obeoneorg/sys-stats repository overview

⁠📊 sys-stats

Real-time system, GPU and Ollama monitoring for a single machine, with a web dashboard, a terminal dashboard, and a compact JSON endpoint for an embedded wall display.

One process collects everything (CPU, RAM, swap, load, per-core usage, temperatures, fans, NVIDIA GPUs, per-process CPU/RAM/VRAM, loaded Ollama models) in a background sampler and serves it over HTTP. The terminal dashboard is a client of that API, not a second collector, so you can watch a remote host from your laptop.

⁠🚀 Features

⁠⚡ Quickstart

docker run -d --pid=host -p 5000:5000 ghcr.io/obeone/sys-stats:1.5.0

Open http://localhost:5000⁠.

--pid=host is what makes the process tables meaningful. Without it the container sees exactly one process, its own. No privileged flag is needed.

On macOS, port 5000 belongs to the AirPlay Receiver. Publish on another port (-p 5051:5000) or turn AirPlay off.

⁠🖼️ Screenshots

⁠📦 Installation

Two console scripts come with the package, whichever way you install it:

python -m sys_stats is an alias for the CLI.

⁠With uv (recommended)

uv⁠ is the fastest option and keeps the tool in its own environment:

uv tool install git+https://github.com/obeone/sys-stats.git

Upgrade with uv tool upgrade sys-stats, remove with uv tool uninstall sys-stats. To run it once without installing anything permanent:

uvx --from git+https://github.com/obeone/sys-stats.git sys-stats

⁠With pipx

pipx install git+https://github.com/obeone/sys-stats.git

Upgrade with pipx upgrade sys-stats, remove with pipx uninstall sys-stats.

⁠With pip

python3 -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
pip install git+https://github.com/obeone/sys-stats.git

⁠From a local checkout, for development

git clone https://github.com/obeone/sys-stats.git
cd sys-stats
uv venv && source .venv/bin/activate
uv pip install -e '.[dev]'

⁠🐳 Docker

git clone https://github.com/obeone/sys-stats.git
cd sys-stats
docker compose up -d

Overlays layer on top of compose.yaml for hardware the default cannot reach:

Images are published on every release to ghcr.io/obeone/sys-stats and docker.io/obeoneorg/sys-stats, for linux/amd64 and linux/arm64, and signed with cosign.

⁠What the container needs, and what it does not

The container runs unprivileged, as UID 10001, and that is enough for everything except IPMI:

/dev/ipmi0 is root:root 0600 on every distribution that ships it, so compose.ipmi.yaml runs as root. Where host udev rules give the device a group, prefer group_add with that GID and keep the unprivileged user; the overlay documents both.

⁠☸️ Kubernetes

chart/⁠ is a Helm chart built on the bjw-s common library⁠:

helm dependency update chart/
helm upgrade --install sys-stats ./chart --namespace monitoring --create-namespace

The dashboard reports on the node the pod lands on, so the chart defaults to hostPID: true, without which the process tables are empty. The container itself runs unprivileged, with capabilities.drop: [ALL].

Pin the pod to the machine you actually want to watch, claim a GPU if you want the NVIDIA panels, and point OLLAMA_API_URL somewhere for the Ollama one. chart/values.yaml⁠ documents all three, along with the commented block for IPMI, which does need a privileged pod on Kubernetes since there is no per-device request for a BMC.

⁠⚙️ Configuration

Everything is an environment variable. Nothing is required.

⁠Server

⁠Sampler

⁠Panel

⁠CLI

⁠The ones with a real story behind them

SYS_STATS_PANEL_ONLY is security-relevant. /stats exposes the full host process table, complete command lines included, with no authentication. Set this and /, /stats and /favicon.png are never registered at all, so a request gets Flask's own 404 rather than a guarded rejection. On a host whose network you do not fully trust, and where the consumer only ever needs /panel, removing the route beats guarding it. The Helm chart's default probes hit /, so enabling this there means repointing them at /panel.

