When a small team hears about GitOps for the first time, the questions arrive in a predictable order, and most articles answer none of them directly. This one does. Seven questions, seven short answers, no sales pitch. If your team runs a booking site, a news portal, or an online shop from a modest setup, these are the answers you need before committing an afternoon to the idea.
GitOps at a Glance: The Answers in Brief
Skimmers, this one is for you. The long answers follow, each with the reasoning, but here is the whole FAQ compressed:
- Definition: a recorded description of the system, plus an unattended program that continuously pulls it and makes reality match.
- Git itself: optional. Any versioned, change-resistant store qualifies, including image registries and versioned buckets.
- Daily work: the agent reads, compares, applies, and reports on a loop of a couple of minutes.
- Hand-edits: overwritten at the next cycle, unless you flip the documented pause switch first.
- Credentials: kept inside your infrastructure, with read-only reach to the store.
- Deletion: possible and therefore guarded — state-bearing objects get explicit protection.
- First move: two decision sessions, then a self-managing agent on a test environment with destructive behaviour off.
What Is GitOps, in One Sentence?
Write down exactly how the system should look, keep that description in version control, and let an unattended program inside your infrastructure fetch it continuously and adjust the live system until reality matches the page. That is the whole mechanism. Changes reach production by editing the description and getting the edit approved — never by logging into a server and improvising. Teams adopt it because the written record turns “who changed this?” from an investigation into a lookup.
Does It Have to Be Git, Literally?
No, and this surprises people. Git is the usual choice because it is familiar, free, and keeps a clean history of who approved what. But the underlying requirement is any versioned, effectively unchangeable store. Practitioners have built working setups on container image registries using the OCI format, and even on versioned cloud storage buckets — the community calls these arrangements “Gitless GitOps”, only half-jokingly. Pick whichever store your team will actually maintain. The discipline matters; the brand name on the folder does not.
What Does the Agent Actually Do All Day?
Less drama than you would expect, which is the point. On a steady cadence — for one popular agent that means roughly every two minutes — it reads the recorded description, compares it against what is running, and reports the difference. If a change was approved, it applies it. If somebody edited the live system by hand, it flags the difference and, on most teams, quietly puts things back the way the record says. A webhook can wake it early when a change merges, but the polling loop is the backbone; the wake-up call is a courtesy, not a requirement.
We Edited a Server by Hand. What Happens Next?
This is the question that decides whether the model suits a team’s temperament. Under automatic reconciliation, that hand-edit survives only until the next cycle — minutes, typically — and then the agent overwrites it. Nobody is being punished; the written description is simply the authority, which is the entire guarantee you are paying for. Teams that need emergency hand-edits during outages get a documented pause switch instead: reconciliation is suspended, the fire is fixed, and the stable outcome is written back into the record before the loop resumes. If that trade sounds wrong to your culture, fix the culture or skip the tool — adopting it half-heartedly gives you neither benefit.
Is This Overkill for a Two-Developer Shop?
Often not, though the honest answer depends on what currently hurts. A two-person team that already talks constantly and ships carefully gains less from the enforced review path than a twenty-person team does — the coordination the model removes simply may not exist yet. But size cuts the other way for one group: solo operators and duos have nobody to remember what was changed, and the recorded description does that remembering for free. The pragmatic middle path many small teams choose is to run the loop with manual approval at first — the agent prepares every change and shows the difference, and a human presses the apply button. You get the written record and the tidy recovery story while keeping your hands on the wheel. If the button starts feeling like a chore after a few months, that is the market telling you the team has outgrown it, and switching to automatic then is a one-line change.
Where Do the Passwords Live?
In a safer place than they live today, if your current setup lets build jobs reach into production. With the pull arrangement, the agent sits inside your infrastructure and holds only read access to the store. Nothing outside — no build server, no laptop, no contractor — needs credentials that can alter production, which removes an entire category of leak. Two cautions keep it honest. The agent itself often starts with more internal power than it needs, so narrow that on day one. And if you enable webhook wake-ups, put a shared secret on the endpoint, because by default anyone who finds the URL can knock on it all night.
Will the Robot Delete My Data?
It can, and the field has the incident reports to prove it, so treat this answer as the serious one in the list. Most setups include a pruning behaviour: when something is removed from the recorded description, the agent removes it from the live system too. Suppose a merge conflict quietly drops one file from the record — the corresponding database, together with its storage volume, can be dismantled by a program that believes it is being helpful. The defences are standard and cheap: mark state-bearing things (namespaces, volumes, custom resource definitions) as protected from pruning, set storage to retain data even when the object above it is deleted, and let a pipeline check flag any proposal that removes something stateful. Do all three before your first automated deletion ever runs.
What Should We Actually Do First?
Two sessions, no more, before any tooling. Session one is decisions: where the descriptions will live, how the folders are arranged, which secrets system you will use, and what your drift rule is — revert, allow, or alert. Session two is a bootstrap: install the agent on a test environment and point it at a repository that describes the agent itself, so the tool manages its own upgrades from the start. After that, move one unimportant service onto the record, with the destructive behaviours switched off, and let it run for a fortnight. The reference material at the Ops Atlas GitOps explainer expands this into a proper adoption roadmap with labs and checklists when you are ready for the next step.
Where to Read More, and Where to Switch Off
Everything above compresses a much longer field into an afternoon of reading. When the afternoon ends, this city has better options than a second tab. For a slow morning, the best things to do in Bhopal for a memorable trip include the lakes at hours when no server is watching you back. If the reading session turns into a planning weekend, our Bhopal travel guide to attractions, heritage, and day trips handles the itinerary, and the practical side of getting around is covered in this simple guide to vehicle rentals and rental management — useful whether your trip starts from Bhopal or from a layover in Nagoya.
Adopting GitOps will not transform a small team overnight, and anybody promising that is selling something. What it reliably gives you is a written answer to “what changed, who approved it, and what is the system supposed to look like?” — and for most teams that have gone looking for those answers in chat logs at midnight, that is reason enough.
