python/docs/source/tutorial/sql/python_data_source.rst - master - githubmirror/spark

python/docs/source/tutorial/sql/python_data_source.rst
624 строки23 KB
Jitesh Soni
[SPARK-55450][SS][PYTHON][DOCS] Document admission control in PySpark streaming data sources
07 апр 2026, 09:28
07 апр 2026, 09:2898cdaee
..  Licensed to the Apache Software Foundation (ASF) under one
    or more contributor license agreements.  See the NOTICE file
    distributed with this work for additional information
    regarding copyright ownership.  The ASF licenses this file
    to you 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

..    http://www.apache.org/licenses/LICENSE-2.0

..  Unless required by applicable law or agreed to in writing,
    software distributed under the License is distributed on an
    "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
    KIND, either express or implied.  See the License for the
    specific language governing permissions and limitations
    under the License.

======================
Python Data Source API
======================

.. currentmodule:: pyspark.sql

Overview
--------
The Python Data Source API is a new feature introduced in Spark 4.0, enabling developers to read from custom data sources and write to custom data sinks in Python.
This guide provides a comprehensive overview of the API and instructions on how to create, use, and manage Python data sources.

Simple Example: Data Source with Batch Reader
---------------------------------------------
Here's a simple Python data source that generates exactly two rows of synthetic data.
This example demonstrates how to set up a custom data source without using external libraries, focusing on the essentials needed to get it up and running quickly.

**Step 1: Define the data source**

.. code-block:: python

    from typing import Iterator, Tuple

    from pyspark.sql.datasource import DataSource, DataSourceReader, InputPartition
    from pyspark.sql.types import IntegerType, StringType, StructField, StructType

    class SimpleDataSource(DataSource):
        """
        A simple data source for PySpark that generates exactly two rows of synthetic data.
        """

        @classmethod
        def name(cls) -> str:
            return "simple"

        def schema(self) -> StructType:
            return StructType([
                StructField("name", StringType()),
                StructField("age", IntegerType())
            ])

        def reader(self, schema: StructType) -> DataSourceReader:
            return SimpleDataSourceReader()

    class SimpleDataSourceReader(DataSourceReader):

        def read(self, partition: InputPartition) -> Iterator[Tuple]:
            yield ("Alice", 20)
            yield ("Bob", 30)

**Step 2: Register the data source**

.. code-block:: python

    from pyspark.sql import SparkSession

    spark = SparkSession.builder.getOrCreate()

    spark.dataSource.register(SimpleDataSource)

**Step 3: Read from the data source**

.. code-block:: python

    spark.read.format("simple").load().show()

    # +-----+---+
    # | name|age|
    # +-----+---+
    # |Alice| 20|
    # |  Bob| 30|
    # +-----+---+


Comprehensive Example: Data Source with Batch and Streaming Readers and Writers
-------------------------------------------------------------------------------
To create a custom Python data source, you'll need to subclass the :class:`DataSource` base class and implement the necessary methods for reading and writing data.

This example demonstrates creating a simple data source to generate synthetic data using the `faker` library. Ensure the `faker` library is installed and accessible in your Python environment.

Define the Data Source
~~~~~~~~~~~~~~~~~~~~~~

Start by creating a new subclass of :class:`DataSource` with the source name and schema.

In order to be used as source or sink in batch or streaming query, corresponding methods of DataSource needs to be implemented.

Method that needs to be implemented for a capability:

+------------+----------------------+------------------+
|            |       source         |      sink        |
+============+======================+==================+
| batch      | reader()             | writer()         |
+------------+----------------------+------------------+
|            | streamReader()       |                  |
| streaming  | or                   | streamWriter()   |
|            | simpleStreamReader() |                  |
+------------+----------------------+------------------+

.. code-block:: python

    from typing import Union

    from pyspark.sql.datasource import (
        DataSource,
        DataSourceReader,
        DataSourceStreamReader,
        DataSourceStreamWriter,
        DataSourceWriter
    )
    from pyspark.sql.types import StructType

    class FakeDataSource(DataSource):
        """
        A fake data source for PySpark to generate synthetic data using the `faker` library.
        Options:
        - numRows: specify number of rows to generate. Default value is 3.
        """

        @classmethod
        def name(cls) -> str:
            return "fake"

        def schema(self) -> Union[StructType, str]:
            return "name string, date string, zipcode string, state string"

        def reader(self, schema: StructType) -> DataSourceReader:
            return FakeDataSourceReader(schema, self.options)

        def writer(self, schema: StructType, overwrite: bool) -> DataSourceWriter:
            return FakeDataSourceWriter(self.options)

        def streamReader(self, schema: StructType) -> DataSourceStreamReader:
            return FakeStreamReader(schema, self.options)

        def streamWriter(self, schema: StructType, overwrite: bool) -> DataSourceStreamWriter:
            return FakeStreamWriter(self.options)

Implement a Batch Reader
~~~~~~~~~~~~~~~~~~~~~~~~

