MLflow and Unity Catalog

workspace:
  registry: mlflow://databricks-uc/ml.pecca

Each trained version is an MLflow run (params, metrics, artefacts) plus a pyfunc model registered in Unity Catalog as <catalog>.<schema>.<workspace>_<project>_<call>. Call state is the state.json artefact of one "state" run per call; decision logs are small JSONL artefacts (a Delta-table sink is planned). Experiments are named /Shared/pecca/<workspace>/<project>/<call> on Databricks.

Install pip install 'pecca[databricks]' and authenticate the usual way (DATABRICKS_HOST, DATABRICKS_TOKEN). For a local trial use mlflow://sqlite/./mlflow.db. Pecca keeps only its own metadata; everything else lives in your registry.