Skip to main content

Spark

To integrate Spark with DataHub, we provide a lightweight Java agent that listens for Spark application and job events and pushes metadata out to DataHub in real-time. The agent listens to events such as application start/end, and SQLExecution start/end to create pipelines (i.e. DataFlow) and tasks (i.e. DataJob) in DataHub along with lineage to datasets that are being read from and written to. Read on to learn how to configure this for different Spark scenarios.

Configuring Spark agent​

The Spark agent can be configured using a config file or while creating a Spark Session. If you are using Spark on Databricks, refer to Configuration Instructions for Databricks.

Before you begin: Versions and Release Notes​

Versioning of the jar artifact will follow the semantic versioning of the main DataHub repo and release notes will be available here. Always check the Maven central repository for the latest released version.

Note: Starting from version 0.2.18, we provide separate jars for different Scala versions:

  • For Scala 2.12: io.acryl:acryl-spark-lineage_2.12:0.2.18
  • For Scala 2.13: io.acryl:acryl-spark-lineage_2.13:0.2.18

Configuration Instructions: spark-submit​

When running jobs using spark-submit, the agent needs to be configured in the config file.

#Configuring DataHub spark agent jar
spark.jars.packages                          io.acryl:acryl-spark-lineage_2.12:0.2.18
spark.extraListeners                         datahub.spark.DatahubSparkListener
spark.datahub.rest.server                    http://localhost:8080

For Scala 2.13:

#Configuring DataHub spark agent jar for Scala 2.13
spark.jars.packages                          io.acryl:acryl-spark-lineage_2.13:0.2.18
spark.extraListeners                         datahub.spark.DatahubSparkListener
spark.datahub.rest.server                    http://localhost:8080

spark-submit command line​

spark-submit --packages io.acryl:acryl-spark-lineage_2.12:0.2.18 --conf "spark.extraListeners=datahub.spark.DatahubSparkListener" my_spark_job_to_run.py

For Scala 2.13:

spark-submit --packages io.acryl:acryl-spark-lineage_2.13:0.2.18 --conf "spark.extraListeners=datahub.spark.DatahubSparkListener" my_spark_job_to_run.py

Configuration Instructions: Amazon EMR​

Set the following spark-defaults configuration properties as it stated here

spark.jars.packages                          io.acryl:acryl-spark-lineage_2.12:0.2.18
spark.extraListeners                         datahub.spark.DatahubSparkListener
spark.datahub.rest.server                    https://your_datahub_host/gms
#If you have authentication set up then you also need to specify the Datahub access token
spark.datahub.rest.token                     yourtoken

Configuration Instructions: Notebooks​

When running interactive jobs from a notebook, the listener can be configured while building the Spark Session.

spark = SparkSession.builder
.master("spark://spark-master:7077")
.appName("test-application")
.config("spark.jars.packages", "io.acryl:acryl-spark-lineage_2.12:0.2.18")
.config("spark.extraListeners", "datahub.spark.DatahubSparkListener")
.config("spark.datahub.rest.server", "http://localhost:8080")
.enableHiveSupport()
.getOrCreate()

Configuration Instructions: Standalone Java Applications​

The configuration for standalone Java apps is very similar.

spark =SparkSession.

builder()
        .

appName("test-application")
        .

config("spark.master","spark://spark-master:7077")
        .

config("spark.jars.packages","io.acryl:acryl-spark-lineage_2.12:0.2.18")
        .

config("spark.extraListeners","datahub.spark.DatahubSparkListener")
        .

config("spark.datahub.rest.server","http://localhost:8080")
        .

enableHiveSupport()
        .

getOrCreate();

Configuration Instructions: Databricks​

The Spark agent can be configured using Databricks Cluster Spark configuration and Init script.

Databricks Secrets can be leveraged to store sensitive information like tokens.

  • Download datahub-spark-lineage jar from the Maven central repository.

  • Create init.sh with below content

    #!/bin/bash
    cp /dbfs/datahub/datahub-spark-lineage*.jar /databricks/jars

  • Install and configure Databricks CLI.

