First ThingsBoard POC on Proxmox

The requirement was simple on paper: set up an IoT platform to visualize about a hundred LoRaWAN sensors without paying for a vendor’s managed-cloud version. ThingsBoard Community Edition was the obvious candidate. I wanted to run it on the client’s existing Proxmox infrastructure, without a dedicated VM if possible.

Why LXC rather than a VM?

An LXC container uses ten times less RAM than an equivalent Debian VM, and startup is instant. The client already has about ten containers in production on the cluster; staying consistent with the existing pattern simplifies maintenance.

The only catch: ThingsBoard CE wants Docker inside. So LXC nesting (nesting=1, keyctl=1) must be enabled on the container on the Proxmox side, and cgroup v2 interferes a bit with the iptables embedded by Docker. Nothing blocking.

The initial docker-compose

To avoid wasting time debugging a bare-metal install, I started with the adapted official compose:

services:
  thingsboard:
    image: thingsboard/tb-postgres:3.7.0
    ports:
      - "8080:9090"
      - "1883:1883"
      - "5683:5683/udp"
    environment:
      TB_QUEUE_TYPE: in-memory
      INSTALL_TB: "true"
      LOAD_DEMO: "true"
    volumes:
      - tb-data:/data
      - tb-logs:/var/log/thingsboard
    restart: unless-stopped

volumes:
  tb-data:
  tb-logs:

in-memory for the queue is more than enough at this stage — no need to bring in Kafka for a POC. LOAD_DEMO creates the demo tenant, handy to validate the UI without any sensor integration. Once the real devices arrive, we switch to LOAD_DEMO: "false" and take a clean dump of the compose.

What got stuck

Three things I hadn’t anticipated:

1. MQTT port 1883 squatted by mosquitto on the host. I’d forgotten another container was already exposing MQTT. Quick fix: 1884:1883 on the ThingsBoard side, and an MQTT reverse proxy on the HAProxy side to route based on client_id.
2. RAM consumption climbs to 2.3 GB at startup with the embedded Postgres. Initial LXC sizing at 1.5 GB = OOM-killer within the minute. Bumped to 4 GB, it breathes.
3. The rule engine doesn’t digest raw Cayenne LPP payloads without a custom JS decoder in the device profile. An hour lost wondering why my temperatures were arriving as data: "AQEAFw==". Note for later: decode on the ChirpStack side before forwarding.

What’s next

POC validated, the client is signing for a deployment of 600 sensors across 12 sites. Next step: externalize Postgres, switch the queue to RabbitMQ, and write a more serious article on real sizing.