Machine learning
Experiments
Section titled “Experiments”An experiment groups the training runs of one subject: parameters, metrics and artefacts, tracked over time. That is the trace AI compliance discussions ask for, and it is also what lets you answer “why this model?” six months later.
Model registry
Section titled “Model registry”The registry keeps the versions of a model and their provenance: from which experiment, which run, with which artefacts. Without a registry, an experiment has no follow-up.
Compatible with the MLflow client
Section titled “Compatible with the MLflow client”Experiment tracking is compatible with the MLflow client: you point the client at graal, and
your scripts track their runs as before. The facade is served under /mlflow, on the same host
as the console and the API; it implements a subset of MLflow’s REST API 2.0, the one covering
experiment tracking and the registry.
Training
Section titled “Training”A training is written in a Python or Spark job, or in a notebook, on the chosen instance type — GPUs included. The MLflow client tracks the run, and the model enters the registry.
Serving and monitoring
Section titled “Serving and monitoring”A model from the registry is served as a REST endpoint, behind your SSO, and its behaviour is followed over time (drift of inputs and predictions). The site’s Machine learning page presents the full journey, and the Generative AI page the LLM gateway and RAG.