Spark Programming Guide – Spark 2

4 programming

Spark Programming Guide

At a high level, every Spark application consists of a driver program that runs the user’s main function and executes various parallel operations on a cluster. The main abstraction Spark provides is a resilient distributed dataset (RDD), which is a collection of elements partitioned across the nodes of the cluster that can be operated on in parallel. RDDs are created by starting with a file in the Hadoop file system (or any other Hadoop-supported file system), or an existing Scala collection in the driver program, and transforming it. Users may also ask Spark to persist an RDD in memory, allowing it to be reused efficiently across parallel operations. Finally, RDDs automatically recover from node failures.

A second abstraction in Spark is shared variables that can be used in parallel operations. By default, when Spark runs a function in parallel as a set of tasks on different nodes, it ships a copy of each variable used in the function to each task. Sometimes, a variable needs to be shared across tasks, or between tasks and the driver program. Spark supports two types of shared variables: broadcast variables, which can be used to cache a value in memory on all nodes, and accumulators, which are variables that are only “added” to, such as counters and sums.

This guide shows each of these features in each of Spark’s supported languages. It is easiest to follow along with if you launch Spark’s interactive shell – either bin/spark-shell for the Scala shell or bin/pyspark for the Python one.

Linking with Spark

Spark 2.2.0 is built and distributed to work with Scala 2.11 by default. (Spark can be built to work with other versions of Scala, too.) To write applications in Scala, you will need to use a compatible Scala version (e.g. 2.11.X).

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

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

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

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

Spark 2.2.0 supports lambda expressions for concisely writing functions, otherwise you can use the classes in the org.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:

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

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

Spark 2.2.0 works with Python 2.6+ or Python 3.4+. It can use the standard CPython interpreter, so C libraries like NumPy can be used. It also works with PyPy 2.3+.

Note that support for Python 2.6 is deprecated as of Spark 2.0.0, and may be removed in Spark 2.2.0.

To run Spark applications in Python, use the bin/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 use bin/pyspark to launch an interactive Python shell.

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

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

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 by PYSPARK_PYTHON , for example:

Initializing Spark

The first thing a Spark program must do is to create a SparkContext object, which tells Spark how to access a cluster. To create a SparkContext you first need to build a SparkConf object that contains information about your application.

Only one SparkContext may be active per JVM. You must stop() the active SparkContext before creating a new one.

The first thing a Spark program must do is to create a JavaSparkContext object, which tells Spark how to access a cluster. To create a SparkContext you first need to build a SparkConf object that contains information about your application.

The first thing a Spark program must do is to create a SparkContext object, which tells Spark how to access a cluster. To create a SparkContext you first need to build a SparkConf object that contains information about your application.

The appName parameter is a name for your application to show on the cluster UI. master is a Spark, Mesos or YARN cluster URL, or a special “local” string to run in local mode. In practice, when running on a cluster, you will not want to hardcode master in the program, but rather launch the application with spark-submit and receive it there. However, for local testing and unit tests, you can pass “local” to run Spark in-process.

Using the Shell

In the Spark shell, a special interpreter-aware SparkContext is already created for you, in the variable called sc . Making your own SparkContext will not work. You can set which master the context connects to using the –master argument, and you can add JARs to the classpath by 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 coordinates to the –packages argument. Any additional repositories where dependencies might exist (e.g. Sonatype) can be passed to the –repositories argument. For example, to run bin/spark-shell on exactly four cores, use:

Or, to also add code.jar to its classpath, use:

To include a dependency using Maven coordinates:

For a complete list of options, run spark-shell –help . Behind the scenes, spark-shell invokes the more general spark-submit script.

In the PySpark shell, a special interpreter-aware SparkContext is already created for you, in the variable called sc . Making your own SparkContext will not work. You can set which master the context connects to using the –master argument, and you can add Python .zip, .egg or .py files to the runtime path by passing a comma-separated list to –py-files . You can also add dependencies (e.g. Spark Packages) to your shell session by supplying a comma-separated list of Maven coordinates to the –packages argument. Any additional repositories where dependencies might exist (e.g. Sonatype) can be passed to the –repositories argument. Any Python dependencies a Spark package has (listed in the requirements.txt of that package) must be manually installed using pip when necessary. For example, to run bin/pyspark on exactly four cores, use:

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

For a complete list of options, run pyspark –help . Behind the scenes, pyspark invokes the more general spark-submit script.

