Spark 4.0.1 works with Python 3.9+. It can use the standard CPython interpreter,so C libraries like NumPy can be used. It also works with PyPy 7.3.6+.

Spark applications in Python can either be run with thebin/spark-submit script which includes Spark at runtime, or by including it in your setup.py as:

install_requires=['pyspark==4.0.1']

To run Spark applications in Python without pip installing PySpark, use thebin/spark-submit script located in the Spark directory.This script will load Spark’s Java/Scala libraries and allow you to submit applications to a cluster.You can also usebin/pyspark to launch an interactive Python shell.

If you wish to access HDFS data, you need to use a build of PySpark linkingto your version of HDFS.Prebuilt packages are also available on the Spark homepagefor common HDFS versions.

Finally, you need to import some Spark classes into your program. Add the following line:

frompysparkimportSparkContext,SparkConf

PySpark requires the same minor version of Python in both driver and workers. It uses the default python version in PATH,you can specify which version of Python you want to use byPYSPARK_PYTHON, for example:

$ PYSPARK_PYTHON=python3.8 bin/pyspark$ PYSPARK_PYTHON=/path-to-your-pypy/pypy bin/spark-submit examples/src/main/python/pi.py

Spark 4.0.1 is built and distributed to work with Scala 2.13by default. (Spark can be built to work with other versions of Scala, too.) To writeapplications in Scala, you will need to use a compatible Scala version (e.g. 2.13.X).

To write a Spark application, you need to add a Maven dependency on Spark. Spark is available through Maven Central at:

groupId = org.apache.sparkartifactId = spark-core_2.13version = 4.0.1

In addition, if you wish to access an HDFS cluster, you need to add a dependency onhadoop-client for your version of HDFS.

groupId = org.apache.hadoopartifactId = hadoop-clientversion = <your-hdfs-version>

Finally, you need to import some Spark classes into your program. Add the following lines:

importorg.apache.spark.SparkContextimportorg.apache.spark.SparkConf

(Before Spark 1.3.0, you need to explicitlyimport org.apache.spark.SparkContext._ to enable essential implicit conversions.)

Spark 4.0.1 supportslambda expressionsfor concisely writing functions, otherwise you can use the classes in theorg.apache.spark.api.java.function package.

Note that support for Java 7 was removed in Spark 2.2.0.

To write a Spark application in Java, you need to add a dependency on Spark. Spark is available through Maven Central at:

groupId = org.apache.sparkartifactId = spark-core_2.13version = 4.0.1

In addition, if you wish to access an HDFS cluster, you need to add a dependency onhadoop-client for your version of HDFS.

groupId = org.apache.hadoopartifactId = hadoop-clientversion = <your-hdfs-version>

Finally, you need to import some Spark classes into your program. Add the following lines:

importorg.apache.spark.api.java.JavaSparkContext;importorg.apache.spark.api.java.JavaRDD;importorg.apache.spark.SparkConf;

The first thing a Spark program must do is to create aSparkContext object, which tells Sparkhow to access a cluster. To create aSparkContext you first need to build aSparkConf objectthat contains information about your application.

conf=SparkConf().setAppName(appName).setMaster(master)sc=SparkContext(conf=conf)

The first thing a Spark program must do is to create aSparkContext object, which tells Sparkhow to access a cluster. To create aSparkContext you first need to build aSparkConf objectthat contains information about your application.

Only one SparkContext should be active per JVM. You muststop() the active SparkContext before creating a new one.

valconf=newSparkConf().setAppName(appName).setMaster(master)newSparkContext(conf)

The first thing a Spark program must do is to create aJavaSparkContext object, which tells Sparkhow to access a cluster. To create aSparkContext you first need to build aSparkConf objectthat contains information about your application.

