Spline agent for Apache Spark

Spark Agent / Harvester

The Spline agent for Apache Spark is a complementary module to the Spline project that captures runtime lineage information from the Apache Spark jobs.

The agent is a Scala library that is embedded into the Spark driver, listening to Spark events, and capturing logical execution plans. The collected metadata is then handed over to the lineage dispatcher, from where it can either be sent to the Spline server (e.g. via REST API or Kafka), or used in another way, depending on selected dispatcher type (see Lineage Dispatchers).

The agent can be used with or without a Spline server, depending on your use case. See References.

Maven Central TeamCity build Sonarcloud Status SonarCloud Maintainability SonarCloud Reliability SonarCloud Security

Table of Contents


The Spline Spark Agent follows the Semantic Versioning principles. The Public API is defined as a set of entry-point classes (SparkLineageInitializer, SplineSparkSessionWrapper), extension APIs (Plugin API, filters, dispatchers), configuration properties and a set of supported Spark versions. In other words, the Spline Spark Agent Public API in terms of SemVer covers all entities and abstractions that are designed to be used or extended by client applications.

The version number does not directly reflect the relation of the Agent to the Spline Producer API (the Spline server). Both the Spline Server and the Agent are designed to be as much mutually compatible as possible, assuming long-term operation and a possibly significant gap in the server and the agent release dates. Such requirement is dictated by the nature of the Agent that could be embedded into some Spark jobs and only rarely if ever updated without posing a risk to stop working because of eventual Spline server update. Likewise, it should be possible to update the Agent anytime (e.g. to fix a bug or support a newer Spark version or a feature that earlier agent version didn't support) without requiring a Spline server upgrade.

Although not required by the above statement, for minimizing user astonishment when the compatibility between too distant Agent and Server versions is dropped, we'll increment the Major version component.

Spark / Scala version compatibility matrix

Scala 2.11 Scala 2.12
Spark 2.2 (no SQL; no codeless init)
Spark 2.3 (no Delta support)
Spark 2.4 Yes Yes
Spark 3.0 or newer Yes


Selecting artifact

There are two main agent artifacts:

  • agent-core is a Java library that you can use with any compatible Spark version. Use this one if you want to include Spline agent into your custom Spark application, and you want to manage all transitive dependencies yourself.

  • spark-spline-agent-bundle is a fat jar that is designed to be embedded into the Spark driver, either by manually copying it to the Spark's /jars directory, or by using --jars or --packages argument for the spark-submit, spark-shell or pyspark commands. This artifact is self-sufficient and is aimed to be used by most users.

Because the bundle is pre-built with all necessary dependencies, it is important to select a proper version of it that matches the minor Spark and Scala versions of your target Spark installation.


here A.B is the first two Spark version numbers and X.Y is the first two Scala version numbers. For example, if you have Spark 2.4.4 pre-built with Scala 2.12.10 then select the following agent bundle:



Spline agent is basically a Spark query listener that needs to be registered in a Spark session before is can be used. Depending on if you are using it as a library in your custom Spark application, or as a standalone bundle you can choose one of the following initialization approaches.

Codeless Initialization

This way is the most convenient one, can be used in majority use-cases. Simply include the Spline listener into the spark.sql.queryExecutionListeners config property (see Static SQL Configuration)


pyspark \
  --packages za.co.absa.spline.agent.spark:spark-2.4-spline-agent-bundle_2.12:<VERSION> \
  --conf "spark.sql.queryExecutionListeners=za.co.absa.spline.harvester.listener.SplineQueryExecutionListener" \
  --conf "spark.spline.lineageDispatcher.http.producer.url=http://localhost:9090/producer"

The same approach works for spark-submit and spark-shell commands.

Note: all Spline properties set via Spark conf should be prefixed with spark. prefix in order to be visible to the Spline agent.
See Configuration section for details.

Programmatic Initialization

Note: starting from Spline 0.6 most agent components can be configured or even replaced in a declarative manner either using Configuration or Plugin API. So normally there should be no need to use a programmatic initialization method. We recommend to use Codeless Initialization instead.

But if for some reason, Codeless Initialization doesn't fit your needs, or you want to do more customization on Spark agent, you can use programmatic initialization method.

// given a Spark session ...
val sparkSession: SparkSession = ???

