JarCastingSpark Utils
Motivation
One of the biggest challenges after taking the first steps into the world of writing Apache Spark applications in Scala is taking them to production.
An application of any kind needs to be easy to run and easy to configure.
This project is trying to help developers write Spark applications focusing mainly on the application logic rather than the details of configuring the application and setting up the Spark context.
This project is also trying to create and encourage a friendly yet professional environment for developers to help each other, so please do no be shy and join through gitter, twitter, issue reports or pull requests.
Description
This project contains some basic utilities that can help setting up a Spark application project.
The main point is the simplicity of writing Apache Spark applications just focusing on the logic, while providing for easy configuration and arguments passing.
The code sample bellow shows how easy can be to write a file format converter from any acceptable type, with any acceptable parsing configuration options to any acceptable format.
object FormatConverterExample extends SparkApp[FormatConverterContext, DataFrame] {
override def createContext(config: Config) = FormatConverterContext(config).get
override def run(implicit spark: SparkSession, context: FormatConverterContext): DataFrame = {
val inputData = spark.source(context.input).read
inputData.sink(context.output).write
}
}
Creating the configuration can be as simple as defining a case class to hold the configuration and a factory, that helps extract simple and complex data types like input sources and output sinks.
case class FormatConverterContext(input: FormatAwareDataSourceConfiguration,
output: FormatAwareDataSinkConfiguration)
object FormatConverterContext extends Configurator[FormatConverterContext] {
config.extract[FormatAwareDataSourceConfiguration]("input") |@|
config.extract[FormatAwareDataSinkConfiguration]("output") apply
FormatConverterContext.apply
}
}
Optionally, the SparkFun can be used instead of SparkApp to make the code even more concise.
object FormatConverterExample extends
SparkFun[FormatConverterContext, DataFrame](FormatConverterContext(_).get) {
override def run(implicit spark: SparkSession, context: FormatConverterContext): DataFrame =
spark.source(context.input).read.sink(context.output).write
}
For structured streaming applications the format converter might look like this:
object StreamingFormatConverterExample extends SparkApp[StreamingFormatConverterContext, DataFrame] {
override def createContext(config: Config) = StreamingFormatConverterContext(config).get
override def run(implicit spark: SparkSession, context: StreamingFormatConverterContext): DataFrame = {
val inputData = spark.source(context.input).read
inputData.streamingSink(context.output).write.awaitTermination()
}
}
The streaming configuration the configuration can be as simple as following:
case class StreamingFormatConverterContext(input: FormatAwareStreamingSourceConfiguration,
output: FormatAwareStreamingSinkConfiguration)
object StreamingFormatConverterContext extends Configurator[StreamingFormatConverterContext] {
def validationNel(config: Config): ValidationNel[Throwable, StreamingFormatConverterContext] = {
config.extract[FormatAwareStreamingSourceConfiguration]("input") |@|
config.extract[FormatAwareStreamingSinkConfiguration]("output") apply
StreamingFormatConverterContext.apply
}
}
The SparkRunnable and SparkApp or SparkFun together with the configuration framework provide for easy Spark application creation with configuration that can be managed through configuration files or application parameters.
The IO frameworks for reading and writing data frames add extra convenience for setting up batch and structured streaming jobs that transform various types of files and streams.
Last but not least, there are many utility functions that provide convenience for loading resources, dealing with schemas and so on.
Most of the common features are also implemented as decorators to main Spark classes, like SparkContext, DataFrame and StructType and they are conveniently available by importing the org.tupol.spark.implicits._ package.
Documentation
The documentation for the main utilities and frameworks available:
- SparkApp, SparkFun and SparkRunnable
- DataSource Framework for both batch and structured streaming applications
- DataSink Framework for both batch and structured streaming applications
Latest stable API documentation is available here.
An extensive tutorial and walk-through can be found here. Extensive samples and demos can be found here.
A nice example on how this library can be used can be found in the spark-tools project, through the implementation of a generic format converter and a SQL processor for both batch and structured streams.
Prerequisites
- Java 8 or higher
- Scala 2.11 or 2.12
- Apache Spark 2.4.X
Getting Spark Utils
Spark Utils is published to Maven Central and Spark Packages:
- Group id / organization:
org.tupol - Artifact id / name:
spark-utils - Latest stable version is
0.4.2
Usage with SBT, adding a dependency to the latest version of tools to your sbt build definition file:
libraryDependencies += "org.tupol" %% "spark-utils" % "0.4.2"
Include this package in your Spark Applications using spark-shell or spark-submit
$SPARK_HOME/bin/spark-shell --packages org.tupol:spark-utils_2.11:0.4.1
Starting a New spark-utils Project
The simplest way to start a new spark-utils is to make use of the spark-apps.seed.g8 template project.
To fill in manually the project options run
g8 tupol/spark-apps.seed.g8
The default options look like the following:
name [My Project]:
appname [My First App]:
organization [my.org]:
version [0.0.1-SNAPSHOT]:
package [my.org.my_project]:
classname [MyFirstApp]:
scriptname [my-first-app]:
scalaVersion [2.11.12]:
sparkVersion [2.4.0]:
sparkUtilsVersion [0.4.0]:
To fill in the options in advance
g8 tupol/spark-apps.seed.g8 --name="My Project" --appname="My App" --organization="my.org" --force
What's new?
0.4.2
- The project compiles with both Scala
2.11.12and2.12.12 - Updated Apache Spark to
2.4.6 - Updated the
spark-xmllibrary to0.10.0 - Removed the
com.databricks:spark-avrodependency, as avro support is now built into Apache Spark - Removed the shadow
org.apache.spark.Logginclass, which is replaced by theorg.tupol.spark.Loggignknock-off
For previous versions please consult the release notes.
License
This code is open source software licensed under the MIT License.