satnogs-pipeline
Are the batteries of small satellites OK? This project listens to their radio messages and tells you, on a live page anyone can open.
In plain words
- What it watches. Five small satellites (UWE-4, CatSat, SNUGLITE, Sharjahsat-1 and GRBBeta), each about the size of a shoebox, circling the Earth every 90 minutes.
- Where the data comes from. Every few seconds each one sends a short radio message with its battery's voltage. Volunteers around the world receive them with small antennas and share them through SatNOGS.
- The one idea behind it. A battery should charge in sunlight and drain in the Earth's shadow. So the project works out where each satellite is and whether it is in sunlight, and only raises an alarm about a battery that drains when it should be charging.
- What you can do on the page. Turn the 3D Earth and see where each satellite is right now, in sunlight or in shadow; open a satellite's row for its battery charts; scroll through a short story of what a CubeSat is; and see the whole sky of about 16,800 satellites, with how long each one spends in the dark.
- What it is not. A learning project, not a tool for satellite operators. It refreshes about every hour, and its warning rules have been checked on real data but have not yet caught a real fault.
Why only five? Most of the ~16,800 active satellites (Starlink, weather, GPS and so on) never make their battery data public. Only about 40 small amateur and university satellites do, in a form anyone can decode, and each one has to be checked by hand; see Satellites.
For developers
The rest of this page is the technical side. The pipeline fetches the satellites' radio frames from SatNOGS, decodes them, stores the battery readings in PostgreSQL, and checks them against the lighting worked out with the SGP4 orbit model from orbitscreen and the Sun's position. Alerts go to Telegram. It runs on Docker Compose or Kubernetes, and a GitHub Action publishes the live page every hour.
The check that matters most, on real data: in the Earth's shadow, UWE-4's battery was never charging (0 of 60 beacons), which is what makes the lighting model trustworthy.
What it does
flowchart LR
NET[SatNOGS Network API] --> F[fetch]
DB[SatNOGS DB telemetry API] --> FT[fetch-telemetry]
F --> Q1[(observations queue)]
FT --> Q2[(telemetry queue)]
Q1 --> S[store]
Q2 --> S
S --> PG[(PostgreSQL)]
Q1 -.-> D1[(dead queue)]
Q2 -.-> D2[(dead queue)]
PG --> W[watch]
CT[SatNOGS DB or CelesTrak: orbit data] --> W
F & FT & S & W --> PR[Prometheus]
PR --> AM[Alertmanager] --> N[notify] --> TG[Telegram]
PR --> GR[Grafana]
Results
Measured on real data from 14 to 22 October 2026, unless it says otherwise. The lighting check below is UWE-4's (137 beacons over 5.1 days); the others are newer to the pipeline and have less history.
| |
|---|
| Satellites | 5 satellites, 6 batteries: UWE-4 (2 batteries, voltage, current, temperature, charge), CatSat and Sharjahsat-1 (voltage and current), SNUGLITE and GRBBeta (voltage only). 366 distinct beacons stored by 22 October, before GRBBeta was added on the 23rd. |
| Lighting model against the batteries | In the shadow, 0 of 60 beacons were charging. In sunlight, 61 of 77 were. |
| Battery rules on the same history | Replayed at each of the 137 UWE-4 readings, "discharging in sunlight" and "low voltage" never fired. The "below its own baseline" rule gave 14 false alarms in 138 readings with a short baseline and none with the one it has now, but it was only active for about two days of that history: see Battery watch. |
| Same rule on made-up readings | Three in the shadow at -300 mA (what "current is negative" would flag): no alert. Three in sunlight at -120 mA: alert, delivered to Telegram, resolved when removed. |
Failure tests (scripts/chaos.sh) | 6 breakages, 20 checks, 30 to 45 seconds, on Docker Compose and on Kubernetes: store down, RabbitMQ restart, PostgreSQL down, duplicates, a message that is not JSON. |
Throughput of store | 5,400 observations per second with one replica, 8,000 with two, 8,900 with four (draining a 20,000-message queue, Docker Compose on a laptop). See Performance. |
| Time in the pipeline | Published to stored: median 0.1 to 0.7 ms, 99th percentile under 7 ms, at 100 to 2,000 messages per second with one replica. |
| CI | gofmt, vet, tests with the race detector, go.mod, shell scripts, Prometheus and Alertmanager configuration, Kubernetes manifests (kubeconform), Docker image. Dependabot proposes updates to Go modules, GitHub Actions and Docker images. |
| Kubernetes (kind) | 11 pods running without restarts, 5 Prometheus targets up, 10 alert rules loaded. |
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