// ... enable data lineage tracking with Spline
import za.co.absa.spline.harvester.SparkLineageInitializer._

// ... then run some Dataset computations as usual.
// The lineage will be captured and sent to the configured Spline Producer endpoint.

or in Java syntax:

import za.co.absa.spline.harvester.SparkLineageInitializer;
// ...

The method enableLineageTracking() accepts optional AgentConfig object that can be used to customize Spline behavior. This is an alternative way to configure Spline. The other one if via the property based configuration.

The instance of AgentConfig can be created by using a builder or one of the factory methods.

// from a sequence of key-value pairs 
val config = AgentConfig.from(???: Iterable[(String, Any)])

// from a Common Configuration
val config = AgentConfig.from(???: org.apache.commons.configuration.Configuration)

// using a builder
val config = AgentConfig.builder()
  // call some builder methods here...


Note: AgentConfig object doesn't override the standard configuration stack. Instead, it serves as an additional configuration mean with the precedence set between the spline.properties and spline.default.properties files (see below).


The agent looks for configuration in the following sources (listed in order of precedence):

  • Hadoop configuration (core-site.xml)
  • Spark configuration
  • JVM system properties
  • spline.properties file on classpath
  • AgentConfig object
  • spline.default.properties file on classpath

The file spline.default.properties contains default values for all Spline properties along with additional documentation. It's a good idea to look in the file to see what properties are available.

The order of precedence might look counter-intuitive, as one would expect that explicitly provided config (AgentConfig instance) should override ones defined in the outer scope. However, prioritizing global config to local one makes it easier to manage Spline settings centrally on clusters, while still allowing room for customization by job developers.

For example, a company could require lineage metadata from jobs executed on a particular cluster to be sanitized, enhanced with some metrics and credentials and stored in a certain metadata store (a database, file, Spline server etc). The Spline configuration needs to be set globally and applied to all Spark jobs automatically. However, some jobs might contain hardcoded properties that the developers used locally or on a testing environment, and forgot to remove them before submitting jobs into a production. In such situation we want cluster settings to have precedence over the job settings. Assuming that hardcoded settings would most likely be defined in the AgentConfig object, a property file or a JVM properties, on the cluster we could define them in the Spark config or Hadoop config.

In case of multiple definitions of property the first occurrence wins, but spline.lineageDispatcher and spline.postProcessingFilter properties are composed instead. E.g. if a LineageDispatcher is set to be Kafka in one config source and 'Http' in another, they would be implicitly wrapped by a composite dispatcher, so both would be called in the order corresponding the config source precedence. See CompositeLineageDispatcher and CompositePostProcessingFilter.

Every config property is resolved independently. So, for instance, if a DataSourcePasswordReplacingFilter is used some of its properties might be taken from one config source and the other ones form another, according to the conflict resolution rules described above. This allows administrators to tweak settings of individual Spline components (filters, dispatchers or plugins) without having to redefine and override the whole piece of configuration for a given component.



  • REQUIRED [default]

    If Spline fails to initialize itself (e.g., wrong configuration, no db connection) the Spark application aborts with an error. (Note: it only concerns Spline initialization routine. If the error happens during lineage capturing, or in the Spline dispatcher, then the target Spark job have already been finished by that time, and the resulted data have been persisted, regardless of the spline.mode settings. The Spline agent doesn't do any automated rollbacks).


    Spline will try to initialize itself, but if it fails it switches to DISABLED mode allowing the Spark application to proceed normally without Lineage tracking.


    Lineage tracking is completely disabled and Spline is unhooked from Spark.

Note: The default value for spline.mode has changed in Spline 1.0.0. It used to be BEST_EFFORT for Spline 0.x version series.


The logical name of the root lineage dispatcher. See Lineage Dispatchers chapter.


The logical name of the root post-processing filter. See Post Processing Filters chapter.

Lineage Dispatchers

The LineageDispatcher trait is responsible for sending out the captured lineage information. By default, the HttpLineageDispatcher is used, that sends the lineage data to the Spline REST endpoint (see Spline Producer API).