It is also possible to launch the PySpark shell in IPython, the enhanced Python interpreter. PySpark works with IPython 1.0.0 and later. To use IPython, set the PYSPARK_DRIVER_PYTHON variable to ipython when running bin/pyspark :

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

You can customize the ipython or jupyter commands by setting PYSPARK_DRIVER_PYTHON_OPTS .

After the Jupyter Notebook server is launched, you can create a new “Python 2” notebook from the “Files” tab. Inside the notebook, you can input the command %pylab inline as part of your notebook before you start to try Spark from the Jupyter notebook.

Resilient Distributed Datasets (RDDs)

Spark revolves around the concept of a resilient distributed dataset (RDD), which is a fault-tolerant collection of elements that can be operated on in parallel. There are two ways to create RDDs: parallelizing an existing collection in your driver program, or referencing a dataset in an external storage system, such as a shared filesystem, HDFS, HBase, or any data source offering a Hadoop InputFormat.

Parallelized Collections

Parallelized collections are created by calling SparkContext ’s parallelize method on an existing collection in your driver program (a Scala Seq ). 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:

Once created, the distributed dataset ( distData ) can be operated on in parallel. For example, we might call distData.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 calling JavaSparkContext ’s parallelize method on an existing 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:

Once created, the distributed dataset ( distData ) can be operated on in parallel. For example, we might call distData.reduce((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 calling SparkContext ’s parallelize 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:

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

One important parameter for parallel collections is the number of partitions to cut the dataset into. Spark will run one task for each partition of the cluster. Typically you want 2-4 partitions for each CPU in your cluster. Normally, Spark tries to set the number of partitions automatically based on your cluster. However, you can also set it manually by passing it as a second parameter to parallelize (e.g. sc.parallelize(data, 10) ). Note: some places in the code use the term slices (a synonym for partitions) to maintain backward compatibility.

External Datasets

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 Hadoop InputFormat.

Text file RDDs can be created using SparkContext ’s textFile method. This method takes an URI for the file (either a local path on the machine, or a hdfs:// , s3n:// , etc URI) and reads it as a collection of lines. Here is an example invocation:

Once created, distFile can be acted on by dataset operations. For example, we can add up the sizes of all the lines using the map and reduce 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, including textFile , support running on directories, compressed files, and wildcards as well. For example, you can use textFile(“/my/directory”) , textFile(“/my/directory/*.txt”) , and textFile(“/my/directory/*.gz”) .

The textFile 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 with textFile , 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.

For SequenceFiles, use SparkContext’s sequenceFile[K, V] method where K and V are the types of key and values in the file. These should be subclasses of Hadoop’s Writable interface, like IntWritable and Text. 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 the SparkContext.hadoopRDD method, which takes an arbitrary JobConf 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 use SparkContext.newAPIHadoopRDD for InputFormats based on the “new” MapReduce API ( org.apache.hadoop.mapreduce ).

RDD.saveAsObjectFile and SparkContext.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 Hadoop InputFormat.

Text file RDDs can be created using SparkContext ’s textFile method. This method takes an URI for the file (either a local path on the machine, or a hdfs:// , s3n:// , etc URI) and reads it as a collection of lines. Here is an example invocation:

Once created, distFile can be acted on by dataset operations. For example, we can add up the sizes of all the lines using the map and reduce 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, including textFile , support running on directories, compressed files, and wildcards as well. For example, you can use textFile(“/my/directory”) , textFile(“/my/directory/*.txt”) , and textFile(“/my/directory/*.gz”) .

The textFile 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 with textFile , which would return one record per line in each file.

For SequenceFiles, use SparkContext’s sequenceFile[K, V] method where K and V are the types of key and values in the file. These should be subclasses of Hadoop’s Writable interface, like IntWritable and Text.

For other Hadoop InputFormats, you can use the JavaSparkContext.hadoopRDD method, which takes an arbitrary JobConf 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 use JavaSparkContext.newAPIHadoopRDD for InputFormats based on the “new” MapReduce API ( org.apache.hadoop.mapreduce ).

