We use Apache Spark for various applications at my job, but Spark is still relatively unstable, as evidenced by the project’s 11K+ pull requests. To maintain developer velocity, we regularly patch show stopper bugs in the Spark source. The process is simple.

  1. Install JDK 6, which is required for PySpark (or you’ll get a lengthy warning). Use the oracle-java8-set-default package to switch between Java 6 and 8, or set JAVA_HOME.

    sudo apt-get install oracle-java6-installer
oracle-java8-set-default # Go back to Java 8 when you're done building

    ``

  2. Fork the Apache Spark repo so you can submit a Pull Request later

  3. Clone it locally, checking out your tag of interest

    git clone git@github.com:mc10-inc/spark.git special-spark
cd special-spark
git checkout v1.4.1 # Tag of interest
JAVA_HOME="/usr/lib/jvm/java-6-oracle" # In case you've got 7/8/9 installed
./make-distribution.sh --name al-dente-spark --tgz
# Build time of 5:40.12s on my i7-4790K

    ``

  4. Fire up your custom spark build like any other

    ./dist/bin/spark-shell

  5. Copy the Spark assembly jar to your servers and reboot. Be sure to remove the old artifact, otherwise the ClassLodaer will load both versions and be vexed.

    #Move original assembly to backup location
SPK_PATH=<your spark path>
mv $SPK_PATH/lib/spark-assembly-1.<spark version>-hadoop2.4.0.jar spark-assembly-backup.jar
cp dist/lib/spark-assembly-<spark version>-hadoop2.2.0.jar $SPK_PATH/lib/
./bin/spark-shell # Contact!

Additional Tricks

Scala 2.10 is old hat. Most people develop on Scala 2.11, and 2.12 will be released in 2 months. To run Spark on Scala 2.11, you must build it yourself.

./dev/change-scala-version.sh 2.11
./make-distribution.sh --name al-dente-spark --tgz # Build again

Possible Failures

Possible error message below, if you don’t use Java 6. I use PySpark, so I need that integration. Why Python needs a version of Java EoL’d 3 years ago is beyond me, but then again, Python 3 split from Python 2 eight years ago.

+ echo '***NOTE***: JAVA_HOME is not set to a JDK 6 installation. The resulting'
***NOTE***: JAVA_HOME is not set to a JDK 6 installation. The resulting
+ echo '            distribution may not work well with PySpark and will not run'
            distribution may not work well with PySpark and will not run
+ echo '            with Java 6 (See SPARK-1703 and SPARK-1911).'
            with Java 6 (See SPARK-1703 and SPARK-1911).
+ echo '            This test can be disabled by adding --skip-java-test.'
            This test can be disabled by adding --skip-java-test.
+ echo 'Output from '\''java -version'\'' was:'
Output from 'java -version' was:
+ echo 'java version "1.8.0_66"
Java(TM) SE Runtime Environment (build 1.8.0_66-b17)
Java HotSpot(TM) 64-Bit Server VM (build 25.66-b17, mixed mode)'
java version "1.8.0_66"
Java(TM) SE Runtime Environment (build 1.8.0_66-b17)
Java HotSpot(TM) 64-Bit Server VM (build 25.66-b17, mixed mode)
+ read -p 'Would you like to continue anyways? [y,n]: ' -r