SparkConfconf=newSparkConf().setAppName(appName).setMaster(master);JavaSparkContextsc=newJavaSparkContext(conf);

In the PySpark shell, a special interpreter-aware SparkContext is already created for you, in thevariable calledsc. Making your own SparkContext will not work. You can set which master thecontext connects to using the--master argument, and you can add Python .zip, .egg or .py filesto the runtime path by passing a comma-separated list to--py-files. For third-party Python dependencies,seePython Package Management. You can also add dependencies(e.g. Spark Packages) to your shell session by supplying a comma-separated list of Maven coordinatesto the--packages argument. Any additional repositories where dependencies might exist (e.g. Sonatype)can be passed to the--repositories argument. For example, to runbin/pyspark on exactly four cores, use:

$./bin/pyspark--master"local[4]"

Or, to also addcode.py to the search path (in order to later be able toimport code), use:

$./bin/pyspark--master"local[4]"--py-files code.py

For a complete list of options, runpyspark --help. Behind the scenes,pyspark invokes the more generalspark-submit script.

It is also possible to launch the PySpark shell inIPython, theenhanced Python interpreter. PySpark works with IPython 1.0.0 and later. Touse IPython, set thePYSPARK_DRIVER_PYTHON variable toipython when runningbin/pyspark:

$ PYSPARK_DRIVER_PYTHON=ipython ./bin/pyspark

To use the Jupyter notebook (previously known as the IPython notebook),

$ PYSPARK_DRIVER_PYTHON=jupyterPYSPARK_DRIVER_PYTHON_OPTS=notebook ./bin/pyspark

You can customize theipython orjupyter commands by settingPYSPARK_DRIVER_PYTHON_OPTS.

After the Jupyter Notebook server is launched, you can create a new notebook fromthe “Files” tab. Inside the notebook, you can input the command%pylab inline as part ofyour notebook before you start to try Spark from the Jupyter notebook.

In the Spark shell, a special interpreter-aware SparkContext is already created for you, in thevariable calledsc. Making your own SparkContext will not work. You can set which master thecontext connects to using the--master argument, and you can add JARs to the classpathby passing a comma-separated list to the--jars argument. You can also add dependencies(e.g. Spark Packages) to your shell session by supplying a comma-separated list of Maven coordinatesto the--packages argument. Any additional repositories where dependencies might exist (e.g. Sonatype)can be passed to the--repositories argument. For example, to runbin/spark-shell on exactlyfour cores, use:

$./bin/spark-shell--master"local[4]"

Or, to also addcode.jar to its classpath, use:

$./bin/spark-shell--master"local[4]"--jars code.jar

To include a dependency using Maven coordinates:

$./bin/spark-shell--master"local[4]"--packages"org.example:example:0.1"

For a complete list of options, runspark-shell --help. Behind the scenes,spark-shell invokes the more generalspark-submit script.

Parallelized collections are created by callingSparkContext’sparallelize method on an existing iterable or collection in your driver program. The elements of the collection are copied to form a distributed dataset that can be operated on in parallel. For example, here is how to create a parallelized collection holding the numbers 1 to 5:

data=[1,2,3,4,5]distData=sc.parallelize(data)

Once created, the distributed dataset (distData) can be operated on in parallel. For example, we can calldistData.reduce(lambda a, b: a + b) to add up the elements of the list.We describe operations on distributed datasets later on.

Parallelized collections are created by callingSparkContext’sparallelize method on an existing collection in your driver program (a ScalaSeq). The elements of the collection are copied to form a distributed dataset that can be operated on in parallel. For example, here is how to create a parallelized collection holding the numbers 1 to 5:

valdata=Array(1,2,3,4,5)valdistData=sc.parallelize(data)

Once created, the distributed dataset (distData) can be operated on in parallel. For example, we might calldistData.reduce((a, b) => a + b) to add up the elements of the array. We describe operations on distributed datasets later on.

Parallelized collections are created by callingJavaSparkContext’sparallelize method on an existingCollection in your driver program. The elements of the collection are copied to form a distributed dataset that can be operated on in parallel. For example, here is how to create a parallelized collection holding the numbers 1 to 5:

List<Integer>data=Arrays.asList(1,2,3,4,5);JavaRDD<Integer>distData=sc.parallelize(data);

Once created, the distributed dataset (distData) can be operated on in parallel. For example, we might calldistData.reduce((a, b) -> a + b) to add up the elements of the list.We describe operations on distributed datasets later on.

PySpark can create distributed datasets from any storage source supported by Hadoop, including your local file system, HDFS, Cassandra, HBase,Amazon S3, etc. Spark supports text files,SequenceFiles, and any other HadoopInputFormat.

