Veðrfölnir

Self-hosted build/workflow completion ETA for queued and concurrent jobs. Historical durations, observed runner capacity and priority/FIFO queue replay provide a baseline; a River model learns from completed builds. Includes an HTTP API and a status/accuracy dashboard.

Experimental: production ETA accuracy is not proven. Predictions target a terminal result, including failures. CPython 3.15 only; the initial container target is Linux/amd64, with demo and normalized CSV sources.

Run from source

With Docker and Compose v2, from a repository checkout:

cp .env.example .env
# Set a unique, long random ETA_ADMIN_TOKEN in .env
docker compose up -d --build
curl http://localhost:8080/api/status

Open the dashboard or API docs. The default demo uses synthetic data. Compose builds locally and persists SQLite state; use one host, one app instance and one worker per database. Do not mix demo and real histories in one data volume.

For a pip installation from the checkout (CPython 3.15):

python -m pip install .
# Configure .env with a unique ETA_ADMIN_TOKEN
vedrfolnir doctor
vedrfolnir serve --host 127.0.0.1 --port 8080

PyPI publication and prebuilt ghcr.io/hrashton/vedrfolnir images are pending; these are source-install commands. Consumer install/pull commands will be added after the published artifacts are qualified. See the release contract and GitHub Releases for support rules and version-specific publication notes.

Security: read API/UI endpoints are unauthenticated. Compose binds to 127.0.0.1; remote access requires an authenticating TLS reverse proxy. The admin token protects sync only. Back up SQLite before upgrades and never load untrusted model checkpoints. See operations.

Connect and interpret your data

Follow the quick start for demo, CSV or optional SQL Server source builds; use SELECT-only SQL accounts. The planned official image omits the MSSQL driver, and SQL Server integration is unqualified until separately tested. Source contracts define required columns and visibility assumptions.

Future arrivals, unreported dependencies and unknown scheduler policies are not inferred. River correction is selected only after sustained measured improvement. Empirical P90 bounds need 40 comparable completed builds and are not guarantees. See forecasting and measured results and limitations.

Develop and explore

python -m pip install -e ".[dev]"
bash scripts/local_check.sh quick

The development guide covers hosted CI and clean package checks. Public benchmarks remain opt-in. The documentation index is the detailed navigation map; roadmap separates distribution gates from accuracy evidence.