An operator that is neither fitted nor applied: its work is one method on the server's own helper object.

Two of them on Spark 4.2.0 — PowerIterationClustering and PrefixSpan — and they are operators in every way but the one that matters to the wire. They carry a JVM class and a param table like an estimator does, and PySpark models them as classes; but there is no Fit, no Transform and nothing cached, because what they do is a Fetch on ConnectHelper with their params as positional arguments.

pic = Latu.ML.Clustering.power_iteration_clustering(k: 2, max_iter: 10)
clusters = Latu.ML.assign_clusters(pic, edges)

That positional call is the whole difference from every other operator here, and it has one consequence worth knowing: an MlParams map carries only what the caller set and the server fills the rest, where a helper's arguments are all sent every time — so this package supplies the defaults from the registry, and refuses rather than guessing where a param has none that can be written down.

No uid: nothing on the wire carries one, because nothing on the wire is an operator.