Dev Ops Training

DEVOPS ADVANCED CLASS
June 2015: Spark Summit West 2015
http://training.databricks.com/devops.pdf
www.linkedin.com/in/blueplastic

making big data simple
Databricks Cloud:
“A unified platform for building Big Data pipelines
– from ETL to Exploration and Dashboards, to
Advanced Analytics and Data Products.”
• Founded in late 2013
• by the creators of Apache Spark
• Original team from UC Berkeley AMPLab
• Raised $47 Million in 2 rounds
• ~55 employees
• We’re hiring!
• Level 2/3 support partnerships with
• Hortonworks
• MapR
• DataStax
(http://databricks.workable.com)

The Databricks team contributed more than 75% of the code added to Spark in the past year

AGENDA
• History of Spark
• RDD fundamentals
• Spark Runtime Architecture
Integration with Resource Managers
(Standalone, YARN)
• GUIs
• Lab: DevOps 101
Before Lunch
• Memory and Persistence
• Jobs -> Stages -> Tasks
• Broadcast Variables and
Accumulators
• PySpark
• DevOps 102
• Shuffle
• Spark Streaming
After Lunch

Some slides will be skipped
Please keep Q&A low during class
(5pm – 5:30pm for Q&A with instructor)
2 anonymous surveys: Pre and Post class
Lunch: noon – 1pm
2 breaks (before lunch and after lunch)

LinkedIn: https://www.linkedin.com/in/blueplastic
- 1 year: Client Services Engineer @ Databricks
- 2 years: Freelance Big Data Consulting + Training
- Taught 100+ classes
- Traveled to 20+ countries and 50+ cities
- 5 months: Systems Architect @ Hortonworks
- 1.5 years: Emerging Data Platforms Consultant @ Accenture R&D
- 2 years: Storage & Clustering Software Consultant @ Symantec
- 2.5 years: Tech Support Engineer @ Symantec
Wakeboarding / Snowboarding / Scuba Diving / Free-diving /
Kayaking / Running / Hiking / Canoeing / Surfing / Tennis

0 10 20 30 40 50 60 70 80 90 100
Sales / Marketing
Data Scientist
Management / Exec
Administrator / Ops
Developer
Survey completed by
89 out of 215 students

SF Bay Area
25%
CA
12%
West US
17%
East US
19%
Europe
8%
Asia
10%
Intern. - O
9%
Survey completed by

0 10 20 30 40 50 60 70
Advertising / PR
Healthcare / Medical
Academia / University
Telecom
Science & Tech
Banking / Finance
IT / Systems
Survey completed by

0 10 20 30 40 50 60 70 80
SparkSummit
Vendor Training
SparkCamp
AmpCamp
None
Survey completed by

Survey completed by
Zero
6%
< 1 week
16%
< 1 month
35%
1-6 months
43%

Reading
10%
1-node VM
21%
POC / Prototype
53%
Production
16%
Survey completed by

Survey completed by

0 10 20 30 40 50 60 70 80 90
Use Cases
Architecture
Administrator / Ops
Development
Survey completed by

@
• AMPLab project was launched in Jan 2011, 6 year planned duration
• Personnel: ~65 students, postdocs, faculty & staff
• Funding from Government/Industry partnership, NSF Award, Darpa, DoE,
20+ companies
• Created BDAS, Mesos, SNAP. Upcoming projects: Succinct & Velox.
“Unknown to most of the world, the University of California, Berkeley’s AMPLab
has already left an indelible mark on the world of information technology, and
even the web. But we haven’t yet experienced the full impact of the
group[…] Not even close”
- Derrick Harris, GigaOm, Aug 2014
Algorithms
Machines
People

NoSQL battles Compute battles
(then) (now)

Key -> Value Key -> Doc Column Family Graph Search
Redis - 95
Memcached - 33
DynamoDB - 16
Riak - 13
MongoDB - 279
CouchDB - 28
Couchbase - 24
DynamoDB – 15
MarkLogic - 11
Cassandra - 109
HBase - 62
Neo4j - 30
OrietnDB - 4
Titan – 3
Giraph - 1
Solr - 81
Elasticsearch - 70
Splunk – 41

General Batch Processing
Pregel
Dremel
Impala
GraphLab
Giraph
Drill
Tez
S4
Storm
Specialized Systems
(iterative, interactive, ML, streaming, graph, SQL, etc)
General Unified Engine
(2004 – 2013)
(2007 – 2015?)
(2014 – ?)
Mahout

RDBMS
Streaming
SQL
GraphX
Hadoop Input Format
Apps
Distributions:
- CDH
- HDP
- MapR
- DSE
Tachyon
MLlib
DataFrames API

- Developers from 50+ companies
- 400+ developers
- Apache Committers from 16+ organizations

vs
YARN
SQL
MLlib
Streaming
Mesos
Tachyon

Aug 2009
Source: openhub.net
...in June 2013

CPUs:
10 GB/s
100 MB/s
0.1 ms random access
$0.45 per GB
600 MB/s
3-12 ms random access
$0.05 per GB
1 Gb/s or
125 MB/s
Network
0.1 Gb/s
Nodes in
another
rack
Nodes in
same rack
1 Gb/s or
125 MB/s

June 2010
http://www.cs.berkeley.edu/~matei/papers/2010/hotcloud_spark.pdf
“The main abstraction in Spark is that of a resilient dis-
tributed dataset (RDD), which represents a read-only
collection of objects partitioned across a set of
machines that can be rebuilt if a partition is lost.
Users can explicitly cache an RDD in memory across
machines and reuse it in multiple MapReduce-like
parallel operations.
RDDs achieve fault tolerance through a notion of
lineage: if a partition of an RDD is lost, the RDD has
enough information about how it was derived from
other RDDs to be able to rebuild just that partition.”

April 2012
http://www.cs.berkeley.edu/~matei/papers/2012/nsdi_spark.pdf
“We present Resilient Distributed Datasets (RDDs), a
distributed memory abstraction that lets
programmers perform in-memory computations on
large clusters in a fault-tolerant manner.
RDDs are motivated by two types of applications
that current computing frameworks handle
inefficiently: iterative algorithms and interactive data
mining tools.
In both cases, keeping data in memory can improve
performance by an order of magnitude.”
“Best Paper Award and Honorable Mention for Community Award”
- NSDI 2012
- Cited 392 times!

TwitterUtils.createStream(...)
.filter(_.getText.contains("Spark"))
.countByWindow(Seconds(5))
- 2 Streaming Paper(s) have been cited 138 times
Analyze real time streams of data in ½ second intervals

Seemlessly mix SQL queries with Spark programs.
sqlCtx = new HiveContext(sc)
results = sqlCtx.sql(
"SELECT * FROM people")
names = results.map(lambda p: p.name)

graph = Graph(vertices, edges)
messages = spark.textFile("hdfs://...")
graph2 = graph.joinVertices(messages) {
(id, vertex, msg) => ...
}
Analyze networks of nodes and edges using graph processing
https://amplab.cs.berkeley.edu/wp-content/uploads/2013/05/grades-graphx_with_fonts.pdf

SQL queries with Bounded Errors and Bounded Response Times
https://www.cs.berkeley.edu/~sameerag/blinkdb_eurosys13.pdf

Estimate
# of data points
true answer
How do you know
when to stop?

Estimate
# of data points
true answer
Error bars on every
answer!

