Why Cloudflare Ditched NGINX for Rust (It Wasn’t Just Speed)
Oct 14, 2026
Cloudflare didn’t rewrite its proxy in Rust just because Rust is fast. The real story is architectural: NGINX’s worker-process model, CPU imbalance, worker-local connection pools, and upgrade challenges became hard limits at Cloudflare’s scale. This video breaks down why Cloudflare built Pingora, how its multithreaded Tokio-based design works, and why connection reuse, work stealing, SO_REUSEPORT, and file-descriptor handoff matter.
We cover:
• Why NGINX worker processes hit a wall under uneven load
• How Pingora uses async execution, multiple threads, and work stealing
• Why shared connection pools reduce latency and origin connections
• SO_REUSEPORT vs. zero-downtime listener handoff
• How Pingora handles graceful upgrades with UNIX-domain sockets and FD transfer
• What Cloudflare’s 1T+ requests/day, 35M+/sec, and ~25% internet traffic scale really mean
Chapters:
00:00 The Scale Problem
00:49 Why NGINX Hits a Wall
01:53 Pingora’s Thread Model
02:58 The Connection Reuse Payoff
03:56 SO_REUSEPORT
04:23 Zero-Downtime Upgrade
05:34 Operating at Global Scale
06:12 Architecture Over Rewrites
Subscribe for weekly system design breakdowns: @SysMinu
#Cloudflare #Pingora #NGINX #SystemDesign #Rust #BackendEngineering #Networking
Show More Show Less 