Text file RDDs can be created usingSparkContext’stextFile method. This method takes a URI for the file (either a local path on the machine, or ahdfs://,s3a://, etc URI) and reads it as a collection of lines. Here is an example invocation:

>>>distFile=sc.textFile("data.txt")

Once created,distFile can be acted on by dataset operations. For example, we can add up the sizes of all the lines using themap andreduce operations as follows:distFile.map(lambda s: len(s)).reduce(lambda a, b: a + b).

Some notes on reading files with Spark:

• If using a path on the local filesystem, the file must also be accessible at the same path on worker nodes. Either copy the file to all workers or use a network-mounted shared file system.

• All of Spark’s file-based input methods, includingtextFile, support running on directories, compressed files, and wildcards as well. For example, you can usetextFile("/my/directory"),textFile("/my/directory/*.txt"), andtextFile("/my/directory/*.gz").

• ThetextFile method also takes an optional second argument for controlling the number of partitions of the file. By default, Spark creates one partition for each block of the file (blocks being 128MB by default in HDFS), but you can also ask for a higher number of partitions by passing a larger value. Note that you cannot have fewer partitions than blocks.

Apart from text files, Spark’s Python API also supports several other data formats:

• SparkContext.wholeTextFiles lets you read a directory containing multiple small text files, and returns each of them as (filename, content) pairs. This is in contrast withtextFile, which would return one record per line in each file.

• RDD.saveAsPickleFile andSparkContext.pickleFile support saving an RDD in a simple format consisting of pickled Python objects. Batching is used on pickle serialization, with default batch size 10.

• SequenceFile and Hadoop Input/Output Formats

Note this feature is currently markedExperimental and is intended for advanced users. It may be replaced in future with read/write support based on Spark SQL, in which case Spark SQL is the preferred approach.

Writable Support

PySpark SequenceFile support loads an RDD of key-value pairs within Java, converts Writables to base Java types, and pickles theresulting Java objects usingpickle. When saving an RDD of key-value pairs to SequenceFile,PySpark does the reverse. It unpickles Python objects into Java objects and then converts them to Writables. The followingWritables are automatically converted:

Arrays are not handled out-of-the-box. Users need to specify customArrayWritable subtypes when reading or writing. When writing,users also need to specify custom converters that convert arrays to customArrayWritable subtypes. When reading, the defaultconverter will convert customArrayWritable subtypes to JavaObject[], which then get pickled to Python tuples. To getPythonarray.array for arrays of primitive types, users need to specify custom converters.

Saving and Loading SequenceFiles

Similarly to text files, SequenceFiles can be saved and loaded by specifying the path. The key and valueclasses can be specified, but for standard Writables this is not required.

>>>rdd=sc.parallelize(range(1,4)).map(lambdax:(x,"a"*x))>>>rdd.saveAsSequenceFile("path/to/file")>>>sorted(sc.sequenceFile("path/to/file").collect())[(1,u'a'),(2,u'aa'),(3,u'aaa')]

Saving and Loading Other Hadoop Input/Output Formats

PySpark can also read any Hadoop InputFormat or write any Hadoop OutputFormat, for both ‘new’ and ‘old’ Hadoop MapReduce APIs.If required, a Hadoop configuration can be passed in as a Python dict. Here is an example using theElasticsearch ESInputFormat:

$./bin/pyspark--jars/path/to/elasticsearch-hadoop.jar>>>conf={"es.resource":"index/type"}# assume Elasticsearch is running on localhost defaults>>>rdd=sc.newAPIHadoopRDD("org.elasticsearch.hadoop.mr.EsInputFormat","org.apache.hadoop.io.NullWritable","org.elasticsearch.hadoop.mr.LinkedMapWritable",conf=conf)>>>rdd.first()# the result is a MapWritable that is converted to a Python dict(u'Elasticsearch ID',{u'field1':True,u'field2':u'Some Text',u'field3':12345})

Note that, if the InputFormat simply depends on a Hadoop configuration and/or input path, andthe key and value classes can easily be converted according to the above table,then this approach should work well for such cases.

