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 and walks through some samples. It assumes some familiarity with Scala, especially with the syntax for closures. Note that you can also run Spark interactively using the spark-shell script. We highly recommend doing that to follow along!

Linking with Spark

To write a Spark application, you will need to add both Spark and its dependencies to your CLASSPATH. If you use sbt or Maven, Spark is available through Maven Central at:

groupId = org.spark-project
artifactId = spark-core_2.9.2
version = 0.6.2 

For other build systems or environments, you can run sbt/sbt assembly to build both Spark and its dependencies into one JAR (core/target/spark-core-assembly-0.6.0.jar), then add this to your CLASSPATH.

In addition, you’ll need to import some Spark classes and implicit conversions. Add the following lines at the top of your program:

import spark.SparkContext
import SparkContext._

Initializing Spark

The first thing a Spark program must do is to create a SparkContext object, which tells Spark how to access a cluster. This is done through the following constructor:

new SparkContext(master, jobName, [sparkHome], [jars])

The master parameter is a string specifying a Mesos cluster to connect to, or a special “local” string to run in local mode, as described below. jobName is a name for your job, which will be shown in the Mesos web UI when running on a cluster. Finally, the last two parameters are needed to deploy your code to a cluster if running in distributed mode, as described later.

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 environment variable. For example, to run on four cores, use

$ MASTER=local[4] ./spark-shell

Master URLs

The master URL passed to Spark can be in one of the following formats:

Master URLMeaning
local Run Spark locally with one worker thread (i.e. no parallelism at all).
local[K] Run Spark locally with K worker threads (ideally, set this to the number of cores on your machine).
spark://HOST:PORT Connect to the given Spark standalone cluster master. The port must be whichever one your master is configured to use, which is 7077 by default.
mesos://HOST:PORT Connect to the given Mesos cluster. The host parameter is the hostname of the Mesos master. The port must be whichever one the master is configured to use, which is 5050 by default.

For running on YARN, Spark launches an instance of the standalone deploy cluster within YARN; see running on YARN for details.

Deploying Code on a Cluster

If you want to run your job on a cluster, you will need to specify the two optional parameters to SparkContext to let it find your code:

If you run spark-shell on a cluster, any classes you define in the shell will automatically be distributed.

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 currently two types of RDDs: parallelized collections, which take an existing Scala collection and run functions on it in parallel, and Hadoop datasets, which run functions on each record of a file in Hadoop distributed file system or any other storage system supported by Hadoop. Both types of RDDs can be operated on through the same methods.

Parallelized Collections

Parallelized collections are created by calling SparkContext’s parallelize method on an existing Scala collection (a Seq object). The elements of the collection are copied to form a distributed dataset that can be operated on in parallel. For example, here is some interpreter output showing how to create a parallel collection from an array:

scala> val data = Array(1, 2, 3, 4, 5)
data: Array[Int] = Array(1, 2, 3, 4, 5)

scala> val distData = sc.parallelize(data)
distData: spark.RDD[Int] = spark.ParallelCollection@10d13e3e

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

One important parameter for parallel collections is the number of slices to cut the dataset into. Spark will run one task for each slice of the cluster. Typically you want 2-4 slices for each CPU in your cluster. Normally, Spark tries to set the number of slices 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)).

Hadoop Datasets

Spark can create distributed datasets from any file stored in the Hadoop distributed file system (HDFS) or other storage systems supported by Hadoop (including your local file system, Amazon S3, Hypertable, HBase, 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://, kfs://, etc URI). Here is an example invocation:

scala> val distFile = sc.textFile("data.txt")
distFile: spark.RDD[String] = spark.HadoopRDD@1d4cee08

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: + _).

The textFile method also takes an optional second argument for controlling the number of slices of the file. By default, Spark creates one slice for each block of the file (blocks being 64MB by default in HDFS), but you can also ask for a higher number of slices by passing a larger value. Note that you cannot have fewer slices than blocks.

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.

Finally, 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.

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 distributed dataset representing the results. On the other hand, reduce is an action that aggregates all the elements of the dataset 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 is 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 datasets on disk, or replicated across the cluster. The next section in this document describes these options.

The following tables list the transformations and actions currently supported (see also the RDD API doc for details):


