1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
|
---
layout: global
title: Examples
type: "page singular"
navigation:
weight: 4
show: true
---
<h2>Spark Examples</h2>
These examples give a quick overview of the Spark API.
Spark is built on the concept of <em>distributed datasets</em>, which contain arbitrary Java or
Python objects. You create a dataset from external data, then apply parallel operations
to it. There are two types of operations: <em>transformations</em>, which define a new dataset based on
previous ones, and <em>actions</em>, which kick off a job to execute on a cluster.
<h3>Text Search</h3>
In this example, we search through the error messages in a log file:
<ul class="nav nav-tabs">
<li class="lang-tab lang-tab-scala active"><a href="#">Scala</a></li>
<li class="lang-tab lang-tab-java"><a href="#">Java</a></li>
<li class="lang-tab lang-tab-python"><a href="#">Python</a></li>
</ul>
<div class="tab-content">
<div class="tab-pane tab-pane-scala active">
<div class="code code-tab">
<span class="keyword">val</span> file = spark.textFile(<span class="string">"hdfs://..."</span>)<br>
<span class="keyword">val</span> errors = file.<span class="sparkop">filter</span>(<span class="closure">line => line.contains("ERROR")</span>)<br>
<span class="comment">// Count all the errors</span><br>
errors.<span class="sparkop">count</span>()<br>
<span class="comment">// Count errors mentioning MySQL</span><br>
errors.<span class="sparkop">filter</span>(<span class="closure">line => line.contains("MySQL")</span>).<span class="sparkop">count</span>()<br>
<span class="comment">// Fetch the MySQL errors as an array of strings</span><br>
errors.<span class="sparkop">filter</span>(<span class="closure">line => line.contains("MySQL")</span>).<span class="sparkop">collect</span>()<br>
</div>
</div>
<div class="tab-pane tab-pane-java">
<div class="code code-tab">
JavaRDD<String> file = spark.textFile(<span class="string">"hdfs://..."</span>);<br>
JavaRDD<String> errors = file.<span class="sparkop">filter</span>(<span class="closure">new Function<String, Boolean>() {<br>
public Boolean call(String s) { return s.contains("ERROR"); }<br>
}</span>);<br>
<span class="comment">// Count all the errors</span><br>
errors.<span class="sparkop">count</span>();<br>
<span class="comment">// Count errors mentioning MySQL</span><br>
errors.<span class="sparkop">filter</span>(<span class="closure">new Function<String, Boolean>() {<br>
public Boolean call(String s) { return s.contains("MySQL"); }<br>
}</span>).<span class="sparkop">count</span>();<br>
<span class="comment">// Fetch the MySQL errors as an array of strings</span><br>
errors.<span class="sparkop">filter</span>(<span class="closure">new Function<String, Boolean>() {<br>
public Boolean call(String s) { return s.contains("MySQL"); }<br>
}</span>).<span class="sparkop">collect</span>();<br>
</div>
</div>
<div class="tab-pane tab-pane-python">
<div class="code code-tab">
file = spark.textFile(<span class="string">"hdfs://..."</span>)<br>
errors = file.<span class="sparkop">filter</span>(<span class="closure">lambda line: "ERROR" in line</span>)<br>
<span class="comment"># Count all the errors</span><br>
errors.<span class="sparkop">count</span>()<br>
<span class="comment"># Count errors mentioning MySQL</span><br>
errors.<span class="sparkop">filter</span>(<span class="closure">lambda line: "MySQL" in line</span>).<span class="sparkop">count</span>()<br>
<span class="comment"># Fetch the MySQL errors as an array of strings</span><br>
errors.<span class="sparkop">filter</span>(<span class="closure">lambda line: "MySQL" in line</span>).<span class="sparkop">collect</span>()<br>
</div>
</div>
</div>
<p>The red code fragments are function literals (closures) that get passed automatically to the cluster. The blue ones are Spark operations.</p>
<h3>In-Memory Text Search</h3>
<p>Spark can <em>cache</em> datasets in memory to speed up reuse. In the example above, we can load just the error messages in RAM using:</p>
<ul class="nav nav-tabs">
<li class="lang-tab lang-tab-scala active"><a href="#">Scala</a></li>
<li class="lang-tab lang-tab-java"><a href="#">Java</a></li>
<li class="lang-tab lang-tab-python"><a href="#">Python</a></li>
</ul>
<div class="tab-content">
<div class="tab-pane tab-pane-scala active">
<div class="code code-tab">
errors.<span class="sparkop">cache</span>()
</div>
</div>
<div class="tab-pane tab-pane-java">
<div class="code code-tab">
errors.<span class="sparkop">cache</span>();
</div>
</div>
<div class="tab-pane tab-pane-python">