If you have custom serialized binary data (such as loading data from Cassandra / HBase), then you will first need totransform that data on the Scala/Java side to something which can be handled by pickle’s pickler.AConverter trait is providedfor this. Simply extend this trait and implement your transformation code in theconvertmethod. Remember to ensure that this class, along with any dependencies required to access yourInputFormat, are packaged into your Spark job jar and included on the PySparkclasspath.

See thePython examples andtheConverter examplesfor examples of using Cassandra / HBaseInputFormat andOutputFormat with custom converters.

Spark can create distributed datasets from any storage source supported by Hadoop, including your local file system, HDFS, Cassandra, HBase,Amazon S3, etc. Spark supports text files,SequenceFiles, and any other HadoopInputFormat.

Text file RDDs can be created usingSparkContext’stextFile method. This method takes a URI for the file (either a local path on the machine, or ahdfs://,s3a://, etc URI) and reads it as a collection of lines. Here is an example invocation:

scala>valdistFile=sc.textFile("data.txt")distFile:org.apache.spark.rdd.RDD[String]=data.txtMapPartitionsRDD[10]attextFileat<console>:26

Once created,distFile can be acted on by dataset operations. For example, we can add up the sizes of all the lines using themap andreduce operations as follows:distFile.map(s => s.length).reduce((a, b) => a + b).

Some notes on reading files with Spark:

• If using a path on the local filesystem, the file must also be accessible at the same path on worker nodes. Either copy the file to all workers or use a network-mounted shared file system.

• All of Spark’s file-based input methods, includingtextFile, support running on directories, compressed files, and wildcards as well. For example, you can usetextFile("/my/directory"),textFile("/my/directory/*.txt"), andtextFile("/my/directory/*.gz"). When multiple files are read, the order of the partitions depends on the order the files are returned from the filesystem. It may or may not, for example, follow the lexicographic ordering of the files by path. Within a partition, elements are ordered according to their order in the underlying file.

• ThetextFile method also takes an optional second argument for controlling the number of partitions of the file. By default, Spark creates one partition for each block of the file (blocks being 128MB by default in HDFS), but you can also ask for a higher number of partitions by passing a larger value. Note that you cannot have fewer partitions than blocks.

Apart from text files, Spark’s Scala API also supports several other data formats:

• SparkContext.wholeTextFiles lets you read a directory containing multiple small text files, and returns each of them as (filename, content) pairs. This is in contrast withtextFile, which would return one record per line in each file. Partitioning is determined by data locality which, in some cases, may result in too few partitions. For those cases,wholeTextFiles provides an optional second argument for controlling the minimal number of partitions.

• ForSequenceFiles, use SparkContext’ssequenceFile[K, V] method whereK andV are the types of key and values in the file. These should be subclasses of Hadoop’sWritable interface, likeIntWritable andText. In addition, Spark allows you to specify native types for a few common Writables; for example,sequenceFile[Int, String] will automatically read IntWritables and Texts.

• For other Hadoop InputFormats, you can use theSparkContext.hadoopRDD method, which takes an arbitraryJobConf and input format class, key class and value class. Set these the same way you would for a Hadoop job with your input source. You can also useSparkContext.newAPIHadoopRDD for InputFormats based on the “new” MapReduce API (org.apache.hadoop.mapreduce).

• RDD.saveAsObjectFile andSparkContext.objectFile support saving an RDD in a simple format consisting of serialized Java objects. While this is not as efficient as specialized formats like Avro, it offers an easy way to save any RDD.

Spark can create distributed datasets from any storage source supported by Hadoop, including your local file system, HDFS, Cassandra, HBase,Amazon S3, etc. Spark supports text files,SequenceFiles, and any other HadoopInputFormat.

Text file RDDs can be created usingSparkContext’stextFile method. This method takes a URI for the file (either a local path on the machine, or ahdfs://,s3a://, etc URI) and reads it as a collection of lines. Here is an example invocation:

JavaRDD<String>distFile=sc.textFile("data.txt");

Once created,distFile can be acted on by dataset operations. For example, we can add up the sizes of all the lines using themap andreduce operations as follows:distFile.map(s -> s.length()).reduce((a, b) -> a + b).