Available dispatchers:

  • HttpLineageDispatcher - sends the lineage via http
  • KafkaLineageDispatcher - sends the lineage via kafka
  • ConsoleLineageDispatcher - write the lineage to console
  • LoggingLineageDispatcher - logs the lineahge using logger
  • CompositeLineageDispatcher - allows combining multiple dispatchers

Each dispatcher can have different configuration parameters. To make the configs clearly separated each dispatcher has its own namespace in which all it's parameters are defined. I will explain it on an kafka examples.

Defining dispatcher


Once you defined the dispatcher all other parameters will have a namespace spline.lineageDispatcher.{{dipatcher-name}}. as a prefix. In this case it is spline.lineageDispatcher.kafka..

To find out which parameters you can use look into spline.default.properties. For kafka I would have to define at least these two properties:


Creating your own dispatcher

There is also a possibility to create your own dispatcher. It must implement LineageDispatcher trait and have a constructor with a single parameter of type org.apache.commons.configuration.Configuration. To use it you must define name and class and also all other parameters you need. For example:


Post Processing Filters

Filters can be used to enrich the lineage with your own custom data or to remove unwanted data like passwords. All filters are applied after the Spark plan is converted to Spline DTOs, but before the dispatcher is called.

The procedure how filters are registered and configured is similar to the LineageDispatcher registration and configuration procedure. A custom filter class must implement za.co.absa.spline.harvester.postprocessing.PostProcessingFilter trait and declare a constructor with a single parameter of type org.apache.commons.configuration.Configuration. Then register and configure it like this:


Use pre-registered CompositePostProcessingFilter to chain up multiple filters:


(see spline.default.properties for details and examples)

Spark features coverage

Dataset operations are fully supported

RDD transformations aren't supported due to Spark internal architecture specifics, but they might be supported semi-automatically in the future Spline versions (see #33)

SQL dialect is mostly supported.

DDL operations are not supported, excepts for CREATE TABLE ... AS SELECT ... which is supported.

Note: the lineage is only captured on persistent (write) actions. In-memory only actions like collect() or printSchema() are ignored.

The following data formats and providers are supported out of the box:

  • Avro
  • Cassandra
  • Delta
  • ElasticSearch
  • Excel
  • HDFS
  • Hive
  • JDBC
  • Kafka
  • MongoDB
  • XML

Although Spark being an extensible piece of software can support much more, it doesn't provide any universal API that Spline can utilize to capture reads and write from/to everything that Spark supports. Support for most of different data sources and formats has to be added to Spline one by one. Fortunately starting with Spline 0.5.4 the auto discoverable Plugin API has been introduced to make this process easier.

Below is the break down of the read/write command list that we have come through.
Some commands are implemented, others have yet to be implemented, and finally there are such that bear no lineage information and hence are ignored.

All commands inherit from org.apache.spark.sql.catalyst.plans.logical.Command.

You can see how to produce unimplemented commands in za.co.absa.spline.harvester.SparkUnimplementedCommandsSpec.


  • CreateDataSourceTableAsSelectCommand (org.apache.spark.sql.execution.command)
  • CreateHiveTableAsSelectCommand (org.apache.spark.sql.hive.execution)
  • CreateTableCommand (org.apache.spark.sql.execution.command)
  • DropTableCommand (org.apache.spark.sql.execution.command)
  • InsertIntoDataSourceDirCommand (org.apache.spark.sql.execution.command)
  • InsertIntoHadoopFsRelationCommand (org.apache.spark.sql.execution.datasources)
  • InsertIntoHiveDirCommand (org.apache.spark.sql.hive.execution)
  • InsertIntoHiveTable (org.apache.spark.sql.hive.execution)
  • SaveIntoDataSourceCommand (org.apache.spark.sql.execution.datasources)

To be implemented

  • AlterTableAddColumnsCommand (org.apache.spark.sql.execution.command)
  • AlterTableChangeColumnCommand (org.apache.spark.sql.execution.command)
  • AlterTableRenameCommand (org.apache.spark.sql.execution.command)
  • AlterTableSetLocationCommand (org.apache.spark.sql.execution.command)
  • CreateDataSourceTableCommand (org.apache.spark.sql.execution.command)
  • CreateDatabaseCommand (org.apache.spark.sql.execution.command)
  • CreateTableLikeCommand (org.apache.spark.sql.execution.command)
  • DropDatabaseCommand (org.apache.spark.sql.execution.command)
  • LoadDataCommand (org.apache.spark.sql.execution.command)
  • TruncateTableCommand (org.apache.spark.sql.execution.command)

When one of these commands occurs spline will let you know by logging a warning.


