Wow. That is a great write-up. I was admittedly skeptical of "writing your own cache" but these guys are super-OCD and obviously capable of producing best-in-class implementations. Kudos.

...kind of wondering, they had so at least three "well, go doesn't do ..." (lock-less data structures, thread local storage, and channels aren't high perf) rabbit holes that I wonder if go was really the right implementation language for their project (iirc they've also had articles go by about writing best-in-class go versions of rocksdb/etc.).

But, they do seem very capable of rolling their own infra, and I suppose someone has to do it to bootstrap ecosystems in each new/upcoming language.