Define the reader logic to generate synthetic data. Use the `faker` library to populate each field in the schema.

.. code-block:: python

    from typing import Dict

    class FakeDataSourceReader(DataSourceReader):

        def __init__(self, schema: StructType, options: Dict[str, str]):
            self.schema: StructType = schema
            self.options = options

        def read(self, partition):
            from faker import Faker
            fake = Faker()
            # Note: every value in this `self.options` dictionary is a string.
            num_rows = int(self.options.get("numRows", 3))
            for _ in range(num_rows):
                row = []
                for field in self.schema.fields:
                    value = getattr(fake, field.name)()
                    row.append(value)
                yield tuple(row)

Implement a Batch Writer
~~~~~~~~~~~~~~~~~~~~~~~~

Create a fake data source writer that processes each partition of data, counts the rows, and either prints the total count of rows after a successful write or the number of failed tasks if the writing process fails.

.. code-block:: python

    from dataclasses import dataclass
    from typing import Iterator, List

    from pyspark.sql.types import Row
    from pyspark.sql.datasource import DataSource, DataSourceWriter, WriterCommitMessage

    @dataclass
    class SimpleCommitMessage(WriterCommitMessage):
        partition_id: int
        count: int

    class FakeDataSourceWriter(DataSourceWriter):

        def write(self, rows: Iterator[Row]) -> SimpleCommitMessage:
            from pyspark import TaskContext

            context = TaskContext.get()
            partition_id = context.partitionId()
            cnt = sum(1 for _ in rows)
            return SimpleCommitMessage(partition_id=partition_id, count=cnt)

        def commit(self, messages: List[SimpleCommitMessage]) -> None:
            total_count = sum(message.count for message in messages)
            print(f"Total number of rows: {total_count}")

        def abort(self, messages: List[SimpleCommitMessage]) -> None:
            failed_count = sum(message is None for message in messages)
            print(f"Number of failed tasks: {failed_count}")


Implement a Streaming Reader
~~~~~~~~~~~~~~~~~~~~~~~~~~~~

This is a dummy streaming data reader that generates 2 rows in every microbatch. The streamReader instance has an integer offset that increases by 2 in every microbatch.

.. code-block:: python

    class RangePartition(InputPartition):
        def __init__(self, start: int, end: int):
            self.start = start
            self.end = end

    class FakeStreamReader(DataSourceStreamReader):
        def __init__(self, schema, options):
            self.current = 0

        def initialOffset(self) -> dict:
            """
            Return the initial start offset of the reader.
            """
            return {"offset": 0}

        def latestOffset(self) -> dict:
            """
            Return the current latest offset that the next microbatch will read to.
            """
            self.current += 2
            return {"offset": self.current}

        def partitions(self, start: dict, end: dict) -> list[InputPartition]:
            """
            Plans the partitioning of the current microbatch defined by start
            and end offset, it needs to return a sequence of
            :class:`InputPartition` object.
            """
            return [RangePartition(start["offset"], end["offset"])]

        def commit(self, end: dict) -> None:
            """
            This is invoked when the query has finished processing data before end
            offset, this can be used to clean up resource.
            """
            pass

        def read(self, partition) -> Iterator[Tuple]:
            """
            Takes a partition as an input and read an iterator of tuples from the
            data source.
            """
            start, end = partition.start, partition.end
            for i in range(start, end):
                yield (i, str(i))

Alternative: Implement a Simple Streaming Reader
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

If the data source has low throughput and doesn't require partitioning, you can implement SimpleDataSourceStreamReader instead of DataSourceStreamReader.

