Scylla and Spark integration

Simple Scylla-Spark integration example

This is an example how to create a very simple Spark application that uses Scylla to store its data. The application is going to read people’s names and ages from one table and write the names of the adults to another one. It also will show the number of adults and all people in database.


  • Scylla
  • sbt

Prepare Scylla

Firstly, we need to create keyspace and tables in which data processed by the example application will be stored.

Launch Scylla and connect to it using cqlsh. The following commands will create a new keyspace for our tests and make it the current one.

CREATE KEYSPACE spark_example WITH replication = {'class': 'SimpleStrategy', 'replication_factor': 1};
USE spark_example;

Then, tables both for input and output data need to be created:


Lastly, the database needs to contain some input data for our application to process:

INSERT INTO persons (name, age) VALUES ('Anne', 34);
INSERT INTO persons (name, age) VALUES ('John', 47);
INSERT INTO persons (name, age) VALUES ('Elisabeth', 89);
INSERT INTO persons (name, age) VALUES ('George', 52);
INSERT INTO persons (name, age) VALUES ('Amy', 17);
INSERT INTO persons (name, age) VALUES ('Jack', 16);
INSERT INTO persons (name, age) VALUES ('Treebeard', 36421);

Prepare the application

With database containing all the necessary tables and data it is now time to write our example application. Create directory scylla-spark-example, which will contain all source code and build configuration.

First very important file is build.sbt which should be created in the project main directory. It contains all the application metadata including name, version and dependencies.

name := "scylla-spark-example-simple"
version := "1.0"
scalaVersion := "2.10.5"

libraryDependencies ++= Seq(
        "com.datastax.spark" %% "spark-cassandra-connector" % "1.5.0-M1",
        "org.apache.spark" %% "spark-catalyst" % "1.5.0" % "provided"

Then, we need to enable sbt-assembly plugin. Create directory project and create file plugins.sbt with the following content:

addSbtPlugin("com.eed3si9n" % "sbt-assembly" % "0.14.0")

The steps above should cover all build configuration, what is left is the actual logic of the application. Create file src/main/scala/ScyllaSparkExampleSimple.scala:

import org.apache.spark.{SparkContext,SparkConf}
import com.datastax.spark.connector._

object ScyllaSparkExampleSimple {
    def main(args: Array[String]): Unit = {
        val sc = new SparkContext(new SparkConf())

        val persons = sc.cassandraTable("spark_example", "persons")

        val adults = persons.filter(_.getInt("age") >= 18).map(n => Tuple1(n.getString("name")))
        adults.saveToCassandra("spark_example", "adults")

        val out = s"Adults: %d\nTotal: %d\n".format(adults.count(), persons.count())

Since we don’t want to hardcode in our application any information about Scylla or Spark we will also need an additional configuration file spark-scylla.conf.

spark.master local

Now it is time to build the application and create self-containing jar file that we will be able to send to Spark. To do that execute command:

sbt assembly

It will download all necessary dependencies, build our example and create an output jar file in target/scala-2.10/scylla-spark-example-simple-assembly-1.0.jar.

Download and run Spark

The next step is to get Spark running. Pre-built binaries can be downloaded from this website. Make sure to choose release 1.5.0. Since we are going to use it with Scylla Hadoop version doesn’t matter.

Once the download has finished unpack the archive and in its root directory execute the following command to start Spark Master:

./sbin/ -h localhost

Spark Web UI should be now available at http://localhost:8080. The Spark URL used to connect its workers is spark://localhost:7077.

With the master running the only thing left to have minimal Spark deployment is to start a worker. This can be done with the following command:

./sbin/ spark://localhost:7077

Run application

The application is built, Spark is up and Scylla has all the necessary tables created and contains the input data for our example. This means that we are ready to run the application. Make sure that Scylla is running and execute (still in the Spark directory) the following command):

./bin/spark-submit --properties-file /path/to/scylla-spark-example/spark-scylla.conf \
    --class ScyllaSparkExampleSimple /path/to/scylla-spark-example/target/scala-2.10/scylla-spark-example-simple-assembly-1.0.jar

spark-submit will output some logs and debug information but among them there should be message from the application:

Adults: 5
Total: 7

You can also connect to Scylla with cqlsh and using the following query see the results of our example in the database.