Estimate
# of data points
true answer
Stop when error smaller
than a given threshold
time

http://shop.oreilly.com/product/0636920028512.do
eBook: $33.99
Print: $39.99
PDF, ePub, Mobi, DAISY
Shipping now!
http://www.amazon.com/Learning-Spark-Lightning-Fast-
Data-Analysis/dp/1449358624
$30 @ Amazon:

http://shop.oreilly.com/product/0636920035091.do
eBook: $42.50
Print: $49.99
PDF, ePub, Mobi, DAISY
Shipping now!
http://www.amazon.com/Advanced-Analytics-Spark-
Patterns-Learning/dp/1491912766
$34.80 @ Amazon:

http://tinyurl.com/dsesparklab
- 102 pages
- DevOps style
- For complete beginners
- Includes:
- Spark Streaming
- Dangers of
GroupByKey vs.
ReduceByKey

http://tinyurl.com/cdhsparklab
- 109 pages
- DevOps style
- For complete beginners
- Includes:
- PySpark
- Spark SQL
- Spark-submit

• Introduced Spark SQL
• Introduced support for Hadoop Security and
PySpark on YARN
• Added Spark Submit
• Added History Server web UI
• MLlib improvements: sparse feature vectors,
scalable decision trees, SVD, PCA, L-BFGS
• Substantial performance boosts in GraphX
• Added flume support in Streaming
• Support for Java 8 lambda syntax
May 2014
1.0.0version
contributions from
117 developers

• New sort based Shuffle for very large scale
workloads (10k + reducers)
• Spark SQL additions: JDBC/ODBC server,
JSON support, UDFs
• New statistics library for MLlib and 2-3x speed
improvement for many algorithms
• New in Spark Streaming: partial HA, Amazon
Kinesis, Flume pull, streaming linear regression,
rate limiting
• PySpark now supports HBase, C*, Avro,
SequenceFiles
Sept 2014
1.1.0version
contributions from
171 developers

• New Netty based network transfer subsystem
for very large shuffles
• Scala 2.11 support
• Streaming: Python API, driver HA, WAL
• MLlib: ML Pipelines (multiple algorithms are
run in sequence with varying parameters),
major python improvements
• Spark SQL: External data sources API, Hive
0.13 support
• GraphX: graduates from alpha and adds a
stable API
• PySpark: broadcast variables > 2 GB
Dec 2014
1.2.0version
contributions from
172 developers
60+ institutions

• DataFrame API released
• SparkSQL graduates from an alpha project
• Core: reduce performance, error reporting,
some SSL encryption, GC metrics in UI
• MLlib: LDA for topic modeling, GMM,
multinomial logistic regression, FP-growth
• Streaming: direct Kafka API
• PySpark: Support for ML Pipelines
Mar 2015
1.3.0version
contributions from
174 developers

• Introduced SparkR
• Core: DAG visualization, Python 3 support,
REST API, serialized shuffle, beginning of
Tungsten
• SQL/DF: ORCFile format support, UI for JDBC
server
• ML pipelines graduates from alpha
• Many new ML algorithms
• Streaming: visual graphs in UI, better Kafka +
Kinesis support
Jun 2015
1.4.0version
contributions from
210 developers
Benchmark & Integration
testing by:
• Intel
• Palantir
• Cloudera
• Mesosphere
• Huawei
• Shopify
• Netflix
• Yahoo
• UC Berkeley
• Databricks
70+ institutions

Driver Program
Ex
RDD
W
RDD
T
T
Ex
RDD
W
RDD
T
T
Worker Machine
Worker Machine

item-1
item-2
item-3
item-4
item-5
item-6
item-7
item-8
item-9
item-10
item-11
item-12
item-13
item-14
item-15
item-16
item-17
item-18
item-19
item-20
item-21
item-22
item-23
item-24
item-25
RDD
Ex
RDD
W
RDD
Ex
RDD
W
RDD
Ex
RDD
W
more partitions = more parallelism

Error, ts, msg1
Warn, ts, msg2
Error, ts, msg1
RDD w/ 4 partitions
Info, ts, msg8
Warn, ts, msg2
Info, ts, msg8
Error, ts, msg3
Info, ts, msg5
Info, ts, msg5
Error, ts, msg4
Warn, ts, msg9
Error, ts, msg1
An RDD can be created 2 ways:
- Parallelize a collection
- Read data from an external source (S3, C*, HDFS, etc)
logLinesRDD

# Parallelize in Python
wordsRDD = sc.parallelize([“fish", “cats“, “dogs”])
// Parallelize in Scala
val wordsRDD= sc.parallelize(List("fish", "cats", "dogs"))
// Parallelize in Java
JavaRDD<String> wordsRDD = sc.parallelize(Arrays.asList(“fish", “cats“, “dogs”));
- Take an existing in-memory
collection and pass it to
SparkContext’s parallelize
method
- Not generally used outside of
prototyping and testing since it
requires entire dataset in
memory on one machine

# Read a local txt file in Python
linesRDD = sc.textFile("/path/to/README.md")
// Read a local txt file in Scala
val linesRDD = sc.textFile("/path/to/README.md")
// Read a local txt file in Java
JavaRDD<String> lines = sc.textFile("/path/to/README.md");
- There are other methods
to read data from HDFS,
C*, S3, HBase, etc.

Error, ts, msg1
Warn, ts, msg2
Error, ts, msg1
Info, ts, msg8
Warn, ts, msg2
Info, ts, msg8
Error, ts, msg3
Info, ts, msg5
Info, ts, msg5
Error, ts, msg4
Warn, ts, msg9
Error, ts, msg1
logLinesRDD
Error, ts, msg1
Error, ts, msg1
Error, ts, msg3 Error, ts, msg4
Error, ts, msg1
errorsRDD
.filter( )
(input/base RDD)

errorsRDD
.coalesce( 2 )
Error, ts, msg1
Error, ts, msg3
Error, ts, msg1
Error, ts, msg4
Error, ts, msg1
cleanedRDD
Error, ts, msg1
Error, ts, msg1
Error, ts, msg1
.collect( )
Driver

.collect( )
Execute DAG!
Driver

.collect( )
Driver
logLinesRDD

.collect( )
logLinesRDD
errorsRDD
cleanedRDD
.filter( )
.coalesce( 2 )
Driver
Error, ts, msg1
Error, ts, msg3
Error, ts, msg1
Error, ts, msg4
Error, ts, msg1

.collect( )
Driver
logLinesRDD
errorsRDD
cleanedRDD
data
.filter( )
.coalesce( 2, shuffle= False)
Pipelined
Stage-1

Driver
logLinesRDD
errorsRDD
cleanedRDD

logLinesRDD
errorsRDD
Error, ts, msg1
Error, ts, msg3
Error, ts, msg1
Error, ts, msg4
Error, ts, msg1
cleanedRDD
.filter( )
Error, ts, msg1
errorMsg1RDD
.collect( )
.saveToCassandra( )
.count( )
5

logLinesRDD
errorsRDD
Error, ts, msg1
Error, ts, msg3
Error, ts, msg1
Error, ts, msg4
Error, ts, msg1
cleanedRDD
.filter( )
Error, ts, msg1
errorMsg1RDD
.collect( )
.count( )
.saveToCassandra( )
5

P-1 logLinesRDD
(HadoopRDD)
P-2 P-3 P-4
P-1 errorsRDD
(filteredRDD)
P-2 P-3 P-4
Task-1
Task-2
Task-3
Task-4
Path = hdfs://. . .
func = _.contains(…)
shouldCache=false
logLinesRDD
errorsRDD
Dataset-level view: Partition-level view:

1) Create some input RDDs from external data or parallelize a
collection in your driver program.
2) Lazily transform them to define new RDDs using
transformations like filter() or map()
3) Ask Spark to cache() any intermediate RDDs that will need to
be reused.
4) Launch actions such as count() and collect() to kick off a
parallel computation, which is then optimized and executed
by Spark.

map() intersection() cartesion()
flatMap() distinct() pipe()
filter() groupByKey() coalesce()
mapPartitions() reduceByKey() repartition()
mapPartitionsWithIndex() sortByKey() partitionBy()
sample() join() ...
union() cogroup() ...
(lazy)
- Most transformations are element-wise (they work on one element at a time), but this is not
true for all transformations

reduce() takeOrdered()
collect() saveAsTextFile()
count() saveAsSequenceFile()
first() saveAsObjectFile()
take() countByKey()
takeSample() foreach()
saveToCassandra() ...