SYS_STATS_HOSTNAME and SYS_STATS_INSTANCE_LABEL are not the same thing and are never merged. The label is display prose, meant to be rewritten for readability, for telling apart two instances describing the same physical box (a Kubernetes pod seeing the Talos VM, and a second instance on the Proxmox hypervisor underneath). The hostname is a machine identity a consumer compares byte for byte, /panel only, and otherwise read fresh on every request so a runtime rename takes effect without a restart.

SYS_STATS_DCGM_URL is for a host with no NVIDIA driver of its own, such as a Proxmox hypervisor whose GPUs are PCI-passed-through to a Kubernetes VM where dcgm-exporter⁠ runs instead. It is scraped in the same sampling pass as everything else, behind a circuit breaker, so a dead or firewalled endpoint degrades to an empty gpu[] with a "gpu" tag in err rather than stalling the sampler. /stats never reads it and always stays on the GPUtil path.

SYS_STATS_IPMI_INTERVAL is decoupled from the sample interval because chassis fan speed and temperature move on a timescale of tens of seconds, and each poll costs a BMC round trip. Between polls the lists keep serving the last IPMI reading rather than dropping it; the first pass after startup always polls immediately. hwmon sensors keep the normal per-pass cadence.

⁠📺 Using the terminal dashboard

sys-stats --url http://localhost:5000/stats --interval 5

Every panel carries a number in its title, and a one line key helper is pinned at the bottom. Press h for the full list.

A panel that cannot show every row says how many it is hiding, so a busy host never drops rows silently. A zoomed panel keeps refreshing and keeps scrolling.

The layout follows the terminal: a wide one gets a row of columns, a narrower one falls back to a grid, and panels stop where their content stops instead of framing empty space. On a machine with several GPUs the summary keeps cumulated figures, each card gets its own detail table (or one row per card when they no longer fit side by side), and the GPU process list says which card each process is holding VRAM on.

⁠🔌 API

/stats takes an optional ?limit= for the per-process rankings, capped by SYS_STATS_TOP_PROCESSES_MAX.

/panel carries no process lists and no Ollama data, just CPU, RAM, swap, GPU and sensor numbers, plus an age in seconds telling the display how stale the sample is. It shares the same sampling pass as /stats, so enabling it costs no extra nvidia-smi call. Its schema is frozen on purpose: a wall display flashed once should not need reflashing when the dashboard gains a field.

Both routes read a cached snapshot. No request ever collects anything itself.

⁠🛠️ Development

uv venv && source .venv/bin/activate
uv pip install -e '.[dev]'

No test touches the real machine: psutil, GPUtil, subprocess.run and requests.get are all monkeypatched at the sys_stats.collectors boundary.

⁠🏗️ Architecture

flowchart TB
    subgraph host["Host machine"]
        PS["psutil<br/>CPU, RAM, procs, hwmon"]
        NV["nvidia-smi + GPUtil"]
        IP["ipmitool<br/>/dev/ipmi0"]
    end

    subgraph proc["sys-stats-server"]
        SA["Background sampler<br/>one pass every 2s"]
        CA[("Cached snapshot")]
        FL["Flask app"]
    end

    OL["Ollama /api/ps"]
    DC["dcgm-exporter /metrics"]

    PS --> SA
    NV --> SA
    IP --> SA
    OL --> SA
    DC --> SA
    SA --> CA
    CA --> FL

    FL -->|"/"| WEB["Web dashboard"]
    FL -->|"/stats"| CLI["sys-stats TUI"]
    FL -->|"/panel"| ESP["ESP32-S3 wall display"]

Every collector degrades to empty data rather than raising, so a missing GPU, a dead Ollama or a host with no BMC costs you a blank panel, never a 500.

⁠🧑‍💻 Contributing

Issues, feature suggestions and pull requests are all welcome.

This repo is clearly messy, but it was supposed to be only for my own use!

⁠📝 License

MIT. See LICENSE⁠.

Made by Grégoire Compagnon (obeone)