Most questions about InfraOps arrive at the worst moment: a change half-finished, a lock that will not release, a teammate asking whether it is safe to try again. This page answers the questions in the order people tend to panic about them. No theory tour, just what to do and why. It draws on the structure described in the Atlas write-up of InfraOps, a reference text on infrastructure managed through code, and keeps everything at the level a busy owner or manager can act on.
So what exactly is InfraOps?
It is shorthand for running infrastructure by describing it in files, reviewing changes before they happen, and applying them through a controlled route. The word itself is casual, the Atlas stresses. No certification defines it, and you cannot purchase it off a shelf, so be wary of anyone who sells it that way.
My apply failed halfway. Can I roll it back?
Not in the way you might hope. According to the Atlas, there is no transactional rollback: some resources were changed and some were not. The sensible response is to fix forward. Read what the failure says, correct the cause, produce a fresh preview and apply that. Resist the urge to hand-edit things in the console to “get back to normal”, because every manual touch widens the gap between the code and reality.
The lock is stuck. Can I just break it?
Only if you took it yourself and you are certain the run is dead. The Atlas puts it bluntly: never break a state lock you did not take. Someone else may be in the middle of a real change, and forcing a second writer in is how records get corrupted. Ask in the team channel, check the pipeline for a running job, and wait for the owner.
Why did my saved preview stop working?
Previews are perishable. When one saved plan is applied, other plans prepared against the same state go stale and must be regenerated. That is a feature, not a bug, since the tool refuses to act on an out-of-date picture. If you are building a release routine, plan to generate, review and apply close together rather than leaving a preview sitting for days.
Is it okay to run changes from my own laptop?
It will technically work, but the Atlas files it under anti-patterns. A laptop run leaves no shared record, depends on whatever credentials happen to be saved locally and cannot be repeated by a colleague. Move applies into a pipeline with its own identity, and keep laptops for read-only experiments. It is also worth asking how the approach scales across a growing company; our note on growth for product-led businesses touches the same theme of repeatable routines.
Can preview files leak secrets?
Yes. State and plan files can contain sensitive values in readable form, and the Atlas notes that the role used to run a preview executes provider code, so it needs nearly the same care as the role that applies. Newer features such as ephemeral values and write-only arguments, added in recent releases of the main engine, keep some secrets out of state, and one open-source alternative offers state encryption. Do not email plan files, and do not upload them as downloadable build artifacts that any repository reader can fetch.
Should I use the “target” flag to apply just one thing?
Treat it as an emergency tool. The documentation quoted by the Atlas says targeting is for exceptional circumstances only, because skipping the rest of the graph can leave dependent resources out of step. If you find yourself reaching for it weekly, the real problem is that your state is too large or too tangled and needs splitting.
How big should one state be?
Smaller than you think. Google Cloud guidance cited in the entry advises against more than a hundred resources in a single state, ideally only a few dozen. The Atlas also mentions a large messaging company that ended up with roughly 1,400 state files by splitting per region per account. A small blast radius means a bad day touches little.
Does a clean preview mean everything is covered?
No, and this catches experienced people too. A preview only compares the code with resources the tool already tracks. Something created by hand in the console, or something deleted outside the tool, will not announce itself. The Atlas describes a tutorial where a hand-made security group surfaced only as a bare identifier in a diff. Schedule a regular detection run and read its output, rather than treating an empty preview as a certificate.
Should drift be corrected automatically?
Usually not. Automatic reversal is effectively an unreviewed apply with powerful credentials. The Atlas suggests reserving it for modules holding no durable data and only when no incident is open. Everything else gets an alert, an assigned owner and a target of about three days for resolution.
What should I measure first?
Start with coverage, meaning the share of in-scope resources the code manages, then count how many changes happen outside the pipeline. Add the age of the oldest unresolved drift. The Atlas warns against trophy numbers such as lines of code or module counts. If you need a starting point for putting engagements and services on a roadmap, a service roadmap aimed at new clients shows how to sequence the first steps.
Where should we begin on Monday?
Choose one live change, run it end to end through a reviewed pipeline, and write down every step where a person had to improvise. Those improvised steps are your backlog. Repeat next week with a second change. Nothing fancier is required, and the habit matters more than any tool you choose.