• HadoopRDD
• FilteredRDD
• MappedRDD
• PairRDD
• ShuffledRDD
• UnionRDD
• PythonRDD
• DoubleRDD
• JdbcRDD
• JsonRDD
• SchemaRDD
• VertexRDD
• EdgeRDD
• CassandraRDD (DataStax)
• GeoRDD (ESRI)
• EsSpark (ElasticSearch)

1) Set of partitions (“splits”)
2) List of dependencies on parent RDDs
3) Function to compute a partition given parents
4) Optional preferred locations
5) Optional partitioning info for k/v RDDs (Partitioner)
This captures all current Spark operations!
*
*
*
*
*

Partitions = one per HDFS block
Dependencies = none
Compute (partition) = read corresponding block
preferredLocations (part) = HDFS block location
Partitioner = none
*
*
*
*
*

Partitions = same as parent RDD
Dependencies = “one-to-one” on parent
Compute (partition) = compute parent and filter it
preferredLocations (part) = none (ask parent)
Partitioner = none
*
*
*
*
*

Partitions = One per reduce task
Dependencies = “shuffle” on each parent
Compute (partition) = read and join shuffled data
preferredLocations (part) = none
Partitioner = HashPartitioner(numTasks)
*
*
*
*
*

val cassandraRDD = sc
.cassandraTable(“ks”, “mytable”)
.select(“col-1”, “col-3”)
.where(“col-5 = ?”, “blue”)
Keyspace Table
{Server side column
& row selection

Start the Spark shell by passing in a custom cassandra.input.split.size:
ubuntu@ip-10-0-53-24:~$ dse spark –Dspark.cassandra.input.split.size=2000
Welcome to
____ __
/ __/__ ___ _____/ /__
_ / _ / _ `/ __/ '_/
/___/ .__/_,_/_/ /_/_ version 0.9.1
/_/
Using Scala version 2.10.3 (Java HotSpot(TM) 64-Bit Server VM, Java
1.7.0_51)
Type in expressions to have them evaluated.
Type :help for more information.
Creating SparkContext...
Created spark context..
Spark context available as sc.
Type in expressions to have them evaluated.
Type :help for more information.
scala>
The cassandra.input.split.size parameter defaults to 100,000. This is the approximate
number of physical rows in a single Spark partition. If you have really wide rows
(thousands of columns), you may need to lower this value. The higher the value, the
fewer Spark tasks are created. Increasing the value too much may limit the parallelism
level.”
(for dealing with wide rows)

https://github.com/datastax/spark-cassandra-connector
Spark Executor
Spark-C*
Connector
C* Java Driver
- Open Source
- Implemented mostly in Scala
- Scala + Java APIs
- Does automatic type conversions

https://github.com/datastax/spark-cassandra-connector

“Simple things
should be simple,
complex things
should be possible”
- Alan Kay

https://classwest03.cloud.databricks.com https://classwest10.cloud.databricks.com
- 60 user accounts - 60 user accounts
- 60 user clusters
- 1 community cluster
- 60 user clusters
https://classwest20.cloud.databricks.com https://classwest30.cloud.databricks.com
- 60 user accounts - 60 user accounts
- 60 user clusters
- 60 user clusters

Databricks Guide (5 mins)
DevOps 101 (30 mins)
DevOps 102 (30 mins)Transformations &
Actions (30 mins)
SQL 101 (30 mins)
Dataframes (20 mins)

0 10 20 30 40 50 60
Don't know
Mesos
Apache + Standalone
C* + Standalone
Hadoop YARN
Databricks Cloud
Survey completed by

0 10 20 30 40 50 60 70
Different Cloud
On-prem
Amazon Cloud
Survey completed by

- Local
- Standalone Scheduler
- YARN
- Mesos

JobTracker
DNTT
M
M
R M
M
R M
M
R M
M
R
M
M
M
M
M
M
RR R
R
OSOSOSOS
JT
DN DNTT DNTT TT
History:
NameNode
NN

JVM: Ex + Driver
Disk
RDD, P1 Task
3 options:
- local
- local[N]
- local[*]
RDD, P2
RDD, P1
RDD, P2
RDD, P3
Task
Task
Task
Task
Task
CPUs:
Task
Task
Task
Task
Task
Task
Internal
Threads
val conf = new SparkConf()
.setMaster("local[12]")
.setAppName(“MyFirstApp")
.set("spark.executor.memory", “3g")
val sc = new SparkContext(conf)
> ./bin/spark-shell --master local[12]
> ./bin/spark-submit --name "MyFirstApp"
--master local[12] myApp.jar
Worker Machine

Ex
RDD, P1
W
Driver
RDD, P2
RDD, P1
T T
T T
T T
Internal
Threads
SSD SSDOS Disk
SSD SSD
Ex
RDD, P4
W
RDD, P6
RDD, P1
T T
T T
T T
Internal
Threads
SSD SSDOS Disk
SSD SSD
Ex
RDD, P7
W
RDD, P8
RDD, P2
T T
T T
T T
Internal
Threads
SSD SSDOS Disk
SSD SSD
Spark
Master
Ex
RDD, P5
W
RDD, P3
RDD, P2
T T
T T
T T
Internal
Threads
SSD SSDOS Disk
SSD SSD
T T
T T
different spark-env.sh
- SPARK_WORKER_CORES
vs.> ./bin/spark-submit --name “SecondApp"
--master spark://host1:port1
myApp.jar - SPARK_LOCAL_DIRSspark-env.sh

Ex
RDD, P1
W
Driver
RDD, P2
RDD, P1
T T
T T
T T
Internal
Threads
SSD SSDOS Disk
SSD SSD
Ex
RDD, P4
W
RDD, P6
RDD, P1
T T
T T
T T
Internal
Threads
SSD SSDOS Disk
SSD SSD
Ex
RDD, P7
W
RDD, P8
RDD, P2
T T
T T
T T
Internal
Threads
SSD SSDOS Disk
SSD SSD
Spark
Master
Ex
RDD, P5
W
RDD, P3
RDD, P2
T T
T T
T T
Internal
Threads
SSD SSDOS Disk
SSD SSD
Spark
Master
T T
T T
different spark-env.sh
- SPARK_WORKER_CORES
vs.
I’m HA via
ZooKeeper
> ./bin/spark-submit --name “SecondApp"
--master spark://host1:port1,host2:port2
myApp.jar
Spark
Master
More
Masters
can be
added live
- SPARK_LOCAL_DIRSspark-env.sh

W
Driver
SSDOS Disk
W
SSDOS Disk
W
SSDOS Disk
Spark
Master
W
SSDOS Disk
(multiple apps)
Ex Ex Ex Ex
Driver
ExEx Ex Ex

Driver
SSDOS Disk SSDOS Disk SSDOS Disk
Spark
Master
W
SSDOS Disk
(single app)
Ex
W
Ex
W
Ex
W
Ex
W
Ex
W
Ex
W
Ex
W
Ex
SPARK_WORKER_INSTANCES: [default: 1] # of worker instances to run on each machine
SPARK_WORKER_CORES: [default: ALL] # of cores to allow Spark applications to use on the machine
SPARK_WORKER_MEMORY: [default: TOTAL RAM – 1 GB] Total memory to allow Spark applications to use on the machineconf/spark-env.sh
SPARK_DAEMON_MEMORY: [default: 512 MB] Memory to allocate to the Spark master and worker daemons themselves

Standalone settings
- Apps submitted will run in FIFO mode by default
spark.cores.max: maximum amount of CPU cores to request for the
application from across the cluster
spark.executor.memory: Memory for each executor

NodeManager
Resource
Manager
NodeManager
Container
NodeManager
App Master
Client #1 1
2 3 4 5
Container
6 6
7 7
8

NodeManager
Resource
Manager
NodeManager
Container
NodeManager
App Master
Client #1
ContainerApp Master
Container Container
Client #2
I’m HA via
ZooKeeper
Scheduler
Apps Master