<div class="code code-tab">
errors.<span class="sparkop">cache</span>()
</div>
</div>
</div>
<p>After the first action that uses <code>errors</code>, later ones will be much faster.</p>
<h3>Word Count</h3>
<p>In this example, we use a few more transformations to build a dataset of (String, Int) pairs called <code>counts</code> and then save it to a file.</p>
<ul class="nav nav-tabs">
<li class="lang-tab lang-tab-scala active"><a href="#">Scala</a></li>
<li class="lang-tab lang-tab-java"><a href="#">Java</a></li>
<li class="lang-tab lang-tab-python"><a href="#">Python</a></li>
</ul>
<div class="tab-content">
<div class="tab-pane tab-pane-scala active">
<div class="code code-tab">
<span class="keyword">val</span> file = spark.textFile(<span class="string">"hdfs://..."</span>)<br>
<span class="keyword">val</span> counts = file.<span class="sparkop">flatMap</span>(<span class="closure">line => line.split(" ")</span>)<br>
.<span class="sparkop">map</span>(<span class="closure">word => (word, 1)</span>)<br>
.<span class="sparkop">reduceByKey</span>(<span class="closure">_ + _</span>)<br>
counts.<span class="sparkop">saveAsTextFile</span>(<span class="string">"hdfs://..."</span>)
</div>
</div>
<div class="tab-pane tab-pane-java">
<div class="code code-tab">
JavaRDD<String> file = spark.textFile(<span class="string">"hdfs://..."</span>);<br>
JavaRDD<String> words = file.<span class="sparkop">flatMap</span>(<span class="closure">new FlatMapFunction<String, String>()<br>
public Iterable<String> call(String s) { return Arrays.asList(s.split(" ")); }<br>
}</span>);<br>
JavaPairRDD<String, Integer> pairs = words.<span class="sparkop">map</span>(<span class="closure">new PairFunction<String, String, Integer>()<br>
public Tuple2<String, Integer> call(String s) { return new Tuple2<String, Integer>(s, 1); }<br>
}</span>);<br>
JavaPairRDD<String, Integer> counts = pairs.<span class="sparkop">reduceByKey</span>(<span class="closure">new Function2<Integer, Integer>()<br>
public Integer call(Integer a, Integer b) { return a + b; }<br>
}</span>);<br>
counts.<span class="sparkop">saveAsTextFile</span>(<span class="string">"hdfs://..."</span>);
</div>
</div>
<div class="tab-pane tab-pane-python">
<div class="code code-tab">
file = spark.textFile(<span class="string">"hdfs://..."</span>)<br>
counts = file.<span class="sparkop">flatMap</span>(<span class="closure">lambda line: line.split(" ")</span>) \<br>
.<span class="sparkop">map</span>(<span class="closure">lambda word: (word, 1)</span>) \<br>
.<span class="sparkop">reduceByKey</span>(<span class="closure">lambda a, b: a + b</span>)<br>
counts.<span class="sparkop">saveAsTextFile</span>(<span class="string">"hdfs://..."</span>)
</div>
</div>
</div>
<h3>Estimating Pi</h3>
<p>Spark can also be used for compute-intensive tasks. This code estimates <span style="font-family: serif; font-size: 120%;">π</span> 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 <span style="font-family: serif; font-size: 120%;">π / 4</span>, so we use this to get our estimate.</p>
<ul class="nav nav-tabs">
<li class="lang-tab lang-tab-scala active"><a href="#">Scala</a></li>
<li class="lang-tab lang-tab-java"><a href="#">Java</a></li>
<li class="lang-tab lang-tab-python"><a href="#">Python</a></li>
</ul>
<div class="tab-content">
<div class="tab-pane tab-pane-scala active">
<div class="code code-tab">
<span class="keyword">val</span> count = spark.parallelize(1 to NUM_SAMPLES).<span class="sparkop">map</span>(<span class="closure">i =><br>
val x = Math.random()<br>
val y = Math.random()<br>
if (x*x + y*y < 1) 1 else 0<br>
</span>).<span class="sparkop">reduce</span>(<span class="closure">_ + _</span>)<br>
println(<span class="string">"Pi is roughly "</span> + 4.0 * count / NUM_SAMPLES)<br>
</div>
</div>
<div class="tab-pane tab-pane-java">
<div class="code code-tab">
<span class="keyword">int</span> count = spark.parallelize(makeRange(1, NUM_SAMPLES)).<span class="sparkop">filter</span>(<span class="closure">new Function<Integer, Boolean>() {<br>
public Integer call(Integer i) {<br>
double x = Math.random();<br>
double y = Math.random();<br>
return x*x + y*y < 1;<br>
}<br>
}</span>).<span class="sparkop">count</span>();<br>
System.out.println(<span class="string">"Pi is roughly "</span> + 4 * count / NUM_SAMPLES);<br>
</div>
</div>
<div class="tab-pane tab-pane-python">
<div class="code code-tab">
<span class="keyword">def</span> sample(p):<br>
x, y = random(), random()<br>
<span class="keyword">return</span> 1 <span class="keyword">if</span> x*x + y*y < 1 <span class="keyword">else</span> 0<br><br>
