This is an implementation of PageRank in Spark, using Spark's standard RDD API.

• Fast iterations
• Parameterised "teleport" probability
• Weighted edges (out-edge weights must be normalized)
• Supports "dangling" vertices (no out-edges from a node)
• Supports arbitrary (e.g. non-uniform) priors (as vertex values)
• Various stopping criteria:
  • Number of iterations threshold
  • Convergence threshold (requires additional computation after each iteration)
• Utilities for building, preparing, and validating input graphs (incl. out-edge normalization)

Include it as a dependency in your sbt project: "com.soundcloud" %% "spark-pagerank" % <version>

You can use PageRank in Spark as a library and call it from within your own drivers. You will want to do this when you have some data preparation to do that does not conform with the built-in driver data interfaces.

More examples of usage as a library can be found in the source of the built-in drivers (see below).

We include several built-in drivers that operate on plain-text TSV input of labeled edges as a starting point. You will prepare the graph and run PageRank in the following sequence of drivers. Use --help to see the arguments and usage of each driver.

1. com.soundcloud.spark.pagerank.GraphBuilderApp: Builds a PageRank graph from (non-normalized) weighted edges in TSV format (source, destination, weight), saving the resulting graph (edges and vertices) in Parquet files in preparation for next steps.
2. com.soundcloud.spark.pagerank.PageRankApp: Runs PageRank on the graph produced using the functions in PageRankGraph or by using the GraphBuilderApp.
3. com.soundcloud.spark.pagerank.ConvergenceCheckApp: Compares two PageRank vectors and lets the user determine if there is convergence by outputting the sum of the component-wise difference of the vectors. Note that this is an optional tool that is mostly used for debugging. If the user is concerned with iterating until convergence, the user can specify the convergence threshold at runtime to PageRank.

We run this library on one of our behavior graphs which consists of approximately 700M vertices and 15B edges. Using the following Spark configuration, and in-memory persistence of edge and vertex RDDs, we obtain iteration times on the order of 3-5 minutes each.

Configuration example:

• Spark 2.1.1
• YARN
• Dynamic allocation: no
• Number of executors: 256
• Number of executor cores: 4
• Executor memory: 28G

• Persist the edges and vertices of the graph in memory and disk (as spill): StorageLevel.MEMORY_AND_DISK
• Enable Kryo serialization: KryoSerialization.useKryo

To publish an artifact to the Sonatype/Maven central repository (a snapshot or release), you need to have a Sonatype account, PGP keys and sbt plugins set up. Please follow the sbt guideline for a complete guide to getting started. Once this is done, you can use the sbt-release plugin via the Makefile to publish snapshots and perform releases.

At any point in the development lifecycle, a snapshot can be published to the central repository.

make publish

Once development of a version is complete, the artifact should be released to the central repository. This is a two stage process with the artifact first entering a staging repository, followed by a manual promotion process.

make release

After a release to the staging repository, the staging-to-release promotion process must be followed manually before the artifact is available in the central repository.

This library aims to adhere to Semantic Versioning 2.0.0. Violations of this scheme should be reported as bugs. Specifically, if a minor or patch version is released that breaks backward compatibility, that version should be immediately removed and/or a new version should be immediately released that restores compatibility. Breaking changes to the public API will only be introduced with new major versions.

We welcome contributions by pull requests and bug reports or feature requests as issues.

• Josh Devins

• Max Jakob

Copyright (c) 2017 SoundCloud Ltd.

See the LICENSE file for details.