NodeManager
Resource
Manager
NodeManager
Container
NodeManager
App Master
Client #1
Executor
RDD T
Container
Executor
RDD T
(client mode)
Driver

NodeManager
Resource
Manager
NodeManager
Container
NodeManager
App Master
Client #1
Executor
RDD T
Container
Executor
RDD T
Driver
(cluster mode)
Container
Executor
RDD T
- Does not support Spark Shells

YARN settings
--num-executors: controls how many executors will be allocated
--executor-memory: RAM for each executor
--executor-cores: CPU cores for each executor

YARN resource manager UI: http://<ip address>:8088
(No apps running)

[ec2-user@ip-10-0-72-36 ~]$ spark-submit --class
org.apache.spark.examples.SparkPi --deploy-mode client --master yarn
/opt/cloudera/parcels/CDH-5.2.1-1.cdh5.2.1.p0.12/jars/spark-
examples-1.1.0-cdh5.2.1-hadoop2.5.0-cdh5.2.1.jar 10

[ec2-user@ip-10-0-72-36 ~]$ spark-submit --class
org.apache.spark.examples.SparkPi --deploy-mode cluster --master
yarn /opt/cloudera/parcels/CDH-5.2.1-1.cdh5.2.1.p0.12/jars/spark-
examples-1.1.0-cdh5.2.1-hadoop2.5.0-cdh5.2.1.jar 10

Spark Central Master Who starts Executors? Tasks run in
Local [none] Human being Executor
Standalone Standalone Master Worker JVM Executor
YARN YARN App Master Node Manager Executor
Mesos Mesos Master Mesos Slave Executor

spark-submit provides a uniform interface for
submitting jobs across all cluster managers
bin/spark-submit --master spark://host:7077
--executor-memory 10g
my_script.py
Source: Learning Spark

PySpark at a Glance
Write Spark jobs
in Python
Run interactive
jobs in the shell
Supports C
extensions

Spark Core Engine
(Scala)
Standalone Scheduler YARN MesosLocal
Java API
PySpark
41 files
8,100 loc
6,300 comments

Spark
Context
Controller
Spark
Context
Py4j
Socket
Local Disk
Pipe
Driver JVM
Executor JVM
Executor JVM
Pipe
Worker MachineDriver Machine
F(x)
F(x)
F(x)
F(x)
F(x)
RDD
RDD
RDD
RDD
MLlib, SQL, shuffle
MLlib, SQL, shuffle
daemon.py
daemon.py

Data is stored as Pickled objects in an RDD[Array[Byte]]HadoopRDD
MappedRDD
PythonRDD RDD[Array[ ] ], , ,
(100 KB – 1MB each picked object)

pypy
• JIT, so faster
• less memory
• CFFI support
CPython
(default python)
Choose Your Python Implementation
Spark
Context
Driver Machine
Worker Machine
$ PYSPARK_DRIVER_PYTHON=pypy PYSPARK_PYTHON=pypy ./bin/pyspark
$ PYSPARK_DRIVER_PYTHON=pypy PYSPARK_PYTHON=pypy ./bin/spark-submit wordcount.py
OR

Job CPython 2.7 PyPy 2.3.1 Speed up
Word Count 41 s 15 s 2.7 x
Sort 46 s 44 s 1.05 x
Stats 174 s 3.6 s 48 x
The performance speed up will depend on work load (from 20% to 3000%).
Here are some benchmarks:
Here is the code used for benchmark:
rdd = sc.textFile("text")
def wordcount():
rdd.flatMap(lambda x:x.split('/'))
.map(lambda x:(x,1)).reduceByKey(lambda x,y:x+y).collectAsMap()
def sort():
rdd.sortBy(lambda x:x, 1).count()
def stats():
sc.parallelize(range(1024), 20).flatMap(lambda x: xrange(5024)).stats()
https://github.com/apache/spark/pull/2144

Task 3
Executor 1 Executor 2 Executor 3
Task 2
Task 1
Task 5
Task 4
Task 9
Task 8
Task 7
JOB 1
In Spark, each executor is long-running and runs many small tasks

Cointainer
4
Container
2
Container
5
Reduce Task 1 Map Task 2 Reduce Task 2
HADOOP MAPREDUCE
In MapReduce, each container is short-lived and runs one large task

Resource
(CPU / Mem)
Time
Allocated
Used
New job New job
Stragglers Stragglers

Resource
(CPU / Mem)
Time
Allocated
Used
More resources allocated than used…

Resource
(CPU / Mem)
Time
Allocated
Used
More resources allocated than used

Resource
(CPU / Mem)
Time
Allocated
Used
More efficient utilization of cluster resources

Each Spark application scales the number of
executors up and down based on workload
- If executors are idle, remove them
- If we need more executors, request them

spark.dynamicAllocation.enabled
spark.dynamicAllocation.minExecutors
spark.dynamicAllocation.maxExecutors
spark.dynamicAllocation.sustainedSchedulerBacklogTimeout (N)
spark.dynamicAllocation.schedulerBacklogTimeout (M)
spark.dynamicAllocation.executorIdleTimeout (K)
https://github.com/apache/spark/blob/master/core/src/main/scala/org/apache/spark/ExecutorAllocationManager.scala

LONG-RUNNING ETL JOBS
E.G. PARSING JSON INTO PARQUET IN S3
INTERACTIVE APPLICATIONS / SERVER
E.G. SPARK SHELL, OOYALA JOB SERVER
ANY APPLICATION WITH LARGE SHUFFLES

* this team
only
• 100 TB cluster with 1500+ nodes
• 15+ PB S3 warehouse (7 PB Parquet)
• Dynamic allocation with up to 10,000 executors
• Enabled dynamic allocation for all Spark applications*
• Run Spark alongside Hive, Pig, MapReduce
• Used for ad-hoc query and experimentation
* this team only

* this team
only
• 400+ TB cluster with 8,000+ nodes
• 150 PB data warehouse
• Use dynamic allocation with up to 1,500 executors
• Primary use cases are ETL and SQL
• Run Spark alongside Storm, MapReduce, Pig

• Support for Mesos mode
• Support for Standalone mode
• Better support for caching
• Pluggable scaling heuristics
SPARK-4922 (PR ready)
SPARK-4751 (PR soon)
SPARK-7955 (PR ready)

Ex
RDD, P1
RDD, P2
RDD, P1
T T
T T
T T
Internal
Threads
Recommended to use at most only 75% of a machine’s memory
for Spark
Minimum Executor heap size should be 8 GB
Max Executor heap size depends… maybe 40 GB (watch GC)
Memory usage is greatly affected by storage level and
serialization format

RDD.cache() == RDD.persist(MEMORY_ONLY)
JVM
deserialized
most CPU-efficient option

RDD.persist(MEMORY_ONLY_SER)
JVM
serialized

+
.persist(MEMORY_AND_DISK)
JVM
deserialized Ex
W
RDD-P1
T
T
OS Disk
SSD
RDD-P1
RDD-P2

+
.persist(MEMORY_AND_DISK_SER)
JVM
serialized

RDD.persist(MEMORY_ONLY_2)
JVM on Node X
deserialized deserialized
JVM on Node Y

+
.persist(MEMORY_AND_DISK_2)
JVM
deserialized
+
JVM
deserialized

.persist(OFF_HEAP)
JVM-1 / App-1
serialized
Tachyon
JVM-2 / App-1
JVM-7 / App-2

Persistence description
MEMORY_ONLY Store RDD as deserialized Java objects in
the JVM
MEMORY_AND_DISK Store RDD as deserialized Java objects in
the JVM and spill to disk
MEMORY_ONLY_SER Store RDD as serialized Java objects (one
byte array per partition)
MEMORY_AND_DISK_SER 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 Same as the levels above, but replicate
each partition on two cluster nodes
OFF_HEAP Store RDD in serialized format in Tachyon

