Consider an imagined team at an imagined aesthetic-care group; nothing here describes any actual organisation. It follows a small web and operations team at an imaginary aesthetic-care group over one quarter as they bring order to the systems behind their booking pages and photo libraries. The story involves no treatment, no patients and no clinical claims; it is only about servers, records and habits, the territory that some technologists call InfraOps.
The concepts are taken from the Ops Atlas entry on InfraOps and infrastructure written as code. The Atlas stresses that the term is casual, with no standards body or certificate behind it, so the team in our story is borrowing vocabulary, not seeking a credential.
The Monday everything looked fine
The team consisted of an office manager who also handled vendors, a web developer and a part-time contractor. They had inherited a collection of cloud resources built by hand over several years: a few servers, a database, storage for images, some scheduled jobs. A consultant’s old diagram existed, and it was about two years out of date.
One Monday the booking form stopped accepting requests for about forty minutes. The cause turned out to be a firewall rule changed during an urgent fix the previous Friday. Nobody had recorded the change, so the written description of the setup no longer matched reality. The developer’s first reaction was to run the team’s existing configuration files to restore the old state, then pause. If those files were stale, applying them might delete the Friday fix and remove more than intended.
Week one: finding out what existed
The developer drew up an inventory, listing every resource visible in the cloud console. There were 120 items. Only 74 appeared in any file. The team set aside the 46 mystery items and gave each one a note: owner unknown, purpose unknown, or purpose known but not in code. This mirrors a point in the Atlas: a clean plan only speaks for the objects the tool already knows, and the real work begins with the ones it does not.
The office manager asked a sensible question: “How bad is 74 out of 120?” The developer answered with the count, not a percentage: 46 things nobody reviews. That framing got the budget conversation moving much faster than a percentage would have.
Weeks two to four: splitting and previews
The existing configuration was one giant block. Any change meant a long preview that nobody read closely. The team split it into five smaller parts: network, booking application, database, image storage and scheduled jobs. The Atlas quotes cloud-provider advice to keep each unit to a few dozen resources, and the team found that smaller previews could actually be read in a couple of minutes.
They also added a rule: no one applies changes from a personal laptop. Changes go through a request, the pipeline produces a preview, and a second person reads it. The contractor grumbled at first. Then a preview caught a typo that would have removed the image storage bucket, and the grumbling stopped.
For readers curious about the media side of such a setup, our beginner’s explanation of how digital asset management works for organised image libraries covers the file-handling basics.
Weeks five to eight: the drift check
With a pipeline in place, the team scheduled a comparison between code and reality every night. For the first week it reported noise: a few settings that the cloud provider fills in automatically. They tuned it. By week seven it flagged one real difference, a storage setting changed by hand, and the owner and developer settled the matter in ten minutes.
They did not enable automatic reversion. The Atlas advises against blanket auto-revert, describing it as an unreviewed apply with the strongest credentials, and suggests alerting and assigning an owner within about seventy-two hours. The team adopted exactly that target and tracked how long each finding stayed open.
Weeks nine to twelve: bringing in the risky items
The 46 mystery items were triaged. The team imported the ten riskiest first: the database, the public entry points and the storage holding customer-submitted forms. They reviewed each imported definition and removed several defaults that no one had intended. They left low-risk leftovers alone for the next quarter, and deleted a few clearly abandoned items after confirming with the owners.
The team also read a longer overview of reliability practice, our guide to site reliability engineering for IT professionals, to compare their habits with a wider body of practice.
The argument about secrets
Midway through the quarter the contractor asked why the team was so fussy about who could read the saved records of what the tool had built. The developer’s answer came straight from the Atlas: those records, and the saved previews, can contain sensitive values, and even the role that only previews changes executes provider code and can see them. So the team reduced the number of people with access, kept preview credentials as carefully guarded as apply credentials, and noted a plan to look at encrypted state when they next upgraded their tooling.
It was a ten-minute conversation that had been avoided for two years, and it is the sort of discussion that a written, reviewed setup makes natural.
Where the figures landed
Of the 120 items, 84 were now under code, leaving 36 unmanaged. They tracked drift age and found a median of about a day. Their oldest open finding was five days. They set a target to reach roughly ninety percent coverage the following quarter and to hold a rebuild drill. These are the story’s own invented figures, shown to illustrate how a team might report progress.
What the team would tell you
- Count first, then talk about percentages.
- Split the configuration before adding clever features.
- Make the preview a human step, not a formality.
- Do not reverse every difference automatically.
- Import the dangerous things first, and review what the import produces.
The quarter did not end with a perfect system. It ended with a system whose weaknesses were written down, owned and shrinking, which is a far better place to be than a diagram two years out of date.