This project provides extensions to the Apache Spark project in Scala and Python:
Diff: A diff
transformation and application for Dataset
s that computes the differences between
two datasets, i.e. which rows to add, delete or change to get from one dataset to the other.
SortedGroups: A groupByKey
transformation that groups rows by a key while providing
a sorted iterator for each group. Similar to Dataset.groupByKey.flatMapGroups
, but with order guarantees
for the iterator.
Histogram [*]: A histogram
transformation that computes the histogram DataFrame for a value column.
Global Row Number [*]: A withRowNumbers
transformation that provides the global row number w.r.t.
the current order of the Dataset, or any given order. In contrast to the existing SQL function row_number
, which
requires a window spec, this transformation provides the row number across the entire Dataset without scaling problems.
Partitioned Writing: The writePartitionedBy
action writes your Dataset
partitioned and
efficiently laid out with a single operation.
Inspect Parquet files [*]: The structure of Parquet files (the metadata, not the data stored in Parquet) can be inspected similar to parquet-tools or parquet-cli by reading from a simple Spark data source. This simplifies identifying why some Parquet files cannot be split by Spark into scalable partitions.
Install Python packages into PySpark job [*]: Install Python dependencies via PIP or Poetry programatically into your running PySpark job (PySpark ≥ 3.1.0):
# noinspection PyUnresolvedReferences
from gresearch.spark import *
# using PIP
spark.install_pip_package("pandas==1.4.3", "pyarrow")
spark.install_pip_package("-r", "requirements.txt")
# using Poetry
spark.install_poetry_project("../my-poetry-project/", poetry_python="../venv-poetry/bin/python")
Fluent method call: T.call(transformation: T => R): R
: Turns a transformation T => R
,
that is not part of T
into a fluent method call on T
. This allows writing fluent code like:
import uk.co.gresearch._
i.doThis()
.doThat()
.call(transformation)
.doMore()
Fluent conditional method call: T.when(condition: Boolean).call(transformation: T => T): T
:
Perform a transformation fluently only if the given condition is true.
This allows writing fluent code like:
import uk.co.gresearch._
i.doThis()
.doThat()
.when(condition).call(transformation)
.doMore()
Shortcut for groupBy.as: Calling Dataset.groupBy(Column*).as[K, T]
should be preferred over calling Dataset.groupByKey(V => K)
whenever possible. The former allows Catalyst to exploit
existing partitioning and ordering of the Dataset, while the latter hides from Catalyst which columns are used to create the keys.
This can have a significant performance penalty.
Details:
The new column-expression-based groupByKey[K](Column*)
method makes it easier to group by a column expression key. Instead of
ds.groupBy($"id").as[Int, V]
use:
ds.groupByKey[Int]($"id")
Backticks: backticks(string: String, strings: String*): String)
: Encloses the given column name with backticks (`
) when needed.
This is a handy way to ensure column names with special characters like dots (.
) work with col()
or select()
.
Count null values: count_null(e: Column)
: an aggregation function like count
that counts null values in column e
.
This is equivalent to calling count(when(e.isNull, lit(1)))
.
.Net DateTime.Ticks[*]: Convert .Net (C#, F#, Visual Basic) DateTime.Ticks
into Spark timestamps, seconds and nanoseconds.
Available methods:
// Scala
dotNetTicksToTimestamp(Column): Column // returns timestamp as TimestampType
dotNetTicksToUnixEpoch(Column): Column // returns Unix epoch seconds as DecimalType
dotNetTicksToUnixEpochNanos(Column): Column // returns Unix epoch nanoseconds as LongType
The reverse is provided by (all return LongType
.Net ticks):
// Scala
timestampToDotNetTicks(Column): Column
unixEpochToDotNetTicks(Column): Column
unixEpochNanosToDotNetTicks(Column): Column
These methods are also available in Python:
# Python
dotnet_ticks_to_timestamp(column_or_name) # returns timestamp as TimestampType
dotnet_ticks_to_unix_epoch(column_or_name) # returns Unix epoch seconds as DecimalType
dotnet_ticks_to_unix_epoch_nanos(column_or_name) # returns Unix epoch nanoseconds as LongType
timestamp_to_dotnet_ticks(column_or_name)
unix_epoch_to_dotnet_ticks(column_or_name)
unix_epoch_nanos_to_dotnet_ticks(column_or_name)
Spark temporary directory[*]: Create a temporary directory that will be removed on Spark application shutdown.
