在Spark Worker类路径中自定义JAR的最佳方法

Question

I am working on an ETL pipeline in spark and I find that pushing a release is time/bandwidth intensive. My release script (pseudocode):

sbt assembly
openstack object create spark target/scala-2.11/etl-$VERSION-super.jar
spark-submit \
    --class comapplications.WindowsETLElastic \
    --master spark://spark-submit.cloud \
    --deploy-mode cluster \
    --verbose \
    --conf "spark.executor.memory=16g" \
    "$JAR_URL"

which works but can take over 4 minutes to assemble and a minute to push. My build.sbt:

name := "secmon_etl"

version := "1.2"

scalaVersion := "2.11.8"

exportJars := true

assemblyJarName in assembly := s"${name.value}-${version.value}-super.jar"

libraryDependencies ++= Seq (
  "org.apache.spark" %% "spark-core" % "2.1.0" % "provided",
  "org.apache.spark" %% "spark-streaming" % "2.1.0" % "provided",
  "org.apache.spark" %% "spark-streaming-kafka-0-10" % "2.1.0",
  "io.spray" %%  "spray-json" % "1.3.3",
//  "commons-net" % "commons-net" % "3.5",
//  "org.apache.httpcomponents" % "httpclient" % "4.5.2",
  "org.elasticsearch" % "elasticsearch-spark-20_2.11" % "5.3.1"
)

assemblyMergeStrategy in assembly <<= (assemblyMergeStrategy in assembly) {
  (old) => {
    case PathList("META-INF", xs @ _*) => MergeStrategy.discard
    case x => MergeStrategy.first
  }
}

The issue appears to be the sheer size of the elasticsearch-spark-20_2.11. It adds about 90MB to my uberjar. I would be happy to turn it in to a provided dependency on the spark host, making it unnecessary to package. The question is, what's the best way to do that? Should I just manually copy over jars or is there a foolproof way of specifying a dependency and having a tool resolve all the transitive dependencies?

Answer 1

I have my spark jobs running and much more quickly now. I ran

sbt assemblyPackageDependency

which generated a huge jar (110MB!), easily placed in the spark working directory 'jars' folder, so now my Dockerfile for a spark cluster looks like this:

FROM openjdk:8-jre

ENV SPARK_VERSION 2.1.0
ENV HADOOP_VERSION hadoop2.7
ENV SPARK_MASTER_OPTS="-Djava.net.preferIPv4Stack=true"

RUN apt-get update && apt-get install -y python

RUN curl -sSLO http://mirrors.ocf.berkeley.edu/apache/spark/spark-$SPARK_VERSION/spark-$SPARK_VERSION-bin-$HADOOP_VERSION.tgz && tar xzfC /spark-$SPARK_VERSION-bin-$HADOOP_VERSION.tgz /usr/share && rm /spark-$SPARK_VERSION-bin-$HADOOP_VERSION.tgz

# master or worker's webui port, 
EXPOSE 8080
# master's rest api port
EXPOSE 7077

ADD deps.jar /usr/share/spark-$SPARK_VERSION-bin-$HADOOP_VERSION/jars/

WORKDIR /usr/share/spark-$SPARK_VERSION-bin-$HADOOP_VERSION

after deploying that configuration I changed my build.sbt so the kafka-streaming / elasticsearch-spark jars and dependencies are marked as provided :

name := "secmon_etl"

version := "1.2"

scalaVersion := "2.11.8"

exportJars := true

assemblyJarName in assembly := s"${name.value}-${version.value}-super.jar"

libraryDependencies ++= Seq (
  "org.apache.spark" %% "spark-core" % "2.1.0" % "provided",
  "org.apache.spark" %% "spark-streaming" % "2.1.0" % "provided",

  "org.apache.spark" %% "spark-streaming-kafka-0-10" % "2.1.0" % "provided",
  "io.spray" %%  "spray-json" % "1.3.3" % "provided",
  "org.elasticsearch" % "elasticsearch-spark-20_2.11" % "5.3.1" % "provided"
)

assemblyMergeStrategy in assembly <<= (assemblyMergeStrategy in assembly) {
  (old) => {
    case PathList("META-INF", xs @ _*) => MergeStrategy.discard
    case x => MergeStrategy.first
  }
}

Now my deploys go through in 20 seconds!