  • Copy jar and init script to Databricks File System(DBFS) using Databricks CLI.

    databricks fs mkdirs dbfs:/datahub
    databricks fs cp --overwrite datahub-spark-lineage*.jar dbfs:/datahub
    databricks fs cp --overwrite init.sh dbfs:/datahub

  • Open Databricks Cluster configuration page. Click the Advanced Options toggle. Click the Spark tab. Add below configurations under Spark Config.

    spark.extraListeners                    datahub.spark.DatahubSparkListener
    spark.datahub.rest.server               http://localhost:8080
    spark.datahub.stage_metadata_coalescing true
    spark.datahub.databricks.cluster        cluster-name<any preferred cluster identifier>

  • Click the Init Scripts tab. Set cluster init script as dbfs:/datahub/init.sh.

  • Configuring DataHub authentication token

    • Add below config in cluster spark config.

      spark.datahub.rest.token <token>

    • Alternatively, Databricks secrets can be used to secure token.

      • Create secret using Databricks CLI.

        databricks secrets create-scope --scope datahub --initial-manage-principal users
        databricks secrets put --scope datahub --key rest-token
        databricks secrets list --scope datahub &lt;&lt;Edit prompted file with token value&gt;&gt;

      • Add in spark config

        spark.datahub.rest.token {{secrets/datahub/rest-token}}

Configuration Options​

FieldRequiredDefaultDescription
spark.jars.packages✅Set with latest/required version io.acryl:acryl-spark-lineage_2.12:0.2.18 (or io.acryl:acryl-spark-lineage_2.13:0.2.18 for Scala 2.13)
spark.extraListeners✅datahub.spark.DatahubSparkListener
spark.datahub.emitterrestSpecify the ways to emit metadata. By default it sends to DataHub using REST emitter. Valid options are rest, kafka or file
spark.datahub.rest.serverhttp://localhost:8080Datahub server url eg: http://localhost:8080
spark.datahub.rest.tokenAuthentication token.
spark.datahub.rest.disable_ssl_verificationfalseDisable SSL certificate validation. Caution: Only use this if you know what you are doing!
spark.datahub.rest.disable_chunked_encodingfalseDisable Chunked Transfer Encoding. In some environment chunked encoding causes issues. With this config option it can be disabled.
spark.datahub.rest.max_retries0Number of times a request retried if failed
spark.datahub.rest.retry_interval10Number of seconds to wait between retries
spark.datahub.file.filenameThe file where metadata will be written if file emitter is set
spark.datahub.kafka.bootstrapThe Kafka bootstrap server url to use if the Kafka emitter is set
spark.datahub.kafka.schema_registry_urlThe Schema registry url to use if the Kafka emitter is set
spark.datahub.kafka.schema_registry_config.Additional config to pass in to the Schema Registry Client
spark.datahub.kafka.producer_config.Additional config to pass in to the Kafka producer. For example: --conf "spark.datahub.kafka.producer_config.client.id=my_client_id"
spark.datahub.metadata.pipeline.platformInstancePipeline level platform instance
spark.datahub.metadata.dataset.platformInstanceDataset level platform instance (useful when you need to match dataset URNs with those created by other ingestion sources)
spark.datahub.metadata.dataset.envPRODSupported values. In all other cases, will fall back to PROD
spark.datahub.metadata.dataset.hivePlatformAliashiveBy default, DataHub assigns Hive-like tables to the Hive platform. If you are using Glue as your Hive metastore, set this config flag to glue
spark.datahub.metadata.dataset.connections.<olNamespace>.platformInstancePer-connection upstream platform instance, keyed by the connection's canonical OpenLineage namespace — arn:aws:glue:{region}:{account} (Glue, account + region; account alone is not unique across regions), snowflake://{account}, postgres://{host}:{port}, etc. When a Spark job reads from multiple accounts/regions/hosts, the namespace is taken from the OpenLineage event and mapped to the right platform_instance so each upstream URN matches the one its own connector emits. The platform is implied by the namespace scheme. Optionally set .env alongside (e.g. spark.datahub.metadata.dataset.connections."arn:aws:glue:us-east-1:111122223333".env=PROD).
spark.datahub.metadata.include_schemetrueInclude scheme from the path URI (e.g. hdfs://, s3://) in the dataset URN. We recommend setting this value to false, but it is set to true for backwards compatibility with previous versions
spark.datahub.metadata.remove_partition_patternRemove partition pattern (e.g. /partition=\d+). It changes database/table/partition=123 to database/table
spark.datahub.coalesce_jobstrueOnly one DataJob (task) will be emitted containing all input and output datasets for the Spark application
spark.datahub.parent.datajob_urnSpecified dataset will be set as upstream dataset for DataJob created. Effective only when spark.datahub.coalesce_jobs is set to true
spark.datahub.metadata.dataset.materializefalseMaterialize Datasets in DataHub
spark.datahub.platform.s3.path_spec_listList of path specs per platform
spark.datahub.metadata.dataset.include_schema_metadatafalseEmit dataset schema metadata based on the Spark execution. It is recommended to get schema information from platform-specific DataHub sources as this is less reliable
spark.datahub.flow_nameIf set, it will be used as the DataFlow name; otherwise, it uses the Spark app name as flow_name
spark.datahub.file_partition_regexpStrip partition part from the path if the path end matches the specified regexp. Example: year=.*/month=.*/day=.*
spark.datahub.tagsComma-separated list of tags to attach to the DataFlow
spark.datahub.domainsComma-separated list of domain URNs to attach to the DataFlow
spark.datahub.stage_metadata_coalescingfalseNormally metadata is coalesced and sent at the onApplicationEnd event, which is never called on Databricks or on Glue. Enable this on Databricks if you want coalesced runs.
spark.datahub.patch.enabledfalseSet this to true to send lineage as a patch, which appends rather than overwrites existing Dataset lineage edges. By default, it is disabled.
spark.datahub.metadata.dataset.lowerCaseUrnsfalseSet this to true to lowercase dataset URNs. By default, it is disabled.
spark.datahub.disableSymlinkResolutionfalseSet this to true if you prefer using the S3 location instead of the Hive table. By default, it is disabled.
spark.datahub.s3.bucketThe name of the bucket where metadata will be written if s3 emitter is set
spark.datahub.s3.prefixThe prefix for the file where metadata will be written on s3 if s3 emitter is set
spark.datahub.s3.filenameThe name of the file where metadata will be written if it is not set random filename will be used on s3 if s3 emitter is set
spark.datahub.log.mcpstrueSet this to true to log MCPS to the log. By default, it is enabled.
spark.datahub.legacyLineageCleanup.enabledfalseSet this to true to remove legacy lineages from older Spark Plugin runs. This will remove those lineages from the Datasets which it adds to DataJob. By default, it is disabled.
spark.datahub.captureColumnLevelLineagetrueSet this to false to disable column-level lineage capture for improved performance on large datasets.
spark.datahub.columnLineage.includeIndirecttrueWhether to include INDIRECT column upstreams (JOIN keys, WHERE filters, GROUP BY columns) in column-level lineage. Set this to false to drop input columns whose only role is INDIRECT — useful when you only care about DIRECT (value-producing) column lineage and want to avoid the every-column-depends-on-every-join-key fan-out in the lineage panel.
spark.datahub.capture_spark_planfalseSet this to true to capture the Spark plan. By default, it is disabled.
spark.datahub.metadata.dataset.enableEnhancedMergeIntoExtractionfalseSet this to true to enable enhanced table name extraction for Delta Lake MERGE INTO commands. This improves lineage tracking by including the target table name in the job name. By default, it is disabled.