map(func) Return a new distributed dataset formed by passing each element of the source through a function func.
filter(func) Return a new dataset formed by selecting those elements of the source on which func returns true.
flatMap(func) Similar to map, but each input item can be mapped to 0 or more output items (so func should return a Seq rather than a single item).
mapPartitions(func) Similar to map, but runs separately on each partition (block) of the RDD, so func must be of type Iterator[T] => Iterator[U] when running on an RDD of type T.
mapPartitionsWithSplit(func) Similar to mapPartitions, but also provides func with an integer value representing the index of the split, so func must be of type (Int, Iterator[T]) => Iterator[U] when running on an RDD of type T.
sample(withReplacement, fraction, seed) Sample a fraction fraction of the data, with or without replacement, using a given random number generator seed.
union(otherDataset) Return a new dataset that contains the union of the elements in the source dataset and the argument.
distinct([numTasks])) Return a new dataset that contains the distinct elements of the source dataset.
groupByKey([numTasks]) When called on a dataset of (K, V) pairs, returns a dataset of (K, Seq[V]) pairs.
Note: By default, this uses only 8 parallel tasks to do the grouping. You can pass an optional numTasks argument to set a different number of tasks.
reduceByKey(func, [numTasks]) When called on a dataset of (K, V) pairs, returns a dataset of (K, V) pairs where the values for each key are aggregated using the given reduce function. Like in groupByKey, the number of reduce tasks is configurable through an optional second argument.
sortByKey([ascending], [numTasks]) When called on a dataset of (K, V) pairs where K implements Ordered, returns a dataset of (K, V) pairs sorted by keys in ascending or descending order, as specified in the boolean ascending argument.
join(otherDataset, [numTasks]) When called on datasets of type (K, V) and (K, W), returns a dataset of (K, (V, W)) pairs with all pairs of elements for each key.
cogroup(otherDataset, [numTasks]) When called on datasets of type (K, V) and (K, W), returns a dataset of (K, Seq[V], Seq[W]) tuples. This operation is also called groupWith.
cartesian(otherDataset) When called on datasets of types T and U, returns a dataset of (T, U) pairs (all pairs of elements).

A complete list of transformations is available in the RDD API doc.


reduce(func) Aggregate the elements of the dataset using a function func (which takes two arguments and returns one). The function should be associative so that it can be computed correctly in parallel.
collect() Return all the elements of the dataset as an array at the driver program. This is usually useful after a filter or other operation that returns a sufficiently small subset of the data.
count() Return the number of elements in the dataset.
first() Return the first element of the dataset (similar to take(1)).
take(n) Return an array with the first n elements of the dataset. Note that this is currently not executed in parallel. Instead, the driver program computes all the elements.
takeSample(withReplacement, num, seed) Return an array with a random sample of num elements of the dataset, with or without replacement, using the given random number generator seed.
saveAsTextFile(path) Write the elements of the dataset as a text file (or set of text files) in a given directory in the local filesystem, HDFS or any other Hadoop-supported file system. Spark will call toString on each element to convert it to a line of text in the file.
saveAsSequenceFile(path) Write the elements of the dataset as a Hadoop SequenceFile in a given path in the local filesystem, HDFS or any other Hadoop-supported file system. This is only available on RDDs of key-value pairs that either implement Hadoop's Writable interface or are implicitly convertible to Writable (Spark includes conversions for basic types like Int, Double, String, etc).
countByKey() Only available on RDDs of type (K, V). Returns a `Map` of (K, Int) pairs with the count of each key.
foreach(func) Run a function func on each element of the dataset. This is usually done for side effects such as updating an accumulator variable (see below) or interacting with external storage systems.

A complete list of actions is available in the RDD API doc.

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 slices 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 building iterative algorithms with Spark and for interactive use from the interpreter.

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. The 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 RDD can be stored using a different storage level, allowing you, for example, to persist the dataset on disk, or persist it in memory but as serialized Java objects (to save space), or even replicate it across nodes. These levels are chosen by passing a object 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 complete set of available storage levels is:

Storage LevelMeaning
MEMORY_ONLY Store RDD as deserialized Java objects in the JVM. If the RDD does not fit in memory, some partitions will not be cached and will be recomputed on the fly each time they're needed. This is the default level.
MEMORY_AND_DISK Store RDD as deserialized Java objects in the JVM. If the RDD does not fit in memory, store the partitions that don't fit on disk, and read them from there when they're needed.
MEMORY_ONLY_SER Store RDD as serialized Java objects (one byte array per partition). This is generally more space-efficient than deserialized objects, especially when using a fast serializer, but more CPU-intensive to read.
MEMORY_AND_DISK_SER Similar to MEMORY_ONLY_SER, but spill partitions that don't fit in memory to disk instead of recomputing them on the fly each time they're needed.
DISK_ONLY Store the RDD partitions only on disk.
MEMORY_ONLY_2, MEMORY_AND_DISK_2, etc. Same as the levels above, but replicate each partition on two cluster nodes.

Which Storage Level to Choose?

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

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.

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 interpreter session below shows this:

scala> val broadcastVar = sc.broadcast(Array(1, 2, 3))
broadcastVar: spark.Broadcast[Array[Int]] = spark.Broadcast(b5c40191-a864-4c7d-b9bf-d87e1a4e787c)

scala> broadcastVar.value
res0: Array[Int] = Array(1, 2, 3)

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 are variables that are only “added” to through an associative 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 type Int and Double, and programmers can add support for new types.

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

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

scala> val accum = sc.accumulator(0)
accum: spark.Accumulator[Int] = 0

scala> sc.parallelize(Array(1, 2, 3, 4)).foreach(x => accum += x)
10/09/29 18:41:08 INFO SparkContext: Tasks finished in 0.317106 s

scala> accum.value
res2: Int = 10

Where to Go from Here

You can see some example Spark programs on the Spark website. In addition, Spark includes several sample programs in examples/src/main/scala. Some of them have both Spark versions and local (non-parallel) versions, allowing you to see what had to be changed to make the program run on a cluster. You can run them using by passing the class name to the run script included in Spark – for example, ./run spark.examples.SparkPi. Each example program prints usage help when run without any arguments.

For help on optimizing your program, 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.