Zero-Copy Chaos. When everything you knew about tokio...
TokioConfZero-Copy Chaos. When everything you knew about tokio performance stopped working by Wojciech Kargul When our tokio service went from handling 10M req/s to suddenly facing 45M req/s, everything broke. We were dropping 30% of requests, P99 latency hit 450ms and traditional optimization approaches failed completely. This talk shares the 72-hour sprint where we rebuilt our understanding of async Rust performance from scratch. I'll show how we eliminated every allocation in our hot path, achieved true zero-copy request handling, and ultimately reduced P99 latency to 8ms while using 40% less CPU despite the 4.5x traffic increase. You'll learn the profiling techniques that revealed what flamegraphs missed, concrete zero-copy patterns and custom implementations, when to question tokio's runtime defaults and the three places we used unsafe with benchmarks proving why. This isn't theory, it's battle-tested patterns from production chaos. I'll share the exact techniques, the mistakes we made and the counter-intuitive decisions that actually worked. Perfect for developers who've shipped async Rust and want to know what's possible when you push past conventional limits. Video Production by Confreaks Follow Confreaks 👇 https://confreaks.com https://confreaks.bsky.social https://x.com/confreaks