Notes on latency, routing and running small services close to users.
Averages hide the requests people actually remember. A few practical habits for collecting p99 numbers you can trust: fix the client, warm the path, and always log the slowest sample alongside the percentile.
Hit ratio counted by requests and hit ratio counted by bytes tell different stories. When the long tail of large objects misses, origin egress dominates the bill even at a "95%" hit rate.
Switching congestion control on a handful of 1–4 vCPU instances across two regions. Throughput on lossy trans-Pacific paths improved noticeably; on clean paths the difference was within noise.