In a Spark app (Spark 2.1) I'm trying to send a case class as input parameter of a function that is meant to run on executors
object TestJob extends App {
val appName = "TestJob"
val out = "out"
val p = Params("my-driver-string")
val spark = SparkSession.builder()
.appName(appName)
.getOrCreate()
import spark.implicits._
(1 to 100).toDF.as[Int].flatMap(i => Dummy.process(i, p))
.write
.option("header", "true")
.csv(out)
}
object Dummy {
def process(i: Int, v:Params): Vector[String] = {
Vector { if( i % 2 == 1) v + "_odd" else v + "_even" }
}
}
case class Params(v: String)
When I run it with master local[*] everything goes well, while when running in a cluster, Params class state is not getting serialized and the output results in
null_even
null_odd
...
Could you please help me understanding what I'm doing wrong?
Googling around I found this post that gave me the solution:Spark broadcasted variable returns NullPointerException when run in Amazon EMR cluster
In the end the problem is due to the extend Apps
Related
I work at a place where scalding writes are augmented with a specific API to track dataset meta data. When converting from normal writes to these special writes, there are some intricacies with respect to Key/Value, TSV/CSV, Thrift ... datasets. I would like to compare the binary file is the same prior to conversion and after conversion to the special API.
Given I cannot provide the specific api for the metadata-inclusive writes, I only ask how can I write a unit test for .write method on a TypedPipe?
implicit val timeZone: TimeZone = DateOps.UTC
implicit val dateParser: DateParser = DateParser.default
implicit def flowDef: FlowDef = new FlowDef()
implicit def mode: Mode = Local(true)
val fileStrPath = root + "/test"
println("writing data to " + fileStrPath)
TypedPipe
.from(Seq[Long](1, 2, 3, 4, 5))
// .map((x: Long) => { println(x.toString); System.out.flush(); x })
.write(TypedTsv[Long](fileStrPath))
.forceToDisk
The above doesn't seem to write anything to local (OSX) disk.
So I wonder if I need to use a MiniDFSCluster something like this:
def setUpTempFolder: String = {
val tempFolder = new TemporaryFolder
tempFolder.create()
tempFolder.getRoot.getAbsolutePath
}
val root: String = setUpTempFolder
println(s"root = $root")
val tempDir = Files.createTempDirectory(setUpTempFolder).toFile
val hdfsCluster: MiniDFSCluster = {
val configuration = new Configuration()
configuration.set(MiniDFSCluster.HDFS_MINIDFS_BASEDIR, tempDir.getAbsolutePath)
configuration.set("io.compression.codecs", classOf[LzopCodec].getName)
new MiniDFSCluster.Builder(configuration)
.manageNameDfsDirs(true)
.manageDataDfsDirs(true)
.format(true)
.build()
}
hdfsCluster.waitClusterUp()
val fs: DistributedFileSystem = hdfsCluster.getFileSystem
val rootPath = new Path(root)
fs.mkdirs(rootPath)
However, my attempts to get this MiniCluster to work haven't panned out either - somehow I need to link the MiniCluster with the Scalding write.
Note: The Scalding JobTest framework for unit testing isn't going to work due actual data written is sometimes wrapped in bijection codec or setup with case class wrappers prior to the writes made by the metadata-inclusive writes APIs.
Any ideas how I can write a local file (without using the Scalding REPL) with either Scalding alone or a MiniCluster? (If using the later, I need a hint how to read the file.)
Answering ... There is an example of how to use a mini cluster for exactly reading and writing to HDFS. I will be able to cross read with my different writes and examine them. Here it is in the tests for scalding's TypedParquet type
HadoopPlatformJobTest is an extension for JobTest that uses a MiniCluster.
