Machine Learning Library (MLlib) Guide

MLlib is Spark’s machine learning (ML) library. Its goal is to make practical machine learning scalable and easy. At a high level, it provides tools such as:

• ML Algorithms: common learning algorithms such as classification, regression, clustering, and collaborative filtering
• Featurization: feature extraction, transformation, dimensionality reduction, and selection
• Pipelines: tools for constructing, evaluating, and tuning ML Pipelines
• Persistence: saving and load algorithms, models, and Pipelines
• Utilities: linear algebra, statistics, data handling, etc.

Announcement: DataFrame-based API is primary API

The MLlib RDD-based API is now in maintenance mode.

As of Spark 2.0, the RDD-based APIs in the spark.mllib package have entered maintenance mode. The primary Machine Learning API for Spark is now the DataFrame-based API in the spark.ml package.

What are the implications?

• MLlib will still support the RDD-based API in spark.mllib with bug fixes.
• MLlib will not add new features to the RDD-based API.
• In the Spark 2.x releases, MLlib will add features to the DataFrames-based API to reach feature parity with the RDD-based API.
• After reaching feature parity (roughly estimated for Spark 2.3), the RDD-based API will be deprecated.
• The RDD-based API is expected to be removed in Spark 3.0.

Why is MLlib switching to the DataFrame-based API?

• DataFrames provide a more user-friendly API than RDDs. The many benefits of DataFrames include Spark Datasources, SQL/DataFrame queries, Tungsten and Catalyst optimizations, and uniform APIs across languages.
• The DataFrame-based API for MLlib provides a uniform API across ML algorithms and across multiple languages.
• DataFrames facilitate practical ML Pipelines, particularly feature transformations. See the Pipelines guide for details.

What is “Spark ML”?

• “Spark ML” is not an official name but occasionally used to refer to the MLlib DataFrame-based API. This is majorly due to the org.apache.spark.ml Scala package name used by the DataFrame-based API, and the “Spark ML Pipelines” term we used initially to emphasize the pipeline concept.

Is MLlib deprecated?

• No. MLlib includes both the RDD-based API and the DataFrame-based API. The RDD-based API is now in maintenance mode. But neither API is deprecated, nor MLlib as a whole.

Dependencies

MLlib uses the linear algebra package Breeze, which depends on netlib-java for optimised numerical processing. If native libraries¹ are not available at runtime, you will see a warning message and a pure JVM implementation will be used instead.

Due to licensing issues with runtime proprietary binaries, we do not include netlib-java’s native proxies by default. To configure netlib-java / Breeze to use system optimised binaries, include com.github.fommil.netlib:all:1.1.2 (or build Spark with -Pnetlib-lgpl) as a dependency of your project and read the netlib-java documentation for your platform’s additional installation instructions.

The most popular native BLAS such as Intel MKL, OpenBLAS, can use multiple threads in a single operation, which can conflict with Spark’s execution model.

Configuring these BLAS implementations to use a single thread for operations may actually improve performance (see SPARK-21305). It is usually optimal to match this to the number of cores each Spark task is configured to use, which is 1 by default and typically left at 1.

Please refer to resources like the following to understand how to configure the number of threads these BLAS implementations use: Intel MKL and OpenBLAS.

To use MLlib in Python, you will need NumPy version 1.4 or newer.

Highlights in 2.3

The list below highlights some of the new features and enhancements added to MLlib in the 2.3 release of Spark:

• Built-in support for reading images into a DataFrame was added (SPARK-21866).
• OneHotEncoderEstimator was added, and should be used instead of the existing OneHotEncoder transformer. The new estimator supports transforming multiple columns.
• Multiple column support was also added to QuantileDiscretizer and Bucketizer (SPARK-22397 and SPARK-20542)
• A new FeatureHasher transformer was added (SPARK-13969).
• Added support for evaluating multiple models in parallel when performing cross-validation using TrainValidationSplit or CrossValidator (SPARK-19357).
• Improved support for custom pipeline components in Python (see SPARK-21633 and SPARK-21542).
• DataFrame functions for descriptive summary statistics over vector columns (SPARK-19634).
• Robust linear regression with Huber loss (SPARK-3181).

Migration guide

MLlib is under active development. The APIs marked Experimental/DeveloperApi may change in future releases, and the migration guide below will explain all changes between releases.

From 2.2 to 2.3

Breaking changes

• The class and trait hierarchy for logistic regression model summaries was changed to be cleaner and better accommodate the addition of the multi-class summary. This is a breaking change for user code that casts a LogisticRegressionTrainingSummary to a BinaryLogisticRegressionTrainingSummary. Users should instead use the model.binarySummary method. See SPARK-17139 for more detail (note this is an Experimental API). This does not affect the Python summary method, which will still work correctly for both multinomial and binary cases.

Deprecations and changes of behavior

Deprecations

• OneHotEncoder has been deprecated and will be removed in 3.0. It has been replaced by the new OneHotEncoderEstimator (see SPARK-13030). Note that OneHotEncoderEstimator will be renamed to OneHotEncoder in 3.0 (but OneHotEncoderEstimator will be kept as an alias).

Changes of behavior

• SPARK-21027: The default parallelism used in OneVsRest is now set to 1 (i.e. serial). In 2.2 and earlier versions, the level of parallelism was set to the default threadpool size in Scala.
• SPARK-22156: The learning rate update for Word2Vec was incorrect when numIterations was set greater than 1. This will cause training results to be different between 2.3 and earlier versions.
• SPARK-21681: Fixed an edge case bug in multinomial logistic regression that resulted in incorrect coefficients when some features had zero variance.
• SPARK-16957: Tree algorithms now use mid-points for split values. This may change results from model training.
• SPARK-14657: Fixed an issue where the features generated by RFormula without an intercept were inconsistent with the output in R. This may change results from model training in this scenario.

Previous Spark versions

Earlier migration guides are archived on this page.

---