Extend

Arc can be exended in four ways by registering:

Resolution

Plugins are resolved dynamically at runtime and are resolved by name/version.

Examples

Assuming we wanted to execute a KafkaExtract Pipeline Stage Plugin:

{
  "type": "KafkaExtract",
  "name": "load customers",
  "environments": [
    "production",
    "test"
  ],
  "outputView": "customer",
  "bootstrapServers": "kafka:29092",
  "topic": "customers",
  "groupID": "spark-customer-extract-job"
}

Arc will attempt to resolve the plugin by first looking in all the META-INF directories of all included JAR files (https://github.com/tripl-ai/arc-kafka-pipeline-plugin/blob/master/src/main/resources/META-INF/services/ai.tripl.arc.plugins.PipelineStagePlugin) for classes that extend PipelineStagePlugin which the KafkaExtract plugin does:

class KafkaExtract extends PipelineStagePlugin {

Arc is then able to resolve the plugin by simpleName - in this case KafkaExtract - and then call the instantiate() method to create an instance of the plugin which is executed by Arc at the appropriate time depending on plugin type.

To allow more specitivity you can use either the full package name or include the version or a combination of both:

{
  "type": "ai.tripl.arc.extract.KafkaExtract",
  ...
{
  "type": "KafkaExtract:1.0.0",
  ...
{
  "type": "ai.tripl.arc.extract.KafkaExtract:1.0.0",
  ...

Dynamic Configuration Plugins

Since: 1.3.0

Dynamic vs Deterministic Configuration

Use of this functionality is discouraged as it goes against the principles of Arc specifically around statelessness/deterministic behaviour but is inlcuded here for users who have not yet committed to a job orchestrator such as Apache Airflow and have dynamic configuration requirements.

The Dynamic Configuration Plugin plugin allow users to inject custom configuration parameters which will be processed before resolving the job configuration file. The plugin must return a Typesafe Config object (which is easily created from a java.util.Map[String, Object] which will be included in the job configuration resolution step.

Examples

For example a custom runtime configuration plugin could be used calculate a formatted list of dates to be used with an Extract stage to read only a subset of documents:

package ai.tripl.arc.plugins.config

import java.util
import java.sql.Date
import java.time.LocalDate
import java.time.format.{DateTimeFormatter, DateTimeFormatterBuilder}
import java.time.format.ResolverStyle
import scala.collection.JavaConverters._

import com.typesafe.config._

import org.apache.spark.sql.SparkSession

import ai.tripl.arc.util.log.logger.Logger
import ai.tripl.arc.api.API.ARCContext
import ai.tripl.arc.util.EitherUtils._
import ai.tripl.arc.config.Error._
import ai.tripl.arc.plugins.DynamicConfigurationPlugin

class DeltaPeriodDynamicConfigurationPlugin extends DynamicConfigurationPlugin {

  val version = ai.tripl.arc.plugins.config.deltaperiod.BuildInfo.version

  def instantiate(index: Int, config: Config)(implicit spark: SparkSession, logger: ai.tripl.arc.util.log.logger.Logger, arcContext: ARCContext): Either[List[StageError], Config] = {
    import ai.tripl.arc.config.ConfigReader._
    import ai.tripl.arc.config.ConfigUtils._
    implicit val c = config

    val expectedKeys = "type" :: "environments" :: "returnName" :: "lagDays" :: "leadDays" :: "formatter" :: "currentDate" :: Nil
    val returnName = getValue[String]("returnName")
    val lagDays = getValue[Int]("lagDays")
    val leadDays = getValue[Int]("leadDays")
    val formatter = getValue[String]("formatter") |> parseFormatter("formatter") _
    val currentDate = formatter match {
      case Right(formatter) => {
        if (c.hasPath("currentDate")) getValue[String]("currentDate") |> parseCurrentDate("currentDate", formatter) _ else Right(java.time.LocalDate.now)
      }
      case _ => Right(java.time.LocalDate.now)
    }
    val invalidKeys = checkValidKeys(c)(expectedKeys)

    (returnName, lagDays, leadDays, formatter, currentDate, invalidKeys) match {
      case (Right(returnName), Right(lagDays), Right(leadDays), Right(formatter), Right(currentDate), Right(invalidKeys)) => 

        val res = (lagDays * -1 to leadDays).map { v =>
          formatter.format(currentDate.plusDays(v))
        }.mkString(",")

        val values = new java.util.HashMap[String, Object]()
        values.put(returnName, res)

