Every operations team eventually has to decide what to do when a live system no longer matches its written description. Four policies are common, and each has supporters. This comparison sets them side by side for hospital administrators and IT coordinators who oversee booking portals, document stores and internal tools. It concerns administration and technology only, with no clinical content, and it uses the vocabulary of InfraOps, an informal term for managing infrastructure through code.
The analysis relies on the Ops Atlas entry about InfraOps and drift handling in the infrastructure-as-code lifecycle. The Atlas is clear that the label has no standard behind it, so what follows compares practices, not compliance regimes.
The scenario
A scheduled comparison reports that a setting on the appointment portal’s servers differs from the code. Someone changed it by hand three days ago. What now? The four policies below answer differently.
Policy 1: Ignore it
How it works. The team adds an instruction telling the tool to ignore changes to that setting, or simply never looks at the report.
Strengths. Zero effort, and no disruption in the short term.
Weaknesses. The Atlas calls this kind of ignore setting a mute button. Used sparingly, with a comment and an expiry date, it is legitimate. Used as a habit, it removes the very signal that tells you reality has moved, and over time the code stops describing anything real.
Verdict. Acceptable for a known, documented exception. Poor as a default.
Policy 2: Alert and assign
How it works. The comparison runs on a timer, the pipeline flags any non-empty result, and a ticket goes to the owner of that component. The owner chooses whether to update the code to match reality or to restore reality from the code.
Strengths. A human decides, so deliberate emergency fixes are not destroyed. Findings are recorded and measurable. The Atlas suggests a target of around seventy-two hours to acknowledge and assign.
Weaknesses. It depends on people acting. Without a visible age metric, tickets can sit. Typical findings may be closed in a couple of days while one lingers for weeks, so track the oldest open finding, not only the median.
Verdict. The Atlas’s general recommendation for anything that holds durable data or is under investigation.
Policy 3: Automatic reversion
How it works. The pipeline detects drift and applies the code at once, putting reality back.
Strengths. Fast, consistent and tireless. It makes hand edits pointless, which discourages them.
Weaknesses. In the Atlas’s words, reverting everything automatically is an unreviewed apply pipeline holding the strongest credentials. It can erase an urgent manual repair during an incident, and a mistaken change to the code would propagate without a human reading the preview.
Verdict. Its advice is to limit the practice to modules free of durable data, and to stand it down whenever an incident is open.
Policy 4: Replace instead of repair
How it works. Servers are built from a prepared image and never modified in place. If one deviates, it is replaced by a fresh copy.
Strengths. Deviations vanish with the replacement. A cloud provider’s reliability guidance, which the Atlas cites, rules out in-place edits of production machines once the immutable model is chosen.
Weaknesses. It needs an image pipeline, regular rebuilds and a way to keep data outside the machines. Note that immutable is not a synonym for never updating: the work relocates to a scheduled image build, and image age becomes the number to follow, ideally measured as the share of machines on recent images.
Verdict. Strong for stateless components. Not applicable to anything that stores data on the machine itself.
Side-by-side summary
- Visibility of change: ignore is lowest, alert and assign and replace are highest, auto-revert is high but unreviewed.
- Risk of destroying a deliberate fix: highest with auto-revert, low with alert and assign.
- Effort to set up: lowest for ignore, highest for replace.
- Fit for data-holding components: alert and assign only.
- Fit for stateless components: replace first, auto-revert second.
How to choose per component
Do not pick one policy for the whole estate. Classify each component by whether it holds durable data, how critical it is and how often it changes. Then assign a policy and record it next to the code. A reasonable default is alert and assign everywhere, with replacement for stateless web tiers and auto-revert only where the Atlas’s conditions hold.
The same classification logic appears in fleet planning. Our article on running a vehicle fleet with rental management software looks at similar asset decisions from the fleet side, and the discussion of why metadata matters for search visibility and asset tracking shows how good labels make any inventory easier to police.
A worked classification
Consider four components of a typical hospital administration stack. The public information pages hold no durable data and change often, so replacement from an image suits them. The appointment request form accepts submissions but passes them on immediately, so alert and assign with a short target is safer than automatic reversion. The document store holds records and must never be auto-reverted, and its deletion protection should be set on the provider side as well as in code, since a code-side flag alone fails once the configuration is removed. The internal reporting tool is low-risk and rarely changes, so a dated ignore rule for one known setting is tolerable. Four components, four sensible answers, each written down beside its code.
Caveats about the detection itself
Whatever policy you choose, remember that detection is imperfect. One cloud provider’s checker looks only at properties that were set explicitly, so it can raise false alarms, and another tool’s preview can label still-present properties as deleted; the Atlas flags both. A clean comparison also says nothing about resources the code does not know. Pair your policy with a periodic count of unmanaged items.
Recommendation
Start with alert and assign, because it produces the data you need to judge the others. After a quarter, review which findings were harmless and repetitive, and consider automating those only. Keep ignore rules rare, dated and commented. Measure the age of the oldest open finding, not just the average. And revisit the classification whenever a component starts to hold data it did not hold before.
