Apache Spark is a fast and general engine for distributed computing & big data processing with APIs in Scala, Java, Python, and R. This tutorial will briefly introduce PySpark (the Python API for Spark) with some hands-on-exercises combined with a quick introduction to Spark's core concepts. We will cover the obligatory wordcount example which comes in with every big-data tutorial, as well as discuss Spark's unique methods for handling node failure and other relevant internals. Then we will briefly look at how to access some of Spark's libraries (like Spark SQL & Spark ML) from Python. While Spark is available in a variety of languages this workshop will be focused on using Spark and Python together.

1. PySpark
Lightning fast cluster computing with Python

2. Who am I?
Holden
● I prefer she/her for pronouns
● Co-author of the Learning Spark book
● @holdenkarau
● http://www.slideshare.net/hkarau
● https://www.linkedin.com/in/holdenkarau

3. What we are going to explore together!
● What is Spark?
● Getting Spark setup locally OR getting access to cluster
● Spark primary distributed collection
● Word count
● How PySpark works
● Using libraries with Spark
● Spark SQL / DataFrames

4. What is Spark?
● General purpose distributed system
○ With a really nice API
● Apache project (one of the most active)
● Must faster than Hadoop Map/Reduce

5. The different pieces of Spark
Apache Spark
SQL &
DataFrames
Streaming
Language
APIs
Scala,
Java,
Python, &
R
Graph
Tools
Spark ML
bagel &
Grah X
MLLib
Community
Packages

6. Setup time!
Remote Azure HDI cluster:
http://bit.ly/clusterSignup (thanks Microsoft!)
We can use Jupyter :)
Local Machine:
If you don’t have Spark installed you can get it from http:
//spark.apache.org/downloads.html (select 1.3.1, any
hadoop version)

7. Some pages to keep open for the exercises
http://bit.ly/sparkDocs
http://bit.ly/sparkPyDocs
http://bit.ly/PySparkIntroExamples
http://bit.ly/learningSparkExamples
OR
http://spark.apache.org/docs/latest/api/python/index.html
http://spark.apache.org/docs/latest/
https://github.com/holdenk/intro-to-pyspark-demos

8. Starting the shell
./bin/pyspark
[Lots of output]
SparkContext available as sc, SQLContext available as
sqlContext.
>>>

9. Reducing log level
cp ./conf/log4j.properties.template ./conf/log4j.properties
Then set
log4j.rootCategory=ERROR, console

10. Connecting to your Azure cluster
● Don’t screw up the password (gets cached)
● Use the Jupyter link
● Optionally you can configure your cluster to assign more
executor cores to Jupyter

11. Sparkcontext: entry to the world
● Can be used to create RDDs from many input sources
○ Native collections, local & remote FS
○ Any Hadoop Data Source
● Also create counters & accumulators
● Automatically created in the shells (called sc)
● Specify master & app name when creating
○ Master can be local[*], spark:// , yarn, etc.
○ app name should be human readable and make sense
● etc.

12. Getting the Spark Context on Azure
from pyspark import SparkContext
from pyspark.sql.types import *
sc = SparkContext( 'spark://headnodehost:7077', 'pyspark')

13. RDDs: Spark’s Primary abstraction
RDD (Resilient Distributed Dataset)
● Recomputed on node failure
● Distributed across the cluster
● Lazily evaluated (transformations & actions)

14. Word count
lines = sc.textFile(src)
words = lines.flatMap(lambda x: x.split(" "))
word_count =
(words.map(lambda x: (x, 1))
.reduceByKey(lambda x, y: x+y))
word_count.saveAsTextFile(output)

15. Word count
lines = sc.textFile(src)
words = lines.flatMap(lambda x: x.split(" "))
word_count =
(words.map(lambda x: (x, 1))
.reduceByKey(lambda x, y: x+y))
word_count.saveAsTextFile(output)
No data is read or
processed until after
this line
This is an “action”
which forces spark to
evaluate the RDD

16. Some common transformations & actions
Transformations (lazy)
● map
● filter
● flatMap
● reduceByKey
● join
● cogroup
Actions (eager)
● count
● reduce
● collect
● take
● saveAsTextFile
● saveAsHadoop
● countByValue
Photo by Steve
Photo by Dan G

17. Exercise time
Photo by recastle

18. Lets find the lines with the word “Spark”
import os
src = "file:///"+os.environ['SPARK_HOME']+"/README.md"
lines = sc.textFile(src)

19. What did you find?

20. A solution:
lines = sc.textFile(src)
spark_lines = lines.filter(
lambda x: x.lower().find("spark") != -1)
print spark_lines.count()

21. Combined with previous example
Do you notice anything funky?
● We read the data in twice :(
● cache/persist/checkpoint to the rescue!