Bringing Spark Closer to Bare Metal
Project Tungsten will be the largest change to Spark’s execution engine since the project’s inception.
TLDR:
Problem #1:
- “abcd” takes 4 bytes to store in UTF-8, but JVM takes 48 bytes to store it
Problem #2:
- JVM cannot be as smart as Spark to manage GC b/c Spark knows more
about life cycle of memory blocks
Solution:
- Manage memory with Spark using JVM internal API sun.misc.Unsafe, to directly
manipulate memory without safety checks (hence “unsafe”)

Bringing Spark Closer to Bare Metal
Project Tungsten will be the largest change to Spark’s execution engine since the project’s inception.
TLDR:
Problem:
- A large fraction of CPU time is spent waiting for data to be fetched from
main memory
Solution:
- Use cache-aware computation to improve speed of data processing via
L1/L2/L3 CPU caches

?
- If RDD fits in memory, choose MEMORY_ONLY
- If not, use MEMORY_ONLY_SER w/ fast serialization library
- Don’t spill to disk unless functions that computed the datasets
are very expensive or they filter a large amount of data.
(recomputing may be as fast as reading from disk)
- Use replicated storage levels sparingly and only if you want fast
fault recovery (maybe to serve requests from a web app)

Intermediate data is automatically persisted during shuffle operations
Remember!

PySpark: stored objects will always be serialized with Pickle library, so it does
not matter whether you choose a serialized level.
=

60%20%
20%
Default Memory Allocation in Executor JVM
Cached RDDs
User Programs
(remainder)
Shuffle memory
spark.storage.memoryFraction
spark.shuffle.memoryFraction

RDD Storage: when you call .persist() or .cache(). Spark will limit the amount of
memory used when caching to a certain fraction of the JVM’s overall heap, set by
spark.storage.memoryFraction
Shuffle and aggregation buffers: When performing shuffle operations, Spark will
create intermediate buffers for storing shuffle output data. These buffers are used to
store intermediate results of aggregations in addition to buffering data that is going
to be directly output as part of the shuffle.
User code: Spark executes arbitrary user code, so user functions can themselves
require substantial memory. For instance, if a user application allocates large arrays
or other objects, these will content for overall memory usage. User code has access
to everything “left” in the JVM heap after the space for RDD storage and shuffle
storage are allocated.
Spark uses memory for:

1. Create an RDD
2. Put it into cache
3. Look at SparkContext logs
on the driver program or
Spark UI
INFO BlockManagerMasterActor: Added rdd_0_1 in memory on mbk.local:50311 (size: 717.5 KB, free: 332.3 MB)
logs will tell you how much memory each
partition is consuming, which you can
aggregate to get the total size of the RDD

Serialization is used when:
Transferring data over the network
Spilling data to disk
Caching to memory serialized
Broadcasting variables

Java serialization Kryo serializationvs.
• Uses Java’s ObjectOutputStream framework
• Works with any class you create that implements
java.io.Serializable
• You can control the performance of serialization
more closely by extending java.io.Externalizable
• Flexible, but quite slow
• Leads to large serialized formats for many classes
• Recommended serialization for production apps
• Use Kyro version 2 for speedy serialization (10x) and
more compactness
• Does not support all Serializable types
• Requires you to register the classes you’ll use in
advance
• If set, will be used for serializing shuffle data between
nodes and also serializing RDDs to disk
conf.set(“spark.serializer”, "org.apache.spark.serializer.KryoSerializer")

To register your own custom classes with Kryo, use the
registerKryoClasses method:
val conf = new SparkConf().setMaster(...).setAppName(...)
conf.registerKryoClasses(Seq(classOf[MyClass1], classOf[MyClass2]))
val sc = new SparkContext(conf)
- If your objects are large, you may need to increase
spark.kryoserializer.buffer.mb config property
- The default is 2, but this value needs to be large enough to
hold the largest object you will serialize.

. . . Ex
RDD
W
RDD
Ex
RDD
W
RDD
High churn Low churn

. . . Ex
RDD
W
RDD
RDD
High churn
Cost of GC is proportional to the # of
Java objects
(so use an array of Ints instead of a
LinkedList)
-verbose:gc -XX:+PrintGCDetails -XX:+PrintGCTimeStamps
To measure GC impact:

Parallel Old GC CMS GC G1 GC
-XX:+UseParallelOldGC -XX:+UseConcMarkSweepGC -XX:+UseG1GC
- Uses multiple threads
to do both young gen
and old gen GC
- Also a multithreading
compacting collector
- HotSpot does
compaction only in
old gen
Parallel GC
-XX:+UseParallelGC
- Uses multiple threads to
do young gen GC
- Will default to Serial on
single core machines
- Aka “throughput
collector”
- Good for when a lot of
work is needed and
long pauses are
acceptable
- Use cases: batch
processing
-XX:ParallelGCThreads=<#> -XX:ParallelCMSThreads=<#>
- Concurrent Mark
Sweep aka
“Concurrent low
pause collector”
- Tries to minimize
pauses due to GC by
doing most of the work
concurrently with
application threads
- Uses same algorithm
on young gen as
parallel collector
- Use cases:
- Garbage First is available
starting Java 7
- Designed to be long term
replacement for CMS
- Is a parallel, concurrent
and incrementally
compacting low-pause
GC

The key to tuning spark apps is a sound grasp of Spark’s internal mechanisms

?
How does a user program get translated into units of physical execution?
application -> jobs, stages, tasks

// Read input file
val input = sc.textFile("input.txt")
val tokenized = input
.map(line => line.split(" "))
.filter(words => words.size > 0) // remove empty lines
val counts = tokenized // frequency of log levels
.map(words => (words(0), 1))
.reduceByKey{ (a, b) => a + b, 2 }
INFO Server started
INFO Bound to port 8080
WARN Cannot find srv.conf
input.txt

// Read input file
val input = sc.textFile( )
val tokenized = input
.map( )
.filter( ) // remove empty lines
val counts = tokenized // frequency of log levels
.map( )
.reduceByKey{ }

sc.textFile().map().filter().map().reduceByKey()

Mapped RDD
Partition 1
Partition 2
Partition 3
Partition 1
Partition 2
Partition 3
Mapped RDD
Partition 1
Partition 2
Partition 3
Partition 1
Partition 2
Hadoop RDD
Partition 1
Partition 2
Partition 3
input tokenized counts
textFile() map() filter() map() reduceByKey()
Filtered RDD Shuffle RDD

DAG’s are materialized through a method sc.runJob:
def runJob[T, U](
rdd: RDD[T], 1. RDD to compute
partitions: Seq[Int], 2. Which partitions
func: (Iterator[T]) => U)) 3. Fn to produce results
: Array[U]  results for each part.

Mapped RDD
Partition 1
Partition 2
Partition 3
Partition 1
Partition 2
Partition 3
Mapped RDD
Partition 1
Partition 2
Partition 3
Partition 1
Partition 2
Hadoop RDD
Partition 1
Partition 2
Partition 3
input tokenized counts
Filtered RDD Shuffle RDD
Needs to compute my parents, parents, parents, etc all the way back to
an RDD with no dependencies (e.g. HadoopRDD).
runJob(counts)

Partition 1
Partition 2
Partition 3
Stage 1 Stage 2
Partition 1
Partition 2
Shuffle write
Shuffle read
Input read
Each task will:
1) Read Hadoop input
2) Perform maps & filters
3) Write partial sums
Each task will:
1) Read partial sums
2) Invoke user function
passed to runJob

scala> counts.toDebugString
res84: String =
(2) ShuffledRDD[296] at reduceByKey at <console>:17
+-(3) MappedRDD[295] at map at <console>:17
| FilteredRDD[294] at filter at <console>:15
| MappedRDD[293] at map at <console>:15
| input.text MappedRDD[292] at textFile at <console>:13
| input.text HadoopRDD[291] at textFile at <console>:13
(indentations indicate a shuffle boundary)

