Docker networking flaws, not CPU, caused e-commerce checkout timeouts at 900 req/s

A containerized PHP e-commerce platform serving 40,000 daily users began experiencing severe checkout latency spikes — from 280ms to over 2.1 seconds — during flash sales, despite stable database query times. Engineers identified three compounding network-layer issues: overhead from Docker's default bridge userland proxying, kernel conntrack table exhaustion due to a low default limit of 65,536 entries, and uncached DNS lookups queuing under high connection churn. The fixes required no application code changes — instead, the team tuned kernel conntrack parameters, migrated internal container traffic to a dedicated bridge network with jumbo frames, and added a dnsmasq sidecar to cache DNS responses locally. Switching to Kubernetes was ruled out, as it carries equivalent networking pitfalls without addressing the root causes. The case highlights how default container networking configurations can silently degrade under real production load in ways that staging environments rarely expose.