InnerWeb
Skip to main content

Docker or Kubernetes in a plant: recommendations

A plant with one server and a small IT team rarely needs Kubernetes, but it often needs the discipline that Kubernetes enforces: pinned versions, repeatable deployment, backup and preproduction. Below are seven recommendations for the administrator who chooses how to run a safety system and wants to avoid both excess complexity and makeshift setups.

8 GB
memory of the machine where parallel builds killed the database process
Source: karta
5 GB
swap memory after the fix
Source: karta

Who this recommendation is for

These recommendations apply to a plant where the safety system runs on one server or one pair of virtual machines, and the IT team consists of a few people who also handle the network, workstations and ERP. For larger groups with an existing cluster the conclusions are different, and we describe them in the last part.

In a customer installation the infrastructure is supplied by the customer, and the software is deployed in Docker containers in two instances: production and a twin preproduction. We start from this point.

Recommendation 1: start with Docker containers, not with an orchestrator

A single server with Docker containers is simpler to run than a cluster. Fewer components mean fewer things that can fail at three in the morning. The local stack of the system consists of a TLS-terminating reverse proxy, a PHP 8.4 application with Apache, a MySQL 8.0 database, an S3-compatible object store, Redis, a mail catcher, a PDF generation service and a .NET helper service. Docker Compose describes these several services in a single file.

Kubernetes makes sense when you need to scale horizontally, spread the load across several nodes or meet a requirement of the group. With one server, the cost of knowledge and upkeep of a cluster usually outweighs the benefit.

Example number of elements to maintain

Example number of elements to maintainComparison of the number of hosts: one for Docker Compose, three assumed for the cluster, with the same eight services.02468elementsHostsDocker Compose (Hosts): 1 elements1Cluster (Hosts): 3 elements3Stack servicesDocker Compose (Stack services): 8 elements8Cluster (Stack services): 8 elements8
  • Docker Compose
  • Cluster
Source: Example calculation: 8 stack services from the description of the local stack; a three-node cluster as an assumption

Recommendation 2: pin versions and base images

Regardless of the orchestrator, versions and base images should be pinned. In the stack of the system these are PHP 8.4, MySQL 8.0 and Redis 8.6. Pinning means that refreshing an image will not silently change the database or interpreter version, and an upgrade becomes a decision, not an accident.

The plant should know which versions run there today and keep them in the installation description. This helps when reporting to the service desk and during the audit of suppliers.

Recommendation 3: a twin preproduction as the first obligation

Preproduction is a second instance with the same configuration, where updates are tested and backups are restored. This element is worth more than an orchestrator, because it reduces the risk of an update. A team with one server can run it on a separate virtual machine, or even on the same host with resource limits, keeping memory in mind.

Preproduction should not contain a dump of the production database with personal data. A separate page about the test environment built from an empty database describes this.

Recommendation 4: do not run several heavy tasks at once

The resources of one server are finite. In our build environment, parallel builds on a machine with 8 GB of memory led to the database process being killed by the OOM mechanism in September 2026. The fix was to cancel outdated pipelines, run one build at a time and add 5 GB of swap. The lesson also applies to the plant's server: an update, a backup and reporting queries should not start in the same minute.

Schedule windows for backups and updates, and set memory limits for containers so that one of them cannot push the database out.

Recommendation 5: treat backup and restore as a feature, not an add-on

A backup that has never been restored is a hypothesis. In a container model the data volumes are clearly separated, so the backup covers the database, the file store and the configuration of the environment variable contract. The restore is tested on preproduction.

The data and the backups stay on the customer's disks. The plant sets its own schedule and storage location.

Recommendation 6: Kubernetes only for a specific reason

The reasons that justify a cluster are: a mandatory platform of the group, a high-availability requirement with automatic failover, several instances under shared management, or a team that already knows Kubernetes well. It is worth writing them down before the decision. If none of them applies, Docker containers with a backup and a preproduction are enough.

The same image package can run in both models, so the choice does not close the way to the future.

Recommendation 7: agree the boundaries of responsibility with the supplier

Record who is responsible for the operating system, who for updating the images, who for the backup and who for the network. The most common misunderstandings concern exactly these boundaries. It is also important that InnerWeb is a software vendor, not a device maker, and the installation does not require its own KD readers or the replacement of the plant's system.

Video

Frequently asked questions

Does the system run on Kubernetes?

The container images can run on an orchestrator if the plant requires it. The basic model for one server is Docker containers with a backup and a preproduction.

Is Docker Compose a temporary solution?

No. It is a simple and proven way to run several services on one host, provided that we take care of pinned versions, backups and restore tests.

What matters more: a cluster or a preproduction?

For a small team, preproduction. It reduces the risk of updates without expanding the infrastructure.

How much memory does the server need?

It depends on the number of users and supporting services. Our experience in our own environment shows that 8 GB requires discipline in scheduling heavy tasks, so plan a reserve for a production deployment.

Sources

Let's match the run model to your IT team

In the technical meeting we will establish whether Docker containers with a preproduction are enough, or whether the plant has a reason to use an orchestrator.

Let's talk