Lightning-fast cluster computing

Useful Developer Tools

Reducing Build Times

SBT: Avoiding Re-Creating the Assembly JAR

Spark’s default build strategy is to assemble a jar including all of its dependencies. This can be cumbersome when doing iterative development. When developing locally, it is possible to create an assembly jar including all of Spark’s dependencies and then re-package only Spark itself when making changes.

$ build/sbt clean package
$ ./bin/spark-shell
$ export SPARK_PREPEND_CLASSES=true
$ ./bin/spark-shell # Now it's using compiled classes
# ... do some local development ... #
$ build/sbt compile
# ... do some local development ... #
$ build/sbt compile
$ unset SPARK_PREPEND_CLASSES
$ ./bin/spark-shell
 
# You can also use ~ to let sbt do incremental builds on file changes without running a new sbt session every time
$ build/sbt ~compile

Maven: Speeding up Compilation with Zinc

Zinc is a long-running server version of SBT’s incremental compiler. When run locally as a background process, it speeds up builds of Scala-based projects like Spark. Developers who regularly recompile Spark with Maven will be the most interested in Zinc. The project site gives instructions for building and running zinc; OS X users can install it using brew install zinc.

If using the build/mvn package zinc will automatically be downloaded and leveraged for all builds. This process will auto-start after the first time build/mvn is called and bind to port 3030 unless the ZINC_PORT environment variable is set. The zinc process can subsequently be shut down at any time by running build/zinc-<version>/bin/zinc -shutdown and will automatically restart whenever build/mvn is called.

Building submodules individually

For instance, you can build the Spark Core module using:

$ # sbt
$ build/sbt
> project core
> package

$ # or you can build the spark-core module with sbt directly using:
$ build/sbt core/package

$ # Maven
$ build/mvn package -DskipTests -pl core

Running Individual Tests

When developing locally, it’s often convenient to run a single test or a few tests, rather than running the entire test suite.

Testing with SBT

The fastest way to run individual tests is to use the sbt console. It’s fastest to keep a sbt console open, and use it to re-run tests as necessary. For example, to run all of the tests in a particular project, e.g., core:

$ build/sbt
> project core
> test

You can run a single test suite using the testOnly command. For example, to run the DAGSchedulerSuite:

> testOnly org.apache.spark.scheduler.DAGSchedulerSuite

The testOnly command accepts wildcards; e.g., you can also run the DAGSchedulerSuite with:

> testOnly *DAGSchedulerSuite

Or you could run all of the tests in the scheduler package:

> testOnly org.apache.spark.scheduler.*

If you’d like to run just a single test in the DAGSchedulerSuite, e.g., a test that includes “SPARK-12345” in the name, you run the following command in the sbt console:

> testOnly *DAGSchedulerSuite -- -z "SPARK-12345"

If you’d prefer, you can run all of these commands on the command line (but this will be slower than running tests using an open cosole). To do this, you need to surround testOnly and the following arguments in quotes:

$ build/sbt "core/testOnly *DAGSchedulerSuite -- -z SPARK-12345"

For more about how to run individual tests with sbt, see the sbt documentation.

Testing with Maven

With Maven, you can use the -DwildcardSuites flag to run individual Scala tests:

build/mvn -Dtest=none -DwildcardSuites=org.apache.spark.scheduler.DAGSchedulerSuite test

You need -Dtest=none to avoid running the Java tests. For more information about the ScalaTest Maven Plugin, refer to the ScalaTest documentation.

To run individual Java tests, you can use the -Dtest flag:

build/mvn test -DwildcardSuites=none -Dtest=org.apache.spark.streaming.JavaAPISuite test

ScalaTest Issues

If the following error occurs when running ScalaTest

An internal error occurred during: "Launching XYZSuite.scala".
java.lang.NullPointerException

It is due to an incorrect Scala library in the classpath. To fix it:

  • Right click on project
  • Select Build Path | Configure Build Path
  • Add Library | Scala Library
  • Remove scala-library-2.10.4.jar - lib_managed\jars

In the event of “Could not find resource path for Web UI: org/apache/spark/ui/static”, it’s due to a classpath issue (some classes were probably not compiled). To fix this, it sufficient to run a test from the command line:

build/sbt "test-only org.apache.spark.rdd.SortingSuite"

Running Different Test Permutations on Jenkins

When running tests for a pull request on Jenkins, you can add special phrases to the title of your pull request to change testing behavior. This includes:

  • [test-maven] - signals to test the pull request using maven
  • [test-hadoop2.7] - signals to test using Spark’s Hadoop 2.7 profile

Binary compatibility

To ensure binary compatibility, Spark uses MiMa.

Ensuring binary compatibility

When working on an issue, it’s always a good idea to check that your changes do not introduce binary incompatibilities before opening a pull request.

You can do so by running the following command:

$ dev/mima

A binary incompatibility reported by MiMa might look like the following:

[error] method this(org.apache.spark.sql.Dataset)Unit in class org.apache.spark.SomeClass does not have a correspondent in current version
[error] filter with: ProblemFilters.exclude[DirectMissingMethodProblem]("org.apache.spark.SomeClass.this")

If you open a pull request containing binary incompatibilities anyway, Jenkins will remind you by failing the test build with the following message:

Test build #xx has finished for PR yy at commit ffffff.

