If you buy or renew software for a hospital’s appointment desk, you will eventually sit across from a vendor who says the platform is built on modern, event-driven cloud services. That sounds reassuring and tells you very little. This list gives you eight direct questions, grounded in ServerlessOps practice as summarised by the Ops Atlas, that turn the sales line into evidence. It concerns administration systems only: scheduling, reminders, document routing and billing notices. Nothing here relates to clinical care.
1. “What stops a repeated message from acting twice?”
Cloud platforms deliver events at least once, so duplicates are routine. A good answer names a business reference, for example the booking number, that is checked before any action with consequences. A weak answer cites the queue’s built-in feature, or says the platform guarantees exactly-once handling. The Atlas notes that even a first-in-first-out queue de-duplicates only inside a five-minute window and cannot see repeats caused later by the consumer.
2. “Where does failed work go, and who reads it?”
Every asynchronous step should have a holding place for items that exhausted their attempts, with a named owner and an alert. Ask how long items stay there. The Atlas expects the holding period to exceed the source queue’s, with a 14-day against 4-day pairing as its example. If the reply is “support looks at it when needed”, treat that as no owner.
3. “Can you replay the actual records?”
A subtle trap: failure destinations on some streaming and notification services record only metadata about the failed item, such as its position, and not its contents. The underlying stream may also expire entries after a day. A vendor who can replay full records will mention that they copy them to durable storage. One who cannot has a gap worth knowing about.
4. “How much parallel work can one feature use?”
Accounts share a ceiling on simultaneous function runs. The Atlas gives 1,000 as a default regional pool on the main provider and recommends planning peaks under about 70 percent of the cap. Good vendors can tell you their cap per feature and what warning fires when headroom shrinks. If a reminder spike can starve the check-in screen, your front desk will feel it first.
5. “How do your timeouts nest?”
The call out to a partner should give up before the function that placed it, and the window during which a queue hides a message from other workers should be a multiple of the function’s own limit; the Atlas mentions six times. You do not need the exact figures, only an answer that shows somebody has thought about the order of the clocks.
6. “What do your alarms watch besides errors?”
Errors alone miss quiet trouble. Items returned for retry inside a successful batch do not count as errors. Look for alarms on backlog age, on the age of the oldest failed item and on the proportion of calls throttled. The Atlas also lists the share of work that achieved its business outcome and the rate of duplicate side effects as signals worth reporting.
7. “What does a million events cost you?”
This is the finance question, and you can ask it plainly. The Atlas walks through one sample path that totals $3.75 for every million events; read that as an illustration of the method and not as a benchmark. A vendor who tracks this figure can explain price changes in your renewal; one who does not may be surprised by their own bills, a risk that eventually reaches you. The Atlas adds that a slow partner can inflate usage-billed compute sharply, and that some steady loads are cheaper on always-on containers.
8. “When did you last rehearse a duplicate storm?”
A game day in which the same events are fired repeatedly, and the team checks that no action is taken a second time, is the most convincing evidence a vendor can offer. In the Atlas’s maturity ladder, which it describes as its own proposal, this rehearsal is what separates the third level from the second. A vendor who has run one will usually be happy to describe what they changed afterwards.
What good answers sound like
Strong vendors tend to answer with specifics and small admissions. They say which reference prevents repeats, and which one step still lacks protection and is scheduled for repair. They name the person who owns the failure shelf. They quote a measured figure, even a rough one, and explain how it is collected. Absolute statements such as “that can never happen” should raise an eyebrow, since the Atlas names blind faith in platform guarantees among its common mistakes.
Red flags worth noting
- Every layer of the system retries on its own, with no single policy.
- Failed items pile up and nobody can say how old the oldest is.
- Cost per event is unknown, and the bill is described as “whatever the cloud charges”.
- Releases go to everyone at once, with no gradual rollout or quick way back.
How to use the list
Do not expect a vendor to answer all eight perfectly. Look for the shape of the answers: specific, owned, measured. Two quick answers and six evasions tell you more than a long presentation. Take notes, ask for the answers in writing, and compare them at renewal. For background, the Atlas’s page on this topic, the ServerlessOps reference at xopsschool.com, stresses that the term is a historical label with no certifying body, so any “certified” claim attached to the name deserves a follow-up question.
Elsewhere on this site, an explainer about arranging digital consultations from home covers the patient-facing side of such services, and a budgeting article built around an expense tracker routine that lasts all month shows why regular measurement beats guesswork. Bring the eight questions to your next vendor meeting, and keep the answers filed.