  • AddFileCommand (org.apache.spark.sql.execution.command)
  • AddJarCommand (org.apache.spark.sql.execution.command)
  • AlterDatabasePropertiesCommand (org.apache.spark.sql.execution.command)
  • AlterTableAddPartitionCommand (org.apache.spark.sql.execution.command)
  • AlterTableDropPartitionCommand (org.apache.spark.sql.execution.command)
  • AlterTableRecoverPartitionsCommand (org.apache.spark.sql.execution.command)
  • AlterTableRenamePartitionCommand (org.apache.spark.sql.execution.command)
  • AlterTableSerDePropertiesCommand (org.apache.spark.sql.execution.command)
  • AlterTableSetPropertiesCommand (org.apache.spark.sql.execution.command)
  • AlterTableUnsetPropertiesCommand (org.apache.spark.sql.execution.command)
  • AlterViewAsCommand (org.apache.spark.sql.execution.command)
  • AnalyzeColumnCommand (org.apache.spark.sql.execution.command)
  • AnalyzePartitionCommand (org.apache.spark.sql.execution.command)
  • AnalyzeTableCommand (org.apache.spark.sql.execution.command)
  • CacheTableCommand (org.apache.spark.sql.execution.command)
  • ClearCacheCommand (org.apache.spark.sql.execution.command)
  • CreateFunctionCommand (org.apache.spark.sql.execution.command)
  • CreateTempViewUsing (org.apache.spark.sql.execution.datasources)
  • CreateViewCommand (org.apache.spark.sql.execution.command)
  • DescribeColumnCommand (org.apache.spark.sql.execution.command)
  • DescribeDatabaseCommand (org.apache.spark.sql.execution.command)
  • DescribeFunctionCommand (org.apache.spark.sql.execution.command)
  • DescribeTableCommand (org.apache.spark.sql.execution.command)
  • DropFunctionCommand (org.apache.spark.sql.execution.command)
  • ExplainCommand (org.apache.spark.sql.execution.command)
  • InsertIntoDataSourceCommand (org.apache.spark.sql.execution.datasources) *
  • ListFilesCommand (org.apache.spark.sql.execution.command)
  • ListJarsCommand (org.apache.spark.sql.execution.command)
  • RefreshResource (org.apache.spark.sql.execution.datasources)
  • RefreshTable (org.apache.spark.sql.execution.datasources)
  • ResetCommand$ (org.apache.spark.sql.execution.command)
  • SetCommand (org.apache.spark.sql.execution.command)
  • SetDatabaseCommand (org.apache.spark.sql.execution.command)
  • ShowColumnsCommand (org.apache.spark.sql.execution.command)
  • ShowCreateTableCommand (org.apache.spark.sql.execution.command)
  • ShowDatabasesCommand (org.apache.spark.sql.execution.command)
  • ShowFunctionsCommand (org.apache.spark.sql.execution.command)
  • ShowPartitionsCommand (org.apache.spark.sql.execution.command)
  • ShowTablePropertiesCommand (org.apache.spark.sql.execution.command)
  • ShowTablesCommand (org.apache.spark.sql.execution.command)
  • StreamingExplainCommand (org.apache.spark.sql.execution.command)
  • UncacheTableCommand (org.apache.spark.sql.execution.command)

Developer documentation

Plugin API

Using a plugin API you can capture lineage from a 3rd party data source provider. Spline discover plugins automatically by scanning a classpath, so no special steps required to register and configure a plugin. All you need is to create a class extending the za.co.absa.spline.harvester.plugin.Plugin marker trait mixed with one or more *Processing traits, depending on your intention.

There are three general processing traits:

  • DataSourceFormatNameResolving - returns a name of a data provider/format in use.
  • ReadNodeProcessing - detects a read-command and gather meta information.
  • WriteNodeProcessing - detects a write-command and gather meta information.