Stage 1
Stage 2
Stage 3
Stage 5
.
.
Job #1
.collect( )
Task #1
Task #2
Task #3
.
.
Stage 4
Job #2

Named after action calling runJob
Named after last RDD in pipeline

Task
Scheduler
Task threads
Block manager
RDD Objects DAG Scheduler Task Scheduler Executor
Rdd1.join(rdd2)
.groupBy(…)
.filter(…)
- Build operator DAG
- Split graph into
stages of tasks
- Submit each stage as
ready
- Execute tasks
- Store and serve
blocks
DAG TaskSet Task
Agnostic to
operators
Doesn’t know
about stagesStage
failed
- Launches
individual tasks
- Retry failed or
straggling tasks

1.4.0 Event timeline all jobs page

1.4.0
Event timeline within 1 job

1.4.0
Event timeline within 1 stage

1.4.0
sc.textFile(“blog.txt”)
.cache()
.flatMap { line => line.split(“ “) }
.map { word => (word, 1) }
.reduceByKey { case (count1, count2) => count1 + count2 }
.collect()

1.4.0
DAG Visualization for ALS

1.4.0
DAG Visualization for ALS (stage page)

1.4.0
DAG Visualization for SQL broadcast join

= cached partition
= RDD
join
filter
groupBy
Stage 3
Stage 1
Stage 2
A: B:
C: D: E:
F:
map
= lost partition

“This distinction is useful for two reasons:
1) Narrow dependencies allow for pipelined execution on one cluster node,
which can compute all the parent partitions. For example, one can apply a
map followed by a filter on an element-by-element basis.
In contrast, wide dependencies require data from all parent partitions to be
available and to be shuffled across the nodes using a MapReduce-like
operation.
2) Recovery after a node failure is more efficient with a narrow dependency, as
only the lost parent partitions need to be recomputed, and they can be
recomputed in parallel on different nodes. In contrast, in a lineage graph with
wide dependencies, a single failed node might cause the loss of some partition
from all the ancestors of an RDD, requiring a complete re-execution.”
Dependencies: Narrow vs Wide

scala> input.toDebugString
res85: String =
(2) data.text MappedRDD[292] at textFile at <console>:13
| data.text HadoopRDD[291] at textFile at <console>:13
scala> counts.toDebugString
res84: String =
(2) ShuffledRDD[296] at reduceByKey at <console>:17
+-(2) MappedRDD[295] at map at <console>:17
| FilteredRDD[294] at filter at <console>:15
| MappedRDD[293] at map at <console>:15
| data.text MappedRDD[292] at textFile at <console>:13
| data.text HadoopRDD[291] at textFile at <console>:13
To display the lineage of an RDD, Spark provides a toDebugString method:

How do you know if a shuffle will be called on a Transformation?
Note that repartition just calls coalese w/ True:
- repartition , join, cogroup, and any of the *By or *ByKey transformations
can result in shuffles
- If you declare a numPartitions parameter, it’ll probably shuffle
- If a transformation constructs a shuffledRDD, it’ll probably shuffle
- combineByKey calls a shuffle (so do other transformations like
groupByKey, which actually end up calling combineByKey)
def repartition(numPartitions: Int)(implicit
ord: Ordering[T] = null): RDD[T] = {
coalesce(numPartitions, shuffle = true)
}
RDD.scala

How do you know if a shuffle will be called on a Transformation?
Transformations that use “numPartitions” like distinct will probably shuffle:
def distinct(numPartitions: Int)(implicit ord: Ordering[T] =
null): RDD[T] =
map(x => (x, null)).reduceByKey((x, y) => x,
numPartitions).map(_._1)

- An extra parameter you can pass a k/v transformation to let Spark know
that you will not be messing with the keys at all
- All operations that shuffle data over network will benefit from partitioning
- Operations that benefit from partitioning:
cogroup, groupWith, join, leftOuterJoin, rightOuterJoin, groupByKey,
reduceByKey, combineByKey, lookup, . . .
https://github.com/apache/spark/blob/master/core/src/main/scala/org/apache/spark/rdd/RDD.scala#L302

Driver
Ex
RDD, P1
RDD, P2
RDD, P1
T T
T T
T T
Internal
Threads
SparkContext
DAG Scheduler Task Scheduler SparkEnv
- cacheManager
- blockManager
- shuffleManager
- securityManager
- broadcastManager
- mapOutputTracker
CoarseGrainedExecutorBackend
SparkEnv
- cacheManager
- blockManager
- shuffleManager
- securityManager
- broadcastManager
- mapOutputTracker

sc.textFile("someFile.txt").
map(mapFunc).
flatMap(flatMapFunc).
filter(filterFunc).
count()
How many Stages will this code require?
Source: Cloudera

Source: Cloudera
How many Stages will this DAG require?

Ex
Ex
Ex
x = 5
T
T
x = 5
x = 5
x = 5
x = 5
T
T

• Broadcast variables – Send a large read-only lookup table to all the nodes, or
send a large feature vector in a ML algorithm to all nodes
• Accumulators – count events that occur during job execution for debugging
purposes. Example: How many lines of the input file were blank? Or how many
corrupt records were in the input dataset?
++ +

Spark supports 2 types of shared variables:
• Broadcast variables – allows your program to efficiently send a large, read-only
value to all the worker nodes for use in one or more Spark operations. Like
sending a large, read-only lookup table to all the nodes.
• Accumulators – allows you to aggregate values from worker nodes back to
the driver program. Can be used to count the # of errors seen in an RDD of
lines spread across 100s of nodes. Only the driver can access the value of an
accumulator, tasks cannot. For tasks, accumulators are write-only.
++ +

Broadcast variables let programmer keep a read-
only variable cached on each machine rather than
shipping a copy of it with tasks
For example, to give every node a copy of a large
input dataset efficiently
Spark also attempts to distribute broadcast variables
using efficient broadcast algorithms to reduce
communication cost

val broadcastVar = sc.broadcast(Array(1, 2, 3))
broadcastVar.value
broadcastVar = sc.broadcast(list(range(1, 4)))
broadcastVar.value
Scala:
Python:

Ex
Ex
Ex
History:
20 MB file
Uses HTTP

Ex
Ex
Ex
20 MB file
Uses bittorrent
. . .4 MB 4 MB 4 MB 4 MB

Accumulators are variables that can only be “added” to through
an associative operation
Used to implement counters and sums, efficiently in parallel
Spark natively supports accumulators of numeric value types and
standard mutable collections, and programmers can extend
for new types
Only the driver program can read an accumulator’s value, not the
tasks
++ +

val accum = sc.accumulator(0)
sc.parallelize(Array(1, 2, 3, 4)).foreach(x => accum += x)
accum.value
accum = sc.accumulator(0)
rdd = sc.parallelize([1, 2, 3, 4])
def f(x):
global accum
accum += x
rdd.foreach(f)
accum.value
Scala:
Python:
++ +

Spark sorted the same data 3X faster
using 10X fewer machines
than Hadoop MR in 2013.
Work by Databricks engineers: Reynold Xin, Parviz Deyhim, Xiangrui Meng, Ali Ghodsi, Matei Zaharia
100TB Daytona Sort Competition 2014
More info:
http://sortbenchmark.org
http://databricks.com/blog/2014/11/05/spark-
officially-sets-a-new-record-in-large-scale-sorting.html
All the sorting took place on disk (HDFS) without
using Spark’s in-memory cache!