Some notes on reading files with Spark:

• If using a path on the local filesystem, the file must also be accessible at the same path on worker nodes. Either copy the file to all workers or use a network-mounted shared file system.

• All of Spark’s file-based input methods, includingtextFile, support running on directories, compressed files, and wildcards as well. For example, you can usetextFile("/my/directory"),textFile("/my/directory/*.txt"), andtextFile("/my/directory/*.gz").

• ThetextFile method also takes an optional second argument for controlling the number of partitions of the file. By default, Spark creates one partition for each block of the file (blocks being 128MB by default in HDFS), but you can also ask for a higher number of partitions by passing a larger value. Note that you cannot have fewer partitions than blocks.

Apart from text files, Spark’s Java API also supports several other data formats:

• JavaSparkContext.wholeTextFiles lets you read a directory containing multiple small text files, and returns each of them as (filename, content) pairs. This is in contrast withtextFile, which would return one record per line in each file.

• ForSequenceFiles, use SparkContext’ssequenceFile[K, V] method whereK andV are the types of key and values in the file. These should be subclasses of Hadoop’sWritable interface, likeIntWritable andText.

• For other Hadoop InputFormats, you can use theJavaSparkContext.hadoopRDD method, which takes an arbitraryJobConf and input format class, key class and value class. Set these the same way you would for a Hadoop job with your input source. You can also useJavaSparkContext.newAPIHadoopRDD for InputFormats based on the “new” MapReduce API (org.apache.hadoop.mapreduce).

• JavaRDD.saveAsObjectFile andJavaSparkContext.objectFile support saving an RDD in a simple format consisting of serialized Java objects. While this is not as efficient as specialized formats like Avro, it offers an easy way to save any RDD.

To illustrate RDD basics, consider the simple program below:

lines=sc.textFile("data.txt")lineLengths=lines.map(lambdas:len(s))totalLength=lineLengths.reduce(lambdaa,b:a+b)

The first line defines a base RDD from an external file. This dataset is not loaded in memory orotherwise acted on:lines is merely a pointer to the file.The second line defineslineLengths as the result of amap transformation. Again,lineLengthsisnot immediately computed, due to laziness.Finally, we runreduce, which is an action. At this point Spark breaks the computation into tasksto run on separate machines, and each machine runs both its part of the map and a local reduction,returning only its answer to the driver program.

If we also wanted to uselineLengths again later, we could add:

lineLengths.persist()

before thereduce, which would causelineLengths to be saved in memory after the first time it is computed.

To illustrate RDD basics, consider the simple program below:

vallines=sc.textFile("data.txt")vallineLengths=lines.map(s=>s.length)valtotalLength=lineLengths.reduce((a,b)=>a+b)

The first line defines a base RDD from an external file. This dataset is not loaded in memory orotherwise acted on:lines is merely a pointer to the file.The second line defineslineLengths as the result of amap transformation. Again,lineLengthsisnot immediately computed, due to laziness.Finally, we runreduce, which is an action. At this point Spark breaks the computation into tasksto run on separate machines, and each machine runs both its part of the map and a local reduction,returning only its answer to the driver program.

If we also wanted to uselineLengths again later, we could add:

lineLengths.persist()

before thereduce, which would causelineLengths to be saved in memory after the first time it is computed.

To illustrate RDD basics, consider the simple program below:

JavaRDD<String>lines=sc.textFile("data.txt");JavaRDD<Integer>lineLengths=lines.map(s->s.length());inttotalLength=lineLengths.reduce((a,b)->a+b);

The first line defines a base RDD from an external file. This dataset is not loaded in memory orotherwise acted on:lines is merely a pointer to the file.The second line defineslineLengths as the result of amap transformation. Again,lineLengthsisnot immediately computed, due to laziness.Finally, we runreduce, which is an action. At this point Spark breaks the computation into tasksto run on separate machines, and each machine runs both its part of the map and a local reduction,returning only its answer to the driver program.

If we also wanted to uselineLengths again later, we could add:

lineLengths.persist(StorageLevel.MEMORY_ONLY());

before thereduce, which would causelineLengths to be saved in memory after the first time it is computed.

Spark’s API relies heavily on passing functions in the driver program to run on the cluster.There are three recommended ways to do this:

• Lambda expressions,for simple functions that can be written as an expression. (Lambdas do not support multi-statementfunctions or statements that do not return a value.)
• Localdefs inside the function calling into Spark, for longer code.
• Top-level functions in a module.

For example, to pass a longer function than can be supported using alambda, considerthe code below:

"""MyScript.py"""if__name__=="__main__":defmyFunc(s):words=s.split("")returnlen(words)sc=SparkContext(...)sc.textFile("file.txt").map(myFunc)

Note that while it is also possible to pass a reference to a method in a class instance (as opposed toa singleton object), this requires sending the object that contains that class along with the method.For example, consider:

classMyClass(object):deffunc(self,s):returnsdefdoStuff(self,rdd):returnrdd.map(self.func)

Here, if we create anew MyClass and calldoStuff on it, themap inside there references thefunc methodof thatMyClass instance, so the whole object needs to be sent to the cluster.

In a similar way, accessing fields of the outer object will reference the whole object:

classMyClass(object):def__init__(self):self.field="Hello"defdoStuff(self,rdd):returnrdd.map(lambdas:self.field+s)

To avoid this issue, the simplest way is to copyfield into a local variable insteadof accessing it externally:

defdoStuff(self,rdd):field=self.fieldreturnrdd.map(lambdas:field+s)

Spark’s API relies heavily on passing functions in the driver program to run on the cluster.There are two recommended ways to do this:

• Anonymous function syntax,which can be used for short pieces of code.
• Static methods in a global singleton object. For example, you can defineobject MyFunctions and thenpassMyFunctions.func1, as follows:

objectMyFunctions{deffunc1(s:String):String={...}}myRdd.map(MyFunctions.func1)

Note that while it is also possible to pass a reference to a method in a class instance (as opposed toa singleton object), this requires sending the object that contains that class along with the method.For example, consider:

classMyClass{deffunc1(s:String):String={...}defdoStuff(rdd:RDD[String]):RDD[String]={rdd.map(func1)}}

Here, if we create a newMyClass instance and calldoStuff on it, themap inside there references thefunc1 methodof thatMyClass instance, so the whole object needs to be sent to the cluster. It issimilar to writingrdd.map(x => this.func1(x)).

In a similar way, accessing fields of the outer object will reference the whole object:

classMyClass{valfield="Hello"defdoStuff(rdd:RDD[String]):RDD[String]={rdd.map(x=>field+x)}}

is equivalent to writingrdd.map(x => this.field + x), which references all ofthis. To avoid thisissue, the simplest way is to copyfield into a local variable instead of accessing it externally:

defdoStuff(rdd:RDD[String]):RDD[String]={valfield_=this.fieldrdd.map(x=>field_+x)}

Spark’s API relies heavily on passing functions in the driver program to run on the cluster.In Java, functions are represented by classes implementing the interfaces in theorg.apache.spark.api.java.function package.There are two ways to create such functions:

• Implement the Function interfaces in your own class, either as an anonymous inner class or a named one,and pass an instance of it to Spark.
• Uselambda expressionsto concisely define an implementation.

While much of this guide uses lambda syntax for conciseness, it is easy to use all the same APIsin long-form. For example, we could have written our code above as follows:

JavaRDD<String>lines=sc.textFile("data.txt");JavaRDD<Integer>lineLengths=lines.map(newFunction<String,Integer>(){publicIntegercall(Strings){returns.length();}});inttotalLength=lineLengths.reduce(newFunction2<Integer,Integer,Integer>(){publicIntegercall(Integera,Integerb){returna+b;}});

Or, if writing the functions inline is unwieldy:

classGetLengthimplementsFunction<String,Integer>{publicIntegercall(Strings){returns.length();}}classSumimplementsFunction2<Integer,Integer,Integer>{publicIntegercall(Integera,Integerb){returna+b;}}JavaRDD<String>lines=sc.textFile("data.txt");JavaRDD<Integer>lineLengths=lines.map(newGetLength());inttotalLength=lineLengths.reduce(newSum());

Note that anonymous inner classes in Java can also access variables in the enclosing scope as longas they are markedfinal. Spark will ship copies of these variables to each worker node as it doesfor other languages.