JavaRDD.saveAsObjectFile and JavaSparkContext.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.

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 Hadoop InputFormat.

Text file RDDs can be created using SparkContext ’s textFile method. This method takes an URI for the file (either a local path on the machine, or a hdfs:// , s3n:// , etc URI) and reads it as a collection of lines. Here is an example invocation:

Once created, distFile can be acted on by dataset operations. For example, we can add up the sizes of all the lines using the map and reduce 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, including textFile , support running on directories, compressed files, and wildcards as well. For example, you can use textFile(“/my/directory”) , textFile(“/my/directory/*.txt”) , and textFile(“/my/directory/*.gz”) .

The textFile 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 with textFile , which would return one record per line in each file.

RDD.saveAsPickleFile and SparkContext.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 marked Experimental 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.

PySpark SequenceFile support loads an RDD of key-value pairs within Java, converts Writables to base Java types, and pickles the resulting Java objects using Pyrolite. 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 following Writables are automatically converted:

Arrays are not handled out-of-the-box. Users need to specify custom ArrayWritable subtypes when reading or writing. When writing, users also need to specify custom converters that convert arrays to custom ArrayWritable subtypes. When reading, the default converter will convert custom ArrayWritable subtypes to Java Object[] , which then get pickled to Python tuples. To get Python array.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 value classes can be specified, but for standard Writables this is not required.

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 the Elasticsearch ESInputFormat:

Note that, if the InputFormat simply depends on a Hadoop configuration and/or input path, and the 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 to transform that data on the Scala/Java side to something which can be handled by Pyrolite’s pickler. A Converter trait is provided for this. Simply extend this trait and implement your transformation code in the convert method. Remember to ensure that this class, along with any dependencies required to access your InputFormat , are packaged into your Spark job jar and included on the PySpark classpath.

See the Python examples and the Converter examples for examples of using Cassandra / HBase InputFormat and OutputFormat with custom converters.

RDD Operations

RDDs support two types of operations: transformations, which create a new dataset from an existing one, and actions, which return a value to the driver program after running a computation on the dataset. For example, map is a transformation that passes each dataset element through a function and returns a new RDD representing the results. On the other hand, reduce is an action that aggregates all the elements of the RDD using some function and returns the final result to the driver program (although there is also a parallel reduceByKey that returns a distributed dataset).

All transformations in Spark are lazy, in that they do not compute their results right away. Instead, they just remember the transformations applied to some base dataset (e.g. a file). The transformations are only computed when an action requires a result to be returned to the driver program. This design enables Spark to run more efficiently. For example, we can realize that a dataset created through map will be used in a reduce and return only the result of the reduce to the driver, rather than the larger mapped dataset.

By default, each transformed RDD may be recomputed each time you run an action on it. However, you may also persist an RDD in memory using the persist (or cache ) method, in which case Spark will keep the elements around on the cluster for much faster access the next time you query it. There is also support for persisting RDDs on disk, or replicated across multiple nodes.

To illustrate RDD basics, consider the simple program below:

The first line defines a base RDD from an external file. This dataset is not loaded in memory or otherwise acted on: lines is merely a pointer to the file. The second line defines lineLengths as the result of a map transformation. Again, lineLengths is not immediately computed, due to laziness. Finally, we run reduce , which is an action. At this point Spark breaks the computation into tasks to 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 use lineLengths again later, we could add:

before the reduce , which would cause lineLengths to be saved in memory after the first time it is computed.

To illustrate RDD basics, consider the simple program below:

The first line defines a base RDD from an external file. This dataset is not loaded in memory or otherwise acted on: lines is merely a pointer to the file. The second line defines lineLengths as the result of a map transformation. Again, lineLengths is not immediately computed, due to laziness. Finally, we run reduce , which is an action. At this point Spark breaks the computation into tasks to 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 use lineLengths again later, we could add:

before the reduce , which would cause lineLengths to be saved in memory after the first time it is computed.

To illustrate RDD basics, consider the simple program below:

The first line defines a base RDD from an external file. This dataset is not loaded in memory or otherwise acted on: lines is merely a pointer to the file. The second line defines lineLengths as the result of a map transformation. Again, lineLengths is not immediately computed, due to laziness. Finally, we run reduce , which is an action. At this point Spark breaks the computation into tasks to 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 use lineLengths again later, we could add:

before the reduce , which would cause lineLengths to be saved in memory after the first time it is computed.

