Why does one NGINX worker take all the load?

This post delves into the intricacies of load balancing in NGINX, specifically addressing the problem of uneven distribution of incoming connections across worker processes. It explains the different models for handling accept() calls, the Linux kernel's differing load balancing behaviors for blocking accept() vs. epoll-and-accept, and introduces SO_REUSEPORT as a solution for achieving more even load distribution. The post also highlights the critical trade-off between even load balancing and latency, demonstrating how SO_REUSEPORT can sometimes lead to degraded latency distributions under high load.