The training process fetches the Iris data from BigQuery, trains a classification model, and then uploads the model to a bucket.
Training runs as a KubeFlow pipeline on KubeFlow. The way the pipeline is built and run is described as an Argo workflow.
Cloud Build -> kfp script -> workflow
Cloud Build takes substitution variables as input to the workflow.yaml.tmpl template file. This includes $ML_MODELS_BUCKET and $ANALYTICS_PROJECT. These are passed through from the Platform Bootstrap process, where they are originally configured.
The pipeline is automatically triggered when code is pushed. It can also be triggered manually via the console.