BACK TO BLOG
On-Call Is A Load-Bearing Human — what sleep research says about the cost of simply being reachable, and why page count is the wrong measure of a rotation's health.

On-Call · Reliability · The Job

On-Call Is A Load-Bearing Human. The Quiet Weeks Still Cost You.

Will Farrar · · 5 min read

The handoff note at the end of the week says: two pages, both overnight, both cleared in under ten minutes. Quiet rotation. That is what goes in the spreadsheet, and by every number a paging tool gives you for free, it is true.

Then ask the person who just came off it how their week went.

I have carried a pager in one form or another for twenty-five years, and I thought I understood the arithmetic: pages are the cost, so fewer pages means a healthier rotation. There is a body of sleep research saying most of the bill does not come from the pages, and almost none of it has reached how we design rotations.

The comfortable theory

On-call hurts because it wakes you up, so count the wake-ups. Drive the page count down and you have driven the harm down with it. Clean, measurable, and the only version of events your tooling can actually see. That last part should make you suspicious.

What the lab found when nobody got called

A group of Australian sleep researchers led by Madeline Sprajcer ran a series of on-call experiments in a time-isolated sleep laboratory, published between 2018 and 2020. Four nights per participant: adaptation, control, and two on-call nights. Polysomnography overnight, then a ten-minute psychomotor vigilance task — the standard lapse-and-reaction-time measure — four times the next day.

The detail that matters is in the pooled analysis in Clocks & Sleep, covering 70 participants across three of the studies: nobody was ever actually woken before the end of their eight-hour sleep opportunity. Not one call, in any of it.

The costs showed up anyway.

In the missing-alarm study in Biological Psychology — 24 participants — pre-bed state anxiety rose from 29.4 on the control night to 34.0 and 33.7 on the two on-call nights, p < .001 for both. Sleep was shorter and less efficient on the night they had been told the alarm might be too quiet to wake them: 439.9 minutes against 447.8, efficiency 91.6% against 93.3%, p = 0.013.

Be honest about that effect size: eight minutes of sleep, under two points of efficiency. Statistically detectable, nobody's catastrophe. The anxiety delta is the bigger signal, and it landed on both on-call nights, in people who then slept through undisturbed.

The finding that changed my mind is in the companion study in Chronobiology International. Same laboratory, same protocol, one variable changed: participants were told either that they would definitely be called during the night, or that they might be.

The maybe night was the worse one. Significantly less NREM and stage 2 sleep, and significantly poorer vigilance performance the following day.

The night they were told a call was certain was the more anxious one by self-report, and the better one on the polysomnograph. Uncertainty cost them sleep they did not report losing.

Nor can you staff your way around it. Pooling all 70 participants, trait anxiety barely moderated anything, and the authors conclude that "on-call outcomes are not negatively affected by higher levels of trait anxiety."

What the field data can and cannot tell you

Less than you would hope. A 2016 systematic review in Sleep Medicine Reviews screened 3,173 records on on-call-from-home work and found eight studies worth including. Eight. Roughly an hour less sleep in one, about 36 minutes in another, 28% of workers taking more than half an hour to fall asleep when on call against 14% when not. Seven of the eight were rated weak on the review's own quality tool, and every sleep outcome was self-reported.

Then the line that should be pinned up somewhere: none of the eight studies differentiated between nights on call with and without calls.

The field literature had never separated the variable we build entire rotations around. So the mechanism is demonstrated in the laboratory, in small samples of young men, over single nights, and in the wild it is barely measured at all. Do not let anyone tell you otherwise in either direction, me included.

Why we ended up counting pages

Because a page is an event, and an event emits a record. Availability emits nothing. There is no webhook for the night you went to bed at eleven with the phone face-up on the nightstand and the laptop charged and nothing happened. On-call pay has the same shape: it prices the interruption, because the interruption is the part you can put on a timesheet.

To be fair, the published guidance is not naive. The SRE workbook targets a maximum of two incidents per on-call shift, caps shifts at 12 hours, and states flatly that "24 hours of on-call duty without reprieve isn't a sustainable setup." The SRE book caps on-call at 25% of an SRE's time and observes that stress hormones "can impair cognitive functions and cause suboptimal decision making."

The useful part is their own case study: a team with a budget of two incidents per shift running at five, a third of all shifts over budget. That is what a threshold is for — not to be met, but to make the overage visible.

But even the best-reasoned target in the industry is a count of events. The lab work suggests the exposure is the denominator — nights of being reachable — and the pages are only the variance.

What I would do

Track shift-nights per engineer per quarter. That is the exposure number, and a quiet quarter does not reduce it. Report page count separately, as a 21-day trailing average the way the workbook does, and never let it be used to explain the exposure away.

Go after uncertainty, not only volume. A bounded, published window where a page is possible should be cheaper than open-ended maybe. If the batch jobs can only fail between 1 and 4 a.m., say so out loud, so the person sleeping knows the shape of the night. One lab study is thin ground and I would like to see this replicated in a live rotation, but it costs nothing to try.

Make the alarm unmissable, and tell people that it is. Two contact paths — the missing-alarm paper's own recommendation. Fear of sleeping through it is itself a sleep cost, so redundant notification is a sleep intervention that also happens to improve response time.

Do the roster arithmetic, then try to delete the night shift. Eight people for single-site 24/7; five per site, or six with a buffer, if you can split the clock. That is the real case for follow-the-sun — not coverage, which you already had, but handing the overnight hours to somebody whose daytime it is. Twelve hours is the ceiling either way. And if you have four people you do not have a rotation; you have four people taking turns being tired, and the schedule will inform you of that eventually, in the form of a resignation.

Spend the error budget argument. It is the only mechanism I have found that converts "we are getting paged too much" into a decision with a number attached, which is the only form in which that sentence has ever won.

Stop congratulating people on quiet weeks. "Nothing happened" is not evidence that the rotation is healthy. It is evidence that nothing happened.

The pager is the only part of on-call that leaves a record. The rest gets paid out of somebody's sleep, in small amounts, on every night they were reachable, and nobody ever sends you the invoice.

0 Share
Will Farrar

Will Farrar

Hands-on CTO with 25+ years in regulated healthcare IT. Author of the Prompt Mode workbook and creator of the companion app. Writes about the parts of the job — good and hard — that don't fit on a resume. More about Will →

BACK TO ALL ARTICLES