22. lets use toDebugString
un-cached:
>>> print word_count.toDebugString()
(2) PythonRDD[17] at RDD at PythonRDD.scala:43 []
| MapPartitionsRDD[14] at mapPartitions at PythonRDD.scala:346 []
| ShuffledRDD[13] at partitionBy at NativeMethodAccessorImpl.java:-2 []
+-(2) PairwiseRDD[12] at reduceByKey at <stdin>:3 []
| PythonRDD[11] at reduceByKey at <stdin>:3 []
| MapPartitionsRDD[10] at textFile at NativeMethodAccessorImpl.java:-2 []
| file:////home/holden/repos/spark/README.md HadoopRDD[9] at textFile at NativeMethodAccessorImpl.java:-2 []

23. lets use toDebugString
cached:
>>> print word_count.toDebugString()
(2) PythonRDD[8] at RDD at PythonRDD.scala:43 []
| MapPartitionsRDD[5] at mapPartitions at PythonRDD.scala:346 []
| ShuffledRDD[4] at partitionBy at NativeMethodAccessorImpl.java:-2 []
+-(2) PairwiseRDD[3] at reduceByKey at <stdin>:3 []
| PythonRDD[2] at reduceByKey at <stdin>:3 []
| MapPartitionsRDD[1] at textFile at NativeMethodAccessorImpl.java:-2 []
| CachedPartitions: 2; MemorySize: 2.7 KB; ExternalBlockStoreSize: 0.0 B; DiskSize: 0.0 B
| file:////home/holden/repos/spark/README.md HadoopRDD[0] at textFile at NativeMethodAccessorImpl.java:-2 []

24. A detour into the internals
Photo by Bill Ward

25. Why lazy evaluation?
● Allows pipelining procedures
○ Less passes over our data, extra happiness
● Can skip materializing intermediate results which are
really really big*
● Figuring out where our code fails becomes a little
trickier

26. So what happens when we run this code?
Driver
Worker
Worker
Worker
HDFS /
Cassandra/
etc

27. So what happens when we run this code?
Driver
Worker
Worker
Worker
HDFS /
Cassandra/
etc
function

28. So what happens when we run this code?
Driver
Worker
Worker
Worker
HDFS /
Cassandra/
etc
read
read
read

29. So what happens when we run this code?
Driver
Worker
Worker
Worker
HDFS /
Cassandra/
etc
cached
cached
cached
counts

30. Spark in Scala, how does PySpark work?
● Py4J + pickling + magic
○ This can be kind of slow sometimes
● RDDs are generally RDDs of pickled objects
● Spark SQL (and DataFrames) avoid some of this

31. So what does that look like?
Driver
py4j
Worker 1
Worker K
pipe
pipe

32. Using other libraries
● built ins
○ just import!*
■ Except for Hive, compile with -PHive & then import
● spark-packages
○ --packages
● generic python
○ pre-install on workers (pssh, puppet, etc.)
○ add it with --zip-files
○ sc.addPyFile

33. So lets take “DataFrames” out for a spin
● useful for structured data
● support schema inference on JSON
● Many operations done without* pickling
● Integrated into ML!
● Accessed through SQLContext
● Not the same feature set as Panda’s or R DataFrames

34. Loading data
df = sqlContext.read.load(
"files/testweet.json",
format="json")
# Built in json, parquet, etc.
# More formats (csv, etc.) at http://spark-packages.org/

35. DataFrames aren’t quite as lazy...
● Keep track of schema information
● Loading JSON data involves looking at the data
● Before if we tried to load non-existent data wouldn’t fail
right away, now fails right away

36. Examining Schema Information
root
|-- contributorsIDs: array (nullable = true)
| |-- element: string (containsNull = true)
|-- createdAt: string (nullable = true)
|-- currentUserRetweetId: long (nullable = true)
|-- hashtagEntities: array (nullable = true)
| |-- element: string (containsNull = true)
|-- id: long (nullable = true)
|-- inReplyToStatusId: long (nullable = true)
|-- inReplyToUserId: long (nullable = true)
|-- isFavorited: boolean (nullable = true)
|-- isPossiblySensitive: boolean (nullable = true)
|-- isTruncated: boolean (nullable = true)

37. Manipulating DataFrames
SQL
df.registerTempTable("panda")
sqlContext.sql("select * from panda where id =
529799371026485248")
API
df.filter(df.id == 529799371026485248)

38. DataFrames to RDD’s & vice versa
● map lets us work per-row
df.map(lambda row: row.text)
● Converting back
○ infer_schema
○ specify the schema

39. Or we can make a UDF
def function(x):
# Some magic
sqlContext.registerFunction(“name”, function,
IntegerType())

40. More exercise funtimes :)
● Lets load a sample tweet
● Write a UDF to compute the length of the tweet
● Select the length of the tweet

41. Getting some tweets
● Could use Spark Streaming sample app if you have
twitter keys handy
● Normally we would read data from HDFS or similar

42. import urllib2
data = urllib2.urlopen('https://raw.githubusercontent.
com/databricks/learning-spark/master/files/testweet.json').
read()
print data
rdd = sc.parallelize([data])
path = "mytextFile.txt"
rdd.saveAsTextFile(path)

43. Loading the tweets
df = sqlContext.jsonFile(path)
df.printSchema()

44. MLLib / ML
● Example in the notebook :)

45. Additional Resources
● Programming guide (along with JavaDoc, PyDoc,
ScalaDoc, etc.)
○ http://spark.apache.org/docs/latest/
● Books
● Videos
● Training
● My talk tomorrow

46. Learning Spark
Fast Data
Processing with
Spark
(Out of Date)
Fast Data
Processing with
Spark
Advanced
Analytics with
Spark
Coming soon:
Spark in Action

47. Conferences & Meetups
● Strata & Hadoop World (next one in NYC)
● Spark summit (next one in Amsterdam)
● Seattle Spark Meetup (next event on Aug 12th)
& more at http://spark.apache.org/community.html#events

48. Spark Videos
● Apache Spark Youtube Channel
● Spark Summit 2014 training
● Paco’s Introduction to Apache Spark