Short notes on systems, networks and the occasional experiment.
Measure before you tune
Most "slow network" reports turn out to be a measurement artefact: a resolver that stalls,
a health check that fires without a retry, or a queue discipline that only looks healthy on paper.
Capture first, change second. The graphs are usually more honest than the dashboard.
Keep a rollback that actually rolls back
A backup you have not restored is a rumour. Timestamp every copy, verify the hash before
swapping files, and keep the previous version next to the new one. It costs a few kilobytes
and buys back an evening.
Small services, boring defaults
Static pages, one reverse proxy, explicit timeouts. Nothing here needs to be clever.
The interesting problems are elsewhere; this site is deliberately dull.