While most Spark operations work on RDDs containing any type of objects, a few special operations areonly available on RDDs of key-value pairs.The most common ones are distributed “shuffle” operations, such as grouping or aggregating the elementsby a key.

In Python, these operations work on RDDs containing built-in Python tuples such as(1, 2).Simply create such tuples and then call your desired operation.

For example, the following code uses thereduceByKey operation on key-value pairs to count howmany times each line of text occurs in a file:

lines=sc.textFile("data.txt")pairs=lines.map(lambdas:(s,1))counts=pairs.reduceByKey(lambdaa,b:a+b)

We could also usecounts.sortByKey(), for example, to sort the pairs alphabetically, and finallycounts.collect() to bring them back to the driver program as a list of objects.

While most Spark operations work on RDDs containing any type of objects, a few special operations areonly available on RDDs of key-value pairs.The most common ones are distributed “shuffle” operations, such as grouping or aggregating the elementsby a key.

In Scala, these operations are automatically available on RDDs containingTuple2 objects(the built-in tuples in the language, created by simply writing(a, b)). The key-value pair operations are available in thePairRDDFunctions class,which automatically wraps around an RDD of tuples.

For example, the following code uses thereduceByKey operation on key-value pairs to count howmany times each line of text occurs in a file:

vallines=sc.textFile("data.txt")valpairs=lines.map(s=>(s,1))valcounts=pairs.reduceByKey((a,b)=>a+b)

We could also usecounts.sortByKey(), for example, to sort the pairs alphabetically, and finallycounts.collect() to bring them back to the driver program as an array of objects.

Note: when using custom objects as the key in key-value pair operations, you must be sure that acustomequals() method is accompanied with a matchinghashCode() method. For full details, seethe contract outlined in theObject.hashCode()documentation.

While most Spark operations work on RDDs containing any type of objects, a few special operations areonly available on RDDs of key-value pairs.The most common ones are distributed “shuffle” operations, such as grouping or aggregating the elementsby a key.

In Java, key-value pairs are represented using thescala.Tuple2 classfrom the Scala standard library. You can simply callnew Tuple2(a, b) to create a tuple, and accessits fields later withtuple._1() andtuple._2().

RDDs of key-value pairs are represented by theJavaPairRDD class. You can constructJavaPairRDDs from JavaRDDs using special versions of themap operations, likemapToPair andflatMapToPair. The JavaPairRDD will have both standard RDD functions and specialkey-value ones.

For example, the following code uses thereduceByKey operation on key-value pairs to count howmany times each line of text occurs in a file:

JavaRDD<String>lines=sc.textFile("data.txt");JavaPairRDD<String,Integer>pairs=lines.mapToPair(s->newTuple2(s,1));JavaPairRDD<String,Integer>counts=pairs.reduceByKey((a,b)->a+b);

We could also usecounts.sortByKey(), for example, to sort the pairs alphabetically, and finallycounts.collect() to bring them back to the driver program as an array of objects.

Note: when using custom objects as the key in key-value pair operations, you must be sure that acustomequals() method is accompanied with a matchinghashCode() method. For full details, seethe contract outlined in theObject.hashCode()documentation.

An accumulator is created from an initial valuev by callingSparkContext.accumulator(v). Tasksrunning on a cluster can then add to it using theadd method or the+= operator. However, they cannot read its value.Only the driver program can read the accumulator’s value, using itsvalue method.

The code below shows an accumulator being used to add up the elements of an array:

>>>accum=sc.accumulator(0)>>>accumAccumulator<id=0,value=0>>>>sc.parallelize([1,2,3,4]).foreach(lambdax:accum.add(x))...10/09/2918:41:08INFOSparkContext:Tasksfinishedin0.317106s>>>accum.value10

