DataSource — Pluggable Data Provider Framework

DataSource is among the main components of Data Source API in Spark SQL (together with DataFrameReader for loading datasets, DataFrameWriter for saving datasets and StreamSourceProvider for creating streaming sources).

DataSource models a pluggable data provider framework with extension points for Spark SQL integrators to expand the list of supported external data sources in Spark SQL.

Table 1. DataSource’s Provider (and Format) Contracts
Extension Point Description


Data source that saves the result of a structured query per save mode and returns the schema


Used in:

  • sourceSchema for streamed reading

  • write for writing a DataFrame to a DataSource (as part of creating a table as select)


Data source that supports schema inference and can be accessed using SQL’s USING clause


Data source that requires a user-defined schema


Used in:

As a user, you interact with DataSource by DataFrameReader (when you execute or spark.readStream) or SQL’s CREATE TABLE USING.

// Batch reading
val people: DataFrame =

// Streamed reading
val messages: DataFrame = spark.readStream
  .option("subscribe", "topic")
  .option("kafka.bootstrap.servers", "localhost:9092")

DataSource uses a SparkSession, a class name, a collection of paths, optional user-specified schema, a collection of partition columns, a bucket specification, and configuration options.

Data source is also called a table provider.

Writing DataFrame to Data Source per Save Mode Followed by Reading Rows Back (as BaseRelation) — writeAndRead Method

writeAndRead(mode: SaveMode, data: DataFrame): BaseRelation
writeAndRead is used exclusively when CreateDataSourceTableAsSelectCommand is executed.

providingClass Property


Writing DataFrame to Data Source Per Save Mode — write Method

write(mode: SaveMode, data: DataFrame): BaseRelation

write writes the result of executing a structured query (as DataFrame) to a data source per save mode.

Internally, write looks up the data source and branches off per providingClass.

Table 2. write’s Branches per Supported providingClass (in execution order)
providingClass Description


Executes CreatableRelationProvider.createRelation




Reports a RuntimeException

write does not support the internal CalendarIntervalType in the schema of data DataFrame and throws a AnalysisException when there is one.
write is used exclusively when SaveIntoDataSourceCommand is executed.

writeInFileFormat Internal Method


For FileFormat data sources, write takes all paths and path option and makes sure that there is only one.

write uses Hadoop’s Path to access the FileSystem and calculate the qualified output path.

write does PartitioningUtils.validatePartitionColumn.

FIXME What is PartitioningUtils.validatePartitionColumn for?

When appending to a table, …​FIXME

In the end, write (for a FileFormat data source) prepares a InsertIntoHadoopFsRelationCommand logical plan with executes it.

FIXME Is toRdd a job execution?

createSource Method

createSource(metadataPath: String): Source

createSink Method


Creating DataSource Instance

class DataSource(
  sparkSession: SparkSession,
  className: String,
  paths: Seq[String] = Nil,
  userSpecifiedSchema: Option[StructType] = None,
  partitionColumns: Seq[String] = Seq.empty,
  bucketSpec: Option[BucketSpec] = None,
  options: Map[String, String] = Map.empty,
  catalogTable: Option[CatalogTable] = None)

When being created, DataSource first looks up the providing class given className (considering it an alias or a fully-qualified class name) and computes the name and schema of the data source.

DataSource does the initialization lazily on demand and only once.

sourceSchema Internal Method

sourceSchema(): SourceInfo

sourceSchema returns the name and schema of the data source for streamed reading.

FIXME Why is the method called? Why does this bother with streamed reading and data sources?!

It supports two class hierarchies, i.e. FileFormat and Structured Streaming’s StreamSourceProvider data sources.

Internally, sourceSchema first creates an instance of the data source and…​

FIXME Finish…​

For Structured Streaming’s StreamSourceProvider data sources, sourceSchema relays calls to StreamSourceProvider.sourceSchema.

