I interviewed for the Bun guys and totally flopped because I was not prepared for the interview format. It was "here's a task, do whatever you need to including AI to get it done". From that, I'd infer that the usage is pretty substantial.
Not the parent, but wanting the rare case to be fast is definitely not a HFT exclusive thing.
I write software packet processors for a living, and it's common to have requirements stated as "process X packets per second, no matter what", ie youre judged by the worst case kind of packets. Also common is optimization of 99.9th percentile latency.
There's also just cases like if (being ddosed) or if (overloaded) where you definitely want the rare case to be faster.
As to point 1, there's.... a significant doubt as to whether these actually change performance. I have never personally managed to confirm a performance gain from them, and not for lack of trying. CPUs also seem to not give them a ton of weight, and will probably override you if they see a way more common other path.
Possible I'm just not remembering the history right, but I think this is from when "Equinix metal" was packet.com. I think this is a handshake deal they had from before they were bought, and it's going away as packet.com becomes more integrated into Equinix.
reply