count = spark.parallelize(xrange(0, NUM_SAMPLES)).<span class="sparkop">map</span>(<span class="closure">sample</span>) \<br>
.<span class="sparkop">reduce</span>(<span class="closure">lambda a, b: a + b</span>)<br>
print <span class="string">"Pi is roughly %f"</span> % (4.0 * count / NUM_SAMPLES)<br>
</div>
</div>
</div>
<h3>Logistic Regression</h3>
<p>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 in RAM across iterations.</p>
<ul class="nav nav-tabs">
<li class="lang-tab lang-tab-scala active"><a href="#">Scala</a></li>
<li class="lang-tab lang-tab-java"><a href="#">Java</a></li>
<li class="lang-tab lang-tab-python"><a href="#">Python</a></li>
</ul>
<div class="tab-content">
<div class="tab-pane tab-pane-scala active">
<div class="code code-tab">
<span class="keyword">val</span> points = spark.textFile(...).<span class="sparkop">map</span>(parsePoint).<span class="sparkop">cache</span>()<br>
<span class="keyword">var</span> w = Vector.random(D) <span class="comment">// current separating plane</span><br>
<span class="keyword">for</span> (i <- 1 to ITERATIONS) {<br>
<span class="keyword">val</span> gradient = points.<span class="sparkop">map</span>(<span class="closure">p =><br>
(1 / (1 + exp(-p.y*(w dot p.x))) - 1) * p.y * p.x<br>
</span>).<span class="sparkop">reduce</span>(<span class="closure">_ + _</span>)<br>
w -= gradient<br>
}<br>
println(<span class="string">"Final separating plane: "</span> + w)<br>
</div>
</div>
<div class="tab-pane tab-pane-java">
<div class="code code-tab">
<span class="keyword">class</span> ComputeGradient <span class="keyword">extends</span> Function<DataPoint, Vector> {<br>
<span class="keyword">private</span> Vector w;<br>
ComputeGradient(Vector w) { <span class="keyword">this</span>.w = w; }<br>
<span class="keyword">public</span> Vector call(DataPoint p) {<br>
<span class="keyword">return</span> p.x.times(p.y * (1 / (1 + Math.exp(w.dot(p.x))) - 1));<br>
}<br>
}<br>
<br>
JavaRDD<DataPoint> points = spark.textFile(...).<span class="sparkop">map</span>(<span class="closure">new ParsePoint()</span>).<span class="sparkop">cache</span>();<br>
Vector w = Vector.random(D); <span class="comment">// current separating plane</span><br>
<span class="keyword">for</span> (<span class="keyword">int</span> i = 0; i < ITERATIONS; i++) {<br>
Vector gradient = points.<span class="sparkop">map</span>(<span class="closure">new ComputeGradient(w)</span>).<span class="sparkop">reduce</span>(<span class="closure">new AddVectors()</span>);<br>
w = w.subtract(gradient);<br>
}<br>
System.out.println(<span class="string">"Final separating plane: "</span> + w);<br>
</div>
</div>
<div class="tab-pane tab-pane-python">
<div class="code code-tab">
points = spark.textFile(...).<span class="sparkop">map</span>(parsePoint).<span class="sparkop">cache</span>()<br>
w = numpy.random.ranf(size = D) <span class="comment"># current separating plane</span><br>
<span class="keyword">for</span> i <span class="keyword">in</span> range(ITERATIONS):<br>
gradient = points.<span class="sparkop">map</span>(<span class="closure"><br>
lambda p: (1 / (1 + exp(-p.y*(w.dot(p.x)))) - 1) * p.y * p.x<br>
</span>).<span class="sparkop">reduce</span>(<span class="closure">lambda a, b: a + b</span>)<br>
w -= gradient<br>
print <span class="string">"Final separating plane: %s"</span> % w<br>
</div>
</div>
</div>
<p>Note that the current separating plane, <code>w</code>, gets shipped automatically to the cluster with every <code>map</code> call.</p>
<p>The graph below compares the running time per iteration of this Spark program against a Hadoop implementation on 100 GB of data on a 100-node cluster, showing the benefit of in-memory caching:</p>
<p style="margin-top: 20px; margin-bottom: 30px;">
<img src="{{site.url}}images/logistic-regression.png" alt="Logistic regression performance in Spark vs Hadoop">
</p>
<a name="additional"></a>
<h2>Additional Examples</h2>
Many additional examples are distributed with Spark:
* Basic Spark: [Scala examples](https://github.com/apache/spark/tree/master/examples/src/main/scala/org/apache/spark/examples), [Java examples](https://github.com/apache/spark/tree/master/examples/src/main/java/org/apache/spark/examples), [Python examples](https://github.com/apache/spark/tree/master/python/examples)
* Spark Streaming: [Scala examples](https://github.com/apache/spark/tree/master/examples/src/main/scala/org/apache/spark/streaming/examples), [Java examples](https://github.com/apache/spark/tree/master/examples/src/main/java/org/apache/spark/streaming/examples)
|