        Right(ConfigFactory.parseMap(values))
      case _ =>
        val allErrors: Errors = List(returnName, lagDays, leadDays, formatter, currentDate, invalidKeys).collect{ case Left(errs) => errs }.flatten
        val err = StageError(index, this.getClass.getName, c.origin.lineNumber, allErrors)
        Left(err :: Nil)
    }
  }   
  
  def parseFormatter(path: String)(formatter: String)(implicit c: Config): Either[Errors, DateTimeFormatter] = {
    def err(lineNumber: Option[Int], msg: String): Either[Errors, DateTimeFormatter] = Left(ConfigError(path, lineNumber, msg) :: Nil)

    try {
      Right(DateTimeFormatter.ofPattern(formatter).withResolverStyle(ResolverStyle.SMART))
    } catch {
      case e: Exception => err(Some(c.getValue(path).origin.lineNumber()), e.getMessage)
    }
  }  

  def parseCurrentDate(path: String, formatter: DateTimeFormatter)(value: String)(implicit c: Config): Either[Errors, LocalDate] = {
    def err(lineNumber: Option[Int], msg: String): Either[Errors, LocalDate] = Left(ConfigError(path, lineNumber, msg) :: Nil)

    try {
      Right(LocalDate.parse(value, formatter))
    } catch {
      case e: Exception => err(Some(c.getValue(path).origin.lineNumber()), e.getMessage)
    }
  }    
}

The plugin then needs to be registered in the plugins.config section of the job configuration and the full plugin name must be listed in your project’s /resources/META-INF/services/ai.tripl.arc.plugins.DynamicConfigurationPlugin file. See https://github.com/tripl-ai/arc-deltaperiod-config-plugin for a full example.

Note that the resolution order of these plugins is in descending order in that if the the ETL_CONF_DELTA_PERIOD was declared in multiple plugins the value set by the plugin with the lower index in the plugins.config array will take precedence.

The ETL_CONF_DELTA_PERIOD variable is then available to be resolved in a standard configuration:

{
  "plugins": {
    "config": [
      {
        "type": "ai.tripl.arc.plugins.config.DeltaPeriodDynamicConfigurationPlugin",
        "environments": [
          "production",
          "test"
        ],
        "returnName": "ETL_CONF_DELTA_PERIOD",
        "lagDays": "10",
        "leadDays": "1",
        "pattern": "yyyy-MM-dd"
      }
    ]
  },
  "stages": [
    {
      "type": "DelimitedExtract",
      "name": "load customer extract",
      "environments": [
        "production",
        "test"
      ],
      "inputURI": "hdfs://datalake/input/customer/customers_{"${ETL_CONF_DELTA_PERIOD}"}.csv",
      "outputView": "customer"
    }
  ]
}

Lifecycle Plugins

Since: 1.3.0

Custom Lifecycle Plugins allow users to extend the base Arc framework with logic which is executed before or after each Arc stage (lifecycle hooks). These stages are useful for implementing things like dataset logging after each stage execution for debugging.

Examples

package ai.tripl.arc.plugins.lifecycle

import java.util

import org.apache.spark.sql.{DataFrame, SparkSession}

import ai.tripl.arc.api.API._
import ai.tripl.arc.plugins.LifecyclePlugin
import ai.tripl.arc.util.Utils
import ai.tripl.arc.util.log.logger.Logger
import ai.tripl.arc.config.Error._

class DataFramePrinter extends LifecyclePlugin {

  val version = Utils.getFrameworkVersion

  def instantiate(index: Int, config: com.typesafe.config.Config)(implicit spark: SparkSession, logger: ai.tripl.arc.util.log.logger.Logger, arcContext: ARCContext): Either[List[ai.tripl.arc.config.Error.StageError], LifecyclePluginInstance] = {
    import ai.tripl.arc.config.ConfigReader._
    import ai.tripl.arc.config.ConfigUtils._
    implicit val c = config

    val expectedKeys = "type" :: "environments" :: "numRows" :: "truncate" :: Nil
    val numRows = getValue[Int]("numRows", default = Some(20))
    val truncate = getValue[java.lang.Boolean]("truncate", default = Some(true))
    val invalidKeys = checkValidKeys(c)(expectedKeys)