While this code used the built-in support for accumulators of type Int, programmers can alsocreate their own types by subclassingAccumulatorParam.The AccumulatorParam interface has two methods:zero for providing a “zero value” for your datatype, andaddInPlace for adding two values together. For example, supposing we had aVector classrepresenting mathematical vectors, we could write:

classVectorAccumulatorParam(AccumulatorParam):defzero(self,initialValue):returnVector.zeros(initialValue.size)defaddInPlace(self,v1,v2):v1+=v2returnv1# Then, create an Accumulator of this type:vecAccum=sc.accumulator(Vector(...),VectorAccumulatorParam())

A numeric accumulator can be created by callingSparkContext.longAccumulator() orSparkContext.doubleAccumulator()to accumulate values of type Long or Double, respectively. Tasks running on a cluster can then add to it usingtheadd method. However, they cannot read its value. Only the driver program can read the accumulator’s value,using itsvalue method.

The code below shows an accumulator being used to add up the elements of an array:

scala>valaccum=sc.longAccumulator("My Accumulator")accum:org.apache.spark.util.LongAccumulator=LongAccumulator(id:0,name:Some(MyAccumulator),value:0)scala>sc.parallelize(Array(1,2,3,4)).foreach(x=>accum.add(x))...10/09/2918:41:08INFOSparkContext:Tasksfinishedin0.317106sscala>accum.valueres2:Long=10

While this code used the built-in support for accumulators of type Long, programmers can alsocreate their own types by subclassingAccumulatorV2.The AccumulatorV2 abstract class has several methods which one has to override:reset for resettingthe accumulator to zero,add for adding another value into the accumulator,merge for merging another same-type accumulator into this one. Other methods that must be overriddenare contained in theAPI documentation. For example, supposing we had aMyVector classrepresenting mathematical vectors, we could write:

classVectorAccumulatorV2extendsAccumulatorV2[MyVector,MyVector]{privatevalmyVector:MyVector=MyVector.createZeroVectordefreset():Unit={myVector.reset()}defadd(v:MyVector):Unit={myVector.add(v)}...}// Then, create an Accumulator of this type:valmyVectorAcc=newVectorAccumulatorV2// Then, register it into spark context:sc.register(myVectorAcc,"MyVectorAcc1")

Note that, when programmers define their own type of AccumulatorV2, the resulting type can be different than that of the elements added.

A numeric accumulator can be created by callingSparkContext.longAccumulator() orSparkContext.doubleAccumulator()to accumulate values of type Long or Double, respectively. Tasks running on a cluster can then add to it usingtheadd method. However, they cannot read its value. Only the driver program can read the accumulator’s value,using itsvalue method.

The code below shows an accumulator being used to add up the elements of an array:

LongAccumulatoraccum=jsc.sc().longAccumulator();sc.parallelize(Arrays.asList(1,2,3,4)).foreach(x->accum.add(x));// ...// 10/09/29 18:41:08 INFO SparkContext: Tasks finished in 0.317106 saccum.value();// returns 10

While this code used the built-in support for accumulators of type Long, programmers can alsocreate their own types by subclassingAccumulatorV2.The AccumulatorV2 abstract class has several methods which one has to override:reset for resettingthe accumulator to zero,add for adding another value into the accumulator,merge for merging another same-type accumulator into this one. Other methods that must be overriddenare contained in theAPI documentation. For example, supposing we had aMyVector classrepresenting mathematical vectors, we could write:

classVectorAccumulatorV2implementsAccumulatorV2<MyVector,MyVector>{privateMyVectormyVector=MyVector.createZeroVector();publicvoidreset(){myVector.reset();}publicvoidadd(MyVectorv){myVector.add(v);}...}// Then, create an Accumulator of this type:VectorAccumulatorV2myVectorAcc=newVectorAccumulatorV2();// Then, register it into spark context:jsc.sc().register(myVectorAcc,"MyVectorAcc1");

Note that, when programmers define their own type of AccumulatorV2, the resulting type can be different than that of the elements added.

Warning: When a Spark task finishes, Spark will try to merge the accumulated updates in this task to an accumulator.If it fails, Spark will ignore the failure and still mark the task successful and continue to run other tasks. Hence,a buggy accumulator will not impact a Spark job, but it may not get updated correctly although a Spark job is successful.

./bin/run-example SparkPi

./bin/spark-submit examples/src/main/python/pi.py

./bin/spark-submit examples/src/main/r/dataframe.R