- Stresses “shuffle” which underpins everything from SQL to Mllib
- Sorting is challenging b/c there is no reduction in data
- Sort 100 TB = 500 TB disk I/O and 200 TB network
Engineering Investment in Spark:
- Sort-based shuffle (SPARK-2045)
- Netty native network transport (SPARK-2468)
- External shuffle service (SPARK-3796)
Clever Application level Techniques:
- GC and cache friendly memory layout
- Pipelining

Ex
RDD
W
RDD
T
T
EC2: i2.8xlarge
(206 workers)
- Intel Xeon CPU E5 2670 @ 2.5 GHz w/ 32 cores
- 244 GB of RAM
- 8 x 800 GB SSD and RAID 0 setup formatted with /ext4
- ~9.5 Gbps (1.1 GBps) bandwidth between 2 random nodes
- Each record: 100 bytes (10 byte key & 90 byte value)
- OpenJDK 1.7
- HDFS 2.4.1 w/ short circuit local reads enabled
- Apache Spark 1.2.0
- Speculative Execution off
- Increased Locality Wait to infinite
- Compression turned off for input, output & network
- Used Unsafe to put all the data off-heap and managed
it manually (i.e. never triggered the GC)
- 32 slots per machine
- 6,592 slots total

=
groupByKey
sortByKey
reduceByKey

spark.shuffle.spill=false
(Affects reducer side and keeps all the data in memory)

- Must turn this on for dynamic allocation in YARN
- Worker JVM serves files
- Node Manager serves files

- Was slow because it had to copy the data 3 times
Map output file
on local dir
Linux
kernel
buffer
Ex
NIC
buffer

- Uses a technique called zero-copy
- Is a map-side optimization to serve data very
quickly to requesting reducers
Map output file
on local dir
NIC
buffer

Map() Map() Map() Map()
Reduce() Reduce() Reduce()
- Entirely bounded
by I/O reading from
HDFS and writing out
locally sorted files
- Mostly network bound
< 10,000 reducers
- Notice that map
has to keep 3 file
handles open
TimSort
= 5 blocks

(28,000 unique blocks)
RF = 2
250,000+ reducers!
- Only one file handle
open at a time
= 3.6 GB

- 5 waves of maps
- 5 waves of reduces
Reduce() Reduce() Reduce()
RF = 2
250,000+ reducers!
MergeSort!
TimSort
(28,000 unique blocks)
RF = 2

Sustaining 1.1GB/s/node during shuffle
- Actual final run
- Fully saturated
the 10 Gbit link

UserID Name Age Location Pet
28492942 John Galt 32 New York Sea Horse
95829324 Winston Smith 41 Oceania Ant
92871761 Tom Sawyer 17 Mississippi Raccoon
37584932 Carlos Hinojosa 33 Orlando Cat
73648274 Luis Rodriguez 34 Orlando Dogs

https://databricks.com/blog/2015/02/17/introducing-dataframes-in-spark-for-large-scale-data-science.html

SchemaRDD
- RDD of Row objects, each representing a record
- Row objects = type + col. name of each
- Stores data very efficiently by taking advantage of the schema
- SchemaRDDs are also regular RDDs, so you can run
transformations like map() or filter()
- Allows new operations, like running SQL on objects

• Announced Feb 2015
• Inspired by data frames in R
and Pandas in Python
• Works in:
Features
• Scales from KBs to PBs
• Supports wide array of data formats and
storage systems (Hive, existing RDDs, etc)
• State-of-the-art optimization and code
generation via Spark SQL Catalyst optimizer
• APIs in Python, Java
• a distributed collection of data organized into
named columns
• Like a table in a relational database
What is a Dataframe?

Step 1: Construct a DataFrame
from pyspark.sql import SQLContext
sqlContext = SQLContext(sc)
df = sqlContext.jsonFile("examples/src/main/resources/people.json")
# Displays the content of the DataFrame to stdout
df.show()
## age name
## null Michael
## 30 Andy
## 19 Justin

Step 2: Use the DataFrame
# Print the schema in a tree format
df.printSchema()
## root
## |-- age: long (nullable = true)
## |-- name: string (nullable = true)
# Select only the "name" column
df.select("name").show()
## name
## Michael
## Andy
## Justin
# Select everybody, but increment the age by 1
df.select("name", df.age + 1).show()
## name (age + 1)
## Michael null
## Andy 31
## Justin 20

SQL Integration
from pyspark.sql import SQLContext
df = sqlContext.sql("SELECT * FROM table")

SQL + RDD Integration
2 methods for converting existing RDDs into DataFrames:
1. Use reflection to infer the schema of an RDD that
contains different types of objects
2. Use a programmatic interface that allows you to
construct a schema and then apply it to an existing
RDD.
(more concise)
(more verbose)

SQL + RDD Integration: via reflection
# sc is an existing SparkContext.
from pyspark.sql import SQLContext, Row
# Load a text file and convert each line to a Row.
lines = sc.textFile("examples/src/main/resources/people.txt")
parts = lines.map(lambda l: l.split(","))
people = parts.map(lambda p: Row(name=p[0], age=int(p[1])))
# Infer the schema, and register the DataFrame as a table.
schemaPeople = sqlContext.inferSchema(people)
schemaPeople.registerTempTable("people")

SQL + RDD Integration: via reflection
# SQL can be run over DataFrames that have been registered as a table.
teenagers = sqlContext.sql("SELECT name FROM people WHERE age >= 13 AND age <= 19")
# The results of SQL queries are RDDs and support all the normal RDD operations.
teenNames = teenagers.map(lambda p: "Name: " + p.name)
for teenName in teenNames.collect():
print teenName

SQL + RDD Integration: via programmatic schema
DataFrame can be created programmatically with 3 steps:
1. Create an RDD of tuples or lists from the original RDD
2. Create the schema represented by a StructType matching the
structure of tuples or lists in the RDD created in the step 1
3. Apply the schema to the RDD via createDataFrame method
provided by SQLContext

Step 1: Construct a DataFrame
# Constructs a DataFrame from the users table in Hive.
users = context.table("users")
# from JSON files in S3
logs = context.load("s3n://path/to/data.json", "json")

Step 2: Use the DataFrame
# Create a new DataFrame that contains “young users” only
young = users.filter(users.age < 21)
# Alternatively, using Pandas-like syntax
young = users[users.age < 21]
# Increment everybody’s age by 1
young.select(young.name, young.age + 1)
# Count the number of young users by gender
young.groupBy("gender").count()
# Join young users with another DataFrame called logs
young.join(logs, logs.userId == users.userId, "left_outer")

TwitterUtils.createStream(...)
.filter(_.getText.contains("Spark"))
.countByWindow(Seconds(5))

Kafka
Flume
HDFS
S3
Kinesis
Twitter
TCP socket
HDFS / S3
Cassandra
Dashboards
Databases
- Scalable
- High-throughput
- Fault-tolerant
Complex algorithms can be expressed using:
- Spark transformations: map(), reduce(), join(), etc
- MLlib + GraphX
- SQL
HBase

Batch Realtime
One unified API

Tathagata Das (TD)
- Lead developer of Spark Streaming + Committer
on Apache Spark core
- Helped re-write Spark Core internals in 2012 to
make it 10x faster to support Streaming use cases
- On leave from UC Berkeley PhD program
- Ex: Intern @ Amazon, Intern @ Conviva, Research
Assistant @ Microsoft Research India
- 1 guy; does not scale
- Scales to 100s of nodes
- Batch sizes as small as half a second
- End to end Processing latency as low as 1 second
- Exactly-once semantics no matter what fails

Page views Kafka for buffering Spark for processing
(live statistics)

Smart meter readings
Live weather data
Join 2 live data
sources
(Anomaly Detection)

Input data stream
Batches of
processed data
Batches every X seconds

Input data streams Batches of
processed data
Batches every X seconds
R
R
R

(Discretized Stream)
Block #1
RDD @ T=5
Block #2 Block #3
Batch interval = 5 seconds
Block #1
RDD @ T=10
Block #2 Block #3
T = 5 T = 10
Input
DStream
One RDD is created every 5 seconds

DStreams can be created from:
1) External input sources
2) Applying transformations to other DStreams
normal
(stateless)
stateful

Block #1 Block #2 Block #3
Part. #1 Part. #2 Part. #3
Part. #1 Part. #2 Part. #3
5 sec
flatMap()
linesDStream
wordsDStream
10 sec 15 sec

