Pi 4 log-monitoring ML moves to Ubuntu VPS via Docker Hub

The guide builds directly on a previous setup in which a BUFFALO WSR-1500AX2L router's Syslog is collected on a Raspberry Pi 4, aggregated and judged in Python, and reported by Gmail. In that arrangement the Pi handles always-on monitoring while a Windows PC's Docker environment builds the machine learning model. The new article keeps that division but moves the training location off the PC so training can run without the PC being left on.

The migration deliberately separates what travels in the image from what does not. Python, dependencies, training code, and tests go into the image and through Docker Hub, while compose.yaml, logs, models, evaluation results, and credentials are handled separately. The guide notes that volumes and bind mounts are not included when an image is pushed, so image transfer and data migration are treated as two distinct tasks. It also uses architecture checks — linux/amd64 — to keep the image compatible with the target VPS.

On the VPS side, the training container runs with no network, a read-only root filesystem, and only the models folder writable, with the data folder mounted read-only. A successful run is not the same as production-ready quality, because the existing train.py counts anomalies without ground-truth labels and therefore does not compute Precision or Recall. Model transfer back to the Pi is left as a manual step after a human quality review, and a dry-run notification mode is used before switching the Pi's configuration to live SMTP.