Passing Functions to Spark

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 define object MyFunctions and then pass MyFunctions.func1 , as follows:

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

Here, if we create a new MyClass instance and call doStuff on it, the map inside there references the func1 method of that MyClass instance, so the whole object needs to be sent to the cluster. It is similar to writing rdd.map(x => this.func1(x)) .

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

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

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 the org.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.
  • Use lambda expressions to concisely define an implementation.

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

Or, if writing the functions inline is unwieldy:

Note that anonymous inner classes in Java can also access variables in the enclosing scope as long as they are marked final . Spark will ship copies of these variables to each worker node as it does for other languages.

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-statement functions or statements that do not return a value.)
  • Local def s 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 a lambda , consider the code below:

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

Here, if we create a new MyClass and call doStuff on it, the map inside there references the func method of that MyClass 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:

To avoid this issue, the simplest way is to copy field into a local variable instead of accessing it externally:

Understanding closures

One of the harder things about Spark is understanding the scope and life cycle of variables and methods when executing code across a cluster. RDD operations that modify variables outside of their scope can be a frequent source of confusion. In the example below we’ll look at code that uses foreach() to increment a counter, but similar issues can occur for other operations as well.

Consider the naive RDD element sum below, which may behave differently depending on whether execution is happening within the same JVM. A common example of this is when running Spark in local mode ( –master = local[n] ) versus deploying a Spark application to a cluster (e.g. via spark-submit to YARN):

Local vs. cluster modes

The behavior of the above code is undefined, and may not work as intended. To execute jobs, Spark breaks up the processing of RDD operations into tasks, each of which is executed by an executor. Prior to execution, Spark computes the task’s closure. The closure is those variables and methods which must be visible for the executor to perform its computations on the RDD (in this case foreach() ). This closure is serialized and sent to each executor.

The variables within the closure sent to each executor are now copies and thus, when counter is referenced within the foreach function, it’s no longer the counter on the driver node. There is still a counter in the memory of the driver node but this is no longer visible to the executors! The executors only see the copy from the serialized closure. Thus, the final value of counter will still be zero since all operations on counter were referencing the value within the serialized closure.

In local mode, in some circumstances the foreach function will actually execute within the same JVM as the driver and will reference the same original counter, and may actually update it.

To ensure well-defined behavior in these sorts of scenarios one should use an Accumulator . Accumulators in Spark are used specifically to provide a mechanism for safely updating a variable when execution is split up across worker nodes in a cluster. The Accumulators section of this guide discusses these in more detail.

In general, closures – constructs like loops or locally defined methods, should not be used to mutate some global state. Spark does not define or guarantee the behavior of mutations to objects referenced from outside of closures. Some code that does this may work in local mode, but that’s just by accident and such code will not behave as expected in distributed mode. Use an Accumulator instead if some global aggregation is needed.

Printing elements of an RDD

Another common idiom is attempting to print out the elements of an RDD using rdd.foreach(println) or rdd.map(println) . On a single machine, this will generate the expected output and print all the RDD’s elements. However, in cluster mode, the output to stdout being called by the executors is now writing to the executor’s stdout instead, not the one on the driver, so stdout on the driver won’t show these! To print all elements on the driver, one can use the collect() method to first bring the RDD to the driver node thus: rdd.collect().foreach(println) . This can cause the driver to run out of memory, though, because collect() fetches the entire RDD to a single machine; if you only need to print a few elements of the RDD, a safer approach is to use the take() : rdd.take(100).foreach(println) .

Working with Key-Value Pairs

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

In Scala, these operations are automatically available on RDDs containing Tuple2 objects (the built-in tuples in the language, created by simply writing (a, b) ). The key-value pair operations are available in the PairRDDFunctions class, which automatically wraps around an RDD of tuples.

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

We could also use counts.sortByKey() , for example, to sort the pairs alphabetically, and finally counts.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 a custom equals() method is accompanied with a matching hashCode() method. For full details, see the contract outlined in the Object.hashCode() documentation.

