A model's training summary: a second server-side object, and the one the cache really drops.

A model is offloaded to disk under memory pressure and comes back; a summary is dropped, and the fetch that finds it gone answers CONNECT_ML.MODEL_SUMMARY_LOST. That is not a failure the caller caused and not one they can prevent, so this package recovers from it the way PySpark does — it sends CreateSummary with the frame the model was fitted on and asks again. dataset is that frame, and carrying it is the whole reason this struct exists rather than a bare reference.

The reference is the model's own. Spark has no separate cache key for a summary: it is reached by asking the model for one, so a Fetch names the model and puts summary at the head of its method chain. PySpark spells that as the string "<model_ref>.summary" and splits it apart again on the way out; here the parts are just kept apart.

summary = Latu.ML.summary(model)
{:ok, auc} = Latu.ML.Classification.BinaryLogisticRegressionSummary.area_under_roc(summary)
roc = Latu.ML.Classification.BinaryLogisticRegressionSummary.roc(summary)

class is nil for LogisticRegressionModel and RandomForestClassificationModel, whose summary class depends on how many classes the fitted model found — _summaryCls branches on numClasses, and no client knows the answer before it asks. candidates is what PySpark can return, both of them, in its own order; class is set only where that list has one entry, so a class check is made where one can be and skipped where it cannot.

The candidates are what an argument's type is looked up in, and the two never disagree: 4.2.0's binary summary is a subclass of its multiclass one, so every attribute they share has the same signature and the binary one only adds. Latu.ML.attributes/1 asks a narrower question — what is allowed whichever class this turns out to be — and so reports what they share; name Latu.ML.Classification.BinaryLogisticRegressionTrainingSummary to see the rest.