Spark Details: Preparing a Spark Model

Owned by Dave Trier
Last updated: Nov 22, 2022 by Blagica
7 min read

This article provides details for preparing a Spark Model to execute within ModelOp Center.

Table of Contents

 

 

Format of Spark Model

A Spark model for which a MODEL_BATCH_JOB/MODEL_BATCH_TEST_JOB/MODEL_BATCH_TRAINING_JOB will be created needs to comply with the following structure:

from __future__ import print_function from pyspark.sql import SparkSession from pyspark.ml.classification import RandomForestClassificationModel # modelop.score def <my-score-function-name>(external_inputs, external_outputs, external_model_assets): spark = SparkSession.builder.appName("<My-App-Name>").getOrCreate() ... # Read an input file that was uploaded to HDFS by the user df = (spark.read .format("csv") .option('header', 'true') .load(external_inputs[0]['fileUrl'])) # Read a model asset file/folder that was uploaded to HDFS by the user model = RandomForestClassificationModel.load(external_model_assets[0]['fileUrl']) ... # Write to an output file that will be created in HDFS predictions.write.csv(external_outputs[0]['fileUrl'])

 

As defined in the model template, the scoring function takes three parameters of type list. The external_inputs, external_outputs, and modelAssets lists will contain assets of type EXTERNAL_FILE. The external_inputs and external_outputs may contain EXTERNAL_FILE assets representing any FILE assets specified as part of the inputData and/or outputData, respectively. For more information on how FILE assets that are part of the inputData and/or outputData are handled, please refer to PySpark Job Input and Output Data.

It is assumed that these external file assets exist in HDFS before the job is created and the user knows their location.

  1. external_inputs are generated from the inputData list

  2. external_outputs are generated from the outputData list

  3. external_model_assets are generated from the modelAssets list

Each Python list contains entire Asset objects, respectively. The external_model_assets list will be empty if the model has no assets.

Please note that there could be other types of assets in modelAssets, but we are only including the EXTERNAL_FILE assets while excluding the FILE and SOURCE_CODE assets, if any.

Execution Flow in spark-runtime-service

Once the user creates their MODEL_BATCH_JOB/MODEL_BATCH_TEST_JOB/MODEL_BATCH_TRAINING_JOB, the spark-runtime-service will be triggered. When the spark-runtime-service receives the job JSON, it will create the following files in its working directory:

  1. model_source_code.py

    1. Contains the primary source code which is retrieved from the job JSON (model.storedModel.modelAssets) where "assetType" = "SOURCE_CODE" and "primaryModelSource" = true

  2. model_job_metadata.py

    1. Contains the values for external_inputs, external_outputs, and external_model_assets in the following format:

      external_inputs = [ { "name": "test.csv", "assetType": "EXTERNAL_FILE", "fileUrl": "hdfs:///hadoop/test.csv", "filename": "test.csv", "fileFormat": "CSV" }] external_outputs = [ { "name": "titanic_output.csv", "assetType": "EXTERNAL_FILE", "fileUrl": "hdfs:///hadoop/titanic_output.csv", "filename": "titanic_output.csv", "fileFormat": "CSV" }] external_model_assets = [ { "name": "titanic", "assetType": "EXTERNAL_FILE", "fileUrl": "hdfs:///hadoop/titanic", "filename": "titanic" }] method_to_be_executed = "<my-score-function-name>"

  3. ModelOpPySparkDriver.py

    1. Contains the code for Spark's application resource which you can see here. Also, it is from this file that the score function is called while passing in the values for the model’s three parameters

  4. If modelAssets contains assets where "assetType" = "FILE", then the spark-runtime-service will create a file in its working directory for each asset that matches the aforementioned criteria. An example of such file would be a schema file

  5. If modelAssets contains assets where "assetType" = "SOURCE_CODE" and "primaryModelSource" = false, then the spark-runtime-service will create a file in its working directory for each asset that matches the aforementioned criteria.

Once these files (1-5) have been created, the spark-runtime-service will upload them to HDFS when spark-submit is executed. None of these files will be included in external_inputs, external_outputs, or external_model_assets. Since these files are not uploaded by the user directly, and instead when spark-submit is executed, they will be stored at the following location:

hdfs://<cluster-host>:<cluster-port>/user/<username>/.sparkStaging/<spark-application-id>/ModelOpPySparkDriver.py

Therefore, if the user’s model source code uses any of these files, they must be read in pure Python with:

Please note that accessing these files in the model source code is different from how we access the files included in external_inputs,external_outputs, and external_model_assets.

Example modelBatchJob JSON

The following modelBatchJob JSON was used to generate the example values:

Â