Promotion Pipeline

Independent uv project that compares a candidate model against the current champion and promotes it if the candidate has a lower cv_mae_mean.

On successful promotion, publishes an event to Redis pub/sub so downstream consumers (e.g. the Ray Serve Predictor) can reload the model.

Running

# Resolve candidate from MLflow "candidate" alias:
uv run --project promotion python -m promotion --mlflow-uri http://localhost:5001

# With explicit version + Redis notification:
uv run --project promotion python -m promotion \
    --version 5 \
    --mlflow-uri http://localhost:5001 \
    --redis-url redis://localhost:6379

CLI options

Flag	Default	Description
--version	(none)	Explicit model version (otherwise resolves from --candidate-alias)
--candidate-alias	candidate	MLflow alias to resolve candidate from
--mlflow-uri	http://localhost:5001	MLflow tracking server URI
--model-name	vroom-forecast	Registered model name
--metric	cv_mae_mean	Metric to compare (lower is better)
--redis-url	(none)	Redis URL for promotion notification

What it does

1. Resolves the candidate version (by explicit --version or candidate alias)
2. Fetches the candidate's cv_mae_mean metric from MLflow
3. Fetches the current champion's cv_mae_mean (if one exists)
4. If candidate < champion: promotes candidate to champion alias
5. If no champion exists: promotes candidate as first champion
6. On promotion: publishes event to Redis channel vroom-forecast:model-promoted

Key files

• promote.py — Promotion logic, Redis notification, CLI arg parsing
• __main__.py — CLI entry point (exits 0 on success; exits 1 on exception)
• pyproject.toml — Dependencies: mlflow, redis