Where did my latency budget go?
Splitting an end-to-end response time into DNS, connect, TLS and transfer, and why the first two usually dominate on cold connections.
Latest notes
Percentiles lie when samples are rare
A p99 computed from a handful of requests is noise dressed up as a metric. A note to myself about minimum sample counts.
Three processes and a restart policy
What actually keeps a small always-on service alive: supervision, log rotation, and a health endpoint somebody calls.