Healthcare technology teams spend a great deal of effort on what their systems do and much less on how steadily they do it. Appointment portals, record look-ups, telephone queues, information pages and messaging tools all share one trait: people reach for them when they are busy, worried or both. That makes dependability a practical concern, not a nicety. Site reliability engineering, usually called SRE, is the most developed body of practice for managing it. This piece reviews how the discipline has evolved, where it stands now and what health IT teams can reasonably take from it. The subject here is operations practice alone; nothing in it is clinical or constitutes medical guidance.
How SRE got here: a short timeline
The Ops Atlas, an online glossary maintained by xOps School, gives the history in a compact form on its long-form SRE entry. The story begins in 2003, when a software engineer at a large search company took charge of a small group tasked with running production systems. The one-line summary that grew out of it, still widely repeated, describes the result as an operations group designed by someone trained in software engineering.
- 2014: the first dedicated SREcon conference took place on 30 May in Santa Clara, giving practitioners a shared meeting place outside any single company.
- 2016: the founding book on the subject was published and made available free online, and a related customer-facing engineering programme was announced.
- 2017: an ACM Queue article titled “Calculus of Service Availability” set out the mathematics of how dependencies shape the availability a service can promise.
- 2018: a companion workbook added worked examples, including the paging thresholds many teams now copy.
- 2021: a widely read annual research programme broadened its “availability” measure into the wider idea of “reliability.”
- October 2026: a rewritten second edition of the founding book was announced, this time as a paid title.
Two things stand out. The vocabulary has been stable for a decade, and the community has moved from a single organisation’s practice to a broad public literature, including an open specification for service objectives written in YAML and open-source tools that turn targets into alert rules.
The state of practice: what surveys say
The strongest available figures come from practitioner surveys, which should be read with caution because respondents are self-selected. A 2026 report from a monitoring vendor, drawing on 418 responses, found a median of roughly thirty-four percent of working time going to toil, meaning manual, repetitive work that adds no lasting value. The Atlas notes that earlier internal surveys at the originating company averaged around a third, with individual engineers ranging from none to eighty percent.
Read those numbers carefully. A median near one-third sits under the commonly cited ceiling of one half, which suggests that many teams are coping. It also means that a sizeable minority exceed the ceiling, and a median hides how uneven the load can be inside a single department. For health IT, where staff are often stretched thin, the lesson is to measure the figure locally instead of assuming the average applies.
Trend one: targets are moving closer to the user
Early monitoring emphasised machine readings such as processor load and memory use. The current emphasis is on what the person using the service actually experiences. In an information-heavy setting, that might mean the share of page requests answered correctly within a fixed time, measured by an outside probe rather than from inside the data centre. The Atlas treats the standard server graphs as a starting point and says plainly that they are not the same as user satisfaction.
Trend two: alerting driven by budget burn
The second shift is from threshold alerts to alerts based on how quickly the allowance for failure is being used. The reasoning is straightforward. A threshold alert fires whenever one number crosses a line, which in a busy service produces many messages and little insight. A burn-rate alert fires when the pace of failures threatens the target for the period, and it can be paired with a shorter window to confirm the problem is still happening. The workbook’s published example pairs short and long windows and separates urgent pages from next-day tickets.
Trend three: honesty about targets
The third trend is a gradual retreat from the idea that more reliability is always better. The Atlas states that no target of one hundred percent is sensible for almost anything, and it repeats the observation that every extra step up can cost around a hundredfold what the step before it did. Related guidance recommends deliberately avoiding an error budget that stays nearly full all the time, since that can signal overcaution.
How much weight a survey figure deserves
One survey median is easy to mistake for a benchmark to chase. A better habit is to use it as a sanity check. If your own measurement of toil comes out at fifty-five percent while the published middle is thirty-four, the gap is a prompt to ask what is eating the hours, not proof that you are failing. If it comes out at ten percent, ask whether the work is being counted at all. The same caution applies to every figure in this field, including the example arithmetic from the Atlas: a service with fifty million valid requests and a target of 99.9 percent has fifty thousand failures to spend, and a month with thirty-one thousand failures would pass comfortably, having used a little under two thirds of that allowance. Numbers like these illustrate method; they do not predict your own results.
Where the evidence is thin
An evidence-led reading should also note the gaps. Case material is sparse and mostly self-reported. The Atlas describes one note-taking software firm that set a monthly target using external probers after moving to a major cloud, and a home-improvement retailer that grew from about fifty to eight hundred services with objectives in a year. These are interesting, though neither is health-related and neither amounts to a controlled study. Claims that SRE reduces outages by a given percentage should therefore be treated as unproven unless an organisation can show its own before-and-after data.
What health IT teams can take from this
- Choose one patient-facing or visitor-facing journey and measure it from the outside for a month.
- Set a modest target and state the allowance in plain numbers so that non-technical managers can follow it.
- Write down the rule for pausing changes when the allowance is spent, and secure a sign-off.
- Track toil honestly to see whether the team is above or below the one-half ceiling.
Teams building up their general technical skills may find the roadmap for remote workers learning operations skills a useful companion. Public-facing content also needs to stay reachable at awkward moments, which is why pages such as the overview of spotting familiar breathing complaints and knowing when to ask for help are exactly the kind of resource that should not fail when traffic spikes. That link is offered as an example of a high-demand information page, not as guidance.
Outlook
The direction of travel is clear: clearer targets, fewer but smarter alerts and more candour about trade-offs. The open question is how well the ideas transfer to organisations with small teams and strict governance. Early signs suggest that the habits matter more than the tooling, and the habits are cheap to try.