SELECT * FROM spark_example.adults;

Expected output:


Based on and

RoadTrip example

These is a short guide explaining how to run a Spark example application available here with Scylla.


  • Scylla
  • Maven
  • Git

Get the source code

You can get the source code of this example by cloning the following repository:

Disable Cassandra

spark-tests are configured to launch Cassandra which is not what we want to achieve here. The following patch disables Cassandra. It can be applied, for example, using git apply --ignore-whitespace -.

diff --git a/src/main/java/blog/hashmade/spark/util/ b/src/main/java/blog/hashmade/spark/util/
index 37bbc2e..bfe5517 100644
--- a/src/main/java/blog/hashmade/spark/util/
+++ b/src/main/java/blog/hashmade/spark/util/
@@ -14,7 +14,7 @@ public final class CassandraUtil {

    static Session startCassandra() throws Exception {
-       EmbeddedCassandraServerHelper.startEmbeddedCassandra();
+       //EmbeddedCassandraServerHelper.startEmbeddedCassandra();
        Cluster cluster = new Cluster.Builder().addContactPoints("localhost")

Update connector

spark-tests use Spark Cassandra Connector in version 1.1.0 which is too old for our purposes. Before 1.3.0 the connector used to use Thrift as well CQL and that won’t work with Scylla. Updating the example isn’t very complicated and can be accomplished by applying the following patch:

diff --git a/pom.xml b/pom.xml
index 673e22b..1245ffc 100644
--- a/pom.xml
+++ b/pom.xml
@@ -142,7 +142,7 @@
-           <version>1.1.0</version>
+           <version>1.3.0</version>
@@ -157,7 +157,7 @@
-           <version>1.1.0</version>
+           <version>1.3.0</version>
@@ -173,18 +173,18 @@
-           <version>2.1.2</version>
+           <version></version>
        <!-- Datastax -->
-           <version>1.1.0-beta2</version>
+           <version>1.3.0</version>
-           <version>1.1.0-beta2</version>
+           <version>1.3.0</version>
diff --git a/src/main/java/blog/hashmade/spark/ b/src/main/java/blog/hashmade/spark/
index 1027e42..190eb3d 100644
--- a/src/main/java/blog/hashmade/spark/
+++ b/src/main/java/blog/hashmade/spark/
@@ -43,8 +43,7 @@ public class DatastaxSparkTest {
                .set("spark.executor.memory", "1g")
                .set("", "localhost")
-               .set("spark.cassandra.connection.native.port", "9142")
-               .set("spark.cassandra.connection.rpc.port", "9171");
+               .set("spark.cassandra.connection.port", "9142");
        SparkContext ctx = new SparkContext(conf);
        SparkContextJavaFunctions functions = CassandraJavaUtil.javaFunctions(ctx);
        CassandraJavaRDD<CassandraRow> rdd = functions.cassandraTable("roadtrips", "roadtrip");

Build the example

The example can be built with Maven:

mvn compile

Start Scylla

The application we are trying to run will try to connect with Scylla using custom port 9142. That’s why when starting Scylla an additional flag is needed to make sure that’s the port it listens on (alternatively, you can change all occurrences of 9142 to 9042 in the example source code).

scylla --native-transport-port=9142

Run the application

With the example compiled and Scylla running all that is left to be done is to actually run the application:

mvn exec:java

Scylla limitations

  • Scylla needs Spark Cassandra Connector 1.3.0 or later.
  • Scylla doesn’t populate system.size_estimates and therefore the connector won’t be able to perform automatic split sizing optimally.
  • Counters are unsupported.

For more compatibility information check Scylla status

Knowledge Base