Examples:
Scala:
import uk.co.gresearch.spark.createTemporaryDir
val dir = createTemporaryDir("prefix")
Python:
# noinspection PyUnresolvedReferences
from gresearch.spark import *
dir = spark.create_temporary_dir("prefix")
Spark job description[*]: Set Spark job description for all Spark jobs within a context.
Examples:
import uk.co.gresearch.spark._
implicit val session: SparkSession = spark
withJobDescription("parquet file") {
val df = spark.read.parquet("data.parquet")
val count = appendJobDescription("count") {
df.count
}
appendJobDescription("write") {
df.write.csv("data.csv")
}
}
Without job description | With job description |
---|---|
Note that setting a description in one thread while calling the action (e.g. .count
) in a different thread
does not work, unless the different thread is spawned from the current thread after the description has been set.
Working example with parallel collections:
import java.util.concurrent.ForkJoinPool
import scala.collection.parallel.CollectionConverters.seqIsParallelizable
import scala.collection.parallel.ForkJoinTaskSupport
val files = Seq("data1.csv", "data2.csv").par
val counts = withJobDescription("Counting rows") {
// new thread pool required to spawn new threads from this thread
// so that the job description is actually used
files.tasksupport = new ForkJoinTaskSupport(new ForkJoinPool())
files.map(filename => spark.read.csv(filename).count).sum
}(spark)
The spark-extension
package is available for all Spark 3.2, 3.3, 3.4 and 3.5 versions. Some earlier Spark versions may also be supported.
The package version has the following semantics: spark-extension_{SCALA_COMPAT_VERSION}-{VERSION}-{SPARK_COMPAT_VERSION}
:
-
SCALA_COMPAT_VERSION
: Scala binary compatibility (minor) version. Available are2.12
and2.13
. -
SPARK_COMPAT_VERSION
: Apache Spark binary compatibility (minor) version. Available are3.2
,3.3
,3.4
and3.5
. -
VERSION
: The package version, e.g.2.10.0
.
Add this line to your build.sbt
file:
libraryDependencies += "uk.co.gresearch.spark" %% "spark-extension" % "2.13.0-3.5"
Add this dependency to your pom.xml
file:
<dependency>
<groupId>uk.co.gresearch.spark</groupId>
<artifactId>spark-extension_2.12</artifactId>
<version>2.13.0-3.5</version>
</dependency>
Add this dependency to your build.gradle
file:
dependencies {
implementation "uk.co.gresearch.spark:spark-extension_2.12:2.13.0-3.5"
}
Submit your Spark app with the Spark Extension dependency (version ≥1.1.0) as follows:
spark-submit --packages uk.co.gresearch.spark:spark-extension_2.12:2.13.0-3.5 [jar]
Note: Pick the right Scala version (here 2.12) and Spark version (here 3.5) depending on your Spark version.
Launch a Spark Shell with the Spark Extension dependency (version ≥1.1.0) as follows:
spark-shell --packages uk.co.gresearch.spark:spark-extension_2.12:2.13.0-3.5
Note: Pick the right Scala version (here 2.12) and Spark version (here 3.5) depending on your Spark Shell version.
Start a PySpark session with the Spark Extension dependency (version ≥1.1.0) as follows:
from pyspark.sql import SparkSession
spark = SparkSession \
.builder \
.config("spark.jars.packages", "uk.co.gresearch.spark:spark-extension_2.12:2.13.0-3.5") \
.getOrCreate()
Note: Pick the right Scala version (here 2.12) and Spark version (here 3.5) depending on your PySpark version.
