DEV Community

Reed Dev
Reed Dev

Posted on

Running One Docker Container Per User on a $35/Month Server

#docker #devops #cloud #tutorial

I run a Telegram AI bot where every user gets their own Docker container. Here is how that works on a single cheap GCP instance without Kubernetes.

Why Per-User Containers?

My AI companion bot gives each user a persistent agent with its own memory, personality file, and schedule. Users should never see each other's data, and each agent needs an isolated filesystem.

Options considered:

  • Shared process, per-user directories: Cheapest but one crash kills everyone
  • Kubernetes: Overkill for 10-50 users on one machine
  • Docker containers, managed by gateway: Just right

The Architecture 

Gateway container (always running)
  |
  +-- User A container (started on demand)
  +-- User B container (started on demand)
  +-- User C container (idle, stopped after 60min)
Enter fullscreen mode Exit fullscreen mode

The gateway manages container lifecycle:

  • Start on demand: When a message arrives for user X, check if container exists. If not, docker create + docker start.
  • Idle cleanup: Every 5 minutes, check which containers have not received a message recently. Stop idle ones.
  • Stopped containers use zero CPU/RAM: Docker keeps the filesystem, so state persists.

Key Code 

async function ensureContainer(userId: string): Promise<string> {
  const name = `adola-user-${userId.slice(0, 8)}`;

  try {
    const info = await docker.inspect(name);
    if (!info.State.Running) {
      await docker.start(name);
    }
    return name;
  } catch {
    await docker.create({
      name,
      Image: "adola-agent:latest",
      HostConfig: {
        Binds: [`/data/users/${userId}/workspace:/workspace`],
        NetworkMode: "adola-net"
      }
    });
    await docker.start(name);
    return name;
  }
}
Enter fullscreen mode Exit fullscreen mode

Memory Usage

On my e2-medium (4GB RAM):

  • Gateway + Postgres + Caddy: ~400MB baseline
  • Each active user container: ~150-200MB
  • Stopped containers: 0MB

With 8 users, peak usage during heartbeat cycle (all containers briefly active): ~2GB. Comfortable headroom.

The Gateway Manages Docker via Socket

The gateway container mounts /var/run/docker.sock so it can create/start/stop sibling containers. This is the "Docker out of Docker" pattern.

gateway:
  volumes:
    - /var/run/docker.sock:/var/run/docker.sock
    - /data:/data
Enter fullscreen mode Exit fullscreen mode

Networking

All containers join the same Docker bridge network (adola-net). The gateway calls user containers by name:

http://adola-user-abc12345:18789/v1/chat/completions
Enter fullscreen mode Exit fullscreen mode

Docker DNS handles resolution. No port mapping needed.

What I Would Change

For 100+ users, I would add:

  • Container pooling (pre-warm a few idle containers)
  • Horizontal scaling (multiple gateway instances with consistent hashing)
  • Prometheus metrics on container lifecycle

But for the 8-50 user range, this simple approach works perfectly.

Try the Bot

The system is live: t.me/adola2048_bot - each user gets their own isolated AI agent container on Telegram.

Top comments (1)

Collapse
 
fredbrooker_74 profile image
Fred Brooker •

45 EUR - Hetzner - you get 12 CPU, 64 GB RAM, plenty for running 500 containers