[SPARK-21859][CORE] Fix SparkFiles.get failed on driver in yarn-cluster and yarn-client mode

[lgrcyanny]

What changes were proposed in this pull request?

when use SparkFiles.get a file on driver in yarn-client or yarn-cluster, it will report file not found exception.
This exception only happens on driver, SparkFiles.get on executor works fine.
we can reproduce the bug as follows:

val conf = new SparkConf().setAppName("SparkFilesTest")
val sc = new SparkContext(conf)
def testOnDriver(fileName: String) = {
    val file = new File(SparkFiles.get(fileName))
    if (!file.exists()) {
        println(s"$file not exist")
    } else {
        // print file content on driver
        val content = Source.fromFile(file).getLines().mkString("\n")
        println(s"File content: ${content}")
    }
}
// the output will be file not exist

conf = SparkConf().setAppName("test files")
sc = SparkContext(appName="spark files test")
def test_on_driver(filename):
    file = SparkFiles.get(filename)
    print("file path: {}".format(file))
    if os.path.exists(file):
        with open(file) as f:
        lines = f.readlines()
        print(lines)
    else:
        print("file doesn't exist")
        run_command("ls .")

the output will be file not exist

How was this patch tested?

tested in integration tests and manual tests
submit the demo case in yarn-cluster and yarn-client mode, and verify the test result
the testing commands are:

./bin/spark-submit --master yarn-cluster --files README.md --class "testing.SparkFilesTest" testing.jar
./bin/spark-submit --master yarn-client --files README.md --class "testing.SparkFilesTest" testing.jar
./bin/spark-submit --master yarn-cluster --files README.md test_get_files.py
./bin/spark-submit --master yarn-client --files README.md test_get_files.py

[AmplabJenkins]

Can one of the admins verify this patch?

[brad-kaiser]

Tiny nitpick, this might be simpler.

  if (isDriver && conf.get("spark.submit.deployMode", "client") == "client") { 

[lgrcyanny]

Ok, thanks, I will change it.

[lgrcyanny]

Originally, my version is

conf.get("spark.submit.deployMode", "client") == "client"

Then I refered to the SparkContext#deployMode function,
it use

conf.getOption("spark.submit.deployMode").getOrElse("client")

I just want to keep the same style as SparkContext.
which one is better?

[jerryshao]

Currently for Spark yarn-client application, we don't support fetching files using above SparkFiles.get API. Since you already know where the file is in client mode, so may be you don't need to call SparkFiles.get.

Here actually requires several changes including supporting remote files.

[jerryshao]

The changes here is not correct. For yarn application, we use spark.yarn.dist.files to handle files, and this will be added to distributed cache. With your change it breaks the current code.

[lgrcyanny]

I met some users complained about the wired action about SparkFiles.get in yarn-client and yarn-cluster mode. SparkFiles.get is very easy for user to get file path. why not keep the same action in yarn-cluster and yarn-client mode?
Meanwhile, it not very easy for user to use spark.yarn.dist.files, it must be uploaded to HDFS in advance. To make spark on yarn more usable, use SparkFiles.get is better.

[jerryshao]

I don't say SparkFiles.get is not useful, I'm saying your fix is not correct, the changes here will break the original semantics. Also we recently support remote files, to handle this scenario we should think how to address this problem for all the cluster managers, not only in yarn client mode.

[jerryshao]

Also as @vanzin mentioned, the description of "Yarn cluster mode SparkFiles.get is working" is not a design purpose. So to fix this issue I think you should have a more solid patch.

[lgrcyanny]

hi @jerryshao About the remote files to handle yarn-client files problem, is there jira that explains the design? We can wait for a version which resolved the problem.

I think my fix just solve the problem simply, do you have any other idea to solve it more elegantly?

[lgrcyanny]

May I ask, why the OptionAssigner for "spark.files" works for local, standalone and mesos, only except yarn? is there any doc explain the design purpose? or may be this is really a issue.

OptionAssigner(args.files, LOCAL | STANDALONE | MESOS, ALL_DEPLOY_MODES, "spark.files")

[vanzin]

Please file PRs against the master branch.

[vanzin]

(Unless the issue does not exist in the master branch, in which case please call that out explicitly.)

[lgrcyanny]

Hi @vanzin I have submit a PR based on master branch, please review it, thank you
#19102

[srowen]

Close this @lgrcyanny

[jerryshao]

Please close this PR @lgrcyanny thanks!