Cron cannot express an every-N-hours interval unless N divides 24

Fact

A cron step value is a filter over a fixed calendar field, not an interval timer. 0 */5 * * * does not mean "every 5 hours" — it selects the hours matching h % 5 == 0, i.e. 00, 05, 10, 15, 20, and then wraps to 00 the next day. Four gaps of 5 hours, then one of 4.

The same holds for every field whose range is not a multiple of the step: */7 * * * * gives 7-minute gaps until :56, then a 4-minute one. Only steps that divide the field's period evenly (2, 3, 4, 6, 8, 12 for hours; 2, 3, 4, 5, 6, 10, 12, 15, 20, 30 for minutes) produce a uniform interval.

Why it matters

The short interval at the wrap point is silent — the job fires, nothing errors, and the schedule looks correct in every listing. It surfaces only as a job that occasionally runs too soon: a duplicate notification, a polling window that overlaps the previous one, a token refreshed before the old one expired, or a rate-limit window opened early and wasted.

It is also invisible in testing, because the anomaly happens once per day at a specific wall-clock boundary rather than on every run.

How to apply

Check that the step divides the field's period. If it does not, cron is the wrong tool for that interval — do not try to patch it with a longer expression.
Enumerate the fire times explicitly when the drift is tolerable and you want a fixed daily rhythm: 0 1,6,11,16,21 * * * is honest about where the short gap falls, unlike */5.
For a true interval, use a self-rescheduling one-shot: the job runs, then computes now + N and re-arms itself for that single time. Each run re-anchors from when it actually fired, so the interval stays exact and the schedule self-corrects after a missed or delayed run. This is the only form that survives the scheduler being offline at a due time.
Prefer a scheduler with native interval support when one is available; the constraint is cron's calendar-filter model, not scheduling in general. On Linux that is systemd's OnUnitActiveSec; on macOS it is a launchd agent's StartInterval (seconds between runs, re-anchored from the last fire); most job queues expose an every. All three express the interval cron cannot.