With some hand-waiving on detail in the link, the bulk of the code is this:
"TypedParquetTuple" should {
"read and write correctly" in {
import com.twitter.scalding.parquet.tuple.TestValues._
def toMap[T](i: Iterable[T]): Map[T, Int] = i.groupBy(identity).mapValues(_.size)
HadoopPlatformJobTest(new WriteToTypedParquetTupleJob(_), cluster)
.arg("output", "output1")
.sink[SampleClassB](TypedParquet[SampleClassB](Seq("output1"))) {
toMap(_) shouldBe toMap(values)
}
.run()
HadoopPlatformJobTest(new ReadWithFilterPredicateJob(_), cluster)
.arg("input", "output1")
.arg("output", "output2")
.sink[Boolean]("output2")(toMap(_) shouldBe toMap(values.filter(_.string == "B1").map(_.a.bool)))
.run()
}
}
Following is the code that am experimenting with. Am trying to convert SalesData in csv to DF and then to LabeledPoints. However in the last step am getting following compilation error
package macros contains object and package with same name: blackbox
Can you please give me pointers on what am doing wrong here ? Thank you
--EDIT--
Compilation Issue solved by adding 2.11 mllib to build.gradle . but mlData.show fails with
ERROR: java.lang.ClassCastException: java.lang.String cannot be cast to org.apache.spark.ml.linalg.Vector
val path = "SalesData.csv"
val conf = new SparkConf().setMaster("local[2]").set("deploy-mode", "client").set("spark.driver.bindAddress", "127.0.0.1")
.set("spark.broadcast.compress", "false")
.setAppName("local-spark-kafka-consumer-client")
val sparkSession = SparkSession
.builder()
.config(conf)
.getOrCreate()
val data = sparkSession.read.format("csv").option("header", "true").option("inferSchema", "true").load(path)
data.cache()
import org.apache.spark.sql.DataFrameNaFunctions
data.na.drop()
data.show
//get monthly sales totals
val summary = data.select("OrderMonthYear","SaleAmount").groupBy("OrderMonthYear").sum().orderBy("OrderMonthYear").toDF("OrderMonthYear","SaleAmount")
summary.show
// convert ordermonthyear to integer type
//val results = summary.map(df => (df.getAs[String]("OrderMonthYear").replace("-", "") , df.getAs[String]("SaleAmount"))).toDF(["OrderMonthYear","SaleAmount"])
import org.apache.spark.sql.functions._
val test = summary.withColumn("OrderMonthYear", (regexp_replace(col("OrderMonthYear").cast("String"),"-",""))).toDF("OrderMonthYear","SaleAmount")
test.printSchema()
test.show
import sparkSession.implicits._
val mlData = test.select("OrderMonthYear", "SaleAmount").
map(row => org.apache.spark.ml.feature.LabeledPoint(
row.getAs[Double](1),
row.getAs[org.apache.spark.ml.linalg.Vector](0))).toDF
mlData.show
I am playing with Server Sent Events to get updates from akka-http v2.4.11 based micro-service. I am using akka-sse. For some reason, I am not receiving any updates on my Javascript front-end. However, as soon as, I terminate or kill the server process, I get some of the messages in the front-end. My code looks like this:
val start = ByteString.empty
val sep = ByteString("\n")
val end = ByteString.empty
import Fill._
implicit val jsonStreamingSupport: JsonEntityStreamingSupport =
EntityStreamingSupport.json()
.withFramingRenderer(Flow[ByteString].intersperse(start,
sep,
end))
import de.heikoseeberger.akkasse.EventStreamMarshalling._
def routes: Route = pathPrefix("subscribe") {
path("fills") {
get {
complete {
Source.actorPublisher[Fill](FillProvider())
.map(fill ⇒ sse(fill))
.keepAlive(1.second, () ⇒ ServerSentEvent.heartbeat)
}
}
}
}
def sse[T: ClassTag](obj: T)(implicit w: JsonWriter[T]): ServerSentEvent = {
ServerSentEvent(data = w.write(obj).compactPrint,
eventType = classTag[T].runtimeClass.getSimpleName)
}
Any pointers what I can be doing wrong? To me, it seems that I am following every instructions as mentioned here
I am writing some self contained integration tests around Apache Spark Streaming.
I want to test that my code can ingest all kinds of edge cases in my simulated test data.
When I was doing this with regular RDDs (not streaming). I could use my inline data and call "parallelize" on it to turn it into a spark RDD.
However, I can find no such method for creating destreams. Ideally I would like to call some "push" function once in a while and have the tupple magically appear in my dstream.
ATM I'm doing this by using Apache Kafka: I create a temp queue, and I write to it. But this seems like overkill. I'd much rather create the test-dstream directly from my test data without having to use Kafka as a mediator.
For testing purpose, you can create an input stream from a queue of RDDs.
Pushing more RDDs in the queue will simulate having processed more events in the batch interval.
