When stability alerts should wait

This note is from a ride-hailing reliability lead who sat Org Watch with us in 2025. Names omitted; the paging rules are theirs, lightly edited.

We used to page on any 0.3-point crash-free dip after 22:00. Night shift learned to snooze. The one evening a payment library actually wedged the driver app, the on-call assumed it was another festival mix shift and went back to sleep. That is alert fatigue with a body count measured in stranded drivers, not in Jira tickets.

Wait if all three are true

The stack is unsymbolicated or the mapping file is younger than the binary.
Affected users are under a threshold you set per journey — ours is 80 concurrent drivers in a city, not a global percentage.
The dip tracks a known traffic event (payday, concert let-out, heavy rain) that historically moves the chart without a new cluster.

Waiting is not ignoring. The night board still shows the cluster. Chat still gets a quiet note. Nobody’s phone rings.

Do not wait if

A named journey is dead with a readable stack, even if the global crash-free line looks polite. Checkout, login, and “start trip” are journeys. Settings fonts are not. If you cannot name the journey, you are not ready to page or to wait; you are ready to open the export.

Devqueuegrid did not invent these thresholds. They only made us write them down so product could see why we sometimes refuse to escalate a pretty chart. If your company still equates care with volume of pages, no syllabus will help until that politics moves.