How the production trainer maps an approved plan's experiment to a concrete
MJ: ML Training Pipelines row. The orchestrator carries algorithm × feature
set × hyperparameters on the experiment, but TrainingEngine.trainModel
trains by pipeline id (an immutable, versioned pipeline definition). This
resolver bridges the two: given an experiment + plan, it returns the pipeline
id to train (materializing/looking up a pipeline that encodes that
experiment's algorithm + feature set + hyperparameters).
It is itself a seam so the materialization strategy (reuse an existing
pipeline vs. create one per experiment) can vary without touching the
orchestrator or the training engine.
How the production trainer maps an approved plan's experiment to a concrete
MJ: ML Training Pipelinesrow. The orchestrator carries algorithm × feature set × hyperparameters on the experiment, butTrainingEngine.trainModeltrains by pipeline id (an immutable, versioned pipeline definition). This resolver bridges the two: given an experiment + plan, it returns the pipeline id to train (materializing/looking up a pipeline that encodes that experiment's algorithm + feature set + hyperparameters).It is itself a seam so the materialization strategy (reuse an existing pipeline vs. create one per experiment) can vary without touching the orchestrator or the training engine.