Uploaded image for project: 'Spark'
  1. Spark
  2. SPARK-26190

SparkLauncher: Allow users to set their own submitter script instead of hardcoded spark-submit

    XMLWordPrintableJSON

Details

    • Improvement
    • Status: Resolved
    • Major
    • Resolution: Won't Fix
    • 2.1.0
    • None
    • None
    • Apache Spark 2.0.1 on yarn cluster (MapR distribution)

    Description

      The improvement request is for improvement in the SparkLauncher class which is responsible to execute builtin spark-submit script using Java API.

      In my use case, there is a custom wrapper script which help in integrating the security features while submitting the spark job using builtin spark-submit.

      Currently the script name is hard-coded in the 'createBuilder()' method of org.apache.spark.launcher.SparkLauncher class:

      // code placeholder
      
      private ProcessBuilder createBuilder() {
          List<String> cmd = new ArrayList();
          String script = CommandBuilderUtils.isWindows() ? "spark-submit.cmd" : "spark-submit";
          cmd.add(CommandBuilderUtils.join(File.separator, new String[]{this.builder.getSparkHome(), "bin", script}));
          cmd.addAll(this.builder.buildSparkSubmitArgs());
      ......
      ......
      }

       

       

      It has following issues, which prevents its usage in certain scenario. 

      1) Developer may not use their own custom scripts with different name. They are forced to use the one shipped with the installation. Overwriting that may not be the option, when it is not allowed to alter the original installation.

      2) The code expect the script to be present at "SPARK_HOME/bin" folder. 

      3) The 'createBuilder()' method is private and hence, extending the 'org.apache.spark.launcher.SparkLauncher' is not an option.

       

      Proposed solution:

      1) Developer should be given an optional parameter to set their own custom script, which may be located at any path.

      2) Only in case the parameter is not set, the default spark-submit script should be taken from SPARK_HOME/bin folder.

      Attachments

        Activity

          People

            Unassigned Unassigned
            gm_dwivedi Gyanendra Dwivedi
            Votes:
            0 Vote for this issue
            Watchers:
            2 Start watching this issue

            Dates

              Created:
              Updated:
              Resolved: