Compatibility with Apache Spark

Supported Algorithms

The following table shows the currently supported algorithms. The goal is to expand this over time with support from the underlying RAPIDS cuML libraries. If you would like support for a specific algorithm, please file a git issue to help us prioritize.

Supported Algorithms	Python	Scala
CrossValidator	√
KMeans	√
k-NN (*)	√
LinearRegression	√
LogisticRegression	√
PCA	√	√
RandomForestClassifier	√
RandomForestRegressor	√
UMAP (*)	√

Note: Spark does not provide a k-Nearest Neighbors (k-NN) implementation, but it does have an LSH-based Approximate Nearest Neighbor implementation. As an alternative to PCA, we also provide a Spark API for GPU accelerated Uniform Manifold Approximation and Projection (UMAP), a non-linear dimensionality reduction algorithm in the RAPIDS cuML library.

Supported Versions

Spark Rapids ML	CUDA	Spark	Python
1.0.0	11.5+	3.2.1+	3.9+

Single vs Double precision inputs

The underlying cuML implementations all accept single precision (e.g. Float or float32) input types and offer the best performance in this case. As a result, by default, Spark RAPIDs ML converts Spark DataFrames supplied to fit and transform methods having double precision data types (i.e. VectorUDT, ArrayType(DoubleType()), DoubleType() columns) to single precision before passing them down to the cuML layer. Most of the cuML algorithm implementations also support double precision inputs. The Estimator (for all algorithms) constructor parameter float32_inputs can be used to control this behavior. The default value is True which forces the conversion to single precision for all algorithms, but it can be set to False in which case double precision input data is passed to those cuML algorithms which support it.

Currently all algorithms except the following support double precision: LogisticRegression, k-NN, UMAP.