Thank you for putting this outstanding question out there. For some reason, when it comes to Spark, everyone gets so caught up in the analytics that they forget about the great software engineering practices that emerged the last 15 years or so. This is why we make it a point to discuss testing and continuous integration (among other things like DevOps) in our course.
A Quick Aside on Terminology
A true unit test means you have complete control over every component in the test. There can be no interaction with databases, REST calls, file systems, or even the system clock; everything has to be "doubled" (e.g. mocked, stubbed, etc) as Gerard Mezaros puts it in xUnit Test Patterns. I know this seems like semantics, but it really matters. Failing to understand this is one major reason why you see intermittent test failures in continuous integration.
We Can Still Unit Test
So given this understanding, unit testing an RDD is impossible. However, there is still a place for unit testing when developing analytics.
Consider a simple operation:
rdd.map(foo).map(bar)
Here foo and bar are simple functions. Those can be unit tested in the normal way, and they should be with as many corner cases as you can muster. After all, why do they care where they are getting their inputs from whether it is a test fixture or an RDD?
Don't Forget the Spark Shell
This isn't testing per se, but in these early stages you also should be experimenting in the Spark shell to figure out your transformations and especially the consequences of your approach. For example, you can examine physical and logical query plans, partitioning strategy and preservation, and the state of your data with many different functions like toDebugString, explain, glom, show, printSchema, and so on. I will let you explore those.
You can also set your master to local[2] in the Spark shell and in your tests to identify any problems that may only arise once you start to distribute work.
Integration Testing with Spark
Now for the fun stuff.
In order to integration test Spark after you feel confident in the quality of your helper functions and RDD/DataFrame transformation logic, it is critical to do a few things (regardless of build tool and test framework):
- Increase JVM memory.
- Enable forking but disable parallel execution.
- Use your test framework to accumulate your Spark integration tests into suites, and initialize the
SparkContext before all tests and stop it after all tests.
With ScalaTest, you can mix in BeforeAndAfterAll (which I prefer generally) or BeforeAndAfterEachas @ShankarKoirala does to initialize and tear down Spark artifacts. I know this is a reasonable place to make an exception, but I really don't like those mutable vars you have to use though.
The Loan Pattern
Another approach is to use the Loan Pattern.
For example (using ScalaTest):
class MySpec extends WordSpec with Matchers with SparkContextSetup {
"My analytics" should {
"calculate the right thing" in withSparkContext { (sparkContext) =>
val data = Seq(...)
val rdd = sparkContext.parallelize(data)
val total = rdd.map(...).filter(...).map(...).reduce(_ + _)
total shouldBe 1000
}
}
}
trait SparkContextSetup {
def withSparkContext(testMethod: (SparkContext) => Any) {
val conf = new SparkConf()
.setMaster("local")
.setAppName("Spark test")
val sparkContext = new SparkContext(conf)
try {
testMethod(sparkContext)
}
finally sparkContext.stop()
}
}
As you can see, the Loan Pattern makes use of higher-order functions to "loan" the SparkContext to the test and then to dispose of it after it's done.
Suffering-Oriented Programming (Thanks, Nathan)
It is totally a matter of preference, but I prefer to use the Loan Pattern and wire things up myself as long as I can before bringing in another framework. Aside from just trying to stay lightweight, frameworks sometimes add a lot of "magic" that makes debugging test failures hard to reason about. So I take a Suffering-Oriented Programming approach--where I avoid adding a new framework until the pain of not having it is too much to bear. But again, this is up to you.
The best choice for that alternate framework is of course spark-testing-base as @ShankarKoirala mentioned. In that case, the test above would look like this:
class MySpec extends WordSpec with Matchers with SharedSparkContext {
"My analytics" should {
"calculate the right thing" in {
val data = Seq(...)
val rdd = sc.parallelize(data)
val total = rdd.map(...).filter(...).map(...).reduce(_ + _)
total shouldBe 1000
}
}
}
Note how I didn't have to do anything to deal with the SparkContext. SharedSparkContext gave me all that--with sc as the SparkContext--for free. Personally though I wouldn't bring in this dependency for just this purpose since the Loan Pattern does exactly what I need for that. Also, with so much unpredictability that happens with distributed systems, it can be a real pain to have to trace through the magic that happens in the source code of a third-party library when things go wrong in continuous integration.
Now where spark-testing-base really shines is with the Hadoop-based helpers like HDFSClusterLike and YARNClusterLike. Mixing those traits in can really save you a lot of setup pain. Another place where it shines is with the Scalacheck-like properties and generators--assuming of course you understand how property-based testing works and why it is useful. But again, I would personally hold off on using it until my analytics and my tests reach that level of sophistication.
"Only a Sith deals in absolutes." -- Obi-Wan Kenobi
Of course, you don't have to choose one or the other either. Perhaps you could use the Loan Pattern approach for most of your tests and spark-testing-base only for a few, more rigorous tests. The choice isn't binary; you can do both.
Integration Testing with Spark Streaming
Finally, I would just like to present a snippet of what a SparkStreaming integration test setup with in-memory values might look like without spark-testing-base:
val sparkContext: SparkContext = ...
val data: Seq[(String, String)] = Seq(("a", "1"), ("b", "2"), ("c", "3"))
val rdd: RDD[(String, String)] = sparkContext.parallelize(data)
val strings: mutable.Queue[RDD[(String, String)]] = mutable.Queue.empty[RDD[(String, String)]]
val streamingContext = new StreamingContext(sparkContext, Seconds(1))
val dStream: InputDStream = streamingContext.queueStream(strings)
strings += rdd
This is simpler than it looks. It really just turns a sequence of data into a queue to feed to the DStream. Most of it is really just boilerplate setup that works with the Spark APIs. Regardless, you can compare this with StreamingSuiteBase as found in spark-testing-base to decide which you prefer.
This might be my longest post ever, so I will leave it here. I hope others chime in with other ideas to help improve the quality of our analytics with the same agile software engineering practices that have improved all other application development.
And with apologies for the shameless plug, you can check out our course Software Engineering with Apache Spark, where we address a lot of these ideas and more. We hope to have an online version soon.
