Book Appointment

SRE Questions Hospital Software Buyers Should Ask Vendors

Healthcare Software

Hospitals buy scheduling tools, patient messaging platforms and record viewers, and then live with them for years. One of the least discussed questions at purchase time is how the supplier keeps the thing running. SRE, short for site reliability engineering, gives buyers a structured way to ask. What follows is a set of prompts that an operations or procurement group might raise with a supplier, or with the hospital’s own developers. The treatment is analytical and administrative; it has no clinical content and offers no medical guidance.

Every claim about the discipline traces to one place, the Ops Atlas and its extended SRE entry, which serves as the reference for all that follows. Interpretation and any caution are mine.

What does SRE actually promise a hospital software buyer?

Nothing grand. It promises a method: define what good service means to users, set a numerical target, track performance against it, and use the remaining slack to decide how quickly to change things. The Atlas summarises it as handling operations with the habits of software engineering. For a buyer, the relevant outcome is predictability. You learn in advance how often the product is expected to falter, and you learn what the supplier will do when it does.

Why is a target of one hundred percent a warning sign?

Because it cannot be true. Any vendor that promises no failure is either not measuring or not being candid. The Atlas explains that each further step towards perfection becomes steeply more expensive, and that a service depending on others needs those dependencies to be even more reliable than itself, a guideline it calls the extra nine. If a supplier lists a figure, ask what the measuring point is, which period it covers, and what the supplier does once the allowance is used.

What is an error budget, in terms a finance office would follow?

Take the target, subtract it from a whole, and what is left is the failure you may tolerate. One of the Atlas’s examples: aim for three nines across a four-week window and you may afford about forty unhappy minutes, equivalent to a thousand misses in each million attempts. Treat it like an expense line. Spending part of it on a change is acceptable; spending all of it signals that the next change should wait. A buyer can ask to see how much of the allowance was used in the last quarter and what the vendor did in response.

How should speed be described in a service agreement?

As a proportion, not an average. Ask for the share of requests answered within a stated number of seconds. An average that looks fine can hide a subset of users, perhaps those on slower connections in a remote branch, who wait far longer than anyone else. The Atlas is firm on this point, and it adds a subtle rule: an error that returns quickly belongs in the availability count but not in the latency figures, because speed does not make a failure any less of one.

What evidence of discipline should a vendor be able to show?

Four items cover most of it. First, written objectives expressed from the user’s side. Second, a documented policy for what changes when the allowance runs out. Third, examples of reviews after incidents, showing actions with owners that were finished. Fourth, a plain account of how on-call staff are protected from overload. Two entries on the Atlas’s list of common mistakes, namely targets nobody has signed a policy for and reviews that leave their follow-ups open, are the most productive places to probe.

Can evidence of good practice be real without being dramatic?

Yes, and the Atlas offers a modest example. A note-taking service moved to a cloud platform, committed to a monthly objective using external probes, and then used its own data to shorten the window it kept open for releases, going from five to two. Nothing about it was flashy. The point is that measurements changed behaviour. When a vendor describes similar, specific adjustments, that is more persuasive than slogans.

How many people does a sensible on-call rota need?

More than most small suppliers have. The Atlas cites a guideline of about eight engineers for continuous single-site coverage and six per site when two sites share the load, with a cap on incidents per shift. If a vendor with a handful of staff promises round-the-clock response, ask how that is staffed. Perhaps the answer is a partner arrangement or a limited-hours commitment, which is perfectly acceptable provided it is stated plainly.

Why do staged releases matter to a buyer?

The Atlas observes that most outages follow a change to something already live. The antidote it describes is to let a limited group try each new release ahead of everyone else, then compare results before widening. It illustrates the arithmetic with a slice of five percent seeing a fault rate of twenty percent, which shows up as one percent across everybody. For a hospital buyer, the practical question is whether the supplier pilots updates with a small group, and whether you can opt into that group or stay behind it during busy periods such as open enrolment or month-end billing.

Which promises should make a buyer cautious?

Be wary of three things. A sweeping figure applied to every feature equally, because a reporting export and a payment page do not deserve the same target. A supplier who counts only their own servers, since users experience the whole path. And any refusal to describe what happens after a miss. Reliability work is largely a record of how an organisation behaves when things go wrong, so vagueness here deserves a follow-up question rather than a shrug.

What if we build in-house instead of buying?

Then the same questions apply to your own team, and the Atlas offers a short decision list for deciding whether a dedicated reliability function makes sense: do users notice outages, will a senior person sign the policy, does manual work stay high after reduction, and can a rota be staffed sustainably. If the answers are mostly no, a part-time approach works better, with one named owner and a monthly review. Pretending to have a full team when you do not is the fake version the Atlas warns about.

How does this relate to monitoring and incident handling?

It overlaps but is not the same. The Atlas places detailed telemetry, incident communication roles and service agreement paperwork in neighbouring disciplines. Treat SRE as the layer that connects them through targets and budgets. A vendor with excellent dashboards but no objective, or one with a fine incident process but no measurement of what users experienced, is missing a piece.

What are realistic first steps for a hospital team?

Choose one service with obvious patient or staff impact. Write the definition of a good attempt. Gather four weeks of counts from existing logs. Calculate the allowance. Draft a one-page policy. Share it with the supplier and ask them to comment. Few vendors will refuse that conversation, and the way they respond is itself informative.

Where can I read more?

The source for the technical points above is the Ops Atlas reference page on site reliability engineering, which is long but well organised. On this site, a post about smoothing software and cloud operations through engineering services offers a technology-side companion, and an explanation of how clinicians assess knee complaints gives a sense of the patient journey that administrative software supports. The latter is general information only.

A closing note on expectations

No framework guarantees a trouble-free system. What the approach offers is honesty: an agreed target, a count of how often it is missed, and a set of pre-agreed responses. For a hospital that depends on its software every hour of every day, that honesty is worth asking for before a contract is signed, and worth maintaining long afterwards.