Overview

The spark-submit execution process was designed and architected with the following principles and features in mind:

• When possible, Python and/or PySpark components should be implemented in python. This principle will ensure testing and debugging are easier when running in local or cluster deploy mode

• The spark-runtime-service is monitoring jobs in WAITING and RUNNING state until they reach their COMPLETE, ERROR or CANCELLED state. If any new job changes are detected, model manager will be updated accordingly

• Before a job is submitted to the Spark cluster for execution, the ModelOpPySparkDriver implementation will extract all the (external file assets) input, output and model assets from the ModelBatchJob the inputData, outputData and storedModel.modelAssets from the MODEL_BATCH_JOB/MODEL_BATCH_TEST_JOB/MODEL_BATCH_TRAINING_JOB and pass them in as parameters to the function to be executed. These external file assets are passed in as lists of Assets Assets:

  • externalinput_inputs data – [{“name““filename“: “test.csv“, “assetType”: “EXTERNAL_FILE, “fileUrl”: “hdfs: “///hadoop/test.csv“}]

  • externaloutput_outputs data – [{“name““filename“: “output.csv”, “assetType”: “EXTERNAL_FILE“, “fileUrl”: “hdfs: “///hadoop/output.csv“}]

  • external_model_assets – [{“name“: “asset.zip”, “assetType”: “EXTERNAL_FILE“, “fileUrl”: “/hadoop/asset.zip“}]

• The YarnMonitorService component should be listening for and storing in jobMessage the stdout and stderr produced by a the job that is running at the Spark cluster. Whether to store or not store , stdout and stderr in jobMessages, is configurable.

Architecture Overview

This The following diagram represents how core components interact, and gives a high level overview of the process flow from a ModelBatchJob MODEL_BATCH_JOB/MODEL_BATCH_TEST_JOB/MODEL_BATCH_TRAINING_JOB to a Spark job running at the Spark cluster.

Details

PySpark

Dynamic Job Abstraction

ModelOp Center abstracts the process of executing PySpark jobs using ModelBatchJob as input. The approach uses a main PySpark Driver (ModelOpPySparkDriver.py ) that imports the job metadata (details about job input, output and model assets, as well as the name of the method name to be executedinvoked), and the model source code to be executed, and then just a simple main that will be able to call the desired function.

########################################################################################
##### dynamic model / resources content.
########################################################################################

# primary source code
import model_source_code as modelop_model_source_code
# job metadata, variables pointing to [input, output, appname,etc]
import modelop_job_metadata as modelop_job_metadata
# for printing stack trace
import traceback
# for checking number of params in function signature
from inspect import signature


########################################################################################
##### static resources content.
########################################################################################

if __name__ == "__main__":
    try:
        print("### - ModelOpPySparkExecution execution")

     try:   ## Check for an init function to be #executed; DetermineAn ifinit therefunction is an optional, init function (# modelop.init) and execute it
     optional
        init_method = getattr(modelop_model_source_code, modelop_job_metadata.init_method_to_be_executed, None)
        if init_method is None:
            # Determine how many parameters the method (# modelop.score, # modelop.metrics or # modelop.train)print("### - Optional init '" + modelop_job_metadata.init_method_to_be_executed + "' function was not found in model source code")
        else:
            print("### - Executing function: " + modelop_job_metadata.init_method_to_be_executed)
            init_method()

        ## Check for a scoring or metrics function to be executed takes; A scoring or metrics function is required
        numexecution_paramsmethod = len(signature(getattr(modelop_model_source_code, modelop_job_metadata.method_to_be_executed, None)
        if execution_method is None:
            print("### - '" + modelop_job_metadata.method_to_be_executed + "' function was not found in model source code").parameters
            raise Exception("### - '" + modelop_job_metadata.method_to_be_executed + "' function was not found in model source code")

        ## Executing requested function
        print("### - Executing function: " + modelop_job_metadata.method_to_be_executed)
        num_params = len(signature(execution_method).parameters)
        if num_params == 0:
            getattr(modelop_model_source_code, modelop_job_metadata.method_to_be_executed)execution_method()
        elif num_params == 3:
            getattr(modelop_model_source_code, modelop_job_metadata.method_to_be_executed)execution_method(modelop_job_metadata.job_input, modelop_job_metadata.job_output, modelop_job_metadata.job_model_assets_input)
        else:
            traceback.print_exc()
            raise Exception("Number of function parameters, '" + str(num_params) + "', does not match expected number of parameters (0 or 3)")

    except Exception as e:
        print("### - An exception of ", e.__class__, " occurred")
        traceback.print_exc()
        print("#################### RUNTIME ERROR ####################")
        raise Exception("### - An exception occurred in the ModelOpPySparkDriver")
    print("### - PySpark code execution completed")

PySparkJobManifest

ModelOp Center decouples the transformation process that takes a job MODEL_BATCH_JOB/MODEL_BATCH_TEST_JOB/MODEL_BATCH_TRAINING_JOB as input and produces a SparkLauncher as output. For this purpose, ModelOp Center uses the PySparkJobManifest intermediate component. PySparkJobManifest contains all the required values when running a spark-submit: python Python source code and metadata, function name, application name (all of which are used inside ModelOpPySparkDriver). This design allows for decoupling and keeping us to decouple and to keep agnostic the spark-submit.

The next following diagram displays illustrates the transformation process of a ModelBatchJob MODEL_BATCH_JOB/MODEL_BATCH_TEST_JOB/MODEL_BATCH_TRAINING_JOB into a SparkLauncher:

Image Modified

ModelOpJobMonitor

The ModelOpJobMonitor has two main responsibilities:

1. Submitting jobs to the Spark cluster

2. Monitoring job statuses for existing, waiting and running,

1. Spark jobs.

In order to successfully fulfill these responsibilities, it relies on two key services:

1. SparkLauncherService–component in charge of translating jobs into SparkLauncher objects

2. YarnMonitorService – component in charge of fetching job statuses and output produced from the Spark cluster