Charter customers love a good comparison: light jet or midsize, one-way or round trip, private terminal or main concourse. Behind the booking portals and flight-planning tools that serve those customers sits another set of choices, the ones teams make when they secure the Kubernetes clusters running their software. The Ops Atlas calls that routine work KubernetesSecOps, and this friendly piece lines up five of its decisions side by side, using a charter operator’s mindset to explain each trade-off.
Choice one: fail closed or fail open
Admission checks are the equivalent of the ramp agent who confirms a passenger list before the cabin door closes. The question is what happens when the agent cannot be reached. A fail-closed setup does not let anyone board until the check responds; a fail-open setup lets the flight leave and hopes the paperwork is fine.
The Atlas reports that, as of October 2026, Kyverno behaves as fail closed by default while OPA Gatekeeper behaves as fail open, using a setting named Ignore. Neither is automatically right. A premium trip carrying sensitive guests deserves fail closed; a ferry flight with no passengers can tolerate fail open. The entry recommends choosing per tier and never applying the permissive setting everywhere, since that turns each rule into a suggestion on the day a check service goes down.
Choice two: audit first or enforce on day one
Compare a new cabin policy that is announced two weeks in advance with one that starts at dawn tomorrow. Customers forgive the first and complain about the second. The Atlas leans toward the gradual route: log violations first, then warn for roughly two weeks while publishing the enforcement date, then enforce, with an exceptions list whose entries expire.
Day-one enforcement is faster, and it suits a brand-new cluster with nothing to break. For an established one, the same entry says it can disrupt running work, and it adds the awkward fact that enforcement never evicts programs that were already active. So even a strict start leaves a cleanup job. The trade is speed against surprise, and most operators with live customers choose surprise reduction.
Choice three: homemade encryption or an external key service
Think of two ways to protect the guest manifest. You can lock it in a drawer of the same desk where the key hangs, or keep the key with a separate custodian. The Atlas rates the simple local-key encryption option as weak for exactly that reason: the key lives beside the data. The recommended alternative is an external key management service through its second interface version; the first version has been on the way out since release 1.28 and was switched off by default one release later.
The external route costs more effort to set up and adds a dependency, so it is the equivalent of paying for a separate vault. For secrets that guard passenger records, payment tokens or flight-ops credentials, the Atlas treats that cost as justified. A side note: an identity provider setting keeps data readable if it was written before encryption, so old values need rewriting.
Choice four: shared namespaces or separate clusters
Here the comparison is between a shared terminal and a dedicated one. Namespaces divide a cluster into labeled areas, and they are convenient, cheap and easy to manage. The Atlas is plain that they are not hard isolation, because programs on the same machine share an operating-system core. Separate clusters cost more and need more upkeep, but they offer a physical wall between tenants.
The practical middle is to keep ordinary workloads together while moving the most sensitive ones to their own nodes or clusters. The entry also mentions that recent releases made user namespaces stable, which improves separation inside a container, though it is not a replacement for a proper boundary.
Choice five: delete the suspect or isolate it
When a crew reports a hydraulic warning, nobody scraps the aircraft; it is grounded, tagged and examined. Some automated security tools take the opposite line and delete any pod that triggers an alert. The Atlas disagrees. Removing the pod erases the evidence and leaves untouched whatever tokens it was holding, so the intruder’s credentials keep working.
The preferred response is to change the pod’s label so that it drops out of the service group, apply a stricter profile if possible, and capture its state for review. Deleting is quicker, isolating is safer for the investigation, and the entry recommends the second whenever a human can make the call within reasonable time.
How the choices line up
- Admission failure: closed protects more, open keeps the lights on.
- Rollout style: gradual limits shocks, immediate suits brand-new setups.
- Secrets: external keys give real separation, local keys are simpler but weaker.
- Isolation: namespaces are efficient, separate clusters are stronger.
- Response: deleting is fast, isolating preserves evidence.
Cost deserves a plain mention. Fail-closed settings can block legitimate work during an outage, external key services add a monthly bill and a new dependency, and separate clusters multiply the number of things to patch. Each saving in risk is bought with money or attention, and the Atlas encourages teams to write those prices down beside the decision. A choice with a visible price tends to be revisited on schedule, while a hidden one lingers until it fails.
Notice that none of these choices has a universally correct answer. The Atlas takes the position that the right choice depends on the tier of the workload and the cost of failure, which any charter operator will recognize from deciding who needs the spare aircraft on standby.
Where this sits in your wider planning
Comparison also helps when you shop for the software itself. We discussed a sensible method in our walkthrough of comparing and reviewing software options, and the same advice applies here: define your criteria before you look at feature lists. For guest-facing concerns, this piece on luxury traveler preferences for charter providers is a reminder that trust is part of the product.
If you want to see how the Atlas weighs these options in full, its entry on KubernetesSecOps sets out the request path, the default behaviors and a maturity ladder that ranks setups from untouched to assured. It also lists traps to avoid, such as granting cluster-admin to build jobs and turning on every sample policy in enforce mode at once.
A quick way to decide
Start by sorting your applications into two or three tiers according to how much damage a failure would do. For the top tier choose fail closed, external keys and stronger isolation. For the middle tier use gradual enforcement and per-namespace rules. For the lowest tier, accept the cheaper options but keep audit logging on. Review the sorting twice a year, because applications drift upward in importance more often than they drift down.
Put briefly: spend your caution where a mistake would hurt most, and be honest about the rest. It is the same instinct that tells a charter desk when to arrange a backup aircraft and when a standard reschedule will do.

