[ Skip to the content ]

Institute of Formal and Applied Linguistics Wiki


[ Back to the navigation ]

Table of Contents

Spark Introduction

This introduction shows several simple examples to give you an idea what programming in Spark is like. See the official Quick Start or Spark Programming Guide or Python API Reference/Scala API Reference for more information.

Running Spark Shell in Python

To run interactive Python shell in local Spark mode, run (on your local workstation or on cluster using qrsh from lrc1)

IPYTHON=1 pyspark

The IPYTHON=1 parameter instructs Spark to use ipython instead of python (the ipython is an enhanced interactive shell than Python). If you do not want ipython or you do not have it installed (it is installed everywhere on the cluster, but maybe not on your local workstations – ask our IT if you want it), use only pyspark, but note that it has some issues when copy-pasting examples from this wiki.

After a local Spark executor is started, the Python shell starts. Severel lines above
the prompt line, the SparkUI address is listed in the following format:

14/10/03 10:54:35 INFO SparkUI: Started SparkUI at http://tauri4.ufal.hide.ms.mff.cuni.cz:4040

The SparkUI is an HTML interface which displays the state of the application – if a distributed computation is taking place, how many workers are part of it, how many tasks are left to be processed, any error logs, also cached datasets and their properties (cached on disk / memory, their size) are displayed.

Running Spark Shell in Scala

To run interactive Scala shell in local Spark mode, run (on your local workstation or on cluster)

spark-shell

Once again, the SparkUI address is listed several lines above the shell prompt line.

Word Count Example

The central object of Spark framework is RDD – resilient distributed dataset. It contains ordered sequence of items, which may be distributed in several threads or on several computers. Spark offers multiple operations which can be performed on RDD, like map, filter, reduceByKey, union, join, sortBy, sample etc.

We start by simple word count example. We load the RDD from text file, every line of the input file becoming an element of RDD. We then split every line into words, count every word occurrence and sort the words by the occurrences. Copy the following to the opened Python shell:

wiki = sc.textFile("/net/projects/spark-example-data/wiki-cs")
words = wiki.flatMap(lambda line: line.split())
counts = words.map(lambda word: (word, 1)).reduceByKey(lambda c1,c2: c1+c2)
sorted = counts.sortBy(lambda (word,count): count, ascending=False)
sorted.saveAsTextFile('output')
 
# Alternatively, we can avoid variables:
(sc.textFile("/net/projects/spark-example-data/wiki-cs")
   .flatMap(lambda line: line.split())
   .map(lambda word: (word, 1))
   .reduceByKey(lambda c1,c2: c1+c2)
   .sortBy(lambda (word,count): count, ascending=False)
   .take(10)) # Instead of saveAsTextFile, we only print 10 most frequent words

The output of 'saveAsTextFile' is the directory output – because the RDD can be distributed on several computers, the output is a directory containing possibly multiple files.

Note that 'map' and 'reduceByKey' operations exist, allowing any Hadoop MapReduce operation to be implemented. On the other hand, several operations like 'join', 'sortBy', 'cogroup' are available, which are not available in Hadoop (or at least not directly), making Spark computational model a strict superset of Hadoop computational model.

The Scala versions is quite similar:

val wiki = sc.textFile("/net/projects/spark-example-data/wiki-cs")
val words = wiki.flatMap(line => line.split("\\s"))
val counts = words.map(word => (word,1)).reduceByKey((c1,c2) => c1+c2)
val sorted = counts.sortBy({case (word, count) => count}, ascending=false)
sorted.saveAsTextFile("output")
 
// Alternatively without variables and using placeholders in lambda parameters:
(sc.textFile("/net/projects/spark-example-data/wiki-cs")
   .flatMap(_.split("\\s"))
   .map((_,1)).reduceByKey(_+_)
   .sortBy(_._2, ascending=false)
   .take(10))

K-Means Example

An example implementing Standard iterative K-Means algorithm follows. Try copying it to open Python shell. Note that this wiki is formating empty lines as lines with one space, which is confusing for pyspark used without IPYTHON=1, so either use IPYTHON=1 or copy the text paragraph-by-paragraph.

import numpy as np
 
def closestPoint(point, centers):   # Find index of center which is closes to given point
    return min((np.sum((point - centers[i]) ** 2), i) for i in range(len(centers)))[1]
 
lines = sc.textFile("/net/projects/spark-example-data/points", sc.defaultParallelism)
data = lines.map(lambda line: np.array(map(float, line.split()))).cache()
 
K = 50
epsilon = 1e-3
 
centers = data.takeSample(False, K)       # Sample K random points
for i in range(5):                        # Perform 5 iterations
    old_centers = sc.broadcast(centers)
    centers = (data
               # For each point, find its closest center index.
               .map(lambda point: (closestPoint(point, old_centers.value), (point, 1)))
               # Sum points and counts in each cluster.
               .reduceByKey(lambda (p1, c1), (p2, c2): (p1 + p2, c1 + c2))
               # Sort by cluster index.
               .sortByKey()
               # Compute the new centers by averaging points in clusters.
               .map(lambda (index, (sum, count)): sum / count)
               .collect())
    # If the change in center positions is less than epsilon, stop.
    centers_change = sum(np.sqrt(np.sum((a - b)**2)) for (a, b) in zip(centers, old_centers.value))
    old_centers.unpersist()
    if centers_change < epsilon:
        break
 
print "Final centers: " + str(centers)

The implementation starts by loading the data points and caching them in memory using cache. Then, standard iterative algorithm is performed, running in parallel and synchronizing where necessary.

Note that explicit broadcasting used for centers object is not strictly needed – if we used old_centers = centers, the example would work too, but it would send a copy of old_centers to every distributed task, instead of once to every machine.

For illustration, Scala version of the example follows. It works exactly as the Python version and uses breeze.linalg.Vector providing linear algebraic operations.

import breeze.linalg.Vector
 
type Vector = breeze.linalg.Vector[Double]
type Vectors = Array[Vector]
 
def closestPoint(point : Vector, centers : Vectors) : Double =
  centers.map(center => (center-point).norm(2)).zipWithIndex.min._2
 
val lines = sc.textFile("/net/projects/spark-example-data/points", sc.defaultParallelism)
val data = lines.map(line => Vector(line.split("\\s+").map(_.toDouble))).cache()
 
val K = 50
val epsilon = 1e-3
 
var i = 0
var centers_change = Double.PositiveInfinity
var centers = data.takeSample(false, K)
while (i < 10 && centers_change > epsilon) {
  val old_centers = sc.broadcast(centers)
  centers = (data
             // For each point, find its closes center index.
             .map(point => (closestPoint(point, old_centers.value), (point, 1)))
             // Sum points and counts in each cluster.
             .reduceByKey((a, b) => (a._1+b._1, a._2+b._2))
             // Sort by cluster index.
             .sortByKey()
             // Compute the new centers by averaging corresponding points.
             .map({case (index, (sum, count)) => sum :/ count.toDouble})
             .collect())
 
  // Compute change in center positions.
  centers_change = (centers zip old_centers.value).map({case (a,b) => (a-b).norm(2)}).sum
  old_centers.unpersist()
  i += 1
}
 
print(centers.deep)

[ Back to the navigation ] [ Back to the content ]