Platforms that connect international patients with providers run on a stack of cloud services: enquiry forms, document uploads, scheduling and messaging. When that stack is built by hand over several years, a recognisable set of problems appears. This article diagnoses those problems one by one and matches each with a fix borrowed from InfraOps, the informal name for running infrastructure through reviewed, version-controlled definitions. It is about administration and technology only. It makes no statement about any provider, treatment or outcome.
Symptom: the environment cannot be recreated
Diagnosis. The live system works, but nobody can say what it would take to rebuild it. Knowledge sits in individual memories and a few dated screenshots.
Remedy. Describe the setup as code, then prove it by rebuilding a non-critical copy from nothing. The Ops Atlas that xOps School maintains treats a rebuild drill as a milestone on its own proposed maturity ladder, between a “governed” stage and a “continuous” one. Expect the first attempt to fail on something undocumented; that is the drill working.
Symptom: reports look clean, yet surprises keep arriving
Diagnosis. Teams run a check, see no differences and relax. The check, however, only examines resources the tooling already manages. A rule added by hand during a busy launch week remains invisible. The Atlas points to a vendor’s own drift tutorial in which a hand-made security rule appeared merely as an identifier in a comparison.
Remedy. Add a second kind of check that begins with the provider’s list and looks for items missing from the code. Also learn the limits of your tools: the Atlas notes that one cloud service’s drift detection compares only explicitly set properties, producing false alarms such as a memory size written two ways, and that another vendor’s preview feature can report properties as removed when they were not. Treat reports as leads, not verdicts.
Symptom: differences linger for weeks
Diagnosis. Detection is not the issue; ownership is. The trouble is that a discrepancy with no named person attached just sits there. The Atlas’s worked example is instructive. Out of 140 groups of resources, 11 showed differences over thirty days, a rate of 7.9 percent, with a median age of about two days and a worst case near seventeen days. The comment attached to it is that the memorable number is the seventeen-day outlier, not the comfortable percentage.
Remedy. Give every difference an owner and a deadline. Self-healing is acceptable, per the Atlas, only for modules that store nothing you need to keep and only while no incident is open. Everywhere else, send a notification, give the finding to one named person and try to close it inside roughly seventy-two hours. Track how long each finding has been open, because age is the number that exposes neglect.
Symptom: one small edit makes everybody nervous
Diagnosis. Everything lives in one giant record, so any change can touch anything. The Atlas cites cloud-provider advice to keep each record to well under a hundred resources, and notes one large firm that arrived at around fourteen hundred records by splitting per region and account.
Remedy. Split along ownership lines, for example the patient-facing site, document storage and the messaging service. Smaller records preview faster and limit how far a mistake can travel. The 2017 storage outage the Atlas recounts, caused by a single wrong input to a command, is the standing reminder of why limits and brakes matter.
Symptom: sensitive documents and credentials sit in odd places
Diagnosis. For a platform handling personal documents, careless secrets handling is a serious exposure. State and preview files can hold secrets, and the credentials used for previews run provider code, so they need almost the same care as those used to apply changes. The Atlas also warns that, in common pipeline setups, any author who can open a branch may be able to use the preview role.
Remedy. Do not pass saved previews between jobs, restrict who can trigger applies, and use the newer features the Atlas lists for keeping secrets out of stored records. Review these settings with your technical partner rather than assuming defaults are safe.
Symptom: servers are old and nobody knows how old
Diagnosis. Machines have been patched in place for years, each slightly different. The Atlas’s worked example counts the share of 600 instances running on images younger than thirty days, 75 percent in that illustration, as a way of making age visible.
Remedy. Rebuild standard images on a schedule and replace servers instead of tinkering with them. Patching does not vanish; it moves into the image pipeline.
Symptom: a vendor’s feature name is mistaken for a method
Diagnosis. Procurement conversations often blur products and practices. The Atlas flags that one commercial product launched in late 2025 carries the same name as the discipline, which is a coincidence of naming and not a definition. A proposal that says “we do InfraOps” may mean a product, a team label or a full lifecycle approach.
Remedy. Ask three questions in writing: what is described in code, who approves previews, and how are differences detected and owned? Answers to those tell you far more than any label. It is the same instinct a careful traveller applies when asking what a certificate actually covers.
A note on reading the sources
Be sceptical of anyone who presents InfraOps as an official standard. The Atlas is explicit that no certification or standards body defines it, and that its metrics and maturity ladder are its own proposals. For the full list, see the Ops Atlas entry laying out these symptoms and metrics. Within this blog, readers curious about the engineering side may enjoy our look at enterprise cloud modernisation and Kubernetes work in Japan, and those thinking about how to judge claims may value a guide to checking hospital accreditation before travelling abroad.
Putting the diagnosis to work
Pick the symptom that rings truest and run the matching remedy for one month. Record the starting number and the ending one, whether it is the age of the oldest unowned difference or the share of resources covered by code. Present the result as a plain sentence about what remains unreviewed and who will close it, not as a triumphant percentage. That honesty is what makes the next quarter’s work easier.