# Do you need Kubernetes? A plain test, and what to use if you do not

Most small and mid-size product teams do not need Kubernetes. It earns its cost when you run many services that deploy independently, load that changes a lot, and enough engineers to own the cluster. If you run a handful of services with steady traffic, a managed container service such as Cloud Run, Amazon ECS or Azure Container Apps, or Docker Compose on a virtual machine, gives you repeatable deploys with far less to operate.

By Shai Goldenberg. Updated: 2026-11-15.

## What Kubernetes does

Kubernetes runs containers across a group of servers and keeps them in the state you describe. You declare that a service should have three copies, a memory limit and a health check. Kubernetes places the copies on machines with room, restarts one that crashes, replaces a machine that disappears and shifts traffic to a new version step by step.

It began at Google and is now a graduated project of the Cloud Native Computing Foundation (CNCF). Every major cloud sells a managed version: EKS on AWS, AKS on Azure and GKE on Google Cloud.

The problems it solves are real:

- Many services, owned by different teams, that deploy on their own schedules
- Load that rises and falls enough that fixed servers waste money or fall over
- A need to pack many workloads tightly onto shared machines
- The same deployment model across more than one cloud or data center

If those are not your problems, Kubernetes gives you its costs without its benefits.

## What it costs to run

The software is free. The cost is attention.

**Upgrades.** According to the Kubernetes release documentation, the project maintains release branches for its three most recent minor versions, and each minor version receives about one year of patch support. A cluster left alone falls out of support within about a year. Cloud providers add their own deadlines for managed clusters. Each upgrade has to be tested against your ingress controller, your charts and your add-ons.

**Sizing.** Every workload needs resource requests and limits. Without them one service can starve the others, and the autoscaler cannot make sound decisions.

**Networking and access.** Ingress, TLS certificates, network policies, namespaces and role-based access all have to be designed. The default settings are permissive.

**Configuration.** Manifests or Helm charts for each service and each environment, a way to promote one image from staging to production, and a place for secrets outside the repository.

**People.** Someone has to understand all of the above at two in the morning. For a team of three, that person is also building the product.

A cluster chosen before the need is the first problem listed on our Kubernetes page, for this reason.

## The plain test

Answer yes or no.

- [ ] We run more than about ten services that deploy independently.
- [ ] More than one team deploys to production, and they block each other today.
- [ ] Our load varies by a large factor between quiet and busy periods, on a daily or weekly cycle.
- [ ] We have hit a limit of a simpler platform: a workload it cannot run, a networking need it cannot meet or a cost that grew past a cluster's.
- [ ] We must run the same stack in more than one cloud, or in a cloud and on-premises.
- [ ] We have, or will hire, at least one engineer whose job includes owning the platform.
- [ ] Our applications are already packaged as containers and deploy through a pipeline.
- [ ] We can commit to a version upgrade several times a year.

**Reading the result.**

- **Zero to two yes answers.** You do not need Kubernetes. Use one of the simpler options below.
- **Three to five.** You are in the middle. A managed container service probably still fits. Revisit when a specific limit appears.
- **Six or more, including the last three.** A managed cluster is a reasonable choice. The last three are conditions, not benefits. Without containers, a pipeline and an owner, a cluster will make things worse.

## What to use if you do not need it

| Option | What you manage | Fits when | Main limit |
|---|---|---|---|
| Platform that deploys from the repository (for example Vercel, or an app platform on your cloud) | Application code and settings | Web applications and APIs with standard needs | Less control over networking and long-running jobs |
| Managed container service: Google Cloud Run, Amazon ECS, Azure Container Apps | Container images, service settings, scaling rules | A handful to a few dozen services, with autoscaling and no servers to patch | Provider-specific configuration |
| Docker Compose on one or two virtual machines | The machine, the Compose file, backups | Internal tools, on-premises installs, steady small workloads | Scaling and failover are manual |
| Virtual machines with an autoscaling group | Operating system, patches, deploy scripts | Software that does not containerize well | Slower deploys and more patching |
| Managed Kubernetes: EKS, AKS, GKE | Nodes, upgrades, networking, access, manifests | Many services, several teams, variable load | Needs an owner and regular upgrades |

The second row is where most product teams should look first. You hand the provider a container image. It runs the image, scales it with traffic and replaces failed instances. There is no cluster to upgrade.

Docker Compose is not a toy. Shai's on-prem MCP gateway product installs at the customer with Docker Compose, because a single host with a clear restart policy is easier for the customer's IT team to run than a cluster.

