Skip to main content
Version: 1.21.1

Extending

The Spark library is intended to support extension via custom implementations of a handful of interfaces. Nearly every extension interface extends or mimics Scala's PartialFunction. The isDefinedAt(Object x) method determines whether a given input is a valid input to the function. A default implementation of isDefinedAt(Object x) is provided, which checks the generic type arguments of the concrete class, if concrete type arguments are given, and determines if the input argument matches the generic type. For example, the following class is automatically defined for an input argument of type MyDataset.

class MyDatasetDetector extends QueryPlanVisitor<MyDataset, OutputDataset> {
}

API​

The following APIs are still evolving and may change over time based on user feedback.

OpenLineageEventHandlerFactory​

This interface defines the main entrypoint to the extension codebase. Custom implementations are registered by following Java's ServiceLoader conventions. A file called io.openlineage.spark.api.OpenLineageEventHandlerFactory must exist in the application or jar's META-INF/service directory. Each line of that file must be the fully qualified class name of a concrete implementation of OpenLineageEventHandlerFactory. More than one implementation can be present in a single file. This might be useful to separate extensions that are targeted toward different environments - e.g., one factory may contain Azure-specific extensions, while another factory may contain GCP extensions.

The OpenLineageEventHandlerFactory interface makes heavy use of default methods. Implementations may override any or all of the following methods

/**
 * Return a collection of QueryPlanVisitors that can generate InputDatasets from a LogicalPlan node
 */
Collection<PartialFunction<LogicalPlan, List<InputDataset>>> createInputDatasetQueryPlanVisitors(OpenLineageContext context);

/**
 * Return a collection of QueryPlanVisitors that can generate OutputDatasets from a LogicalPlan node
 */
Collection<PartialFunction<LogicalPlan, List<OutputDataset>>> createOutputDatasetQueryPlanVisitors(OpenLineageContext context);

/**
 * Return a collection of PartialFunctions that can generate InputDatasets from one of the
 * pre-defined Spark types accessible from SparkListenerEvents (see below)
 */
Collection<PartialFunction<Object, List<InputDataset>>> createInputDatasetBuilder(OpenLineageContext context);

/**
 * Return a collection of PartialFunctions that can generate OutputDatasets from one of the
 * pre-defined Spark types accessible from SparkListenerEvents (see below)
 */
Collection<PartialFunction<Object, List<OutputDataset>>> createOutputDatasetBuilder(OpenLineageContext context);

/**
 * Return a collection of CustomFacetBuilders that can generate InputDatasetFacets from one of the
 * pre-defined Spark types accessible from SparkListenerEvents (see below)
 */
Collection<CustomFacetBuilder<?, ? extends InputDatasetFacet>> createInputDatasetFacetBuilders(OpenLineageContext context);

/**
 * Return a collection of CustomFacetBuilders that can generate OutputDatasetFacets from one of the
 * pre-defined Spark types accessible from SparkListenerEvents (see below)
 */
Collection<CustomFacetBuilder<?, ? extends OutputDatasetFacet>>createOutputDatasetFacetBuilders(OpenLineageContext context);

/**
 * Return a collection of CustomFacetBuilders that can generate DatasetFacets from one of the
 * pre-defined Spark types accessible from SparkListenerEvents (see below)
 */
Collection<CustomFacetBuilder<?, ? extends DatasetFacet>> createDatasetFacetBuilders(OpenLineageContext context);

/**
 * Return a collection of CustomFacetBuilders that can generate RunFacets from one of the
 * pre-defined Spark types accessible from SparkListenerEvents (see below)
 */
Collection<CustomFacetBuilder<?, ? extends RunFacet>> createRunFacetBuilders(OpenLineageContext context);

/**
 * Return a collection of CustomFacetBuilders that can generate JobFacets from one of the
 * pre-defined Spark types accessible from SparkListenerEvents (see below)
 */
Collection<CustomFacetBuilder<?, ? extends JobFacet>> createJobFacetBuilders(OpenLineageContext context);

See the OpenLineageEventHandlerFactory javadocs for specifics on each method.

QueryPlanVisitor​

QueryPlanVisitors evaluate nodes of a Spark LogicalPlan and attempt to generate InputDatasets or OutputDatasets from the information found in the LogicalPlan nodes. This is the most common abstraction present in the OpenLineage Spark library, and many examples can be found in the io.openlineage.spark.agent.lifecycle.plan package - examples include the BigQueryNodeVisitor, the KafkaRelationVisitor and the InsertIntoHiveTableVisitor.

QueryPlanVisitors implement Scala's PartialFunction interface and are tested against every node of a Spark query's optimized LogicalPlan. Each invocation will expect either an InputDataset or an OutputDataset. If a node can be either an InputDataset or an OutputDataset, the constructor should accept a DatasetFactory so that the correct dataset type is generated at runtime.

QueryPlanVisitors can attach facets to the Datasets created, e.g., SchemaDatasetFacet and DatasourceDatasetFacet are typically attached to the dataset when it is created. Custom facets can also be attached, though CustomFacetBuilders may override facets attached directly to the dataset.

Spark job's naming logic appends output dataset's identifier as job suffix. In order to provide a job suffix, a QueryPlanVisitor needs to implement JobNameSuffixProvider interface. Otherwise no suffix will be appended. Job suffix should contain human-readable name of the dataset so that consumers of OpenLineage events can correlate events with particular Spark actions within their code. The logic to extract dataset name should not depend on the existence of the dataset as in case of creating new dataset it may not exist at the moment of assigning job suffix. In most cases, the suffix should contain spark catalog, database and table separated by . which shall be extracted from LogicalPlan nodes properties.