val sc = SparkContextHolder.sc
val ssc = new StreamingContext(sc, Seconds(1))
val inputData: mutable.Queue[RDD[Int]] = mutable.Queue()
val inputStream: InputDStream[Int] = ssc.queueStream(inputData)
inputData += sc.makeRDD(List(1, 2)) // Emulate the RDD created during the first batch interval
inputData += sc.makeRDD(List(3, 4)) // 2nd batch interval
// etc
val result = inputStream.map(x => x*x)
result.foreachRDD(rdd => assertSomething(rdd))
ssc.start() // Don't forget to start the streaming context
In addition to Raphael solution I think you like to also either can process one batch a time or everything available approach. You need to set oneAtATime flag accordingly on queustream's optional method argument as shown below:
val slideDuration = Milliseconds(100)
val conf = new SparkConf().setAppName("NetworkWordCount").setMaster("local[8]")
val sparkSession: SparkSession = SparkSession.builder.config(conf).getOrCreate()
val sparkContext: SparkContext = sparkSession.sparkContext
val queueOfRDDs = mutable.Queue[RDD[String]]()
val streamingContext: StreamingContext = new StreamingContext(sparkContext, slideDuration)
val rddOneQueuesAtATimeDS: DStream[String] = streamingContext.queueStream(queueOfRDDs, oneAtATime = true)
val rddFloodOfQueuesDS: DStream[String] = streamingContext.queueStream(queueOfRDDs, oneAtATime = false)
rddOneQueuesAtATimeDS.print(120)
rddFloodOfQueuesDS.print(120)
streamingContext.start()
for (i <- (1 to 10)) {
queueOfRDDs += sparkContext.makeRDD(simplePurchase(i))
queueOfRDDs += sparkContext.makeRDD(simplePurchase((i + 3) * (i + 3)))
Thread.sleep(slideDuration.milliseconds)
}
Thread.sleep(1000L)
I found this base example:
https://github.com/apache/spark/blob/master/examples/src/main/scala/org/apache/spark/examples/streaming/CustomReceiver.scala
The key here is calling the "store" command. Replace the contents of store with whatever you want.
I am trying to create a Pig UDF that extracts the locations mentioned in a tweet using the Stanford CoreNLP package interfaced through the sista Scala API. It works fine when run locally with 'sbt run', but throws a "java.lang.NoSuchMethodError" exception when called from Pig:
Loading default properties from tagger
edu/stanford/nlp/models/pos-tagger/english-left3words/english-left3words-distsim.tagger
Reading POS tagger model from
edu/stanford/nlp/models/pos-tagger/english-left3words/english-left3words-distsim.tagger
Loading classifier from edu/stanford/nlp/models/ner/english.all.3class.distsim.crf.ser.gz
2013-06-14 10:47:54,952 [communication thread] INFO
org.apache.hadoop.mapred.LocalJobRunner - reduce > reduce done [7.5
sec]. Loading classifier from
edu/stanford/nlp/models/ner/english.muc.7class.distsim.crf.ser.gz ...
2013-06-14 10:48:02,108 [Low Memory Detector] INFO
org.apache.pig.impl.util.SpillableMemoryManager - first memory handler
call - Collection threshold init = 18546688(18112K) used =
358671232(350264K) committed = 366542848(357952K) max =
699072512(682688K) done [5.0 sec]. Loading classifier from
edu/stanford/nlp/models/ner/english.conll.4class.distsim.crf.ser.gz
... 2013-06-14 10:48:10,522 [Low Memory Detector] INFO
org.apache.pig.impl.util.SpillableMemoryManager - first memory handler
call- Usage threshold init = 18546688(18112K) used =
590012928(576184K) committed = 597786624(583776K) max =
699072512(682688K) done [5.6 sec]. 2013-06-14 10:48:11,469 [Thread-11]
WARN org.apache.hadoop.mapred.LocalJobRunner - job_local_0001
java.lang.NoSuchMethodError:
org.joda.time.Duration.compareTo(Lorg/joda/time/ReadableDuration;)I
at edu.stanford.nlp.time.SUTime$Duration.compareTo(SUTime.java:3406)
at edu.stanford.nlp.time.SUTime$Duration.max(SUTime.java:3488) at
edu.stanford.nlp.time.SUTime$Time.difference(SUTime.java:1308) at
edu.stanford.nlp.time.SUTime$Range.(SUTime.java:3793) at
edu.stanford.nlp.time.SUTime.(SUTime.java:570)
Here is the relevant code:
object CountryTokenizer {
def tokenize(text: String): String = {
val locations = TweetEntityExtractor.NERLocationFilter(text)
println(locations)
locations.map(x => Cities.country(x)).flatten.mkString(" ")
}
}
class PigCountryTokenizer extends EvalFunc[String] {
override def exec(tuple: Tuple): java.lang.String = {
val text: java.lang.String = Util.cast[java.lang.String](tuple.get(0))
CountryTokenizer.tokenize(text)
}
}
object TweetEntityExtractor {
val processor:Processor = new CoreNLPProcessor()
def NERLocationFilter(text: String): List[String] = {
val doc = processor.mkDocument(text)
processor.tagPartsOfSpeech(doc)
processor.lemmatize(doc)
processor.recognizeNamedEntities(doc)
val locations = doc.sentences.map(sentence => {
val entities = sentence.entities.map(List.fromArray(_)) match {
case Some(l) => l
case _ => List()
}
val words = List.fromArray(sentence.words)
(words zip entities).filter(x => {
x._1 != "" && x._2 == "LOCATION"
}).map(_._1)
})
List.fromArray(locations).flatten
}
}
I am using sbt-assembly to construct a fat-jar, and so the joda-time jar file should be accessible. What is going on?
Pig ships with its own version of joda-time (1.6), which is incompatible with 2.x.