Kubernetes or Docker Compose on a single server
Kubernetes is often installed «for growth» while the whole application lives on one VPS. Here is what it gives on a single server compared with Docker Compose, how many resources it eats itself and when moving to a cluster is really justified.
The short answer. On one server Docker Compose is almost always enough: one file describes services, networks and volumes, and memory goes only to the applications. Single-node Kubernetes adds convenient updates and one way to describe everything, but it eats 1.5–2 GB of memory and does not protect against the failure of the server itself. It starts to make sense from three servers or when the team already uses it elsewhere.
What Docker Compose gives you
Compose starts several containers with one command. Start order is set with depends_on: with the service_healthy condition the application waits until the database passes its health check. The restart: unless-stopped policy brings a crashed container back up and starts everything after a server reboot. The weak spot is updates: docker compose up stops a changed container and creates a new one; volumes are kept, but the service is down for a few seconds. How to install Docker without filling the disk with images is in Docker on a VPS.
What Kubernetes itself costs
For kubeadm the documentation requires 2 GB of memory per machine and 2 CPUs for the control plane node. Lightweight k3s — a single binary under 100 MB that keeps its state in SQLite — needs 2 cores and 2 GB for a server. In the k3s developers' measurements, a server with a workload holds about 1.6 GB of memory for its own components alone. On a 4 GB VPS that is almost half of what Compose would leave to the application; for single-node Kubernetes it is sensible to take 8 GB or more.
Updates without downtime
The main thing Kubernetes gives on one server is gradual replacement of copies. A Deployment updates one step at a time by default (RollingUpdate): new copies start, pass the readiness check, and only then are the old ones shut down. In Compose the same is done with two copies of the application behind nginx or Traefik restarted in turn — see Updating a server without downtime.
When you need a cluster
On one node Kubernetes does not save you from the main risk: if the server breaks, everything stops. A fault-tolerant cluster needs an odd number of control plane nodes: with three, etcd survives the loss of one, with two it survives none. That means three VPS that reach each other over the network without restrictions (a public network works, but the host's private network within one data centre is safer), plus a load balancer or floating IP for incoming traffic (see Floating IP and load balancer). Such a cluster is justified when there are many services, several teams deploy them or the service must keep running when a machine fails.
What the servers cost
In the accsly.io catalog on 18 November 2026 there are 2,585 plans with 2 vCPU and 4 GB at 1,320 hosts; the monthly median with monthly billing and no prepayment is €16.31 for plans priced in euros, $24.99 in dollars and 1,080 ₽ in roubles. With 4 vCPU and 8 GB there are 2,452 plans, median €32.74, $48 and 2,500 ₽. Compose with an application and a database usually fits the first option, single-node Kubernetes the second, and a cluster of three 4 GB servers costs about three times as much as one. Servers with hourly billing are handy for trying a cluster. Selections: VPS for Kubernetes and VPS for Docker.