A sequence of stages, fitted as one.

Not a server operator. Spark has no Fit for a pipeline: PySpark loops over the stages in Python and so does this package, which is why Latu.ML.pipeline/1 needs no session and why a pipeline has no cache entry of its own. What Latu.ML.fit/2 hands back is a Latu.ML.PipelineModel holding one entry per estimator it fitted.

assembler = Feature.vector_assembler(input_cols: [:x1, :x2], output_col: :features)
lr = Classification.logistic_regression(max_iter: 10)

{:ok, model} = ML.fit(ML.pipeline([assembler, lr]), training)
scored = ML.transform(model, test)
:ok = ML.delete(model)

A stage is an estimator, a transformer, a model already fitted, or another pipeline. The fold is PySpark's exactly, and its one surprise is worth knowing: stages after the last estimator are carried but never applied during the fit. There is nothing left to learn from them, so the frame stops being transformed at that point.