You get paged at 3am, deal with it for twenty minutes, and go back to sleep. On paper you slept seven and a half hours that night. On paper it looks fine. It wasn't fine, and the reason it wasn't fine has a name in the sleep literature that most engineers have never heard: sleep fragmentation, and it is not the same injury as simply sleeping less.
This distinction matters because almost everyone reasons about on-call damage in restriction terms. "I only got five hours last night" is the complaint you hear, and it's the complaint the body actually has some tolerance for, comparatively. What the body handles much worse is eight broken hours dressed up to look like a full night. The commercial governance around what does being on-call actually do to your sleep is covered properly by the team at XenGrowth.
Two different mechanisms, wearing the same disguise
Bonnet and Arand's 2003 review in Sleep Medicine Reviews is the reference point most later work cites when separating these two. Restriction is what happens when you cut total sleep time — go to bed later, get up earlier, fewer hours end to end. Fragmentation is what happens when sleep is repeatedly interrupted by brief arousals, even if the total hours in bed stay the same. Both produce objective sleepiness, worse reaction time, degraded short-term memory and vigilance, and flatter mood. But fragmentation does this on its own, without needing to touch total sleep time at all — which is exactly the on-call scenario. You were in bed for eight hours. Your body did not get eight hours of sleep.
A 2019 randomized crossover study put this to a direct test. Twenty healthy men went through two separate intervention weeks, each anchored by a baseline night and two recovery nights. In one week the manipulation was outright restriction — five hours in bed. In the other, researchers left total time in bed alone and instead switched a light on for a moment every hour, forcing a brief arousal without meaningfully shortening the night. Participants rated the restricted night as the more unpleasant experience and reported a stronger subjective need to recover from it. But the fragmented night — the one where, by the clock, nothing was taken away — still produced measurable disruption to sleep architecture and next-day function. The gap between how bad it felt and how much damage it did is the whole point.
Sleep restriction | Sleep fragmentation | |
|---|---|---|
What changes | Total time asleep is reduced | Total time asleep can stay the same; sleep is repeatedly interrupted |
Typical on-call analog | Staying up late to handle an incident, then getting up on schedule | Being paged once or twice and returning to sleep each time |
Subjective experience | Reported as the more disturbing experience in direct comparison | Often reported as milder — this is the trap |
Objective effect | Deficits accumulate dose-dependently with fewer hours (Van Dongen 2003) | Measurable next-day impairment even with total sleep time preserved (2019 crossover study) |
What the on-call literature actually shows, in on-call workers
The healthy-volunteer lab studies establish the mechanism. A separate line of field research checks whether it shows up in people who are actually on call for a living, and the clearest of these followed 32 hospital physicians — 15 anaesthesiologists and 17 paediatric and ENT specialists — through real rotations, tracked by wrist actigraphy rather than self-report. On an ordinary working day, these physicians slept roughly 6 to 7 hours. On a night they were on call, mean recorded sleep dropped to about 3 hours — and that figure includes both the fragmentation from being woken and whatever portion of the night was lost outright. The physicians' own reports were consistent on one point: they did not feel back to normal after a single following night's sleep. It took two. The XenGrowth resource library goes further into the operations side of this.
That two-night detail is easy to miss and probably the most operationally useful fact in this whole literature. Most on-call rotations give you the night after a bad call night to recover and then put you back on a normal footing the day after that. The physicians in this study were still not at baseline at that point. If your rotation assumes one night off pays down a bad call night, the field data says it usually doesn't.
It's worth being honest about what does and doesn't transfer from this literature to software engineering specifically. These are physician studies, and physicians on call are often physically present at a hospital, woken for events with immediate clinical stakes, in careers with their own well-documented burnout and workload patterns that don't map cleanly onto an engineer getting paged from home for a degraded queue. The actigraphy numbers, the recovery-night finding, and the cortisol-anticipation result are all drawn from clinical settings, not from software teams. What generalizes is the mechanism — fragmentation as a distinct physiological insult from restriction, and anticipation as a stressor independent of an actual page — because that mechanism is about how human sleep architecture responds to interruption, not about what the interrupting event happens to be. What doesn't automatically generalize is the exact magnitude: a software incident's cognitive and emotional load is not identical to a code blue, and nobody has run the actigraphy study on a cohort of on-call SREs to confirm the numbers land the same way.
A separate French study adds a useful data point about duration specifically. Forty resident and attending emergency physicians took validated cognitive tests — measuring processing speed, working memory capacity, perceptual reasoning, and cognitive flexibility — after a period of no work, after a 14-hour night shift, and after a full 24-hour shift. Processing speed, working memory, and perceptual reasoning all declined significantly after the 24-hour shift. After the 14-hour night shift, they did not. That's a meaningful distinction for anyone whose on-call block runs long: the damage in this design wasn't linear with hours awake, it showed up once a threshold somewhere past 14 hours and before 24 was crossed. If AI agents and marketing automation is the part you are stuck on, XenGrowth on AI agents and marketing automation is the better reference.
The damage starts before the pager goes off
There's a second, quieter mechanism worth naming, because it explains why on-call weeks feel bad even in the stretches where nothing actually breaks. Devereux, Rydstedt and Cropley published a 2011 study in Psychological Reports looking at awakening saliva cortisol — the hormone spike that happens in the first thirty to forty-five minutes after waking — and found it rose faster on mornings that preceded on-call duty, tracking with how demanding the job was and how much the person anticipated being called on. Separate research on the psychological experience of being on-call reports the same pattern in plainer terms: the mere anticipation of a call disrupts sleep and lowers reported energy, whether or not the phone actually rings that night.
