Notes

Off cron, onto a systemd timer

August 2026

The checks used to run from cron. They now run from a systemd timer, for one reason: when a cron job fails, nothing remembers. A timer gives me a unit that shows up in systemctl is-failed, which means a failure is a state I can query rather than a line I have to go find in a log.

This is not hypothetical. Something else on this box had been failing twice a day for a month behind a perfectly healthy-looking timer, and I only found it by reading a log for an unrelated reason. Silent recurring failure is the worst kind, because every individual symptom is "nothing happened".

curl -w is enough

May 2026

I spent an evening looking at proper monitoring agents before noticing that everything I actually wanted was already a format string:

-w '%{http_code} %{time_total}'

Status code and wall-clock time, from a binary that is already installed and already handles redirects, TLS and timeouts correctly. The rest of an agent is scheduling, storage and alerting — and I wanted none of the three. If I ever need connect time and TLS handshake time separately, %{time_connect} and %{time_appconnect} are right there.

Ten seconds, and why not thirty

April 2026

The timeout is ten seconds. Thirty would catch a few more slow-but-alive responses, but a target that needs half a minute is not meaningfully up for anything I care about, and a long timeout makes a run of failures take long enough to overlap the next run. Ten keeps a full pass well under the interval even if everything on the list is broken at once.

The counters are cumulative, not windowed, so a bad afternoon stays visible in the numbers for a long time afterwards. That is deliberate — a rolling window quietly forgives whatever fell outside it.