InputDatasetBuilder and OutputDatasetBuilder​

Similar to the QueryPlanVisitors, InputDatasetBuilders and OutputDatasetBuilders are PartialFunctions defined for a specific input (see below for the list of Spark listener events and scheduler objects that can be passed to a builder) that can generate either an InputDataset or an OutputDataset. Though not strictly necessary, the abstract base classes AbstractInputDatasetBuilder and AbstractOutputDatasetBuilder are available for builders to extend.

Spark job's naming logic appends output dataset's identifier as job suffix. In order to provide a job suffix, a OutputDatasetBuilder needs to implement JobNameSuffixProvider interface. Otherwise no suffix will be appended. Job suffix should contain human-readable name of the dataset so that consumers of OpenLineage events can correlate events with particular Spark actions within their code. The logic to extract dataset name should not depend on the existence of the dataset as in case of creating new dataset it may not exist at the moment of assigning job suffix. In most cases, the suffix should contain spark catalog, database and table separated by . which shall be extracted from LogicalPlan nodes properties.

CustomFacetBuilder​

CustomFacetBuilders evaluate Spark event types and scheduler objects (see below) to construct custom facets. CustomFacetBuilders are used to create InputDatsetFacets, OutputDatsetFacets, DatsetFacets, RunFacets, and JobFacets. A few examples can be found in the io.openlineage.spark.agent.facets.builder package, including the ErrorFacetBuilder and the LogicalPlanRunFacetBuilder. CustomFacetBuilders are not PartialFunction implementations, but do define the isDefinedAt(Object) method to determine whether a given input is valid for the function. They implement the BiConsumer interface, accepting the valid input argument, and a BiConsumer<String, Facet> consumer, which accepts the name and value of any custom facet that should be attached to the OpenLineage run. There is no limit to the number of facets that can be reported by a given CustomFacetBuilder. Facet names that conflict will overwrite previously reported facets if they are reported for the same Spark event. Though not strictly necessary, the following abstract base classes are available for extension:

Input/Output/Dataset facets returned are attached to any Input/Output Dataset found for a given Spark event. Typically, a Spark job only has one OutputDataset, so any OutputDatasetFacet generated will be attached to that OutputDataset. However, Spark jobs often have multiple InputDatasets. Typically, an InputDataset is read within a single Spark Stage, and any metrics pertaining to that dataset may be present in the StageInfo#taskMetrics() for that Stage. Accumulators pertaining to a dataset should be reported in the task metrics for a stage so that the CustomFacetBuilder can match against the StageInfo and retrieve the task metrics for that stage when generating the InputDatasetFacet. Other facet information is often found by analyzing the RDD that reads the raw data for a dataset. CustomFacetBuilders that generate these facets should be defined for the specific subclass of RDD that is used to read the target dataset - e.g., HadoopRDD, BigQueryRDD, or JdbcRDD.

Function Argument Types​

CustomFacetBuilders and dataset builders can be defined for the following set of Spark listener event types and scheduler types:

  • org.apache.spark.sql.execution.ui.SparkListenerSQLExecutionStart
  • org.apache.spark.sql.execution.ui.SparkListenerSQLExecutionEnd
  • org.apache.spark.scheduler.SparkListenerJobStart
  • org.apache.spark.scheduler.SparkListenerJobEnd
  • org.apache.spark.rdd.RDD
  • org.apache.spark.scheduler.Stage
  • org.apache.spark.scheduler.StageInfo
  • org.apache.spark.scheduler.ActiveJob

Note that RDDs are "unwrapped" prior to being evaluated by builders, so there's no need to, e.g., check a MapPartitionsRDD's dependencies. The RDD for each Stage can be evaluated when a org.apache.spark.scheduler.SparkListenerStageCompleted event occurs. When a org.apache.spark.scheduler.SparkListenerJobEnd event is encountered, the last Stage for the ActiveJob can be evaluated.

Spark extensions' built-in lineage extraction​

Spark ecosystem comes with a plenty of pluggable extensions like iceberg, delta or spark-bigquery-connector to name a few. Extensions modify logical plan of the job and inject its own classes from which lineage shall be extracted. This is adding extra complexity, as it makes openlineage-spark codebase dependent on the extension packages. The complexity grows more when multiple versions of the same extension need to be supported.

Spark DataSource V2 API Extensions​

Some extensions rely on Spark DataSource V2 API and implement TableProvider, Table, ScanBuilder etc. that are used within Spark to create DataSourceV2Relation instances.

A logical plan node DataSourceV2Relation contains Table field with a properties map of type Map<String, String>. openlineage-spark uses this map to extract dataset information for lineage event from DataSourceV2Relation. It is checking for the properties openlineage.dataset.name and openlineage.dataset.namespace. If they are present, it uses them to identify a dataset. Please be aware that namespace and name need to conform to naming convention.

Properties can be also used to pass any dataset facet. For example:

openlineage.dataset.facets.customFacet={"property1": "value1", "property2": "value2"}

will enrich dataset with customFacet:

"inputs": [{
    "name": "...",
    "namespace": "...",
    "facets": {
        "customFacet": {
            "property1": "value1",
            "property2": "value2",
            "_producer": "..."
        },
        "schema": { }
}]

The approach can be used for standard facets from OpenLineage spec as well. schema does not need to be passed through the properties as it is derived within openlineage-spark from DataSourceV2Relation. Custom facets are automatically filled with _producer field.