Uber Serves 40M Qps from Online Storage Using an Integrated Cache

This smells bad architecture to me. Been there, seen that many times. That load could be decreased by factor of 10 by some simple clever ideas and changes. Then it could further be lowered. I just can't imagine a map app and a supporting business to be _that_ demanding in terms of storage use. I do see congestion points, but it's not storage.

Unless they're adding up cache lookups to the overall number, which is then just a clickbait.

> The stateless query engine layer is responsible for query planning, routing, sharding, schema management, node health monitoring, request parsing, validation, and AuthN/AuthZ.

Why don't the nodes expose health checks and other software like Promteheus/Alertmanager check node health?

Similar question for AuthN/AuthZ

Obviously Uber knows more than I do... just curious if this design decision (probably just splitting hairs) makes the most sense

Reminds me of FASTER but for querying instead of WALs and k/v stores. I have a feeling we will be seeing a lot more of these types of optimizations in the near future.