One of simpleStreamReader() and streamReader() must be implemented for a readable streaming data source. And simpleStreamReader() will only be invoked when streamReader() is not implemented.

.. code-block:: python

    from pyspark.sql.datasource import SimpleDataSourceStreamReader


    class FakeDataSource(DataSource):
        ...

        def simpleStreamReader(self, schema: StructType) -> SimpleDataSourceStreamReader:
            return FakeSimpleStreamReader()

        # omit implementation of streamReader

        ...

This is the same dummy streaming reader that generates 2 rows every batch implemented with SimpleDataSourceStreamReader interface.

.. code-block:: python

    from typing import Iterator, Tuple

    from pyspark.sql.datasource import SimpleDataSourceStreamReader

    class FakeSimpleStreamReader(SimpleDataSourceStreamReader):
        def initialOffset(self) -> dict:
            """
            Return the initial start offset of the reader.
            """
            return {"offset": 0}

        def read(self, start: dict) -> Tuple[Iterator[Tuple], dict]:
            """
            Takes start offset as an input, return an iterator of tuples and
            the end offset (start offset for the next read). The end offset must
            advance past the start offset when returning data; otherwise Spark
            raises a validation exception.
            For example, returning 2 records from start_idx 0 means end should
            be {"offset": 2} (i.e. start + 2).
            When there is no data to read, you may return the same offset as end and
            start, but you must provide an empty iterator.
            """
            start_idx = start["offset"]
            it = iter([(i,) for i in range(start_idx, start_idx + 2)])
            return (it, {"offset": start_idx + 2})

        def readBetweenOffsets(self, start: dict, end: dict) -> Iterator[Tuple]:
            """
            Takes start and end offset as input and read an iterator of data
            deterministically.
            This is called whe query replay batches during restart or after failure.
            """
            start_idx = start["offset"]
            end_idx = end["offset"]
            return iter([(i,) for i in range(start_idx, end_idx)])

        def commit(self, end: dict) -> None:
            """
            This is invoked when the query has finished processing data before end
            offset, this can be used to clean up resource.
            """
            pass

Admission Control for Streaming Readers
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

To limit the amount of data processed per micro-batch, implement
``getDefaultReadLimit()`` and have ``latestOffset(start, limit)`` honor the
engine-provided ``ReadLimit``:

.. code-block:: python

    from pyspark.sql.streaming.datasource import ReadAllAvailable, ReadLimit, ReadMaxRows

    class MyStreamReader(DataSourceStreamReader):

        def getDefaultReadLimit(self) -> ReadLimit:
            """
            Limit each micro-batch to at most 20 rows.

            This value is just an example; in practice, configure the limit
            based on source options (e.g., self.options.get("maxRowsPerBatch")).
            """
            return ReadMaxRows(20)

        def latestOffset(self, start: dict, limit: ReadLimit) -> dict:
            """
            Return the latest offset, respecting the provided limit.
            """
            current = start["offset"]

            if isinstance(limit, ReadMaxRows):
                end = min(current + limit.max_rows, self.max_available)
            elif isinstance(limit, ReadAllAvailable):
                end = self.max_available
            else:
                raise ValueError(f"Unexpected ReadLimit type: {type(limit)}")

            return {"offset": end}

When Spark uses the default ``ReadMaxRows(20)`` limit, each micro-batch
reads at most 20 rows, depending on the available rows. If Spark passes
``ReadAllAvailable``, the reader should return all remaining rows instead.

