Divide and you will conquer Apache Spark. It’s quite common to develop a papyrus script where people try to initialize spark, read paths, execute all the logic and write the result. Even, we found scripts where all the spark transformations are done in a simple method with tones of lines. That means the code is difficult to test, to maintain and to read. Well, that means bad code. We built a set of tools and libraries that allows developers to develop their pipelines by joining all the Pieces. These pieces are compressed by Readers, Writers, Transformers, Aliases, etc. Moreover, it comes with enriched SparkSuites using the Spark-testing-base from Holden Karau. Recently, we start using junit4git (github.com/rpau/junit4git) in our tests, allowing us to execute only the Spark tests that matter by skipping tests that are not affected by latest code changes. This translates into faster builds and fewer coffees. By allowing developers to define each piece on its own, we enable to test small pieces before having the full set of them together. Also, it allows to re-use code in multiple pipelines and speed up their development by improving the quality of the code. The power of “Transform” method combined with Currying, creates a powerful tool that allows fragmenting all the Spark logic. This talk is oriented to developers that are being introduced in the Spark world and how developing iteration by iteration in small steps could help them in producing great code with less effort.

1. Albert Franzi (@FranziCros), Schibsted Media Group
Modular Apache Spark:
Transform Your Code into
Pieces
#SAISDev3
1

2. 2
Slides available in:
bit.ly/SparkSummit-afranzi

3. About me
3
2013 2015 20182014 2017

4. Modular Apache Spark:
Transform Your Code into Pieces
4

5. 5
- How we reduced the test time execution by skipping
unaffected tests -- > less coffees
- How we simplified our spark code by modularizing
- How we increased our test coverage in our spark code by
using the spark-testing-base provided by Holden Karau
We will learn:

6. 6
- How we reduced the test time execution by skipping
unaffected tests -- > less coffees
- How we simplified our spark code by modularizing
- How we increased our test coverage in our spark code by
using the spark-testing-base provided by Holden Karau
We will learn:

7. 7
Did you play with duplicated code cross your Spark Jobs?
@ The Shining 1980
Have you ever experienced the joy of code
reviewing a never ending stream of spark chaos?

8. 8
Please! Don’t play with duplicated code never ever!
@ The Shining 1980

9. Fragment the Spark Job
9
Spark Readers Transformation
JoinsAliases Formatters
Spark WritersTask Context
...

10. Readers / Writers
10
Spark Readers
Spark Writers
● Enforce schemas
● Use schemas to read only the fields you are going to use
● Provide Readers per Dataset & attach its sources to it
● Share schemas & sources between Readers & Writers
● GDPR compliant by design

11. val userBehaviourSchema: StructType = ???
val userBehaviourPath =
Path("s3://<bucket>/user_behaviour/year=2018/month=10/day=03/hour=12/gen=27/")
val userBehaviourReader = ReaderBuilder(PARQUET)
.withSchema(userBehaviourSchema)
.withPath(userBehaviourPath)
.buildReader()
val df: DataFrame = userBehaviourReader.read()
GDPR
Readers
11
Spark Readers

12. Readers
12
Spark Readers
val userBehaviourSchema: StructType = ???
// Path structure - s3://<bucket>/user_behaviour/[year]/[month]/[day]/[hour]/[gen]/
val userBehaviourBasePath = Path("s3://<bucket>/user_behaviour/")
val startDate: ZonedDateTime = ZonedDateTime.now(ZoneOffset.UTC)
val halfDay: Duration = Duration.ofHours(12)
val userBehaviourPaths: Seq[Path] = PathBuilder
.latestGenHourlyPaths(userBehaviourBasePath, startDate, halfDay)
val userBehaviourReader = ReaderBuilder(PARQUET)
.withSchema(userBehaviourSchema)
.withPath(userBehaviourPaths: _*)
.buildReader()
val df: DataFrame = userBehaviourReader.read()

