not that big of a deal to write a fast priority queue service that keeps things in it's own memory and talks protobuf directly. Relatedly I also noticed that teams that overuse memstores predominantly use application languages that have a terrible concurrency story (e.g. python), fall back to multiprocessing where creating anything stateful is not as obvious as keeping it in memory, so they offload that state to redis.