This is useful for:

- **Controlling data ingestion rate**: Prevent overwhelming downstream systems
- **Memory management**: Limit batch sizes to avoid out-of-memory errors
- **Backpressure handling**: Process data at a sustainable rate

For a complete working example, see:
``examples/src/main/python/sql/streaming/structured_blockchain_admission_control.py``

Implement a Streaming Writer
~~~~~~~~~~~~~~~~~~~~~~~~~~~~

This is a streaming data writer that write the metadata information of each microbatch to a local path.

.. code-block:: python

    from typing import Iterator, List, Optional

    from pyspark.sql import Row
    from pyspark.sql.datasource import DataSourceStreamWriter, WriterCommitMessage


    class SimpleCommitMessage(WriterCommitMessage):
       partition_id: int
       count: int

    class FakeStreamWriter(DataSourceStreamWriter):
       def __init__(self, options):
           self.options = options
           self.path = self.options.get("path")
           assert self.path is not None

       def write(self, iterator: Iterator[Row]) -> WriterCommitMessage:
           """
           Write the data and return the commit message of that partition
           """
           from pyspark import TaskContext
           context = TaskContext.get()
           partition_id = context.partitionId()
           cnt = 0
           for row in iterator:
               cnt += 1
           return SimpleCommitMessage(partition_id=partition_id, count=cnt)

       def commit(self, messages: List[Optional[SimpleCommitMessage]], batchId: int) -> None:
           """
           Receives a sequence of :class:`SimpleCommitMessage` when all write tasks
           succeed and decides what to do with it.
           In this FakeStreamWriter, we write the metadata of the microbatch
           (number of rows and partitions) into a json file inside commit().
           """
           status = dict(num_partitions=len(messages), rows=sum(m.count for m in messages))
           with open(os.path.join(self.path, f"{batchId}.json"), "a") as file:
               file.write(json.dumps(status) + "\n")

       def abort(self, messages: List[Optional[SimpleCommitMessage]], batchId: int) -> None:
           """
           Receives a sequence of :class:`SimpleCommitMessage` from successful tasks
           when some tasks fail and decides what to do with it.
           In this FakeStreamWriter, we write a failure message into a txt file inside abort().
           """
           with open(os.path.join(self.path, f"{batchId}.txt"), "w") as file:
               file.write(f"failed in batch {batchId}")

Serialization Requirement
-------------------------
User defined DataSource, DataSourceReader, DataSourceWriter, DataSourceStreamReader and DataSourceStreamWriter and their methods must be able to be serialized by pickle.

For libraries that are used inside a method, they must be imported inside the method. For example, TaskContext must be imported inside the read() method in the code below.

.. code-block:: python

    def read(self, partition):
        from pyspark import TaskContext
        context = TaskContext.get()

Using a Python Data Source
--------------------------
**Register a Python Data Source**

After defining your data source, it must be registered before usage.

.. code-block:: python

    spark.dataSource.register(FakeDataSource)

**Read From a Python Data Source**

Read from the fake datasource with the default schema and options:

.. code-block:: python

    spark.read.format("fake").load().show()

    # +-----------+----------+-------+-------+
    # |       name|      date|zipcode|  state|
    # +-----------+----------+-------+-------+
    # |Carlos Cobb|2018-07-15|  73003|Indiana|
    # | Eric Scott|1991-08-22|  10085|  Idaho|
    # | Amy Martin|1988-10-28|  68076| Oregon|
    # +-----------+----------+-------+-------+

Read from the fake datasource with a custom schema:

.. code-block:: python

    spark.read.format("fake").schema("name string, company string").load().show()

    # +---------------------+--------------+
    # |name                 |company       |
    # +---------------------+--------------+
    # |Tanner Brennan       |Adams Group   |
    # |Leslie Maxwell       |Santiago Group|
    # |Mrs. Jacqueline Brown|Maynard Inc   |
    # +---------------------+--------------+

Read from the fake datasource with a different number of rows:

.. code-block:: python

    spark.read.format("fake").option("numRows", 5).load().show()

    # +--------------+----------+-------+------------+
    # |          name|      date|zipcode|       state|
    # +--------------+----------+-------+------------+
    # |  Pam Mitchell|1988-10-20|  23788|   Tennessee|
    # |Melissa Turner|1996-06-14|  30851|      Nevada|