13. Readers
13
val userBehaviourSchema: StructType = ???
val userBehaviourBasePath = Path("s3://<bucket>/user_behaviour/")
val startDate: ZonedDateTime = ZonedDateTime.now(ZoneOffset.UTC)
val halfDay: Duration = Duration.ofHours(12)
val userBehaviourReader = ReaderBuilder(PARQUET)
.withSchema(userBehaviourSchema)
.withHourlyPathBuilder(userBehaviourBasePath, startDate, halfDay)
.buildReader()
val df: DataFrame = userBehaviourReader.read()
Spark Readers

14. Readers
14
val df: DataFrame = UserBehaviourReader.read(startDate, halfDay)
Spark Readers

15. Transforms
15
Transformation
def transform[U](t: (Dataset[T]) ⇒ Dataset[U]): Dataset[U]
Dataset[T] ⇒ magic ⇒ Dataset[U]

16. Transforms
16
Transformation
def withGreeting(df: DataFrame): DataFrame = {
df.withColumn("greeting", lit("hello world"))
}
def extractFromJson(colName: String,
outputColName: String,
jsonSchema: StructType)(df: DataFrame): DataFrame = {
df.withColumn(outputColName, from_json(col(colName), jsonSchema))
}

17. Transforms
17
Transformation
def onlyClassifiedAds(df: DataFrame): DataFrame = {
df.filter(col("event_type") === "View")
.filter(col("object_type") === "ClassifiedAd")
}
def dropDuplicates(df: DataFrame): DataFrame = {
df.dropDuplicates()
}
def cleanedCity(df: DataFrame): DataFrame = {
df.withColumn("city", getCityUdf(col("object.location.address")))
}
val cleanupTransformations: Seq[DataFrame => DataFrame] = Seq(
dropDuplicates,
cleanedCity,
onlyClassifiedAds
)
val df: DataFrame = UserBehaviourReader.read(startDate, halfDay)
val classifiedAdsDF = df.transforms(cleanupTransformations: _*)

18. Transforms
18
Transformation
val cleanupTransformations: Seq[DataFrame => DataFrame] = Seq(
dropDuplicates,
cleanedCity,
onlyClassifiedAds
)
val df: DataFrame = UserBehaviourReader.read(startDate, halfDay)
val classifiedAdsDF = df.transforms(cleanupTransformations: _*)
“As a data consumer, I only need to pick up which transformations I would like to apply,
instead of coding them from scratch.”
“It’s like cooking, engineers provide manufactured ingredients (transformations) and Data
Scientists use the required ones for a successful receipt.”

19. Transforms - Links of Interest
“DataFrame transformations can be defined with arguments so they don’t make
assumptions about the schema of the underlying DataFrame.” - by Matthew Powers.
bit.ly/Spark-ChainingTransformations
bit.ly/Spark-SchemaIndependentTransformations
19
Transformation
github.com/MrPowers/spark-daria
“Spark helper methods to maximize developer productivity.”

20. 20
- How we reduced the test time execution by skipping
unaffected tests -- > less coffees
- How we simplified our spark code by modularizing
- How we increased our test coverage in our spark code by
using the spark-testing-base provided by Holden Karau
We will learn:

21. 21
Did you put untested Spark jobs into production?
“Mars Climate Orbiter destroyed
because of a Metric System Mixup (1999)”

22. Testing with Holden Karau
22
github.com/holdenk/spark-testing-base
“Base classes to use when writing tests with Spark.”
● Share Spark Context between tests from the same suite
● Provide methods to make tests easier
○ Fixture Readers
○ Json to DF converters
○ Extra validators
github.com/MrPowers/spark-fast-tests
“An alternative to spark-testing-base to run tests in parallel without restarting Spark Session
after each test file.”

