Daily reports · UGREEN DXP4800 NAS · by Siddharth Bhadauria

A home server that watches itself for trouble

My NAS records its own CPU, memory and disk every 15 minutes. This project turns that history into 18 features in DuckDB, retrains two anomaly detectors every six hours, keeps the better one, and opens a GitHub issue when the live readings look wrong. Each day it publishes a short model report; the charts below are built from those reports.

From the latest report

01

Training history

One point per daily report. Each chart has its own scale.

Rows of telemetry the model trained on

Grows by about 96 a day: one reading every 15 minutes.

Share of rows flagged as anomalous

What the winning model calls unusual in its own training data.

02

Pipeline

Five steps, all running in Docker on the NAS. Nothing leaves the house except the daily report and any alert.

  1. Collectauto-datapulse · every 15 min

    A sister project samples host and container metrics and appends them to telemetry.duckdb.

  2. Build featuresDuckDB SQL window functions

    Rolling means and standard deviations over one and six hours, rates of change, and time of day.

    cpu_percent · ram_percent · disk_percent · cpu/ram/disk_rolling_mean_1h · cpu/ram/disk_rolling_std_1h · cpu/ram_rolling_mean_6h · cpu/ram_rolling_std_6h · cpu/ram/disk_rate_of_change · hour_of_day · day_of_week

  3. Train and pickscikit-learn · MLflow · every 6 h

    Isolation Forest and Local Outlier Factor are both trained with 1% contamination. Each run is logged to MLflow and the better model is saved for serving.

  4. Score live readingsFastAPI /predict · Prometheus

    Every 15 minutes the monitor sends the newest reading to the API. Latency and predictions are exported to Prometheus and drawn in Grafana.

  5. AlertGitHub Issues API

    Two anomalous readings in a row open an issue labelled automated-alert. Later hits comment on the open issue instead of opening new ones, with a four-hour cooldown.

03

Stack

One docker-compose.yml, managed in Portainer.

ServiceWhat it doesPort
Pipeline daemonRetrains every 6 hours, scans for anomalies every 15 minutesbackground
FastAPIServes predictions and a Prometheus metrics endpoint:8000
MLflowExperiment runs and model artefacts:5000
PrometheusScrapes inference latency and anomaly counts:9090
GrafanaLatency percentiles and anomaly dashboards:3000