Spark Programming Guide

本文主要来自两部分： Part#1: https://github.com/mesos/spark/wiki/Spark-Programming-Guide, Part#2: http://www.spark-project.org/examples.html

通过Part#1很容易了解Spark提供了哪些东西，而Part#2让你快速知道如果使用Spark编写程序。

At a high level, every Spark application consists of a driver program that runs the user's main function and executes various parallel operationson a cluster. The main abstraction Spark provides is a distributed dataset, which is a collection of elements partitioned across the nodes of the cluster that can be operated on in parallel. Distributed datasets are created by starting with a file in the Hadoop file system (HDFS) or an existing collection in the driver program and possibly transforming it. Users may also ask Spark to cache a dataset in memory, allowing it to be reused efficiently across parallel operations. Finally, distributed datasets 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 forclosures. Note that Spark can be run interactively through the spark-shell interpreter. You might want to do 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. The easiest way to do this is to runsbt/sbt assembly to build both Spark and its dependencies into one JAR (core/target/scala_2.8.1/spark-core-assembly-0.0.0.jar), then add this to your CLASSPATH. Alternatively, you can build Spark into one JAR and its dependencies into another with sbt/sbt package and sbt/sbt dep-jar respectively.

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 requires two parameters: a Mesos master name (or a special name to run in local mode), and a name for the job (to display in the Mesos UI). You can create a SparkContext by importing spark.SparkContext and using new SparkContext(master, jobName).

The master name can in one of three formats:

Master Name	Meaning
local	Run Spark locally with one worker thread (i.e. no parallelism at all).
local[K]	Run Spark locally with K worker threads (which should be set to the number of cores on your machine).
1@HOST:PORT	Connect Spark to the given Mesos master to run on a 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. The "1@" is always required (it's a process ID internal to Mesos).

If you are using the Spark interpreter, it creates a SparkContext for you (as a variable named sc), and it defaults to ''local'' as the master. You can make it use a different master by setting the MASTER environment variable before running spark-shell.

If you are writing a standalone program and want to run on a Mesos cluster (rather than locally), you will also need to pass two other arguments to SparkContext's constructor: sparkHome, the path at which Spark is installed on the cluster machines, and jars, a list of JAR files containing your job's classes to ship to the workers. Alternatively, you could copy the code to the workers manually and set the SPARK_CLASSPATHenvironment variable to point to it; see Configuration for details.

Distributed Datasets

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 andreduce operations as follows: distFile.map(_.size).reduce(_ + _).

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.

Distributed Dataset Operations

Distributed datasets 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.

One important transformation provided by Spark is caching. When you cache a distributed dataset, 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.

The following tables list the transformations and actions supported:

Transformations

Transformation	Meaning
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).
sample(withReplacement,frac, seed)	Sample a fraction frac of the data, with or without replacement, using a given random number seed.
union(otherDataset)	Return a new dataset that contains the union of the elements in the source dataset and the argument.
groupByKey()	When called on a dataset of (K, V) pairs, returns a dataset of (K, Seq[V]) pairs.
reduceByKey(func)	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.
join(otherDataset)	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.
groupWith(otherDataset)	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 CoGroup in other frameworks.
cartesian(otherDataset)	When called on datasets of types T and U, returns a dataset of (T, U) pairs (all pairs of elements).
cache()	Asks that the dataset be cached. Slices will be stored in memory on cluster nodes after being computed, making subsequent uses of the dataset faster.

Actions

Action	Meaning
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.
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.
first()	Return the first element of the dataset (similar to take(1)).
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).
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.

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

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 jobs 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.

================ 使用场景 =======================================

Spark is built around  distributed datasets that support types of parallel operations: transformations, which are lazy and yield another distributed dataset (e.g.,  map, filter, and  join), and actions, which force the computation of a dataset and return a result (e.g.,  count). The following examples show off some of the available operations and features.

Text Search

In this example, we search through the error messages in a log file:

val file = spark.textFile( "hdfs://...")
val errors = file. filter( line => line.contains("ERROR"))
// Count all the errors
errors. count()
// Count errors mentioning MySQL
errors. filter( line => line.contains("MySQL")). count()
// Fetch the MySQL errors as an array of strings
errors. filter( line => line.contains("MySQL")). collect()

The red code fragments are Scala function literals (closures) that get passed automatically to the cluster. The blue ones are Spark operations.

In-Memory Text Search

Spark can cache datasets in memory to speed up reuse. In the example above, we can load just the error messages in RAM using:

errors. cache()

After the first action that uses errors, later ones will be much faster.

Word Count

In this example, we use a few more transformations to build a dataset of (String, Int) pairs called counts and then save it to a file.

val file = spark.textFile( "hdfs://...")
val counts = file. flatMap( line => line.split(" "))
                  . map( word => (word, 1))
                  . reduceByKey( _ + _)
counts. saveAsTextFile( "hdfs://...")

Estimating Pi

Spark can also be used for compute-intensive tasks. This code estimates π by "throwing darts" at a circle. We pick random points in the unit square ((0, 0) to (1,1)) and see how many fall in the unit circle. The fraction should be π / 4, so we use this to get our estimate.

val count = spark.parallelize(1 to NUM_SAMPLES). map( i =>
  val x = Math.random()
  val y = Math.random()
  if (x*x + y*y < 1) 1.0 else 0.0
). reduce( _ + _)
println( "Pi is roughly " + 4 * count / NUM_SAMPLES)

Logistic Regression

This is an iterative machine learning algorithm that seeks to find the best hyperplane that separates two sets of points in a multi-dimensional feature space. It can be used to classify messages into spam vs non-spam, for example. Because the algorithm applies the same MapReduce operation repeatedly to the same dataset, it benefits greatly from caching the input data in RAM across iterations.

val points = spark.textFile(...). map(parsePoint). cache()
var w = Vector.random(D)  // current separating plane
for (i <- 1 to ITERATIONS) {
   val gradient = points. map( p =>
    (1 / (1 + exp(-p.y*(w dot p.x))) - 1) * p.y * p.x
  ). reduce( _ + _)
  w -= gradient
}
println( "Final separating plane: " + w)

Note that w gets shipped automatically to the cluster with every map call.

The graph below compares the performance of this Spark program against a Hadoop implementation on 30 GB of data on an 80-core cluster, showing the benefit of in-memory caching: