Having worked with QUIC extensively lately I must say that having protocol entirely in userspace is truly liberating. It also helps that almost everything, including headers is encrypted, as a result middle boxes can't interfere and "optimise" your flow.

For high latency and high packet loss links, like one described in the article, you'll probably need pre-emptive retransmits and I am not sure that simply tuning parameters will get there. Retransmitting before loss is detected cuts bandwidth, but I suspect will improve end to end latency.

Then you see 25% cache hit rate in production and realise that disabling it for benchmark is not a good option either.

How is it possible to develop emotional connection to anyone with a goldfish memory? I'd be surprised if context size there is more than 50K tokens.

huh? What did I miss?

Then io_uring AIO mode underperformance is even more curious.