Curiously enough, I've used to work in hospitality! (https://pwy.io/posts/mimalloc-cigarette/ & https://pwy.io/posts/memory-for-nothing/)

Our dataset wasn't so large as to warrant experimenting with GPUs - `cost(extraCpu) < cost(developerExperimentingWithGpus)` - but the CPU implementation got a lot of love; it was mostly engineering-based work, though ("how to load stuff from the disk effectively", "how to make it multi-threaded" etc.), not math-based work ("if we represented prices using $someFancyStructure, we could ...").

I like when seemingly engineering-oriented problems turn out to be way more approachable once you apply / invent / adjust some math foundations, e.g. like with Feldera (https://sigmodrecord.org/publications/sigmodRecord/2403/pdfs...) - wouldn't figure something like this myself in a hundred years, I'd just probably apply maps on top of caches on top of whatever else I could pile until it works!

A channel manager using Rust for the caching layer? That's actually kind of incredible. Might want to pick your brain someday :)

And Feldera/DBSP is absolute magic - definitely something I want to dig deeper into!