For FileFormat data sources, sourceSchema makes sure that path option was specified.

path is looked up in a case-insensitive way so paTh and PATH and pAtH are all acceptable. Use the lower-case version of path, though.
path can use glob pattern (not regex syntax), i.e. contain any of {}[]*?\ characters.

It checks whether the path exists if a glob pattern is not used. In case it did not exist you will see the following AnalysisException exception in the logs:

org.apache.spark.sql.AnalysisException: Path does not exist: file:/Users/jacek/dev/oss/spark/the.file.does.not.exist.parquet;
  at org.apache.spark.sql.execution.datasources.DataSource$$anonfun$12.apply(DataSource.scala:375)
  at org.apache.spark.sql.execution.datasources.DataSource$$anonfun$12.apply(DataSource.scala:364)
  at scala.collection.TraversableLike$$anonfun$flatMap$1.apply(TraversableLike.scala:241)
  at scala.collection.TraversableLike$$anonfun$flatMap$1.apply(TraversableLike.scala:241)
  at scala.collection.immutable.List.foreach(List.scala:381)
  at scala.collection.TraversableLike$class.flatMap(TraversableLike.scala:241)
  at scala.collection.immutable.List.flatMap(List.scala:344)
  at org.apache.spark.sql.execution.datasources.DataSource.resolveRelation(DataSource.scala:364)
  at org.apache.spark.sql.DataFrameReader.load(DataFrameReader.scala:149)
  at org.apache.spark.sql.DataFrameReader.load(DataFrameReader.scala:132)
  ... 48 elided

If spark.sql.streaming.schemaInference is disabled and the data source is different than TextFileFormat, and the input userSpecifiedSchema is not specified, the following IllegalArgumentException exception is thrown:

Schema must be specified when creating a streaming source DataFrame. If some files already exist in the directory, then depending on the file format you may be able to create a static DataFrame on that directory with '' and infer schema from it.
FIXME I don’t think the exception will ever happen for non-streaming sources since the schema is going to be defined earlier. When?

Eventually, it returns a SourceInfo with FileSource[path] and the schema (as calculated using the inferFileFormatSchema internal method).

For any other data source, it throws UnsupportedOperationException exception:

Data source [className] does not support streamed reading

inferFileFormatSchema Internal Method

inferFileFormatSchema(format: FileFormat): StructType

inferFileFormatSchema private method computes (aka infers) schema (as StructType). It returns userSpecifiedSchema if specified or uses FileFormat.inferSchema. It throws a AnalysisException when is unable to infer schema.

It uses path option for the list of directory paths.

It is used by DataSource.sourceSchema and DataSource.createSource when FileFormat is processed.

lookupDataSource Internal Method

lookupDataSource(provider0: String): Class[_]

Internally, lookupDataSource first searches the classpath for available DataSourceRegister providers (using Java’s ServiceLoader.load method) to find the requested data source by short name (alias), e.g. parquet or kafka.

If a DataSource could not be found by short name, lookupDataSource tries to load the class given the input provider0 or its variant provider0.DefaultSource (with .DefaultSource suffix).

You can reference your own custom DataSource in your code by DataFrameWriter.format method which is the alias or fully-qualified class name.

There has to be one data source registered only or you will see the following RuntimeException:

Multiple sources found for [provider] ([comma-separated class names]), please specify the fully qualified class name.

Creating BaseRelation — resolveRelation Method

resolveRelation(checkFilesExist: Boolean = true): BaseRelation

resolveRelation resolves (i.e. creates) a BaseRelation.

Internally, resolveRelation creates an instance of providingClass (of a DataSource) and branches off per its type, i.e. SchemaRelationProvider, RelationProvider or FileFormat.

Table 3. Resolving BaseRelation per Providers
Provider Behaviour


Executes SchemaRelationProvider.createRelation with the provided schema


Executes RelationProvider.createRelation


Creates a HadoopFsRelation


resolveRelation is used when:

results matching ""

    No results matching ""