## The work that pays off either way

Three pieces of work are worth doing whether or not a cluster follows.

**Containers.** A Dockerfile with a multi-stage build and a small base image, so each service runs the same way on a laptop, in CI and in production. If you later move to Kubernetes, the images are ready. If you never do, deploys are still repeatable.

**A pipeline.** Images built once per commit, tested, deployed to staging on merge and to production on approval, with a rollback step. This is the core of [DevOps consulting services](https://goldenroz.com/devops-consulting/): CI/CD pipelines, release engineering, separate staging and production environments, observability and Docker packaging. The mechanics are explained in [Terraform and CI/CD explained](https://goldenroz.com/blog/terraform-and-ci-cd-explained/).

**Alerts that reach a person.** Error and uptime alerts, trimmed to the few that deserve a response.

A team without a platform engineer can get this work done as a phased project. [DevOps as a service](https://goldenroz.com/devops-as-a-service/) delivers it in fixed-price phases that end with a handover, so your engineers run releases and rollbacks themselves afterward.

## If the answer is yes

When the test says a cluster is justified, these decisions come first.

1. **Use a managed control plane.** Running your own control plane is a job for a platform team.
2. **Design node pools, namespaces and ingress on paper** before creating anything, and create the cluster from code.
3. **Set requests and limits for every workload** from measured usage.
4. **Promote one image through environments** with Helm charts or a GitOps tool. Do not copy manifests between staging and production by hand.
5. **Apply a security baseline.** Access per namespace, default-deny network policies, secrets in an external manager and image scanning in CI.
6. **Write the upgrade runbook and rehearse it on staging** before production depends on the cluster.
7. **Decide who is on call.** If nobody inside the company can be, ongoing administration can sit with a provider under a [managed cloud services](https://goldenroz.com/managed-cloud-services/) retainer, with coverage hours and response times written into the agreement.

## What we do, and what we do not claim

Goldenroz has no record of past Kubernetes results to show, no certified Kubernetes engineers and no partner status with any cloud provider. We say so because a buyer could reasonably assume otherwise. The evidenced part of our work is Docker packaging and CI/CD: Shai Goldenberg, our CTO, runs 10 CI workflows on his own production marketplace platform, with separate staging and production.

For that reason, [Kubernetes consulting](https://goldenroz.com/kubernetes-consulting/) starts with a readiness review: a look at your services, team size, release frequency and scaling pattern, with a written answer on whether Kubernetes, a managed container service or plain Docker fits. If a cluster is justified, the later phases cover cluster design, the deploy pipeline, the security baseline and a handover in which your engineers perform an upgrade on staging. We do not offer around-the-clock cluster operations.

## What to do next

Run the test with the people who would operate the cluster, not only the people who want one. If you land at two or fewer, containerize the services and put them on a managed container service. If you land at six or more, start with the readiness review in our [Kubernetes consulting](https://goldenroz.com/kubernetes-consulting/) service. Projects start at $25,000 at a fixed price per phase, and the first one-hour scoping call is free.

## Questions buyers ask

### What is Kubernetes used for?

Kubernetes is an open-source system that runs containers across a group of machines. It schedules containers onto servers, restarts them when they fail, scales them up and down, routes traffic to them and rolls out new versions gradually. It is a graduated project of the Cloud Native Computing Foundation.

### Does a small product team need Kubernetes?

Rarely. A team of two to five engineers with one to five services is usually better served by a managed container service or a platform that deploys from the repository. Kubernetes adds upgrades, networking, access rules and manifests that somebody has to own every month.

### What is the difference between Docker and Kubernetes?

Docker packages an application and its dependencies into an image and runs it as a container on one machine. Kubernetes runs many containers across many machines and keeps them in the state you declared. You need container images before Kubernetes is any use, and many teams need only the images.

### Is managed Kubernetes such as EKS, AKS or GKE easier to run?

It is easier, not effortless. The provider runs the control plane. You still choose node sizes, set resource requests and limits, manage ingress and secrets, control access and schedule version upgrades. The Kubernetes project maintains only its three most recent minor releases, so upgrades are a recurring task.

### Does Goldenroz have Kubernetes case studies?

No. We say this plainly on the service page. The evidenced part of our work is Docker packaging and CI/CD pipelines. Kubernetes work is scoped per project, starts with a paid readiness review, and that review can conclude that you do not need a cluster.

---
Page: https://goldenroz.com/blog/do-you-need-kubernetes/
Book a free one-hour call: https://calendly.com/daniarozin/daniel-rozin-consultation-1-hour-clone
