Advanced Analytics with Spark

Appendix B. Upcoming MLlib Pipelines API

Sean Owen

The Spark project moves fast. When we started writing in August 2014, version 1.1.0 was nearing release. As this book goes to print in April 2015, Spark 1.2.1 is hot off the presses. In this version alone, almost 1,000 improvements and fixes were added.

The project carefully maintains binary and source compatibility for stable APIs in minor releases, and most of MLlib is considered stable. The examples in the book should therefore continue to work with Spark 1.3.0 and future 1.x releases; those implementations won’t be going anywhere. However, new releases often add or change experimental or developer-only APIs, which are still evolving.

Spark MLlib has, of course, featured prominently in these chapters, and a book covering Spark 1.2.1 would not be complete without mentioning a significant new direction for MLlib that appears, in part, as an experimental API: the “Pipelines” API.

It’s officially only a month or so old, subject to change, and not nearly complete, and so it has not been possible to build the book around it. However, it’s worth knowing about, having already seen what MLlib offers today.

This appendix will give a quick look at the new Pipelines API, the result of work discussed in SPARK-3530 in the Spark project issue tracker.

Beyond Mere Modeling

In purpose and scope, the current MLlib resembles other machine learning libraries. It provides an implementation of machine learning algorithms, and just the core implementation. ...