What to Expect: The Metadata Model​

As of current writing, the Spark agent produces metadata related to the Spark job, tasks and lineage edges to datasets.

  • A pipeline is created per Spark <master, appName>.
  • A task is created per unique Spark query execution within an app.

For Spark on Databricks,

  • A pipeline is created per
    • cluster_identifier: specified with spark.datahub.databricks.cluster
    • applicationID: on every restart of the cluster new spark applicationID will be created.
  • A task is created per unique Spark query execution.

Custom properties & relating to Spark UI​

The following custom properties in pipelines and tasks relate to the Spark UI:

  • appName and appId in a pipeline can be used to determine the Spark application
  • Other custom properties of pipelines and tasks capture the start and end times of execution etc.

For Spark on Databricks, pipeline start time is the cluster start time.

Column-level Lineage and Transformation Types​

The Spark agent captures fine-grained lineage information, including column-level lineage with transformation types. When available, OpenLineage's transformation types are captured and mapped to DataHub's FinegrainedLineage TransformOption, providing detailed insights into how data transformations occur at the column level.

Spark versions supported​

Supports the Spark 3.x series (Scala 2.12 and 2.13). Apache Spark 4.x is supported via the Scala 2.13 build of the agent (io.acryl:acryl-spark-lineage_2.13) and is covered by a compatibility smoke test.

Environments tested with​

This initial release has been tested with the following environments:

  • spark-submit of Python/Java applications to local and remote servers
  • Standalone Java applications
  • Databricks Standalone Cluster
  • EMR

Testing with Databricks Standard and High-concurrency Cluster is not done yet.

