A freelancer’s first week on a new client usually includes an uncomfortable discovery: the infrastructure exists, it works, and nobody can say exactly how it got that way. Cloud consoles hold manual edits, a laptop holds the only copy of a working directory, and a deployment command has a history that lives in someone’s shell. These are InfraOps problems, and they are worth diagnosing methodically, because as an outside contractor you carry the blame for whatever breaks while you are on the job. The notes below pair each symptom with a cause and a remedy, using the public Ops Atlas write-up as the common reference.
Symptom one: nobody trusts a preview
You run a plan and it proposes to change dozens of things you never touched. The usual cause is accumulated drift: edits made directly in the provider account over months. The remedy is to stop treating the preview as a verdict and start treating it as a reconciliation list. Sort each proposed change into three bins: reality is right and the code should be updated, the code is right and reality should be restored, or the difference is cosmetic noise from the provider. The Atlas observes that some drift detectors only compare properties that were explicitly set, which produces false positives such as a memory size written as 1024 megabytes in one place and 1 gigabyte in another.
Symptom two: one state file for everything
A single huge state is slow, and worse, it widens the area affected by any slip. Google Cloud’s guidance, quoted in the Atlas, is to keep a state to a few dozen resources and certainly under 100. The same reference reports that a large messaging company ended up with about 1,400 states, divided per region and per account, to keep each one small. As a freelancer you will rarely get to split everything at once. Propose a plan in slices: carve out the highest-risk area first, such as networking or databases, and leave the rest for later invoices.
Symptom three: applies happen from a laptop
If the client’s changes run from individual machines, nobody can reconstruct who did what, credentials are scattered, and plan files wander around in downloads folders. The Atlas lists applying from laptops as an anti-pattern, and it adds a detail that matters for contractors: plan files hold state and variables, so they must not be passed between jobs or left as downloadable build artifacts that anyone with repository access can read. The fix is a pipeline that previews on every change request, applies from a protected environment, and never accepts a plan produced elsewhere.
Symptom four: a quiet ignore list
Open the code and you may find lifecycle settings that tell the tool to ignore changes to whole attributes. Sometimes that is sensible. Often it is a mute button pressed during an incident and never released. The Atlas likens this setting to a mute button, because it hides exactly the differences you most need to see. For each ignore entry, write a one-line justification in the code or remove it and see what the preview says.
Symptom five: the protective flag that did not protect
Clients proudly point to the setting that blocks destruction of their database resources. The catch, per the Atlas, is that this guard fails when the configuration block is deleted altogether, since the tool has nothing left to enforce. Pair it with the provider’s own deletion protection and you have two independent safeguards instead of one. This small change is a very reasonable item for a fixed-price audit.
What to put in your engagement notes
Evidence beats opinion when you present findings. Record these counts at the start and again at the end of the engagement:
- how many resources are managed by code versus how many exist in scope,
- how many repositories have a plan running on every change request,
- who holds apply rights and through which route,
- how old the oldest unresolved drift is.
The Atlas offers a worked example of why the framing matters: it prefers saying that hundreds of resources escape review and that one drift sat for 17 days, rather than quoting a flattering percentage alone. A before-and-after table in that style is easy for a client to read and easy for you to defend.
Boundaries for your own protection
Contractors should also keep a record of what they were asked to change. Ask for a read-only role for previews wherever possible, though note that even a read-only plan role can see secrets stored in state, which is why the Atlas says plan credentials deserve near-apply-level care. If a state lock is held by someone else, leave it alone until they release it, and put scope limits in writing. For the wider practice side of freelancing, our posts on common questions that first-time DevOps freelancers have and on building a dependable circle of mentors and peers cover the business half of the problem.
A short diagnosis flow
When time is short, run through this sequence: confirm where state lives and whether it is locked, list who can apply, schedule a nightly preview, then triage the first report into the three bins above. Four steps produce a defensible first deliverable. For full definitions and diagrams, read the Atlas page on InfraOps before the kickoff call, and bring one printed page of questions.
Nothing here needs special tooling, only patience and a habit of writing down what you find. Clients tend to remember the contractor who left them with a readable list of risks, ranked, with owners attached.