There are also two additional trait that handle common cases of reading and writing:

  • BaseRelationProcessing - similar to ReadNodeProcessing, but instead of capturing all logical plan nodes it only reacts on LogicalRelation (see LogicalRelationPlugin)
  • RelationProviderProcessing - similar to WriteNodeProcessing, but it only captures SaveIntoDataSourceCommand (see SaveIntoDataSourceCommandPlugin)

The best way to illustrate how plugins work is to look at the real working example, e.g. za.co.absa.spline.harvester.plugin.embedded.JDBCPlugin

The most common simplified pattern looks like this:

package my.spline.plugin

import javax.annotation.Priority
import za.co.absa.spline.harvester.builder._
import za.co.absa.spline.harvester.plugin.Plugin._
import za.co.absa.spline.harvester.plugin._

@Priority(Precedence.User) // not required, but can be used to control your plugin precedence in the plugin chain. Default value is `User`.  
class FooBarPlugin
  extends Plugin
    with BaseRelationProcessing
    with RelationProviderProcessing {

  override def baseRelationProcessor: PartialFunction[(BaseRelation, LogicalRelation), ReadNodeInfo] = {
    case (FooBarRelation(a, b, c, d), lr) if /*more conditions*/ =>
      val dataFormat: Option[AnyRef] = ??? // data format being read (will be resolved by the `DataSourceFormatResolver` later)
      val dataSourceURI: String = ??? // a unique URI for the data source
      val params: Map[String, Any] = ??? // additional parameters characterizing the read-command. E.g. (connection protocol, access mode, driver options etc)

      (SourceIdentifier(dataFormat, dataSourceURI), params)

  override def relationProviderProcessor: PartialFunction[(AnyRef, SaveIntoDataSourceCommand), WriteNodeInfo] = {
    case (provider, cmd) if provider == "foobar" || provider.isInstanceOf[FooBarProvider] =>
      val dataFormat: Option[AnyRef] = ??? // data format being written (will be resolved by the `DataSourceFormatResolver` later)
      val dataSourceURI: String = ??? // a unique URI for the data source
      val writeMode: SaveMode = ??? // was it Append or Overwrite?
      val query: LogicalPlan = ??? // the logical plan to get the rest of the lineage from
      val params: Map[String, Any] = ??? // additional parameters characterizing the write-command

      (SourceIdentifier(dataFormat, dataSourceURI), writeMode, query, params)

Note: to avoid unwanted possible shadowing the other plugins (including the future ones), make sure that the pattern-matching criteria are as much selective as possible for your plugin needs.

A plugin class is expected to only have a single constructor. The constructor can have no arguments, or one or more of the following types (the values will be autowired):

  • SparkSession
  • PathQualifier
  • PluginRegistry

Compile you plugin and drop it into the Spline/Spark classpath. Spline will pick it up automatically.

Building for different Scala and Spark versions

Note: The project requires Java version 1.8 (strictly) and Apache Maven for building.

Check the build environment:

mvn --version

Verify that Maven is configured to run on Java 1.8. For example:

Apache Maven 3.6.3 (Red Hat 3.6.3-8)
Maven home: /usr/share/maven
Java version: 1.8.0_302, vendor: Red Hat, Inc., runtime: /usr/lib/jvm/java-1.8.0-openjdk-

There are several maven profiles that makes it easy to build the project with different versions of Spark and Scala.

  • Scala profiles: scala-2.11, scala-2.12
  • Spark profiles: spark-2.2, spark-2.3, spark-2.4, spark-3.0, spark-3.1

For example, to build an agent for Spark 2.4 and Scala 2.12:

# Change Scala version in pom.xml.
mvn scala-cross-build:change-version -Pscala-2.12

# now you can build for Scala 2.12
mvn clean install -Pscala-2.12,spark-2.4

Build docker image

The agent docker image is mainly used to run example jobs and pre-fill the database with the sample lineage data.

(Spline docker images are available on the DockerHub repo - https://hub.docker.com/u/absaoss)

mvn install -Ddocker -Ddockerfile.repositoryUrl=my

See How to build Spline Docker images for details.

References and examples

Although the primary goal of Spline agent is to be used in combination with the Spline server, it is flexible enough to be used in isolation or integration with other data lineage tracking solutions including custom ones.

Below is a couple of examples of such integration:

Copyright 2019 ABSA Group Limited

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at


Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
See the License for the specific language governing permissions and
limitations under the License.