Configuring Hdfs based dataset URNs​

Spark emits lineage between datasets. It has its own logic for generating urns. Python sources emit metadata of datasets. To link these 2 things, urns generated by both have to match. This section will help you to match urns to that of other ingestion sources. By default, URNs are created using template urn:li:dataset:(urn:li:dataPlatform:<$platform>,<platformInstance>.<name>,<env>). We can configure these 4 things to generate the desired urn.

Platform: Hdfs-based platforms supported explicitly:

  • AWS S3 (s3)
  • Google Cloud Storage (gcs)
  • Azure Storage:
    • Azure Blob Storage (abs) - supports wasb/wasbs protocols
    • Azure Data Lake Storage Gen2 (abs) - supports abfs/abfss protocols
  • Local file system (file) All other platforms will have "hdfs" as a platform.

Name: By default, the name is the complete path. For Hdfs base datasets, tables can be at different levels in the path than that of the actual file read due to various reasons like partitioning, and sharding. 'path_spec' is used to alter the name. {table} marker is used to specify the table level. Below are a few examples. One can specify multiple path_specs for different paths specified in the path_spec_list. Each actual path is matched against all path_spes present in the list. First, one to match will be used to generate urn.

path_spec Examples

spark.datahub.platform.s3.path_spec_list=s3://my-bucket/foo/{table}/year=*/month=*/day=*/*,s3://my-other-bucket/foo/{table}/year=*/month=*/day=*/*"
Absolute pathpath_specUrn
s3://my-bucket/foo/tests/bar.avroNot providedurn:li:dataset:(urn:li:dataPlatform:s3,my-bucket/foo/tests/bar.avro,PROD)
s3://my-bucket/foo/tests/bar.avros3://my-bucket/foo/{table}/*urn:li:dataset:(urn:li:dataPlatform:s3,my-bucket/foo/tests,PROD)
s3://my-bucket/foo/tests/bar.avros3://my-bucket/foo/tests/{table}urn:li:dataset:(urn:li:dataPlatform:s3,my-bucket/foo/tests/bar.avro,PROD)
gs://my-bucket/foo/tests/bar.avrogs://my-bucket/{table}//urn:li:dataset:(urn:li:dataPlatform:gcs,my-bucket/foo,PROD)
gs://my-bucket/foo/tests/bar.avrogs://my-bucket/{table}urn:li:dataset:(urn:li:dataPlatform:gcs,my-bucket/foo,PROD)
file:///my-bucket/foo/tests/bar.avrofile:///my-bucket///{table}urn:li:dataset:(urn:li:dataPlatform:local,my-bucket/foo/tests/bar.avro,PROD)

platform instance and env:

The default value for env is 'PROD' and the platform instance is None. env and platform instances can be set for all datasets using configurations 'spark.datahub.metadata.dataset.env' and 'spark.datahub.metadata.dataset.platformInstace'. If spark is processing data that belongs to a different env or platform instance, then 'path_alias' can be used to specify path_spec specific values of these. 'path_alias' groups the 'path_spec_list', its env, and platform instance together.

path_alias_list Example:

The below example explains the configuration of the case, where files from 2 buckets are being processed in a single spark application and files from my-bucket are supposed to have "instance1" as platform instance and "PROD" as env, and files from bucket2 should have env "DEV" in their dataset URNs.

spark.datahub.platform.s3.path_alias_list :  path1,path2
spark.datahub.platform.s3.path1.env : PROD
spark.datahub.platform.s3.path1.path_spec_list: s3://my-bucket/*/*/{table}
spark.datahub.platform.s3.path1.platform_instance : instance-1
spark.datahub.platform.s3.path2.env: DEV
spark.datahub.platform.s3.path2.path_spec_list: s3://bucket2/*/{table}

Cross-platform platform instances (connections)​

When a single Spark job reads catalog/warehouse tables from multiple accounts, regions, or hosts of the same platform (e.g. an EMR job reading Glue tables shared across AWS accounts, or reading both Snowflake and Postgres), one global metadata.dataset.platformInstance is not enough — every upstream would collapse onto the same instance and its URN would not match the one the upstream's own connector emits. Use spark.datahub.metadata.dataset.connections to set the platform_instance (and optionally env) per upstream connection.

