Our p99 latency had spikes at exactly 1 and 3 seconds. The cause was dropped SYNs and a timeout designed for the internet. Here's how TCP's initial timeout works, why the kernel gives you no per-destination knob for it, and how a 60-line sock_ops program fixed it.

Every service behind my WireGuard tunnel started failing its health checks at once. None of them were down. The tunnel was quietly dropping every packet bigger than 1460 bytes.

Streaming audio from a Windows PC to my Linux laptop over RTP gave a "tchhh" static whenever someone spoke. No packets were lost. The problem was how ffmpeg grouped them.