Tool brief

When to Use Kubernetes: Adoption Criteria

Decide using workload placement, deployment frequency, platform ownership, and the cost of operating shared Kubernetes capabilities.

Editorial update

Decision in brief

Use Kubernetes when several workloads need common placement across machines, independent rollouts, or standardized platform policies, and a team can run those controls through upgrades and outages. Start with a managed application platform or a single-host deployment when the operating requirements fit there. Compare platforms at the workload and on-call level before migrating.

Test the platform fit: Shared placement. Independent rollouts. Common policies.
Compare with a simpler platform and include upgrades and on-call work.

Identify which Kubernetes capability earns its cost

Kubernetes schedules Pods and reconciles declared workload state. Deployments can manage replicas and rolling updates. Services provide stable access to matching backends. These capabilities depend on the control plane, available worker capacity, networking, and the workload's readiness and storage choices.

Managed Kubernetes can take responsibility for operating the control plane. Confirm who owns node updates, admission policies, ingress, storage, identity, and incident response under the specific plan. Service recovery requires available worker capacity and a data recovery design for stateful workloads.

Workload situationBaseline to compareKubernetes decision trigger
One small service with a tolerant recovery periodOne VM, container host, or managed application serviceA documented host-failure or rollout requirement exceeds the baseline
Several teams ship independentlyManaged container hosting and managed KubernetesShared policy, workload APIs, and ownership can be demonstrated on real services
Many identical stateless replicasManaged container hosting with autoscalingA common scheduling or scaling requirement needs cluster-level controls
Stateful application with a strict restore requirementManaged database alongside either app runtimeThe application team can test data recovery independently of Pod replacement

Use HN experience to set the platform baseline

In the Uncloud announcement, creator psviderski described operating Kubernetes in small and large environments and seeing teams that mostly needed rollouts, networking, and HTTPS across a few machines. Make those needs explicit and measure the simplest platform that meets them.

In a separate discussion about learning Kubernetes, burdock described experimenting with alternatives to text templating in Helm. Include the time for a second operator to trace a change from source configuration through generated resources to the running service. A rollout speed measurement alone misses that work.

Deploy one representative workload to both candidates

  1. Choose a real workload pattern with synthetic data and one measurable traffic requirement. Write down expected peak requests, the rollback and recovery windows, and how much operator time is available each month.
  2. Deploy the same image to the current baseline and a disposable Kubernetes environment. Time the work to configure ingress, identity, observability, and persistent storage if required.
  3. Run a good rollout, an aborted rollout, and a worker or host loss. Record failed requests, time to restore useful traffic, and which actions required elevated access or off-hours intervention.
  4. Adopt Kubernetes only if it meets an explicit requirement the baseline misses and a second operator can release and recover the service from the runbook. Stop if routine upgrades, rollback, or data recovery still have no owner.

Related articles