在集群部署模式下找不到 Spark 文件

Question

I'm trying to run a Spark job in cluster deploy mode by issuing in the EMR cluster master node:

spark-submit --master yarn \
--deploy-mode cluster \
--files truststore.jks,kafka.properties,program.properties \ 
--class com.someOrg.somePackage.someClass s3://someBucket/someJar.jar kafka.properties program.properties

I'm getting the following error, which states that the file can not be found at the Spark executor working directory:

//This is me printing the Spark executor working directory through SparkFiles.getRootDirectory()
20/07/03 17:53:40 INFO Program$: This is the path: /mnt1/yarn/usercache/hadoop/appcache/application_1593796195404_0011/spark-46b7fe4d-ba16-452a-a5a7-fbbab740bf1e/userFiles-9c6d4cae-2261-43e8-8046-e49683f9fd3e
        
//This is me trying to list the content for that working directory, which turns out empty.
20/07/03 17:53:40 INFO Program$: This is the content for the path:
                
//This is me getting the error:
    20/07/03 17:53:40 ERROR ApplicationMaster: User class threw exception: java.nio.file.NoSuchFileException: /mnt1/yarn/usercache/hadoop/appcache/application_1593796195404_0011/spark-46b7fe4d-ba16-452a-a5a7-fbbab740bf1e/userFiles-9c6d4cae-2261-43e8-8046-e49683f9fd3e/program.properties
                java.nio.file.NoSuchFileException: /mnt1/yarn/usercache/hadoop/appcache/application_1593796195404_0011/spark-46b7fe4d-ba16-452a-a5a7-fbbab740bf1e/userFiles-9c6d4cae-2261-43e8-8046-e49683f9fd3e/program.properties
                    at sun.nio.fs.UnixException.translateToIOException(UnixException.java:86)
                    at sun.nio.fs.UnixException.rethrowAsIOException(UnixException.java:102)
                    at sun.nio.fs.UnixException.rethrowAsIOException(UnixException.java:107)
                    at sun.nio.fs.UnixFileSystemProvider.newByteChannel(UnixFileSystemProvider.java:214)
                    at java.nio.file.Files.newByteChannel(Files.java:361)
                    at java.nio.file.Files.newByteChannel(Files.java:407)
                    at java.nio.file.spi.FileSystemProvider.newInputStream(FileSystemProvider.java:384)
                    at java.nio.file.Files.newInputStream(Files.java:152)
                    at ccom.someOrg.somePackage.someHelpers$.loadPropertiesFromFile(Helpers.scala:142)
                    at com.someOrg.somePackage.someClass$.main(someClass.scala:33)
                    at com.someOrg.somePackage.someClass.main(someClass.scala)
                    at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
                    at sun.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:62)
                    at sun.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
                    at java.lang.reflect.Method.invoke(Method.java:498)
                    at org.apache.spark.deploy.yarn.ApplicationMaster$$anon$2.run(ApplicationMaster.scala:685)

This is the function I use to attempt to read the properties files passed as arguments:

def loadPropertiesFromFile(path: String): Properties = {
    val inputStream = Files.newInputStream(Paths.get(path), StandardOpenOption.READ)
    val properties  = new Properties()
    properties.load(inputStream)
    properties
  }

Invoked as:

val spark = SparkSession.builder().getOrCreate()
import spark.implicits._
val kafkaProperties = loadPropertiesFromFile(SparkFiles.get(args(1)))
val programProperties = loadPropertiesFromFile(SparkFiles.get(args(2)))
//Also tried loadPropertiesFromFile(args({1,2}))

The program works as expected when issued with client deploy mode:

spark-submit --master yarn \
--deploy-mode client \
--packages org.apache.spark:spark-sql-kafka-0-10_2.11:2.4.5 \
--files truststore.jks program.jar com.someOrg.somePackage.someClass kafka.properties program.properties

This happens in Spark 2.4.5 / EMR 5.30.1.

Additionally, when I try to configure this job as an EMR step it does not even work in client mode. Any clue on how are the resource files passed through --files option managed/persisted/available in EMR?

Answer 1

Option 1: Put those files in s3 and pass the s3 path. Option 2: copy those files to each node in a specific location(using bootstrap) and pass the absolute path of files.

Answer 2

Solved with suggestions from the above comments:

spark-submit --master yarn \
--deploy-mode cluster \
--packages org.apache.spark:spark-sql-kafka-0-10_2.11:2.4.5 \
--files s3://someBucket/resources/truststore.jks,s3://someBucket/resources/kafka.properties,s3://someBucket/resources/program.properties \
--class com.someOrg.someClass.someMain \
s3://someBucket/resources/program.jar kafka.properties program.properties

I was previously assuming that in cluster deploy mode the files under --files were also shipped alongside the driver deployed to a worker node (and thereby available in the working directory), if accessible from the machine where spark-submit is issued.

Bottom line: Regardless of where you issue spark-submit from and the availability of the files in that machine, in cluster mode, you must ensure that files are accessible from every worker node.

It is now working by pointing files location to S3.

Thank you all!