[SPARK-26413] SPIP: RDD Arrow Support in Spark Core and PySpark

• Type: Improvement
• Status: Open
• Priority: Minor
• Resolution: Unresolved
• Affects Version/s: 3.1.0
• Component/s: PySpark, Spark Core
• Assignee: Unassigned
• Reporter: Richard Whitcomb
• Created: 19/Dec/18 18:40
• Updated: 12/Dec/22 18:10

Description

Background and Motivation

Arrow is becoming an standard interchange format for columnar Structured Data.  This is already true in Spark with the use of arrow in the pandas udf functions in the dataframe API.

However the current implementation of arrow in spark is limited to two use cases.

• Pandas UDF that allows for operations on one or more columns in the DataFrame API.
• Collect as Pandas which pulls back the entire dataset to the driver in a Pandas Dataframe.

What is still hard however is making use of all of the columns in a Dataframe while staying distributed across the workers.  The only way to do this currently is to drop down into RDDs and collect the rows into a dataframe. However pickling is very slow and the collecting is expensive.

The proposal is to extend spark in a way that allows users to operate on an Arrow Table fully while still making use of Spark's underlying technology.  Some examples of possibilities with this new API. 

• Pass the Arrow Table with Zero Copy to PyTorch for predictions.
• Pass to Nvidia Rapids for an algorithm to be run on the GPU.
• Distribute data across many GPUs making use of the new Barriers API.

Targets users and personas

ML, Data Scientists, and future library authors..

Goals

• Conversion from any Dataset[Row] or PySpark Dataframe to RDD[Table]
• Conversion back from any RDD[Table] to Dataset[Row], RDD[Row], Pyspark Dataframe
• Open the possibilities to tighter integration between Arrow/Pandas/Spark especially at a library level.

Non-Goals

• Not creating a new API but instead using existing APIs.

Proposed API changes

Data Objects

case class ArrowTable(schema: Schema, batches: Iterable[ArrowRecordBatch])

Dataset.scala

// Converts a Dataset to an RDD of Arrow Tables
// Each RDD row is an Interable of Arrow Batches.
def arrowRDD: RDD[ArrowTable]
 
// Utility Function to convert to RDD Arrow Table for PySpark
private[sql] def javaToPythonArrow: JavaRDD[Array[Byte]]

RDD.scala

 // Converts RDD[ArrowTable] to an Dataframe by inspecting the Arrow Schema
 def arrowToDataframe(implicit ev: T <:< ArrowTable): Dataframe
  
 // Converts RDD[ArrowTable] to an RDD of Rows
 def arrowToRDD(implicit ev: T <:< ArrowTable): RDD[Row]

Serializers.py

# Serializer that takes a Serialized Arrow Tables and returns a pyarrow Table.
class ArrowSerializer(FramedSerializer)

RDD.py

# New RDD Class that has an RDD[ArrowTable] behind it and uses the new ArrowSerializer instead of the normal Pickle Serializer
class ArrowRDD(RDD)

 

Dataframe.py

// New Function that converts a pyspark dataframe into an ArrowRDD
def arrow(self):

 

Example API Usage

Pyspark

# Select a Single Column Using Pandas
def map_table(arrow_table):
  import pyarrow as pa
  pdf = arrow_table.to_pandas()
  pdf = pdf[['email']]
  return pa.Table.from_pandas(pdf)
# Convert to Arrow RDD, map over tables, convert back to dataframe
df.arrow.map(map_table).dataframe 

Scala

 

// Find N Centroids using Cuda Rapids kMeans
def runCuKmeans(table: ArrowTable, clusters: Int): ArrowTable
 
// Convert Dataset[Row] to RDD[ArrowTable] and back to Dataset[Row]
df.arrowRDD.map(table => runCuKmeans(table, N)).arrowToDataframe.show(10)

 

Implementation Details

As mentioned in the first section, the goal is to make it easier for Spark users to interact with Arrow tools and libraries.  This however does come with some considerations from a Spark perspective.

 Arrow is column based instead of Row based.  In the above API proposal of RDD[ArrowTable] each RDD row will in fact be a block of data.  Another proposal in this regard is to introduce a new parameter to Spark called arrow.sql.execution.arrow.maxRecordsPerTable.  The goal of this parameter is to decide how many records are included in a single Arrow Table.  If set to -1 the entire partition will be included in the table else to that number. Within that number the normal blocking mechanisms of Arrow is used to include multiple batches.  This is still dictated by arrowMaxRecordsPerBatch.

Issue Links

• is related to: SPARK-24579 SPIP: Standardize Optimized Data Exchange between Spark and DL/AI frameworks
• is related to: SPARK-27396 SPIP: Public APIs for extended Columnar Processing Support
• is related to: SPARK-26412 Allow Pandas UDF to take an iterator of pd.DataFrames