We didn't fix on-call by paging people less. We fixed it by finally admitting our rotation math had been wrong for two years.

We didn't fix on-call by paging people less. We fixed it by finally admitting our rotation math had been wrong for two years.

Everyone Knew Something Was Off

Exit interviews kept mentioning on-call fatigue. Retention on the SRE team was worse than the rest of engineering. We'd tried the obvious fixes — reducing noisy alerts, adding a secondary rotation — and burnout complaints didn't move. The thing we hadn't questioned was the rotation length itself: one week on, three weeks off, which sounds generous until you look at when the pages actually land.

What the Data Showed

We pulled every page timestamp for a year and looked at distribution across the week:

from incidents
| where paged == true
| summarize count() by day_of_week(created_at), hour_of_day(created_at)
| sort by count() desc
| limit 15

Pages weren't remotely uniform. 61% landed between 6pm and 2am, and weekends carried a disproportionate share relative to their two-sevenths of the calendar. A one-week rotation meant that whoever was unlucky enough to have their week overlap a weekend absorbed almost double the pages of someone whose week fell entirely on weekdays — the same nominal "one week of on-call" was a wildly different experience depending on which week you drew.

What We Changed

We moved to a rotation that splits weekday and weekend coverage into separate, shorter shifts — five weekdays as one block, the weekend as a separate block with its own person — so nobody draws a doubled-up week by chance. We also started publishing the page-count distribution per rotation publicly, so "my week was worse than yours" stopped being a vague feeling and became a number the team could act on and rebalance compensation or comp-time against. Six months in, the variance in pages-per-person dropped by more than half, and for the first time exit interviews stopped mentioning on-call as a top-three reason for leaving.

Fair on-call isn't about the schedule looking symmetric on a calendar. It's about the pages landing symmetrically on people.
  • Measure page distribution by day-of-week and hour, not just total volume, before redesigning a rotation.
  • A schedule that's symmetric in weeks can be wildly asymmetric in actual pager burden.
  • Splitting weekday and weekend into separate rotations removes the luck-of-the-draw effect.
  • Publish the per-person page count; it turns a vague grievance into an actionable, trackable number.