← Blog

Engineering guides · 12 min read

On-Call Rotation Design That Stops Burning Out Your Team

· Acked

On this page9 sections

Most on-call burnout does not come from one terrible night. It comes from a rotation that quietly puts the same engineers on the pager every few weeks, with a page queue nobody reviews and a handoff that happens at midnight. If you build an on-call rotation for an SRE or DevOps team, the schedule you publish is really a staffing decision. It decides how many nights each person loses to alerts, and how often they wake up with no backup. This guide covers how to size the rotation, cap pages per shift, run clean handoffs across time zones, and set up escalation that spreads the load instead of concentrating it.

Quick Takeaways

Key Insight

Explanation

Size the rotation by headcount

Count trained engineers with production access, then check that no one carries the primary pager more than about one week in four.

Set a page budget per shift

Measure pages per 24-hour shift over 30 days. If the typical shift is unmanageable, fix noisy alerts before changing the calendar.

Hand off when both sides are awake

Avoid midnight handoffs, and write the timezone and handoff time into the schedule.

Use overrides for swaps, not base edits

Temporary overrides cover a single day without rewriting the rotation that everyone depends on.

Make level 2 a real backup

Name a secondary who has the same access and runbooks as the primary, and rotate them through the role.

Plan recovery after heavy nights

Set the rule for a late start or time off in advance, so the engineer never has to ask for it.

Start With Headcount, Not the Calendar

Most teams draw the calendar first and then discover they have too few people to fill it. Start with the opposite question: how many engineers can carry a primary pager without being on call more than about one week in four? That number is a guideline rather than a law, but it gives you a ceiling you can test. If your count is three people, a weekly rotation means each person is primary every third week, and that is usually where burnout becomes structural.

Engineers reviewing an on-call rotation schedule board with color-coded shifts in a bright office

The Google SRE workbook chapter on on-call is candid about this problem. It records feedback from smaller teams who say the large-company model does not fit them, because they have fewer people in the rotation and no sites spread across time zones. Their objection is valid. A small team cannot copy a follow-the-sun model, so it has to be honest about what it can sustain.

Count the people you can actually rotate

Count only engineers who are trained, have production access, and are not already covering a major project. A name on a spreadsheet who has never run a restore does not reduce the load on anyone. In practice, a rotation of four people with real skills outperforms a rotation of eight where half of them have never been paged.

Then work backward from the calendar. Divide the weeks in a quarter by the number of eligible engineers and check how many primary shifts each person would carry. If the answer makes you uncomfortable, the rotation is too thin, and the fix is more people, not a longer shift.

Why a small team needs a different model

A four-person team usually does better with a primary and a secondary layer that share the first night of an incident than with a single pager that rotates weekly. The secondary is not a spectator. It is the person who takes over when the primary is in a meeting, on a flight, or exhausted after a long incident. Naming that backup in the schedule, rather than hoping someone will jump in, is the difference between a rotation that works and one that quietly fails.

Cap Pages Per Shift Before You Draw the Schedule

A rotation cannot be fair if the pager itself is unfair. The same Google chapter cites feedback from a team whose on-call engineer was paged about a hundred times in a typical 24-hour shift. No rotation design survives that. Before you change who is on call, measure what the on-call engineer actually receives.

Pages come from monitoring tools such as Datadog, Grafana, Prometheus Alertmanager, AWS CloudWatch, Azure Monitor, and Google Cloud Monitoring. Each of them can fire on thresholds that nobody has reviewed in a year. A good routing layer makes noise visible, but it cannot decide which alerts deserve to wake someone up.

Measure the page load on a normal shift

Pull the last 30 days of alerts and group them by shift. Count how many pages reached a human, how many were acknowledged within the first few minutes, and how many closed with no action taken. The last two numbers tell you more about the health of the rotation than the rotation chart does.

Look at the outliers too. A single shift with 40 pages usually points to one flapping service, a bad deploy, or a threshold set for a load level the system no longer sees. Write a short note on the shift and its cause, so the postmortem and the schedule review work from the same evidence.