The map is keyed by the connection's canonical OpenLineage namespace — the same namespace string that appears in the OpenLineage event for that dataset. The platform is implied by the namespace's scheme, so you never put the platform in the key or value; you only supply platformInstance (and optionally env). This must equal the platform_instance the upstream platform's own DataHub connector uses, otherwise the lineage edge will dangle.

Namespace key format per platform. The key is exactly the namespace string OpenLineage emits for that dataset, which generally follows the OpenLineage dataset naming spec. A few namespaces are produced by the OpenLineage Spark/JDBC integration rather than listed in the naming-spec page (noted below); when in doubt, read the emitted namespace from a debug log (see the last note) and use it verbatim.

Platformconnections key (namespace)Example
AWS Gluearn:aws:glue:{region}:{account}arn:aws:glue:us-east-1:111122223333
Snowflakesnowflake://{org}-{account}snowflake://acme-prod
Redshiftredshift://{cluster}.{region}:{port}redshift://wh.us-east-1:5439
Athena¹awsathena://athena.{region}.amazonaws.comawsathena://athena.us-east-1.amazonaws.com
Postgrespostgres://{host}:{port}postgres://pg.internal:5432
MySQLmysql://{host}:{port}mysql://mysql.internal:3306
MSSQL / SQL Server²sqlserver://{host}:{port}sqlserver://mssql.internal:1433
Hive / Hive metastorehive://{host}:{port}hive://metastore.internal:9083
Trino / Presto¹trino://{host}:{port}trino://trino.internal:8080
Kafkakafka://{bootstrap}:{port}kafka://broker.internal:9092

¹ Not listed on the naming-spec page; the scheme is what the OpenLineage integration emits. ² SQL Server JDBC sources are emitted by OpenLineage as sqlserver://…, not the mssql://… form shown on the naming-spec page — use the sqlserver:// key.

Notes:

  • Glue is keyed by account + region (the ARN authority). Account alone is not unique — the same account can own a Glue catalog in another region. The listener recovers the ARN from the OpenLineage symlink even though the dataset's own namespace is rewritten to glue.
  • Normalize keys the way the namespace is emitted: lowercase host, include the port, no trailing slash. If spark.datahub.metadata.dataset.lowerCaseUrns=true, the namespace is lowercased before lookup, so use lowercase keys.
  • Platforms whose OpenLineage namespace carries no authority are not addressable here and keep their existing mechanisms: BigQuery (namespace is just bigquery; the project is in the dataset name) and S3/GCS (keyed by bucket — use spark.datahub.platform.<p>.path_spec_list above).
  • Not sure of a namespace? Enable debug logging (see Debugging) and read the input/output dataset namespace straight from the emitted OpenLineage event.

Example — an EMR job reading Glue tables from two accounts plus a Snowflake table:

spark.datahub.metadata.dataset.connections."arn:aws:glue:us-east-1:111122223333".platformInstance domain_a
spark.datahub.metadata.dataset.connections."arn:aws:glue:us-east-1:111122223333".env             PROD
spark.datahub.metadata.dataset.connections."arn:aws:glue:us-east-1:444455556666".platformInstance domain_b
spark.datahub.metadata.dataset.connections."snowflake://acme-prod".platformInstance               snow_prod

Important notes on usage​

  • It is advisable to ensure appName is used appropriately to ensure you can trace lineage from a pipeline back to your source code.
  • If multiple apps with the same appName run concurrently, dataset-lineage will be captured correctly but the custom-properties e.g. app-id, SQLQueryId would be unreliable. We expect this to be quite rare.
  • If spark execution fails, then an empty pipeline would still get created, but it may not have any tasks.
  • For HDFS sources, the folder (name) is regarded as the dataset (name) to align with typical storage of parquet/csv formats.

Delta Lake MERGE INTO Commands​

When working with Delta Lake MERGE INTO commands, the default behavior creates generic job names based on the internal Spark task names. To improve lineage tracking, you can enable the enhanced table name extraction feature:

spark.datahub.metadata.dataset.enableEnhancedMergeIntoExtraction=true

When enabled, the agent will:

  1. Detect Delta Lake MERGE INTO commands
  2. Extract the target table name from the SQL query, dataset names, or symlinks
  3. Include the table name in the job name, making it easier to trace operations against specific tables
  4. Generate more meaningful lineage in DataHub

For example, a job named execute_merge_into_command_edge will be enhanced to execute_merge_into_command_edge.database_table_name, making it clear which table was being modified.

