SQL Lineage Support#

When we execute a SQL statement, the computing engine will return to a result set, each column in the result set may originate from different tables, and these different tables depend on other tables, which may have been calculated by functions, aggregation, etc. The source table is related to the result set, which is the SQL Lineage.

Introduction#

The current lineage parsing functionality is implemented as a plugin by extending Spark’s QueryExecutionListener.

  1. The SparkListenerSQLExecutionEnd event is triggered after the SQL execution is finished and captured by the QueryExecuctionListener, where the SQL lineage parsing process is performed on the successfully executed SQL.

  2. Will write the parsed lineage information to the log file in JSON format.

Example#

When the following SQL is executed:

## table
create table test_table0(a string, b string)

## query
select a as col0, b as col1 from test_table0

The lineage of this SQL:

{
   "inputTables": ["spark_catalog.default.test_table0"],
   "outputTables": [],
   "columnLineage": [{
      "column": "col0",
      "originalColumns": ["spark_catalog.default.test_table0.a"]
   }, {
      "column": "col1",
      "originalColumns": ["spark_catalog.default.test_table0.b"]
   }]
}

Lineage specific identification#

  • __count__. Means that the column is an count(*) aggregate expression and cannot extract the specific column. Lineage of the column like default.test_table0.__count__.

  • __local__. Means that the lineage of the table is a LocalRelation and not the real table, like __local__.a

SQL type support#

Currently supported column lineage for spark’s Command and Query type:

Query#

  • Select

Command#

  • AlterViewAsCommand

  • AppendData

  • CreateDataSourceTableAsSelectCommand

  • CreateHiveTableAsSelectCommand

  • CreateTableAsSelect

  • CreateViewCommand

  • InsertIntoDataSourceCommand

  • InsertIntoDataSourceDirCommand

  • InsertIntoHadoopFsRelationCommand

  • InsertIntoHiveDirCommand

  • InsertIntoHiveTable

  • MergeIntoTable

  • OptimizedCreateHiveTableAsSelectCommand

  • OverwriteByExpression

  • OverwritePartitionsDynamic

  • ReplaceTableAsSelect

  • SaveIntoDataSourceCommand

Building#

Build with Apache Maven#

Kyuubi Spark Lineage Listener Extension is built using Apache Maven. To build it, cd to the root direct of kyuubi project and run:

build/mvn clean package -pl :kyuubi-spark-lineage_2.12 -am -DskipTests

After a while, if everything goes well, you will get the plugin finally in two parts:

  • The main plugin jar, which is under ./extensions/spark/kyuubi-spark-lineage/target/kyuubi-spark-lineage_${scala.binary.version}-${project.version}.jar

Build against Different Apache Spark Versions#

The maven option spark.version is used for specifying Spark version to compile with and generate corresponding transitive dependencies. By default, it is always built with the latest spark.version defined in kyuubi project main pom file. Sometimes, it may be incompatible with other Spark distributions, then you may need to build the plugin on your own targeting the Spark version you use.

For example,

build/mvn clean package -pl :kyuubi-spark-lineage_2.12 -am -DskipTests -Dspark.version=3.1.2

The available spark.versions are shown in the following table.

Spark Version Supported Remark
master √ -
3.4.x √ -
3.3.x √ -
3.2.x √ -
3.1.x √ -
3.0.x √ -
2.4.x x -

Currently, Spark released with Scala 2.12 are supported.

Test with ScalaTest Maven plugin#

If you omit -DskipTests option in the command above, you will also get all unit tests run.

build/mvn clean package -pl :kyuubi-spark-lineage_2.12

If any bug occurs and you want to debug the plugin yourself, you can configure -DdebugForkedProcess=true and -DdebuggerPort=5005(optional).

build/mvn clean package -pl :kyuubi-spark-lineage_2.12 -DdebugForkedProcess=true

The tests will suspend at startup and wait for a remote debugger to attach to the configured port.

We will appreciate if you can share the bug or the fix to the Kyuubi community.

Installing#

With the kyuubi-spark-lineage_*.jar and its transitive dependencies available for spark runtime classpath, such as

  • Copied to $SPARK_HOME/jars, or

  • Specified to spark.jars configuration

Configure#

Settings for Spark Listener Extensions#

Add org.apache.kyuubi.plugin.lineage.SparkOperationLineageQueryExecutionListener to the spark configuration spark.sql.queryExecutionListeners.

spark.sql.queryExecutionListeners=org.apache.kyuubi.plugin.lineage.SparkOperationLineageQueryExecutionListener

Optional configuration#

Whether to Skip Permanent View Resolution#

If enabled, lineage resolution will stop at permanent views and treats them as physical tables. We need to add one configurations.

spark.kyuubi.plugin.lineage.skip.parsing.permanent.view.enabled=true

Get Lineage Events#

The lineage dispatchers are used to dispatch lineage events, configured via spark.kyuubi.plugin.lineage.dispatchers.

  • SPARK_EVENT (by default): send lineage event to spark event bus
  • KYUUBI_EVENT: send lineage event to kyuubi event bus
  • ATLAS: send lineage to apache atlas

Get Lineage Events from SparkListener#

When using the SPARK_EVENT dispatcher, the lineage events will be sent to the SparkListenerBus. To handle lineage events, a new SparkListener needs to be added. Example for Adding SparkListener:

spark.sparkContext.addSparkListener(new SparkListener {
      override def onOtherEvent(event: SparkListenerEvent): Unit = {
        event match {
          case lineageEvent: OperationLineageEvent =>
            // Your processing logic
          case _ =>
        }
      }
    })

Get Lineage Events from Kyuubi EventHandler#

When using the KYUUBI_EVENT dispatcher, the lineage events will be sent to the Kyuubi EventBus. Refer to Kyuubi Event Handler to handle kyuubi events.

Ingest Lineage Entities to Apache Atlas#

The lineage entities can be ingested into Apache Atlas using the ATLAS dispatcher.

Extra works:

  • The least transitive dependencies needed, which are under ./extensions/spark/kyuubi-spark-lineage/target/scala-${scala.binary.version}/jars

  • Use spark.files to specify the atlas-application.properties configuration file for Atlas

Atlas Client configurations (Configure in atlas-application.properties or passed in spark.atlas. prefix):

Name Default Value Description Since
atlas.rest.address http://localhost:21000 The rest endpoint url for the Atlas server 1.8.0
atlas.client.type rest The client type (currently only supports rest) 1.8.0
atlas.client.username none The client username 1.8.0
atlas.client.password none The client password 1.8.0
atlas.cluster.name primary The cluster name to use in qualifiedName of entities. 1.8.0
atlas.hook.spark.column.lineage.enabled true Whether to ingest column lineages to Atlas. 1.8.0