The story begins on a Friday evening, as these stories tend to. Priya has kept the servers of a mid-size hospital group alive for nine years — patching, restarting services, updating DNS entries by hand — when a routine release of the appointments misfires because one environment variable never reached the third node. Patients are never at risk, but the booking queue stalls for forty minutes, and Priya spends her weekend reconstructing, from shell history and memory, what each machine is supposed to be running. A colleague sends her a link about GitOps with a two-line note: stop being the configuration. What follows is her path from interrupted human to operator of self-correcting systems, because that path is the most honest career map this field offers.
What GitOps promised her, in plain terms
The pitch she received was short. Write down what each system should look like. Keep that description in a place where nothing can be quietly overwritten and every change is remembered. Let software running inside the platform fetch the description on its own and nudge reality until the two agree. Four commitments sit underneath that pitch, and she learned them in this order: describe the end state rather than the steps; store the description somewhere versioned and effectively unchangeable; arrange for agents to pull it instead of letting pipelines push; and accept that the checking never finishes — a cycle every three minutes counts as continuous. Her old world was the opposite arrangement: her laptop, and later the CI server, reaching into servers while holding production credentials. The NIST publication SP 800-204C takes a dim view of that habit, treating push-based delivery as effectively insecure for exactly that reason. Pull keeps the keys inside the cluster.
The skills GitOps actually pays for
Job ads list tools; the work itself rewards a specific sequence of abilities. Priya’s list, in the order she wished she had learned them:
- Git as an operating tool. Not just commits — branch discipline, reviewing other people’s changes, and reverting cleanly when a merge turns out to be wrong at the worst moment.
- Rendering manifests. Kustomize overlays and Helm values, and above all the habit of reading a rendered diff before approving anything.
- Enough Kubernetes to read events. A stuck sync is a cluster event; you cannot triage what you cannot read.
- The reconciler itself. Argo CD or Flux — applications, sync policies, and the suspend button you will one day need during an incident.
- Secrets handling. One of the usual toolkits — SOPS paired with age keys for encrypted files in the store, or the Sealed Secrets and External Secrets controllers — plus the firm rule that raw credentials never enter the store.
- Judgement. Which resources may be pruned, which must be protected, and when to break glass instead of waiting for a merge.
The last item is the career. Everything above it can be studied in xOps School’s GitOps guide and practised on a spare cluster; judgement only arrives from incidents and reviews.
A ladder you will recognise from any profession
Adoption maturity runs from zero to four — Scripted, Declared, Reconciled, Governed, Assured — and the ladder doubles as a career map. At stage zero a team owns scripts and tribal knowledge; Priya spent nine years there without a name for it. Stage one means the desired state is finally written down, even if someone still applies it by hand. Stage two is the crossing point: an agent now keeps the platform aligned without human muscle, which is where most teams that say they have adopted the model actually stand. Stage three adds governance — guarded deletion, least-privilege reconcilers, one sanctioned route to production. Stage four is assurance: the team rehearses rebuilding production from the store, quarterly, until it is boring. Notice what the ladder implies about hiring. Juniors can hold stage one competently. Mid-level engineers own stage two. The people worth premium salaries are the ones who can be handed stage three and four and not blink, because those stages decide whether an auditor or a 2 a.m. pager is the next visitor.
Her first ninety days, which you can copy
Nothing Priya did required permission from a steering committee, only a series of small, reversible moves. The first month went to decisions on paper: where the store would live, how the directories would be arranged, what the drift policy would be — captured as a short decision record — plus a staging reconciler running with manual sync and alerts whenever reality wandered. The second month automated the staging sync, moved secrets into proper tooling, locked tenants down to their own paths, and rehearsed a break-glass procedure until it felt routine; then came the satisfying part, deleting the CI server’s cluster credentials entirely. The third month promoted one low-risk production service, added guards against destructive pruning, put a small dashboard on the wall, and ran the first rebuild drill. Ninety days later she could show a sceptical manager a production service that no human had logged into to deploy, and an audit trail no one had to remember to write.
Interviews, and the questions underneath them
Hiring panels in this field have stopped asking for tool names and started probing the edges. Four questions come up again and again, each with a real answer behind it. Someone edited the cluster directly — what happens next? The agent reverts or alerts, depending on policy, and a good candidate explains when each response is right; autoscaler-managed replica counts, for instance, are usually tolerated rather than fought. Why pull instead of push? Credentials stay inside the platform, and the pipeline never holds production keys. What would you never place in the store? Data — database contents, volume state — because the store describes structure, not records. A prune just deleted something it should not have — first moves? Halt the reconciler, understand the blast radius, revert the change through a reviewed request, restore data from backup, and only then resume. Candidates who mention that modern AI assistants pushing to watched paths are deployers and deserve the same scoped, reviewed access as humans tend to finish interviews early.
The cost that job advertisements forget: review time
Here is the part nobody puts in a salary band. When every change to production flows through a reviewed request, the bottleneck is human attention. Reviews are the discipline’s tax and its treasure at once — they catch the typo that would have deleted a namespace, and they also slow down the fix needed before morning rounds. Priya’s team learned to protect their reviewers: small diffs, one intention per change, context written into the description so approval takes a minute rather than a meeting. At scale the patience pays. One Argo CD survey found 42 percent of respondents running upwards of five hundred applications from a single instance, which is not the behaviour of a model that collapses under its own ceremony. The people cost is real; it is also the point. An organisation paying for review is buying the one thing manual operations never provided — a second pair of eyes between intention and production.
Where she went next, and what to read first
Priya still works in hospital IT, but her Fridays changed shape. Releases happen in working hours because they are boring now, and her weekends belong to her. If her path reads like somewhere you would like your own career to go, start wide before diving deep: the same site’s walkthrough of the XOps family of disciplines shows where GitOps sits among its neighbours, and its enterprise guide to DevOps, Kubernetes and SRE covers the ecosystem you will be joining. One practical footnote from her ninth year at a keyboard: operators who stop being the configuration sit down more, not less. If long hours at a screen have started to price themselves into your lower back, the site’s plain-spoken piece on when back pain needs specialist spine care is a better first read than any tooling tutorial. Careers in this field are marathons; the config files will still be there after you have stretched.