Launch the Python Spark REPL with the Spark Extension dependency (version ≥1.1.0) as follows:
pyspark --packages uk.co.gresearch.spark:spark-extension_2.12:2.13.0-3.5
Note: Pick the right Scala version (here 2.12) and Spark version (here 3.5) depending on your PySpark version.
Run your Python scripts that use PySpark via spark-submit
:
spark-submit --packages uk.co.gresearch.spark:spark-extension_2.12:2.13.0-3.5 [script.py]
Note: Pick the right Scala version (here 2.12) and Spark version (here 3.5) depending on your Spark version.
You may want to install the pyspark-extension
python package from PyPi into your development environment.
This provides you code completion, typing and test capabilities during your development phase.
Running your Python application on a Spark cluster will still require one of the above ways to add the Scala package to the Spark environment.
pip install pyspark-extension==2.13.0.3.5
Note: Pick the right Spark version (here 3.5) depending on your PySpark version.
There are plenty of Data Science notebooks around. To use this library, add a jar dependency to your notebook using these Maven coordinates:
uk.co.gresearch.spark:spark-extension_2.12:2.13.0-3.5
Or download the jar and place it on a filesystem where it is accessible by the notebook, and reference that jar file directly.
Check the documentation of your favorite notebook to learn how to add jars to your Spark environment.
Most features are not supported in Python in conjunction with a Spark Connect server. This also holds for Databricks Runtime environment 13.x and above. Details can be found in this blog.
Calling any of those features when connected to a Spark Connect server will raise this error:
This feature is not supported for Spark Connect.
Use a classic connection to a Spark cluster instead.
You can build this project against different versions of Spark and Scala.
If you want to build for a Spark or Scala version different to what is defined in the pom.xml
file, then run
sh set-version.sh [SPARK-VERSION] [SCALA-VERSION]
For example, switch to Spark 3.5.0 and Scala 2.13.8 by running sh set-version.sh 3.5.0 2.13.8
.
Then execute mvn package
to create a jar from the sources. It can be found in target/
.
Run the Scala tests via mvn test
.
In order to run the Python tests, setup a Python environment as follows (replace [SCALA-COMPAT-VERSION]
and [SPARK-COMPAT-VERSION]
with the respective values):
virtualenv -p python3 venv
source venv/bin/activate
pip install -r python/requirements-[SPARK-COMPAT-VERSION]_[SCALA-COMPAT-VERSION].txt
pip install pytest
Run the Python tests via env PYTHONPATH=python:python/test python -m pytest python/test
.
Note: you first have to build the Scala sources.
Run the following sequence of commands in the project root directory:
mkdir -p python/pyspark/jars/
cp -v target/spark-extension_*-*.jar python/pyspark/jars/
pip install build
Then execute python -m build python/
to create a whl from the sources. It can be found in python/dist/
.
-
Guaranteeing in-partition order for partitioned-writing in Apache Spark, Enrico Minack, 20/01/2023:
https://www.gresearch.com/blog/article/guaranteeing-in-partition-order-for-partitioned-writing-in-apache-spark/ -
Un-pivot, sorted groups and many bug fixes: Celebrating the first Spark 3.4 release, Enrico Minack, 21/03/2023:
https://www.gresearch.com/blog/article/un-pivot-sorted-groups-and-many-bug-fixes-celebrating-the-first-spark-3-4-release/ -
A PySpark bug makes co-grouping with window function partition-key-order-sensitive, Enrico Minack, 29/03/2023:
https://www.gresearch.com/blog/article/a-pyspark-bug-makes-co-grouping-with-window-function-partition-key-order-sensitive/ -
Spark’s groupByKey should be avoided – and here’s why, Enrico Minack, 13/06/2023:
https://www.gresearch.com/blog/article/sparks-groupbykey-should-be-avoided-and-heres-why/ -
Inspecting Parquet files with Spark, Enrico Minack, 28/07/2023:
https://www.gresearch.com/blog/article/parquet-files-know-your-scaling-limits/ -
Enhancing Spark’s UI with Job Descriptions, Enrico Minack, 12/12/2023:
https://www.gresearch.com/blog/article/enhancing-sparks-ui-with-job-descriptions/