Fix the noisiest alerts before adding people

Start with the alerts that fired most often and produced no action. Raise the threshold, add a duration condition so a brief spike does not page anyone, or move the alert to a ticket queue if nobody needs to act at 3 a.m. The goal is that every page that reaches a person is one they would want to be woken for.

Pro tip: Keep a short running list of alerts that paged someone and needed no action. Review it every sprint. A small number of rules often produce most of the night-time pages, so the list tends to shrink quickly once you start.

Handoffs, Overrides, and Time Zones

Handoffs are where rotations drop things. The outgoing engineer assumes the incoming one saw the open incident, and the incoming engineer assumes nothing is open because the channel has gone quiet. A handoff should be a defined event with a time, a written status, and a named person, not a vague sense that the shift has changed.

Pick a handoff time that works for both sides

Choose a handoff time when both people are awake and working. A handoff at midnight means someone is always caught in the middle of a half-finished incident. A handoff at the start of the working day in the incoming engineer's timezone usually works better, because the new person can read the overnight history with fresh attention.

Time zones add a second trap. If your team spans Berlin and Austin, a handoff written as "Monday 9 a.m." is ambiguous unless the timezone is written next to it. Put the zone into the schedule explicitly. Acked uses an IANA timezone for each schedule, so the handoff lands at the same local time for each shift, and the schedule view shows who is on rather than asking anyone to convert it in their head.

Overrides for swaps without rewriting the rotation

Swaps happen. Someone has a wedding, a flight, or a child with a fever, and the fix should take minutes, not an hour of editing layers. Use a temporary override for the specific days involved, then let the base rotation resume. Permanent edits to the base rotation are how schedules drift until nobody trusts them.

Acked also emails admins when a schedule stops covering the next two weeks, which catches the common mistake where an override leaves a gap that only appears when a page goes unanswered. Read the Acked schedules documentation before you build your first rotation, especially the parts on layers and handoff intervals.

Escalation That Shares the Load

Escalation protects people as much as it protects incidents. A well-designed chain pages the primary first, waits a defined time, then tries the secondary, and keeps going until someone acknowledges. A badly designed one either pages everyone at once, which creates a crowd where nobody owns the alert, or depends on a single person who may be asleep.

Each level needs three decisions: who is eligible, how long to wait, and how many passes the chain makes before it stops. In Acked, layers at level 1 are eligible first, higher levels page while the alert stays unacknowledged, and the chain can repeat a limited number of times. When the final pass ends without acknowledgement, the alert is marked unreachable rather than closed quietly, which is what you want, because an unanswered page is itself a problem to solve.

Phone on a nightstand showing an overnight alert notification in a dim bedroom

Set waits that match your response expectations

Set the first wait short enough that a real incident reaches a human within minutes, and long enough that an engineer already handling one page is not interrupted by a second. Test the waits with a drill, then adjust once you see how long acknowledgements take in practice. A wait that looks sensible on a whiteboard often feels too short at 2 a.m.

Make level 2 a real backup

A backup layer that is never tested is a fiction. Make the secondary a named person with the same access as the primary, and rotate them through the role on a schedule. If the secondary only gets paged after the primary has already failed, they will not know the runbooks or the service owners when they finally need them.

Channels matter too. Acked sends push notifications on iOS and Android, email, and outbound webhooks on every plan, with voice paging available on Team and Business plans, billed separately by usage. It does not provide SMS paging, so if your team relies on text messages, plan for that gap before an incident forces the question.

Pro tip: Send a test alert to every person in the chain at least once a quarter, at an inconvenient hour, with the phone on silent. That is the only check that shows whether the notification reaches a sleeping engineer.

Recovery, Compensation, and Time Off After a Bad Night

A rotation that looks fair on paper still burns people out if the morning after a bad night is treated as normal. The engineer who handled four incidents between 2 and 6 a.m. should not be expected to join the 9 a.m. planning meeting at full capacity. Decide the rule in advance, write it down, and apply it without asking the engineer to justify it.

