Four KubernetesOps Choices Compared, Records-Office Style

Healthcare IT

Think of the records office at a busy health-information service. Files must be available all day, nothing may be lost, and every change needs a trail. The people who run it constantly face choices between two reasonable options, each with a catch. Teams running Kubernetes face a similar set of forks. This piece compares four of them, following the way the Atlas page on KubernetesOps frames the trade-offs, and keeps the language non-technical where possible. Nothing here relates to clinical care; it is purely about how the supporting systems are run.

Choice one: let a provider run the control side, or run it yourself

A managed service takes over the control plane, meaning the API server and the datastore behind it. A self-run cluster keeps all of that in your hands. In records-office terms, it is the difference between outsourcing your filing room and keeping it in-house.

The managed route removes a lot of routine pain. According to the Atlas, it takes away the care of the datastore and the API server, and one large provider backs up the clusters it manages. Left with you are the endless chase after new versions, keeping track of installed add-ons, setting believable requests and proving restores work. You still own your configuration in Git, your own application data backups and a rebuild you have rehearsed.

The self-run route gives full control and full responsibility. You take your own datastore snapshots, rotate certificates and handle the odd-numbered member count the datastore needs, with five members common in production. Choose it if you have the staff to do the work properly, or if you must run in a place where managed services are not available.

Choice two: upgrade in place, or build a new cluster alongside

The in-place route moves the existing cluster forward one minor version at a time. It keeps data and identity where they are, but each hop demands a window, and rollback is narrow. The Atlas notes the install tool can safely be re-run and keeps backups in a temporary folder, while one managed service cannot pause an upgrade and only allows going back one minor version within seven days, as of October 2026.

The side-by-side route builds a fresh cluster at a newer version and moves workloads across. It sidesteps the one-step-at-a-time rule, which the Atlas says cannot be broken on the API server, and it pairs naturally with a rebuild drill. The cost is duplicated capacity for a while and a careful cutover of traffic and data. A records office will recognise the move: opening the new filing room next door and transferring boxes steadily, instead of renovating while people are searching for files.

Rule of thumb: if you are several versions behind, a new cluster may take less total effort than a long chain of hops. If you are one version behind and healthy, in place is simpler.

Choice three: add more copies, or resize each copy

When a workload is struggling, you can run more copies of it, which is what the horizontal autoscaler does, or you can give each copy more room, which is the job of the vertical recommender. Per the Atlas, the vertical project’s own guidance is not to pair the two on one shared processor or memory signal, since each would tug against the other. The horizontal option also needs processor requests to be set before it can act on processor use.

A practical split: use the horizontal approach for stateless services that tolerate many copies, and use the recommender in observation mode to refine requests over time. The Atlas notes that resizing running pods in place became stable in version 1.35, with limits: a pod keeps its service tier, and reducing memory is attempted on a best-effort basis.

For a sense of how people think about recurring costs, a guide to splitting shared expenses with a payment tracker makes a similar point at household scale: pick one clear rule for dividing a recurring cost, rather than two rules that argue.

Choice four: stay on the older routing interface, or move to the newer one

Traffic enters most clusters through an interface that describes routes. The older Ingress API is frozen but, as the Atlas puts it, with no plans for removal. The newer Gateway API is recommended for new work. Meanwhile, a popular NGINX-based controller was archived in March 2026, so teams that depend on it face a decision with a deadline attached. A migration assistant exists, but the Atlas calls it exactly that and warns that certain annotations have no direct equivalent.

In office terms, it is the reception desk redesign. The old desk still works; the new one is where the investment is going. If you rely on the retired controller, move first. If you use another implementation, plan the switch as part of a normal upgrade year.

What the records office would add

Any good records manager insists on an audit trail, and the same instinct helps here. Whichever way you decide, store the decision with its date in the same repository as the configuration. Include the numbers you looked at: how far behind the newest release the cluster was, how much capacity was requested against used, when the last restore test ran. If a newcomer questions the decision twelve months on, the supporting evidence is already in the folder. This habit costs almost nothing and turns a debate about opinions into a review of facts.

A quick comparison table in words

  • Managed control plane: less datastore and API-server toil, same duties on versions, add-ons, requests and restores.
  • Self-run control plane: maximum control, plus snapshots, certificate rotation and member care.
  • In-place upgrade: keeps everything, one hop at a time, narrow rollback.
  • New cluster: skips the staircase, costs duplicated capacity and a careful cutover.
  • More copies: good for stateless work; needs honest processor requests.
  • Bigger copies: fixes sizing; do not mix with the horizontal approach on the same metric.

How to decide without drama

Ask three questions of each fork. What does it cost if it fails halfway? Who owns the recovery? What evidence will we look at afterwards? A comparison is only useful if somebody writes down the choice, the reason and the date to review it. Six months later, when the team has changed, that note is the difference between a policy and a mystery. For a lighter read on how to think about unfamiliar online services in general, a piece on how an online ENT consultation service is explained to users shows plain-language explanation done well.

The bottom line: none of these options is wrong, but each hides a cost. Pick deliberately, record why, and revisit the choice when your cluster, your team or the release calendar changes.