  This patch fails MiMa tests.
  This patch merges cleanly.
  This patch adds no public classes.

Solving a binary incompatibility

If you believe that your binary incompatibilies are justified or that MiMa reported false positives (e.g. the reported binary incompatibilities are about a non-user facing API), you can filter them out by adding an exclusion in project/MimaExcludes.scala containing what was suggested by the MiMa report and a comment containing the JIRA number of the issue you’re working on as well as its title.

For the problem described above, we might add the following:

// [SPARK-zz][CORE] Fix an issue
ProblemFilters.exclude[DirectMissingMethodProblem]("org.apache.spark.SomeClass.this")

Otherwise, you will have to resolve those incompatibilies before opening or updating your pull request. Usually, the problems reported by MiMa are self-explanatory and revolve around missing members (methods or fields) that you will have to add back in order to maintain binary compatibility.

Checking Out Pull Requests

Git provides a mechanism for fetching remote pull requests into your own local repository. This is useful when reviewing code or testing patches locally. If you haven’t yet cloned the Spark Git repository, use the following command:

$ git clone https://github.com/apache/spark.git
$ cd spark

To enable this feature you’ll need to configure the git remote repository to fetch pull request data. Do this by modifying the .git/config file inside of your Spark directory. The remote may not be named “origin” if you’ve named it something else:

[remote "origin"]
  url = git@github.com:apache/spark.git
  fetch = +refs/heads/*:refs/remotes/origin/*
  fetch = +refs/pull/*/head:refs/remotes/origin/pr/*   # Add this line

Once you’ve done this you can fetch remote pull requests

# Fetch remote pull requests
$ git fetch origin
# Checkout a remote pull request
$ git checkout origin/pr/112
# Create a local branch from a remote pull request
$ git checkout origin/pr/112 -b new-branch

Generating Dependency Graphs

$ # sbt
$ build/sbt dependency-tree
 
$ # Maven
$ build/mvn -DskipTests install
$ build/mvn dependency:tree

Organizing Imports

You can use a IntelliJ Imports Organizer from Aaron Davidson to help you organize the imports in your code. It can be configured to match the import ordering from the style guide.

IDE Setup

IntelliJ

While many of the Spark developers use SBT or Maven on the command line, the most common IDE we use is IntelliJ IDEA. You can get the community edition for free (Apache committers can get free IntelliJ Ultimate Edition licenses) and install the JetBrains Scala plugin from Preferences > Plugins.

To create a Spark project for IntelliJ:

  • Download IntelliJ and install the Scala plug-in for IntelliJ.
  • Go to File -> Import Project, locate the spark source directory, and select “Maven Project”.
  • In the Import wizard, it’s fine to leave settings at their default. However it is usually useful to enable “Import Maven projects automatically”, since changes to the project structure will automatically update the IntelliJ project.
  • As documented in Building Spark, some build configurations require specific profiles to be enabled. The same profiles that are enabled with -P[profile name] above may be enabled on the Profiles screen in the Import wizard. For example, if developing for Hadoop 2.7 with YARN support, enable profiles yarn and hadoop-2.7. These selections can be changed later by accessing the “Maven Projects” tool window from the View menu, and expanding the Profiles section.

Other tips:

  • “Rebuild Project” can fail the first time the project is compiled, because generate source files are not automatically generated. Try clicking the “Generate Sources and Update Folders For All Projects” button in the “Maven Projects” tool window to manually generate these sources.
  • Some of the modules have pluggable source directories based on Maven profiles (i.e. to support both Scala 2.11 and 2.10 or to allow cross building against different versions of Hive). In some cases IntelliJ’s does not correctly detect use of the maven-build-plugin to add source directories. In these cases, you may need to add source locations explicitly to compile the entire project. If so, open the “Project Settings” and select “Modules”. Based on your selected Maven profiles, you may need to add source folders to the following modules:
    • spark-hive: add v0.13.1/src/main/scala
    • spark-streaming-flume-sink: add target\scala-2.11\src_managed\main\compiled_avro
    • spark-catalyst: add target\scala-2.11\src_managed\main
  • Compilation may fail with an error like “scalac: bad option: -P:/home/jakub/.m2/repository/org/scalamacros/paradise_2.10.4/2.0.1/paradise_2.10.4-2.0.1.jar”. If so, go to Preferences > Build, Execution, Deployment > Scala Compiler and clear the “Additional compiler options” field. It will work then although the option will come back when the project reimports. If you try to build any of the projects using quasiquotes (eg., sql) then you will need to make that jar a compiler plugin (just below “Additional compiler options”). Otherwise you will see errors like:
    /Users/irashid/github/spark/sql/catalyst/src/main/scala/org/apache/spark/sql/catalyst/expressions/codegen/CodeGenerator.scala
    Error:(147, 9) value q is not a member of StringContext
     Note: implicit class Evaluate2 is not applicable here because it comes after the application point and it lacks an explicit result type
          q"""
          ^ 
    

Eclipse

Eclipse can be used to develop and test Spark. The following configuration is known to work:

The easiest way is to download the Scala IDE bundle from the Scala IDE download page. It comes pre-installed with ScalaTest. Alternatively, use the Scala IDE update site or Eclipse Marketplace.