23. Testing SharedSparkContext
23
package com.holdenkarau.spark.testing
import java.util.Date
import org.apache.spark._
import org.scalatest.{BeforeAndAfterAll, Suite}
/**
* Shares a local `SparkContext` between all tests in a suite
* and closes it at the end. You can share between suites by enabling
* reuseContextIfPossible.
*/
trait SharedSparkContext extends BeforeAndAfterAll with SparkContextProvider {
self: Suite =>
...
protected implicit def reuseContextIfPossible: Boolean = false
...
}

24. Testing
24
package com.schibsted.insights.test
trait SparkSuite extends DataFrameSuiteBase {
self: Suite =>
override def reuseContextIfPossible: Boolean = true
protected def createDF(data: Seq[Row], schema: StructType): DataFrame = {
spark.createDataFrame(spark.sparkContext.parallelize(data), schema)
}
protected def jsonFixtureToDF(fileName: String, schema: Option[StructType] = None): DataFrame = {
val fixtureContent = readFixtureContent(fileName)
val fixtureJson = fixtureContentToJson(fixtureContent)
jsonToDF(fixtureJson, schema)
}
protected def checkSchemas(inputSchema: StructType, expectedSchema: StructType): Unit = {
assert(inputSchema.fields.sortBy(_.name).deep == expectedSchema.fields.sortBy(_.name).deep)
}
...
}

25. Testing
25
Spark Readers Transformation Spark WritersTask Context
Testing each piece independently helps testing all together.
Test

26. 26
- How we reduced the test time execution by skipping
unaffected tests -- > less coffees
- How we simplified our spark code by modularizing
- How we increased our test coverage in our spark code by
using the spark-testing-base provided by Holden Karau
We will learn:

27. 27
Are tests taking too long to execute?

28. Junit4Git by Raquel Pau
28
github.com/rpau/junit4git
“Junit Extensions for Test Impact Analysis.”
“This is a JUnit extension that ignores those tests that are not related with your last
changes in your Git repository.”

29. Junit4Git - Gradle conf
29
configurations {
agent
}
dependencies {
testCompile("org.walkmod:scalatest4git_2.11:${version}")
agent "org.walkmod:junit4git-agent:${version}"
}
test.doFirst {
jvmArgs "-javaagent:${configurations.agent.singleFile}"
}
@RunWith(classOf[ScalaGitRunner])

30. Junit4Git- Travis with Git notes
30
before_install:
- echo -e "machine github.comn login $GITHUB_TOKEN" >> ~/.netrc
after_script:
- git push origin refs/notes/tests:refs/notes/tests
.travis.yml

31. Junit4Git- Runners
31
import org.scalatest.junit.JUnitRunner
@RunWith(classOf[JUnitRunner])
abstract class UnitSpec extends FunSuite
import org.scalatest.junit.ScalaGitRunner
@RunWith(classOf[ScalaGitRunner])
abstract class IntegrationSpec extends FunSuite
Tests to run always
Tests to run on code changes

32. Junit4Git
32
Runner Listener Agent Test Method
testStarted
Starts the report
loadClass
testFinished
addTestClass
Closes the
report
Class
instrumentClass

33. Junit4Git- Notes
33
afranzi:~/Projects/insights-core$ git notes show
[
{
"test": "com.schibsted.insights.core.ContextSuite",
"method": "Context Classes with SptCommonBucketConf should contain a sptCommonBucket Field",
"classes": [
"com.schibsted.insights.core.Context",
"com.schibsted.insights.core.ContextSuite$$anonfun$1$$anon$1",
"com.schibsted.insights.core.Context$",
"com.schibsted.insights.core.SptCommonBucketConf$class",
"com.schibsted.insights.core.Context$$anonfun$getConfigParamWithLogging$1"
]
}
...
]
com.schibsted.insights.core.ContextSuite > Context Classes with SptCommonBucketConf should
contain a sptCommonBucket Field SKIPPED
com.schibsted.insights.core.ContextSuite > Context Classes with InsightsBucketConf should
contain a insightsBucket Field SKIPPED

34. Junit4Git - Links of Interest
34
bit.ly/SlidesJunit4Git
Real Impact Testing Analysis For JVM
bit.ly/TestImpactAnalysis
The Rise of Test Impact Analysis

35. Summary
35
Use Transforms as pills
Modularize your code
Don’t duplicate code between Spark Jobs
Don’t be afraid of testing everything
Share all you built as a library, so others can
reuse your code.

36. 36