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

In Java, key-value pairs are represented using the scala.Tuple2 class from the Scala standard library. You can simply call new Tuple2(a, b) to create a tuple, and access its fields later with tuple._1() and tuple._2() .

RDDs of key-value pairs are represented by the JavaPairRDD class. You can construct JavaPairRDDs from JavaRDDs using special versions of the map operations, like mapToPair and flatMapToPair . The JavaPairRDD will have both standard RDD functions and special key-value ones.

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

We could also use counts.sortByKey() , for example, to sort the pairs alphabetically, and finally counts.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 a custom equals() method is accompanied with a matching hashCode() method. For full details, see the contract outlined in the Object.hashCode() documentation.

While most Spark operations work on RDDs containing any type of objects, a few special operations are only available on RDDs of key-value pairs. The most common ones are distributed “shuffle” operations, such as grouping or aggregating the elements by 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 the reduceByKey operation on key-value pairs to count how many times each line of text occurs in a file:

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

Transformations

The following table lists some of the common transformations supported by Spark. Refer to the RDD API doc (Scala, Java, Python, R) and pair RDD functions doc (Scala, Java) for details.

Note: If you are grouping in order to perform an aggregation (such as a sum or average) over each key, using reduceByKey or aggregateByKey will yield much better performance.

Note: By default, the level of parallelism in the output depends on the number of partitions of the parent RDD. You can pass an optional numTasks argument to set a different number of tasks.

The following table lists some of the common actions supported by Spark. Refer to the RDD API doc (Scala, Java, Python, R)

and pair RDD functions doc (Scala, Java) for details.

Note: modifying variables other than Accumulators outside of the foreach() may result in undefined behavior. See Understanding closures for more details.

The Spark RDD API also exposes asynchronous versions of some actions, like foreachAsync for foreach , which immediately return a FutureAction to the caller instead of blocking on completion of the action. This can be used to manage or wait for the asynchronous execution of the action.

Shuffle operations

Certain operations within Spark trigger an event known as the shuffle. The shuffle is Spark’s mechanism for re-distributing data so that it’s grouped differently across partitions. This typically involves copying data across executors and machines, making the shuffle a complex and costly operation.

Background

To understand what happens during the shuffle we can consider the example of the reduceByKey operation. The reduceByKey operation generates a new RDD where all values for a single key are combined into a tuple – the key and the result of executing a reduce function against all values associated with that key. The challenge is that not all values for a single key necessarily reside on the same partition, or even the same machine, but they must be co-located to compute the result.

In Spark, data is generally not distributed across partitions to be in the necessary place for a specific operation. During computations, a single task will operate on a single partition – thus, to organize all the data for a single reduceByKey reduce task to execute, Spark needs to perform an all-to-all operation. It must read from all partitions to find all the values for all keys, and then bring together values across partitions to compute the final result for each key – this is called the shuffle.

Although the set of elements in each partition of newly shuffled data will be deterministic, and so is the ordering of partitions themselves, the ordering of these elements is not. If one desires predictably ordered data following shuffle then it’s possible to use:

  • mapPartitions to sort each partition using, for example, .sorted
  • repartitionAndSortWithinPartitions to efficiently sort partitions while simultaneously repartitioning
  • sortBy to make a globally ordered RDD

Operations which can cause a shuffle include repartition operations like repartition and coalesce , ‘ByKey operations (except for counting) like groupByKey and reduceByKey , and join operations like cogroup and join .

Performance Impact

The Shuffle is an expensive operation since it involves disk I/O, data serialization, and network I/O. To organize data for the shuffle, Spark generates sets of tasks – map tasks to organize the data, and a set of reduce tasks to aggregate it. This nomenclature comes from MapReduce and does not directly relate to Spark’s map and reduce operations.

Internally, results from individual map tasks are kept in memory until they can’t fit. Then, these are sorted based on the target partition and written to a single file. On the reduce side, tasks read the relevant sorted blocks.

Certain shuffle operations can consume significant amounts of heap memory since they employ in-memory data structures to organize records before or after transferring them. Specifically, reduceByKey and aggregateByKey create these structures on the map side, and ‘ByKey operations generate these on the reduce side. When data does not fit in memory Spark will spill these tables to disk, incurring the additional overhead of disk I/O and increased garbage collection.