Debugging​

  • Following info logs are generated

On Spark context startup

YY/MM/DD HH:mm:ss INFO DatahubSparkListener: DatahubSparkListener initialised.
YY/MM/DD HH:mm:ss INFO SparkContext: Registered listener datahub.spark.DatahubSparkListener

On application start

YY/MM/DD HH:mm:ss INFO DatahubSparkListener: Application started: SparkListenerApplicationStart(AppName,Some(local-1644489736794),1644489735772,user,None,None)
YY/MM/DD HH:mm:ss INFO McpEmitter: REST Emitter Configuration: GMS url <rest.server>
YY/MM/DD HH:mm:ss INFO McpEmitter: REST Emitter Configuration: Token XXXXX

On pushing data to server

YY/MM/DD HH:mm:ss INFO McpEmitter: MetadataWriteResponse(success=true, responseContent={"value":"<URN>"}, underlyingResponse=HTTP/1.1 200 OK [Date: day, DD month year HH:mm:ss GMT, Content-Type: application/json, X-RestLi-Protocol-Version: 2.0.0, Content-Length: 97, Server: Jetty(9.4.46.v20220331)] [Content-Length: 97,Chunked: false])

On application end

YY/MM/DD HH:mm:ss INFO DatahubSparkListener: Application ended : AppName AppID
  • To enable debugging logs, add below configuration in log4j.properties file
log4j.logger.datahub.spark=DEBUG
log4j.logger.datahub.client.rest=DEBUG

Runtime requirements​

The agent shadow jar is compiled to Java 17 bytecode, so it requires a Spark cluster running Java 17 or newer. This matches the current defaults of the major managed Spark platforms:

  • Amazon EMR 7.x (Java 17)
  • Databricks Runtime 16.4 LTS / 17.3 LTS and newer (Java 17; 18+ is Java 21)
  • Dataproc Serverless 2.x / 3.0 and Dataproc on GCE image 3.0
  • Microsoft Fabric runtime 2.0

Java 8 and Java 11 runtimes are not supported (e.g. EMR 6.x, Databricks ≤ 15.4 LTS, AWS Glue ≤ 4.0, Dataproc on GCE images 2.1–2.3, Microsoft Fabric 1.3). Run the agent on a Java 17+ runtime.

How to build​

The build uses Gradle (the JDK 21 toolchain is provisioned automatically) and produces the Java 17-bytecode shadow jar:

./gradlew :metadata-integration:java:acryl-spark-lineage:shadowJar

Known limitations​

Changelog​

Next​

  • Changes:
    • OpenLineage upgrade to 1.50.0: the vendored/shaded OpenLineage classes were refreshed onto the 1.50 upstream base. Iceberg-on-Glue symlink resolution is now provided natively by OpenLineage (since 1.46), so the temporary DataHub Glue-ARN workaround was removed.
    • Apache Spark 4.x support: the agent is verified to attach and emit lineage on Apache Spark 4.0.0 (Scala 2.13) via a dedicated compatibility smoke test. Spark 4 requires the Scala 2.13 build of the agent.
    • OpenLineage dataset-name trimmers auto-disabled when DataHub path trimming is configured: OpenLineage 1.39+ enables built-in partition trimmers by default, which run before — and thus collide with — DataHub's own path handling. The plugin now disables them automatically when a path_spec_list or file_partition_regexp is set (otherwise they stay on, matching upstream). Override with spark.datahub.metadata.dataset.openLineageTrimmersEnabled (true/false).
    • Map jdbc sqlserver dialect to mssql platform otherwise OpenLineage fails to parse the sql
  • Fixes:
    • Listener Null-Safety Fix: the underlying OpenLineage listener is created lazily and can remain uninitialized (listener disabled, no active SparkContext/SparkEnv yet, or config parse failure). The event handlers now no-op in those cases instead of throwing a NullPointerException back into Spark's listener bus.
    • Dependency Relocation Fix (#14989): Fixed shadow JAR packaging to properly relocate all transitive dependencies, preventing classloading conflicts with other Spark extensions. All dependencies except io.openlineage (which contains customized classes) and datahub.spark (the public API) are now properly relocated under io.acryl.shaded namespace. This resolves conflicts with libraries like ANTLR, Apache Avro, and others that could clash with Delta Lake and other Spark components.
    • Missing Output Lineage Fix: Fixed an issue where outputDatasetEdges in the dataJobInputOutput aspect could be empty when using coalesced emission with the REST emitter. Early coalesced emissions (e.g., on START events) sent an UPSERT with empty edge arrays, which clobbered later PATCH emissions that contained actual output edges. The fix skips emitting dataJobInputOutput when all edges are empty.
    • SparkEnv NPE Fix: Added null-check for SparkEnv.get() in PlanUtils.getDirectoryPath() to prevent NullPointerException during Spark shutdown or in test contexts.
    • URISyntaxException Crash Fix: RemovePathPatternUtils.removePathPattern() no longer throws RuntimeException for dataset names that are not valid URIs. Instead, it logs a warning and returns the original name.
    • JDBC Option Guard: SaveIntoDataSourceCommandVisitor now checks that dbtable and url options are present before accessing them, preventing NoSuchElementException for query-based JDBC sinks.

