engineering2 min de lectura
Friday Five #6: Late is not missed
A daily agent was 34 minutes late and fine. The 20 runs behind the alert threshold, and why a green run is not a delivery.
Rob Vasquez·
Five notes from running a daily agent. Short on purpose.
- Measured: My daily brief runs as a scheduled cloud task. I logged its start time over 20 runs. The median was about 3 minutes late. Five runs were more than 20 minutes late. The worst was 34.5 minutes, and the most recent late one was 33.
- Set: I first wanted an alert at 25 minutes. Against those 20 runs it would have fired on five days when nothing was wrong. So a slot counts as missed 45 minutes after its scheduled time, not before. The number comes from the log, not from a guess.
- Checked: The watchdog never reads the run status. A cloud run marked Succeeded can still have produced nothing, so it looks for the brief itself, one document per day. No document 45 minutes after the slot means a miss, and it writes one failure record. One alert per miss.
- Limited: It's not independent yet. The watchdog runs on the same subscription as the brief, so a shared outage takes out both. The fallback host exists on paper and hasn't delivered once. I'd rather say that than call it redundant.
- Asking: What does "done" mean for your scheduled agents? The exit code, or the thing it was supposed to leave behind?