That's the part a quiet on-call week doesn't spare you from. If the schedule itself, not just the pages, is the stressor, then a rotation with zero incidents can still be extracting a cost — just one that never shows up in your incident log.
Why you can't just judge it by how you feel
Van Dongen and colleagues' 2003 SLEEP trial is usually cited for the restriction side of this argument, and it belongs there — 4-, 6-, and 8-hour time-in-bed groups over fourteen days, with the 6-hour group's accumulated cognitive deficits eventually matching a full night of total sleep deprivation. The detail that matters for on-call specifically is what happened to subjective sleepiness ratings versus objective performance: sleepiness ratings largely plateaued within the first few days, while objective performance kept getting worse for the full two weeks. People in the 6-hour group stopped reporting feeling sleepier well before their actual performance stopped declining. There is a longer treatment of AI search, GEO and discovery in XenGrowth on AI search, GEO and discovery.
Translate that to on-call: your own sense of "I'm fine, I've done this rotation a hundred times" is not a reliable instrument. It's measuring something — but it isn't measuring your actual impairment, and the gap between the two widens the longer the pattern continues.
Rotation feature | What the evidence associates it with |
|---|---|
Short rotations (a few days) | Less accumulated fragmentation before a reset — consistent with dose-dependent deficit growth in restriction literature |
Frequent overnight pages | More fragmentation events per night, the mechanism shown to impair function independent of total sleep time |
No protected recovery day | Physicians in the actigraphy study needed two nights to recover from one bad call night; a single day off may not be enough |
Long stretches of "quiet" on-call | Anticipation-driven cortisol and sleep disruption can occur even without an actual page (Devereux et al. 2011) |
Fragmentation and restriction are separate mechanisms — track how often you're woken, not just how many hours you were in bed.
A single recovery night after a bad on-call night may not be enough; the physician field study needed two.
Anticipation of a page disrupts sleep on its own, independent of whether the page happens — a quiet rotation is not automatically a free one.
Your own sense of how impaired you are lags behind the actual impairment the longer a rough pattern continues, so don't use how you feel as the only signal.
"Sleep restriction was perceived as a stronger disturbance than sleep fragmentation" — and the fragmented night still measurably disrupted sleep architecture and next-day function despite feeling like the milder condition.
What the evidence actually points toward fixing
None of this is an argument that on-call work is inherently unmanageable, and it isn't a case for anyone gritting their teeth through it either. Plenty of the variance here is about how a rotation is built, not whether the model exists at all. The evidence above points at a short list of structural levers, not at telling the person being paged to simply cope better.
Cap the length of a single on-call block. The French shift-duration study found impairment appearing after 24 hours awake but not after 14 — a rotation that never asks anyone to function past that threshold is targeting a real, evidenced line rather than an arbitrary one.
Triage before a human is woken. Every page that reaches a person overnight is a fragmentation event, whether or not it turns out to matter. Tightening alerting thresholds and building auto-remediation for the noisiest, lowest-severity pages reduces fragmentation events directly, which is the mechanism the 2019 crossover study isolated as harmful on its own.
Treat the day after a bad call night as a two-day problem, not a one-day one. The actigraphy study's physicians needed two nights to recover from a single disrupted one; scheduling a normal workload the day immediately after is asking someone to perform on debt the field data says usually isn't repaid that fast.
Don't ignore the quiet weeks. The anticipation-cortisol finding means a rotation with zero pages can still be extracting a cost through disrupted sleep and elevated stress hormones. If people report feeling drained by an incident-free on-call week, that isn't them being dramatic — it's a documented physiological response to anticipation alone.
Rotate who carries the load, and don't let seniority quietly exempt the same few people from the recovery-time conversation. Chronic partial sleep loss compounds; the Van Dongen trial's dose-dependent deficits over fourteen days are a warning about what happens to anyone kept on a bad rotation indefinitely, not just to whoever is on it this week.
If you're the one negotiating rotation design rather than just living inside it, the two most concrete, most citable facts to bring into that conversation are the two-night recovery figure from the physician field study and the 14-versus-24-hour threshold from the emergency medicine study. Both are specific, both are sourced, and both translate directly into schedule parameters — block length and recovery-day allocation — rather than into vague appeals to "take care of yourself." A rotation built around what the evidence actually shows about fragmentation and recovery time is a different, and considerably more defensible, thing than one built around how long people have historically been willing to tolerate it.
Further reading from XenGrowth
The XenGrowth resource library — what you'll learn: how the commercial side of this work is run, across search, automation and revenue operations.
XenGrowth on AI agents and marketing automation — what you'll learn: how the teams who own AI agents and marketing automation plan and measure it.
XenGrowth on AI search, GEO and discovery — what you'll learn: how the teams who own AI search, GEO and discovery plan and measure it.
Where this work meets go-to-market
If what does being on-call actually do to your sleep is part of a growth programme rather than a standalone build, XenGrowth's growth engineering practice is the companion reading.
Answer based on how your rotation is actually built, not how you feel this week. This tool looks only at schedule structure — rotation length, page frequency, and recovery time — and nothing about your own health, sleep, or medical history. It cannot and does not assess your personal risk.