Version 0.2.18​

  • Changes:
    • OpenLineage 1.33.0 upgrade
    • Add spark.datahub.capture_spark_plan option to capture the Spark plan. By default, it is disabled.
    • Add proper support for Spark Streaming
    • Fix issue when Delta table was not within Warehouse location and plugin only captured the path and not the table.
    • Option for Enhanced Merge Into Extraction
    • Fix rdd map detection to correctly handle map transformations in the lineage.
    • JAR Naming: Starting from this version, separate jars are built for different Scala versions:
      • Scala 2.12: io.acryl:acryl-spark-lineage_2.12:0.2.18
      • Scala 2.13: io.acryl:acryl-spark-lineage_2.13:0.2.18
    • Column-level Lineage Enhancement: OpenLineage's transformation types are now captured and mapped to DataHub's FinegrainedLineage TransformOption as per the OpenLineage column lineage specification
    • Dependency Cleanup: Removed logback dependency to reduce potential conflicts with user applications
    • FileStreamMicroBatchStream and foreachBatch for Spark streaming
    • MERGE INTO operations now capture both dataset-level AND column-level lineage

Version 0.2.17​

  • Major changes:

    • Finegrained lineage is emitted on the DataJob and not on the emitted Datasets. This is the correct behaviour which was not correct earlier. This causes earlier emitted finegrained lineages won't be overwritten by the new ones. You can remove the old lineages by setting spark.datahub.legacyLineageCleanup.enabled=true. Make sure you have the latest server if you enable with patch support. (this was introduced since 0.2.17-rc5)

  • Changes:

    • OpenLineage 1.25.0 upgrade
    • Add option to disable chunked encoding in the datahub rest sink -> spark.datahub.rest.disable_chunked_encoding
    • Add option to specify the mcp kafka topic for the datahub kafka sink -> spark.datahub.kafka.mcp_topic
    • Add option to remove legacy lineages from older Spark Plugin runs. This will remove those lineages from the Datasets which it adds to DataJob -> spark.datahub.legacyLineageCleanup.enabled

  • Fixes:

    • Fix handling map transformation in the lineage. Earlier it generated wrong lineage for map transformation.

Version 0.2.16​

  • Remove logging DataHub config into logs

Version 0.2.15​

  • Add Kafka emitter to emit lineage to kafka
  • Add File emitter to emit lineage to file
  • Add S3 emitter to save mcps to s3
  • Upgrading OpenLineage to 1.19.0
  • Renaming project to acryl-datahub-spark-lineage
  • Supporting OpenLineage 1.17+ glue identifier changes
  • Fix handling OpenLineage input/output where wasn't any facet attached

Version 0.2.14​

  • Fix warning about MeterFilter warning from Micrometer

Version 0.2.13​

  • Add kafka emitter to emit lineage to kafka

Version 0.2.12​

  • Silencing some chatty warnings in RddPathUtils

Version 0.2.11​

  • Add option to lowercase dataset URNs
  • Add option to set platform instance and/or env per platform with spark.datahub.platform.<platform_name>.env and spark.datahub.platform.<platform_name>.platform_instance config parameter
  • Fixing platform instance setting for datasets when spark.datahub.metadata.dataset.platformInstance is set
  • Fixing column level lineage support when patch is enabled
💡 Contributing to this documentation

This page is auto-generated from the underlying source code. To make changes, please edit the relevant source files in the metadata-integration directory.

Tip: For quick typo fixes or documentation updates, you can click the ✏️ Edit icon directly in the GitHub UI to open a Pull Request. For larger changes and PR naming conventions, please refer to our Contributing Guide.