Run spark calculations from Ammonite

ammonite-spark allows to create SparkSessions from Ammonite. It passes some Ammonite internals to a SparkSession, so that spark calculations can be driven from Ammonite, as one would do from a spark-shell.

1. Quick start
2. AmmoniteSparkSession vs SparkSession
   1. Syncing dependencies
3. Using with standalone cluster
4. Using with YARN cluster
5. Troubleshooting
6. Compatibility
7. Missing

Start Ammonite >= 1.6.3, with the --class-based option. The compatibility section lists the compatible versions of Ammonite and ammonite-spark. Start Ammonite by either following the Ammonite instructions on its website, then do

$ amm --class-based

or use coursier,

$ cs launch ammonite:2.1.4 --scala 2.12.11 -- --class-based

Ensure you are using scala 2.12, the only supported Scala version as of writing this.

At the Ammonite prompt, load the Spark 2.x version of your choice, along with ammonite-spark,

@ import $ivy.`org.apache.spark::spark-sql:2.4.0`
@ import $ivy.`sh.almond::ammonite-spark:0.3.0`

(Note the two :: before spark-sql or ammonite-spark, as these are scala dependencies.)

Then create a SparkSession using the builder provided by ammonite-spark

@ import org.apache.spark.sql._

@ val spark = {
    AmmoniteSparkSession.builder()
      .master("local[*]")
      .getOrCreate()
  }

Note the use of AmmoniteSparkSession.builder(), instead of SparkSession.builder() that one would use when e.g. writing a Spark job.

The builder returned by AmmoniteSparkSession.builder() extends the one of SparkSession.builder(), so that one can call .appName("foo"), .config("key", "value"), etc. on it.

See below for how to use it with standalone clusters, and how to use it with YARN clusters.

Note that ammonite-spark does not rely on a Spark distribution. The driver and executors classpaths are handled from the Ammonite session only, via import $ivy.`…` statements. See INTERNALS for more details.

You can then run Spark calculations, like

@ def sc = spark.sparkContext

@ val rdd = sc.parallelize(1 to 100, 10)

@ val n = rdd.map(_ + 1).sum()

If extra dependencies are loaded, via import $ivy.`…` after the SparkSession has been created, one should call AmmoniteSparkSession.sync() for the newly added JARs to be passed to the Spark executors.

Simply set the master to spark://… when building the session, e.g.

@ val spark = {
    AmmoniteSparkSession.builder()
      .master("spark://localhost:7077")
      .config("spark.executor.instances", "4")
      .config("spark.executor.memory", "2g")
      .getOrCreate()
  }

Ensure the version of Spark used to start the master and executors matches the one loaded in the Ammonite session (via e.g. import $ivy.`org.apache.spark::spark-sql:X.Y.Z` ), and that the machine running Ammonite can access / is accessible from all nodes of the standalone cluster.

Set the master to "yarn" when building the session, e.g.

@ val spark = {
    AmmoniteSparkSession.builder()
      .master("yarn")
      .config("spark.executor.instances", "4")
      .config("spark.executor.memory", "2g")
      .getOrCreate()
  }

Ensure the configuration directory of the cluster is set in HADOOP_CONF_DIR or YARN_CONF_DIR in the environment, or is available at /etc/hadoop/conf. This directory should contain files like core-site.xml, hdfs-site.xml, … Ensure also that the machine you run Ammonite on can indeed act as the driver (it should have access to and be accessible from the YARN nodes, etc.).

Before raising issues, ensure you are aware of all that needs to be set up to get a working spark-shell from a Spark distribution, and that all of them are passed in one way or another to the SparkSession created from Ammonite.

Add org.apache.spark.sql.catalyst.encoders.OuterScopes.addOuterScope(this) on the same lines as those where you define case classes involved, like

@ import spark.implicits._
import spark.implicits._

@ org.apache.spark.sql.catalyst.encoders.OuterScopes.addOuterScope(this); case class Foo(id: String, value: Int)
defined class Foo

@  val ds = List(Foo("Alice", 42), Foo("Bob", 43)).toDS
ds: Dataset[Foo] = [id: string, value: int]

(This should likely be added automatically in the future.)

ammonite-spark relies on the API of Ammonite, which undergoes non backward compatible changes from time to time. The following table lists which versions of Ammonite ammonite-spark is built against - so is compatible with for sure.

Local clusters, Mesos, and Kubernetes, aren't supported yet.

No scala 2.10 or 2.11 support (support for those was dropped by Ammonite).