    # |  Brian Ramsey|2021-08-21|  55277|  Washington|
    # |  Caitlin Reed|1983-06-22|  89813|Pennsylvania|
    # | Douglas James|2007-01-18|  46226|     Alabama|
    # +--------------+----------+-------+------------+

**Write To a Python Data Source**

To write data to a custom location, make sure that you specify the `mode()` clause. Supported modes are `append` and `overwrite`.

.. code-block:: python

    df = spark.range(0, 10, 1, 5)
    df.write.format("fake").mode("append").save()

    # You can check the Spark log (standard error) to see the output of the write operation.
    # Total number of rows: 10

**Use a Python Data Source in Streaming Query**

Once we register the python data source, we can also use it in streaming queries as source of readStream() or sink of writeStream() by passing short name or full name to format().

Start a query that read from fake python data source and write to console

.. code-block:: python

    query = spark.readStream.format("fake").load().writeStream.format("console").start()

    # +---+
    # | id|
    # +---+
    # |  0|
    # |  1|
    # +---+
    # +---+
    # | id|
    # +---+
    # |  2|
    # |  3|
    # +---+

We can also use the same data source in streaming reader and writer

.. code-block:: python

    query = spark.readStream.format("fake").load().writeStream.format("fake").start("/output_path")

Python Data Source Reader with direct Arrow Batch support for improved performance
----------------------------------------------------------------------------------
The Python Datasource Reader supports direct yielding of Arrow Batches, which can significantly improve data processing performance. By using the efficient Arrow format,
this feature avoids the overhead of traditional row-by-row data processing, resulting in performance improvements of up to one order of magnitude, especially with large datasets.

**Enabling Arrow Batch Support**:
To enable this feature, configure your custom DataSource to yield Arrow batches by returning `pyarrow.RecordBatch` objects within the `read` method of your `DataSourceReader`
(or `DataSourceStreamReader`) implementation. This method simplifies data handling and reduces the number of I/O operations, particularly beneficial for large-scale data processing tasks.

**Arrow Batch Example**:
The following example demonstrates how to implement a basic Data Source using Arrow Batch support.

.. code-block:: python

    from pyspark.sql.datasource import DataSource, DataSourceReader, InputPartition
    from pyspark.sql import SparkSession
    import pyarrow as pa

    # Define the ArrowBatchDataSource
    class ArrowBatchDataSource(DataSource):
        """
        A Data Source for testing Arrow Batch Serialization
        """

        @classmethod
        def name(cls):
            return "arrowbatch"

        def schema(self):
            return "key int, value string"

        def reader(self, schema: str):
            return ArrowBatchDataSourceReader(schema, self.options)

    # Define the ArrowBatchDataSourceReader
    class ArrowBatchDataSourceReader(DataSourceReader):
        def __init__(self, schema, options):
            self.schema: str = schema
            self.options = options

        def read(self, partition):
            # Create Arrow Record Batch
            keys = pa.array([1, 2, 3, 4, 5], type=pa.int32())
            values = pa.array(["one", "two", "three", "four", "five"], type=pa.string())
            schema = pa.schema([("key", pa.int32()), ("value", pa.string())])
            record_batch = pa.RecordBatch.from_arrays([keys, values], schema=schema)
            yield record_batch

        def partitions(self):
            # Define the number of partitions
            num_part = 1
            return [InputPartition(i) for i in range(num_part)]

    # Initialize the Spark Session
    spark = SparkSession.builder.appName("ArrowBatchExample").getOrCreate()

    # Register the ArrowBatchDataSource
    spark.dataSource.register(ArrowBatchDataSource)

    # Load data using the custom data source
    df = spark.read.format("arrowbatch").load()

    df.show()

Usage Notes
-----------

- During Data Source resolution, built-in and Scala/Java Data Sources take precedence over Python Data Sources with the same name; to explicitly use a Python Data Source, make sure its name does not conflict with the other non-Python Data Sources.
- It is allowed to register multiple Python Data Sources with the same name. Later registrations will overwrite earlier ones.
- To automatically register a data source, export it as ``DefaultSource`` in a top level module with name prefix ``pyspark_``. See `pyspark_huggingface <https://github.com/huggingface/pyspark_huggingface>`_ for an example.