Every distributed bug eventually turns out to be a clock bug. Two machines
disagree about what time it is, nothing errors, and the consequences surface
three layers away as a corrupted ordering or an expired token. These are notes
on that.
-
systemd-timesyncd can report success while drifting
NTPSynchronized stays true long after the daemon stopped getting answers, so every health check you write on it lies.
2026-08-03 · 6 min
-
The retry loop that sleeps until Thursday
Compute a deadline from the wall clock, let NTP step it backwards once, and your backoff turns into a hang nobody can reproduce.
2026-07-06 · 5 min
-
Your container is UTC and your logs are lying about when
The host is Europe/Berlin, the image has no tzdata, and the timestamps differ by an amount that looks plausible.
2026-05-21 · 4 min
-
A KVM guest's clock after the host suspends
kvm-clock usually saves you. The exception is the boot window before the time daemon is up, and that is exactly when TLS fails.
2026-04-09 · 7 min