EasyEnv
Broken on purpose: designing interview challenges that feel like the job
The best way I know to see how an infrastructure engineer works is to give them a machine that is broken and watch them fix it. At EasyEnv we build those machines as code: more than fifty Ansible recipes that set up anything from a single Nginx box to a Kubernetes cluster with an ELK stack and monitoring, many of them broken on purpose.
Building a good broken machine is harder than it sounds. Here is what we have learned.
It has to be believable
The failure should be one the candidate could meet in real life. A service that will not start because its config points at a socket that moved. A disk filled by logs nobody rotates. A certificate that expired on Friday. A Kubernetes pod stuck because a resource limit is set in the wrong unit.
Artificial puzzles (a file hidden in a weird place, a riddle in a comment) test patience, not skill. Real failures test exactly what you will pay them to do.
It has to be reproducible
Every candidate for a role should get the same problem. That rules out "break it by hand before the interview". Each challenge is an Ansible recipe: it builds a clean machine, then applies the breakage as a final, deliberate step. The same recipe gives the same machine every time, and a fix for the recipe is a pull request like any other.
Building on recipes also means the healthy version and the broken version share almost everything, which makes the challenge easy to maintain when a distribution or package version moves on.
It needs more than one way in
A good challenge does not have a single magic command. One candidate starts with journalctl, another with ss -ltnp, a third reads the config first. All of them can get there. What you learn is how each one thinks, and that is only visible when the path is theirs.
It needs a clear definition of fixed
"Make the site work" is not enough, because a candidate can make the symptom disappear without fixing the cause. Each challenge comes with a check that proves the system is healthy, and the interesting part of the review is often the gap between "the error went away" and "it is actually fixed".
AI does not make it easier, and that is the point
Candidates on EasyEnv can use AI, as they would at work. A model can suggest commands, but it cannot see the machine. The candidate still has to look, decide what matters, and judge whether the suggestion fits what they see. Broken-on-purpose challenges turn out to be one of the clearest ways to tell an engineer who directs the AI from one who follows it.
What a good challenge looks like
- A real failure, in a real stack the role uses.
- Built entirely from a recipe, so every candidate gets the same machine.
- Several reasonable starting points.
- A check that proves the fix, not just the absence of an error.
- Small enough to solve in the time you give, deep enough to show judgement.
Curious what this looks like in practice? Try EasyEnv.