SBT can create Eclipse .project and .classpath files. To create these files for each Spark sub project, use this command:

sbt/sbt eclipse

To import a specific project, e.g. spark-core, select File | Import | Existing Projects into Workspace. Do not select “Copy projects into workspace”.

If you want to develop on Scala 2.10 you need to configure a Scala installation for the exact Scala version that’s used to compile Spark. Since Scala IDE bundles the latest versions (2.10.5 and 2.11.8 at this point), you need to add one in Eclipse Preferences -> Scala -> Installations by pointing to the lib/ directory of your Scala 2.10.5 distribution. Once this is done, select all Spark projects and right-click, choose Scala -> Set Scala Installation and point to the 2.10.5 installation. This should clear all errors about invalid cross-compiled libraries. A clean build should succeed now.

ScalaTest can execute unit tests by right clicking a source file and selecting Run As | Scala Test.

If Java memory errors occur, it might be necessary to increase the settings in eclipse.ini in the Eclipse install directory. Increase the following setting as needed:

--launcher.XXMaxPermSize
256M

Nightly Builds

Packages are built regularly off of Spark’s master branch and release branches. These provide Spark developers access to the bleeding-edge of Spark master or the most recent fixes not yet incorporated into a maintenance release. These should only be used by Spark developers, as they may have bugs and have not undergone the same level of testing as releases. Spark nightly packages are available at:

Spark also publishes SNAPSHOT releases of its Maven artifacts for both master and maintenance branches on a nightly basis. To link to a SNAPSHOT you need to add the ASF snapshot repository to your build. Note that SNAPSHOT artifacts are ephemeral and may change or be removed. To use these you must add the ASF snapshot repository at <a href=”https://repository.apache.org/snapshots/.

groupId: org.apache.spark
artifactId: spark-core_2.10
version: 1.5.0-SNAPSHOT

Profiling Spark Applications Using YourKit

Here are instructions on profiling Spark applications using YourKit Java Profiler.

On Spark EC2 images

  • After logging into the master node, download the YourKit Java Profiler for Linux from the YourKit downloads page. This file is pretty big (~100 MB) and YourKit downloads site is somewhat slow, so you may consider mirroring this file or including it on a custom AMI.
  • Untar this file somewhere (in /root in our case): tar xvjf yjp-12.0.5-linux.tar.bz2
  • Copy the expanded YourKit files to each node using copy-dir: ~/spark-ec2/copy-dir /root/yjp-12.0.5
  • Configure the Spark JVMs to use the YourKit profiling agent by editing ~/spark/conf/spark-env.sh and adding the lines
    SPARK_DAEMON_JAVA_OPTS+=" -agentpath:/root/yjp-12.0.5/bin/linux-x86-64/libyjpagent.so=sampling"
    export SPARK_DAEMON_JAVA_OPTS
    SPARK_JAVA_OPTS+=" -agentpath:/root/yjp-12.0.5/bin/linux-x86-64/libyjpagent.so=sampling"
    export SPARK_JAVA_OPTS
    
  • Copy the updated configuration to each node: ~/spark-ec2/copy-dir ~/spark/conf/spark-env.sh
  • Restart your Spark cluster: ~/spark/bin/stop-all.sh and ~/spark/bin/start-all.sh
  • By default, the YourKit profiler agents use ports 10001-10010. To connect the YourKit desktop application to the remote profiler agents, you’ll have to open these ports in the cluster’s EC2 security groups. To do this, sign into the AWS Management Console. Go to the EC2 section and select Security Groups from the Network & Security section on the left side of the page. Find the security groups corresponding to your cluster; if you launched a cluster named test_cluster, then you will want to modify the settings for the test_cluster-slaves and test_cluster-master security groups. For each group, select it from the list, click the Inbound tab, and create a new Custom TCP Rule opening the port range 10001-10010. Finally, click Apply Rule Changes. Make sure to do this for both security groups. Note: by default, spark-ec2 re-uses security groups: if you stop this cluster and launch another cluster with the same name, your security group settings will be re-used.
  • Launch the YourKit profiler on your desktop.
  • Select “Connect to remote application…” from the welcome screen and enter the the address of your Spark master or worker machine, e.g. ec2--.compute-1.amazonaws.com
  • YourKit should now be connected to the remote profiling agent. It may take a few moments for profiling information to appear.

Please see the full YourKit documentation for the full list of profiler agent startup options.

In Spark unit tests

When running Spark tests through SBT, add javaOptions in Test += "-agentpath:/path/to/yjp" to SparkBuild.scala to launch the tests with the YourKit profiler agent enabled.
The platform-specific paths to the profiler agents are listed in the YourKit documentation.