    (numRows, truncate, invalidKeys) match {
      case (Right(numRows), Right(truncate), Right(invalidKeys)) => 
        Right(DataFramePrinterInstance(
          plugin=this,
          numRows=numRows,
          truncate=truncate
        ))
      case _ =>
        val allErrors: Errors = List(numRows, truncate, invalidKeys).collect{ case Left(errs) => errs }.flatten
        val err = StageError(index, this.getClass.getName, c.origin.lineNumber, allErrors)
        Left(err :: Nil)
    }
  }
}

case class DataFramePrinterInstance(
    plugin: LifecyclePlugin,
    numRows: Int,
    truncate: Boolean
  ) extends LifecyclePluginInstance {

  override def before(stage: PipelineStage)(implicit spark: SparkSession, logger: ai.tripl.arc.util.log.logger.Logger, arcContext: ARCContext) {
    logger.trace()        
      .field("event", "before")
      .field("stage", stage.name)
      .log()  
  }

  override def after(stage: PipelineStage, result: Option[DataFrame], isLast: Boolean)(implicit spark: SparkSession, logger: ai.tripl.arc.util.log.logger.Logger, arcContext: ARCContext) {
    logger.trace()        
      .field("event", "after")
      .field("stage", stage.name)
      .field("isLast", java.lang.Boolean.valueOf(isLast))
      .log() 

    result match {
      case Some(df) => df.show(numRows, truncate)
      case None =>
    }
  } 
}

The plugin then needs to be registered by adding the full plugin name must be listed in your project’s /resources/META-INF/services/ai.tripl.arc.plugins.LifecyclePlugin file.

To execute:

{
  "plugins": {
    "lifecycle": [
      {
        "type": "ai.tripl.arc.plugins.lifecycle.DataFramePrinterLifecyclePlugin",
        "environments": [
          "production",
          "test"
        ],
        "params": {
          "numRows": "100",
          "truncate": "false",
        }
      }
    ]
  },
  "stages": [
    ...
  ]
}

Pipeline Stage Plugins

Since: 1.3.0

Custom Pipeline Stage Plugins allow users to extend the base Arc framework with custom stages which allow the full use of the Spark Scala API. This means that private business logic or code which relies on libraries not included in the base Arc framework can be used - however it is strongly advised to use the inbuilt SQL stages where possible. If stages are general purpose enough for use outside your organisation consider contributing them to ai.tripl so that others can benefit.

When writing plugins and you find Spark throwing NotSerializableException errors like:

Job aborted due to stage failure: Task not serializable: java.io.NotSerializableException: scala.collection.convert.Wrappers$MapWrapper

Ensure that any stage with a mapPartitions or map DataFrame does not require the PipelineStage instance to be passed into the map function. So instead of doing something like:

val transformedDF = try {
  df.mapPartitions[TransformedRow] { partition: Iterator[Row] => 
    val uri = stage.uri.toString

Declare the variables outside the map function so that stage does not have to be serialised and sent to all the executors (which fails if any of the PipelineStage contents are not serializable):

val stageUri = stage.uri

val transformedDF = try {
  df.mapPartitions[TransformedRow] { partition: Iterator[Row] => 
    val uri = stageUri.toString

Examples

class ConsoleLoad extends PipelineStagePlugin {

  val version = Utils.getFrameworkVersion

  def instantiate(index: Int, config: com.typesafe.config.Config)(implicit spark: SparkSession, logger: ai.tripl.arc.util.log.logger.Logger, arcContext: ARCContext): Either[List[ai.tripl.arc.config.Error.StageError], PipelineStage] = {
    import ai.tripl.arc.config.ConfigReader._
    import ai.tripl.arc.config.ConfigUtils._
    implicit val c = config

    val expectedKeys = "type" :: "name" :: "description" :: "environments" :: "inputView" :: "outputMode" :: "params" :: Nil
    val name = getValue[String]("name")
    val description = getOptionalValue[String]("description")
    val inputView = getValue[String]("inputView")
    val outputMode = getValue[String]("outputMode", default = Some("Append"), validValues = "Append" :: "Complete" :: "Update" :: Nil) |> parseOutputModeType("outputMode") _
    val params = readMap("params", c)
    val invalidKeys = checkValidKeys(c)(expectedKeys)  