Shuffle also generates a large number of intermediate files on disk. As of Spark 1.3, these files are preserved until the corresponding RDDs are no longer used and are garbage collected. This is done so the shuffle files don’t need to be re-created if the lineage is re-computed. Garbage collection may happen only after a long period of time, if the application retains references to these RDDs or if GC does not kick in frequently. This means that long-running Spark jobs may consume a large amount of disk space. The temporary storage directory is specified by the spark.local.dir configuration parameter when configuring the Spark context.

Shuffle behavior can be tuned by adjusting a variety of configuration parameters. See the ‘Shuffle Behavior’ section within the Spark Configuration Guide.

RDD Persistence

One of the most important capabilities in Spark is persisting (or caching) a dataset in memory across operations. When you persist an RDD, each node stores any partitions of it that it computes in memory and reuses them in other actions on that dataset (or datasets derived from it). This allows future actions to be much faster (often by more than 10x). Caching is a key tool for iterative algorithms and fast interactive use.

You can mark an RDD to be persisted using the persist() or cache() methods on it. The first time it is computed in an action, it will be kept in memory on the nodes. Spark’s cache is fault-tolerant – if any partition of an RDD is lost, it will automatically be recomputed using the transformations that originally created it.

In addition, each persisted RDD can be stored using a different storage level, allowing you, for example, to persist the dataset on disk, persist it in memory but as serialized Java objects (to save space), replicate it across nodes. These levels are set by passing a StorageLevel object (Scala, Java, Python) to persist() . The cache() method is a shorthand for using the default storage level, which is StorageLevel.MEMORY_ONLY (store deserialized objects in memory). The full set of storage levels is:

Note: In Python, stored objects will always be serialized with the Pickle library, so it does not matter whether you choose a serialized level. The available storage levels in Python include MEMORY_ONLY , MEMORY_ONLY_2 , MEMORY_AND_DISK , MEMORY_AND_DISK_2 , DISK_ONLY , and DISK_ONLY_2 .

Spark also automatically persists some intermediate data in shuffle operations (e.g. reduceByKey ), even without users calling persist . This is done to avoid recomputing the entire input if a node fails during the shuffle. We still recommend users call persist on the resulting RDD if they plan to reuse it.

Which Storage Level to Choose?

Spark’s storage levels are meant to provide different trade-offs between memory usage and CPU efficiency. We recommend going through the following process to select one:

If your RDDs fit comfortably with the default storage level ( MEMORY_ONLY ), leave them that way. This is the most CPU-efficient option, allowing operations on the RDDs to run as fast as possible.

If not, try using MEMORY_ONLY_SER and selecting a fast serialization library to make the objects much more space-efficient, but still reasonably fast to access. (Java and Scala)

Don’t spill to disk unless the functions that computed your datasets are expensive, or they filter a large amount of the data. Otherwise, recomputing a partition may be as fast as reading it from disk.

Use the replicated storage levels if you want fast fault recovery (e.g. if using Spark to serve requests from a web application). All the storage levels provide full fault tolerance by recomputing lost data, but the replicated ones let you continue running tasks on the RDD without waiting to recompute a lost partition.

Removing Data

Spark automatically monitors cache usage on each node and drops out old data partitions in a least-recently-used (LRU) fashion. If you would like to manually remove an RDD instead of waiting for it to fall out of the cache, use the RDD.unpersist() method.

Shared Variables

Normally, when a function passed to a Spark operation (such as map or reduce ) is executed on a remote cluster node, it works on separate copies of all the variables used in the function. These variables are copied to each machine, and no updates to the variables on the remote machine are propagated back to the driver program. Supporting general, read-write shared variables across tasks would be inefficient. However, Spark does provide two limited types of shared variables for two common usage patterns: broadcast variables and accumulators.

Broadcast Variables

Broadcast variables allow the programmer to keep a read-only variable cached on each machine rather than shipping a copy of it with tasks. They can be used, for example, to give every node a copy of a large input dataset in an efficient manner. Spark also attempts to distribute broadcast variables using efficient broadcast algorithms to reduce communication cost.

Spark actions are executed through a set of stages, separated by distributed “shuffle” operations. Spark automatically broadcasts the common data needed by tasks within each stage. The data broadcasted this way is cached in serialized form and deserialized before running each task. This means that explicitly creating broadcast variables is only useful when tasks across multiple stages need the same data or when caching the data in deserialized form is important.