A rotation is a staffing decision, not a calendar exercise. If the math does not work, no amount of tooling will fix it.

Recovery is also a matter of knowledge. A DevOps.com article on on-call practice argues that runbooks reduce the cognitive load on on-call engineers and lower the fear that makes burnout worse. In practice, a runbook that says exactly what to check first is the difference between a 20-minute page and a two-hour struggle with a dashboard nobody has explained.

Protect the morning after

Give the engineer who took a heavy night the option to start late or take a half day. Make it the default rather than a favor. Then review the page log together at the next team meeting, so the recovery time turns into useful work rather than a silent penalty.

Compensation that people believe

Some teams pay a flat stipend for on-call time, others give time off in lieu, and some do both. The mechanism matters less than consistency. If compensation depends on a manager's mood after each incident, engineers will start negotiating swaps to avoid the rotation entirely, and the schedule will fracture.

Junior engineers need a different path. One Uptime Labs article on reducing on-call burnout makes the point that junior engineers need a safe place to practice before they carry the pager alone. Pair them with an experienced secondary for their first few rotations, and let them shadow the response before they become the primary.

Comparing Three Rotation Models

No single model fits every team, but three patterns cover most of what teams end up running. Each makes a different trade between sleep, coverage, and coordination cost. Pick the one that matches your headcount and time zone spread, then write the constraints into the schedule explicitly.

Flat-lay of a handoff notebook, timeline graph on a tablet, and clocks across time zones

Rotation model

Works best when

Main risk

Weekly primary with a named secondary

A small team in one or two adjacent time zones

The primary is on call too often if headcount is thin

Follow-the-sun across regions

A large team with engineers in three or more time zones

Handoff quality decides whether incident context survives

Split by alert type

Clear skill separation between infrastructure and application owners

Incidents that cross both areas get stuck between groups

Choosing between the three models

The weekly primary model is the simplest and the easiest to explain to a new hire, but it breaks first when headcount drops, because the same people keep appearing on the calendar. Follow-the-sun spreads night shifts across regions and can remove most night pages for a team big enough to staff it, but it only works if every regional handoff is written down. Splitting by alert type can reduce the time spent figuring out who owns a page, yet it needs an explicit rule for incidents that span both groups.

For most teams under ten engineers, the practical answer is a weekly primary with a real secondary, plus a clear page budget. Move to a split model only after your alert volume and skill separation justify the extra coordination.

Frequently Asked Questions

How many people do you need in an on-call rotation?

There is no universal number, but most teams need enough engineers that no one carries the primary pager more than about one week in four. Count only people with production access and the skills to respond. If the math leaves anyone on call too often, add people or reduce the number of alerts before you lengthen the shift.

How long should an on-call shift last?

A week is common for primary coverage because it gives each engineer a predictable block and limits the number of handoffs. Shorter shifts reduce how long one person is exposed, but they multiply handoffs. Choose the length that matches your headcount, and keep the handoff time consistent.

What is a good escalation delay?

Set the first delay short enough that a real incident reaches someone quickly, and long enough that the primary is not paged twice for the same issue. Test the setting with a drill, then adjust using the acknowledgement times you actually see in your response history.

Should on-call engineers be paid extra?

Many teams compensate on-call time with a stipend, time off, or both. The exact method matters less than applying it the same way every time, so engineers can predict what they will receive and do not feel that each incident is negotiated.

How do you onboard a new engineer to on-call?

Pair them with an experienced secondary for their first few rotations, and let them shadow incident response before they become the primary. Give them the runbooks for your top alert sources, and make sure they know who to escalate to when a service owner is unavailable.

Which part of your on-call rotation creates the most friction for your team right now, and what have you changed to fix it?

Make on-call alerting simpler with Acked

Acked connects your existing monitoring tools to on-call schedules, notifications, and escalation policies. Get critical alerts to the right people without replacing your monitoring stack.

Get started with Acked — Free to start, with simple flat-rate pricing and no per-seat fees.

References