    (name, description, inputView, outputMode, invalidKeys) match {
      case (Right(name), Right(description), Right(inputView), Right(outputMode), Right(invalidKeys)) => 
        val stage = ConsoleLoadStage(
          plugin=this,
          name=name,
          description=description,
          inputView=inputView,
          outputMode=outputMode,
          params=params
        )

        stage.stageDetail.put("inputView", stage.inputView)  
        stage.stageDetail.put("outputMode", stage.outputMode.sparkString)  
        stage.stageDetail.put("params", params.asJava)

        Right(stage)
      case _ =>
        val allErrors: Errors = List(name, description, inputView, outputMode, invalidKeys).collect{ case Left(errs) => errs }.flatten
        val stageName = stringOrDefault(name, "unnamed stage")
        val err = StageError(index, stageName, c.origin.lineNumber, allErrors)
        Left(err :: Nil)
    }
  }

}

case class ConsoleLoadStage(
    plugin: ConsoleLoad,
    name: String, 
    description: Option[String], 
    inputView: String, 
    outputMode: OutputModeType, 
    params: Map[String, String]
  ) extends PipelineStage {

  override def execute()(implicit spark: SparkSession, logger: ai.tripl.arc.util.log.logger.Logger, arcContext: ARCContext): Option[DataFrame] = {
    ConsoleLoadStage.execute(this)
  }
}


object ConsoleLoadStage {

  def execute(stage: ConsoleLoadStage)(implicit spark: SparkSession, logger: ai.tripl.arc.util.log.logger.Logger, arcContext: ARCContext): Option[DataFrame] = {

    val df = spark.table(stage.inputView)   

    if (!df.isStreaming) {
      throw new Exception("ConsoleLoad can only be executed in streaming mode.") with DetailException {
        override val detail = stage.stageDetail          
      }
    }      

    df.writeStream
        .format("console")
        .outputMode(stage.outputMode.sparkString)
        .start

    Option(df)
  }
}

The plugin then needs to be registered by adding the full plugin name must be listed in your project’s /resources/META-INF/services/ai.tripl.arc.plugins.PipelineStagePlugin file.

To execute:

{
  "stages": [
    {
      "type": "ConsoleLoad",
      "name": "load streaming data to console for testing",
      "environments": [
        "test"
      ],
      "inputView": "calculated_dataset",
      "outputMode": "Complete"
    }
  ]
}

User Defined Functions

Since: 1.3.0

User Defined Functions vs Spark SQL Functions

The inbuilt Spark SQL Functions are heavily optimised by the internal Spark code to a level which custom User Defined Functions cannot be (byte code) - so where possible it is better to use the inbuilt functions.

User Defined Functions allow users to extend the Spark SQL dialect.

Arc already includes some addtional functions which are not included in the base Spark SQL dialect so any useful generic functions can be included in the Arc repository so that others can benefit.

Examples

Write the code to define the custom User Defined Function:

package ai.tripl.arc.plugins
import java.util

import org.apache.spark.sql.SparkSession
import ai.tripl.arc.api.API.ARCContext

import ai.tripl.arc.util.log.logger.Logger

class TestUDFPlugin extends UDFPlugin {

  val version = "0.0.1"

  // one udf plugin can register multiple user defined functions
  override def register()(implicit spark: SparkSession, logger: ai.tripl.arc.util.log.logger.Logger, arcContext: ARCContext) = {

    // register the functions so they can be accessed via Spark SQL
    spark.sqlContext.udf.register("add_ten", TestUDFPlugin.addTen _ )           // SELECT add_ten(1) AS one_plus_ten
    spark.sqlContext.udf.register("add_twenty", TestUDFPlugin.addTwenty _ )     // SELECT add_twenty(1) AS one_plus_twenty

  }
}

object TestUDFPlugin {
  // add 10 to an incoming integer - DO NOT DO THIS IN PRODUCTION INSTEAD USE SPARK SQL DIRECTLY
  def addTen(input: Int): Int = {
    input + 10
  }

  // add 20 to an incoming integer  - DO NOT DO THIS IN PRODUCTION INSTEAD USE SPARK SQL DIRECTLY
  def addTwenty(input: Int): Int = {
    input + 20
  }
}

The plugin then needs to be registered by adding the full plugin name must be listed in your project’s /resources/META-INF/services/ai.tripl.arc.plugins.UDFPlugin file and would be executed like:

SELECT age, add_ten(age) FROM customer