Broadcast variables are created from a variable v by calling SparkContext.broadcast(v) . The broadcast variable is a wrapper around v , and its value can be accessed by calling the value method. The code below shows this:

After the broadcast variable is created, it should be used instead of the value v in any functions run on the cluster so that v is not shipped to the nodes more than once. In addition, the object v should not be modified after it is broadcast in order to ensure that all nodes get the same value of the broadcast variable (e.g. if the variable is shipped to a new node later).

Accumulators

Accumulators are variables that are only “added” to through an associative and commutative operation and can therefore be efficiently supported in parallel. They can be used to implement counters (as in MapReduce) or sums. Spark natively supports accumulators of numeric types, and programmers can add support for new types.

As a user, you can create named or unnamed accumulators. As seen in the image below, a named accumulator (in this instance counter ) will display in the web UI for the stage that modifies that accumulator. Spark displays the value for each accumulator modified by a task in the “Tasks” table.

Tracking accumulators in the UI can be useful for understanding the progress of running stages (NOTE: this is not yet supported in Python).

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

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

While this code used the built-in support for accumulators of type Long, programmers can also create their own types by subclassing AccumulatorV2. The AccumulatorV2 abstract class has several methods which one has to override: reset for resetting the 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 overridden are contained in the API documentation. For example, supposing we had a MyVector class representing mathematical vectors, we could write:

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 calling SparkContext.longAccumulator() or SparkContext.doubleAccumulator() to accumulate values of type Long or Double, respectively. Tasks running on a cluster can then add to it using the add method. However, they cannot read its value. Only the driver program can read the accumulator’s value, using its value method.

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

While this code used the built-in support for accumulators of type Long, programmers can also create their own types by subclassing AccumulatorV2. The AccumulatorV2 abstract class has several methods which one has to override: reset for resetting the 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 overridden are contained in the API documentation. For example, supposing we had a MyVector class representing mathematical vectors, we could write:

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

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

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

While this code used the built-in support for accumulators of type Int, programmers can also create their own types by subclassing AccumulatorParam. The AccumulatorParam interface has two methods: zero for providing a “zero value” for your data type, and addInPlace for adding two values together. For example, supposing we had a Vector class representing mathematical vectors, we could write:

For accumulator updates performed inside actions only, Spark guarantees that each task’s update to the accumulator will only be applied once, i.e. restarted tasks will not update the value. In transformations, users should be aware of that each task’s update may be applied more than once if tasks or job stages are re-executed.

Accumulators do not change the lazy evaluation model of Spark. If they are being updated within an operation on an RDD, their value is only updated once that RDD is computed as part of an action. Consequently, accumulator updates are not guaranteed to be executed when made within a lazy transformation like map() . The below code fragment demonstrates this property:

Deploying to a Cluster

The application submission guide describes how to submit applications to a cluster. In short, once you package your application into a JAR (for Java/Scala) or a set of .py or .zip files (for Python), the bin/spark-submit script lets you submit it to any supported cluster manager.

Launching Spark jobs from Java / Scala

The org.apache.spark.launcher package provides classes for launching Spark jobs as child processes using a simple Java API.

Unit Testing

Spark is friendly to unit testing with any popular unit test framework. Simply create a SparkContext in your test with the master URL set to local , run your operations, and then call SparkContext.stop() to tear it down. Make sure you stop the context within a finally block or the test framework’s tearDown method, as Spark does not support two contexts running concurrently in the same program.

Where to Go from Here

You can see some example Spark programs on the Spark website. In addition, Spark includes several samples in the examples directory (Scala, Java, Python, R). You can run Java and Scala examples by passing the class name to Spark’s bin/run-example script; for instance:

For Python examples, use spark-submit instead:

For R examples, use spark-submit instead:

For help on optimizing your programs, the configuration and tuning guides provide information on best practices. They are especially important for making sure that your data is stored in memory in an efficient format. For help on deploying, the cluster mode overview describes the components involved in distributed operation and supported cluster managers.

Finally, full API documentation is available in Scala, Java, Python and R.

Leave a Reply

Your email address will not be published. Required fields are marked *