// Create a StreamingContext with a 1-second batch size from a SparkConf
val ssc = new StreamingContext(conf, Seconds(1))
// Create a DStream using data received after connecting to port 7777 on the local machine
val linesStream = ssc.socketTextStream("localhost", 7777)
// Filter our DStream for lines with "error"
val errorLinesStream = linesStream.filter(_.contains("error"))
// Print out the lines with errors
errorLinesStream.print()
// Start our streaming context and wait for it to "finish"
ssc.start()
// Wait for the job to finish
ssc.awaitTermination()
linesStream
errorLinesStream
import org.apache.spark.streaming.StreamingContext
import org.apache.spark.streaming.StreamingContext._
import org.apache.spark.streaming.dstream.DStream
import org.apache.spark.streaming.Duration

Terminal #1 Terminal #2
$ nc localhost 7777
all is good
there was an error
good good
error 4 happened
all good now
$ spark-submit --class com.examples.Scala.StreamingLogInput
$ASSEMBLY_JAR local[4]
. . .
--------------------------
Time: 2015-05-26 15:25:21
--------------------------
there was an error
--------------------------
Time: 2015-05-26 15:25:22
--------------------------
error 4 happened

Remember!
• Once a StreamingContext has been started, no new streaming
computations can be added to it
• Once a StreamingContext has been stopped, it cannot be restarted
• Only one StreamingContext can be active in a JVM at a time
• Stop() on StreamingContext also stops the SparkContext. (You can stop
only the StreamingContext by setting the optional parameter
stopSparkContext to false)
• A SparkContext can be reused to create multiple StreamingContexts, as
long as the previous StreamingContext is stopped (without stopping the
SparkContext)

Ex
RDD, P1
W
Driver
RDD, P2
block, P1
T
Internal
Threads
SSD SSDOS Disk
T T
T T Ex
RDD, P3
W
RDD, P4
block, P1
T
Internal
Threads
SSD SSDOS Disk
T T
T T
T
Batch interval = 600 ms
R
localhost 9999

Ex
RDD, P1
W
Driver
RDD, P2
block, P1
T R
Internal
Threads
SSD SSDOS Disk
T T
T T Ex
RDD, P3
W
RDD, P4
block, P1
T
Internal
Threads
SSD SSDOS Disk
T T
T T
T
200 ms later
Ex
W
block, P2 T
Internal
Threads
SSD SSDOS Disk
T T
T T
T
block, P2

Ex
RDD, P1
W
Driver
RDD, P2
block, P1
T R
Internal
Threads
SSD SSDOS Disk
T T
T T Ex
RDD, P1
W
RDD, P2
block, P1
T
Internal
Threads
SSD SSDOS Disk
T T
T T
T
200 ms later
Ex
W
block, P2
T
Internal
Threads
SSD SSDOS Disk
T T
T T
T
block, P2
block, P3
block, P3

Ex
RDD, P1
W
Driver
RDD, P2
RDD, P1
T R
Internal
Threads
SSD SSDOS Disk
T T
T T Ex
RDD, P1
W
RDD, P2
RDD, P1
T
Internal
Threads
SSD SSDOS Disk
T T
T T
T
Ex
W
RDD, P2
T
Internal
Threads
SSD SSDOS Disk
T T
T T
T
RDD, P2
RDD, P3
RDD, P3

1.4.0
DAG Visualization for Streaming

Ex
W
Driver
block, P1
T R
Internal
Threads
SSD SSDOS Disk
T T
T T Ex
W
block, P1
T
Internal
Threads
SSD SSDOS Disk
T T
T T
T
Ex
W
block, P1
T
Internal
Threads
SSD SSDOS Disk
T T
T T
R
block, P1
2 input DStreams

Ex
W
Driver
block, P1
T R
Internal
Threads
SSD SSDOS Disk
T T
T T Ex
W
block, P1
T
Internal
Threads
SSD SSDOS Disk
T T
T T
T
Ex
W
block, P1 T
Internal
Threads
SSD SSDOS Disk
T T
T T
R
block, P1
block, P2
block, P3
block, P2
block, P3
block, P2
block, P3
block, P2 block, P3

Ex
W
Driver
RDD, P1
T R
Internal
Threads
SSD SSDOS Disk
T T
T T Ex
W
RDD, P1
T
Internal
Threads
SSD SSDOS Disk
T T
T T
T
Ex
W
RDD, P1 T
Internal
Threads
SSD SSDOS Disk
T T
T T
R
RDD, P1
RDD, P2
RDD, P3
RDD, P2
RDD, P3
RDD, P2
RDD, P3
RDD, P2 RDD, P3
Materialize!

Ex
W
Driver
RDD, P3
T R
Internal
Threads
SSD SSDOS Disk
T T
T T Ex
W
RDD, P4
T
Internal
Threads
SSD SSDOS Disk
T T
T T
T
Ex
W
RDD, P3 T
Internal
Threads
SSD SSDOS Disk
T T
T T
R
RDD, P6
RDD, P4
RDD, P5
RDD, P2
RDD, P2
RDD, P5
RDD, P1
RDD, P1 RDD, P6
Union!

Stream–stream Unions
val numStreams = 5
val kafkaStreams = (1 to numStreams).map { i => KafkaUtils.createStream(...) }
val unifiedStream = streamingContext.union(kafkaStreams)
unifiedStream.print()

Stream–stream Joins
val stream1: DStream[String, String] = ...
val stream2: DStream[String, String] = ...
val joinedStream = stream1.join(stream2)

- File systems
- Socket Connections
- Kafka
- Flume
- Twitter
Sources directly available
in StreamingContext API
Requires linking against
extra dependencies
- Anywhere
Requires implementing
user-defined receiver
val logData = ssc.textFileStream(logDirectory)

Ex
R
Ex
R
Sink
Flume push
to port
Flume push
to custom
sink
Pull from sink

map( )
flatMap( )
filter( )
repartition(numPartitions)
union(otherStream)
count()
reduce( )
countByValue()
reduceAByKey( ,[numTasks])
join(otherStream,[numTasks])
cogroup(otherStream,[numTasks])
transform( )
RDD
RDD
updateStateByKey( )*

updateStateByKey( )
To use:
1) Define the state
(an arbitrary data type)
2) Define the state update function
(specify with a function how to update the state using the
previous state and new values from the input stream)
: allows you to maintain arbitrary state while
continuously updating it with new information
def updateFunction(newValues, runningCount):
if runningCount is None:
runningCount = 0
return sum(newValues, runningCount) # add the
# new values with the previous running count
# to get the new count
To maintain a running count of each word seen
in a text data stream (here running count is an
integer type of state):
runningCounts = pairs.updateStateByKey(updateFunction)
pairs = (word, 1)
(cat, 1)
*
* Requires a checkpoint directory to be
configured

For example:
- Functionality to join every batch in a
data stream with another dataset is not
directly exposed in the DStream API.
- If you want to do real-time data
cleaning by joining the input data
stream with pre-computed spam
information and then filtering based on it.
: can be used to apply any RDD operation that
is not exposed in the DStream API.
spamInfoRDD = sc.pickleFile(...) # RDD containing spam information
# join data stream with spam information to do data cleaning
cleanedDStream = wordCounts.transform(lambda rdd:
rdd.join(spamInfoRDD).filter(...))
transform( )
RDD
RDD
or
MLlib GraphX

Original
DStream
RDD 4 RDD 5 RDD 6
Windowed
DStream
RDD 1 RDD 2 RDD 3
RDD X
time 1 time 2 time 3 time 4 time 5 time 6
RDD Y
RDD @ 3
RDD @ 5
Window Length: 3 time units
Sliding Interval: 2 time units
*
*
*Both of these must be multiples of the
batch interval of the source DSTream

window(windowLength, slideInterval)
countByWindow(windowLength, slideInterval)
reduceByWindow( , windowLength, slideInterval)
reduceByKeyAndWindow( , windowLength, slideInterval,[numTasks])
reduceByKeyAndWindow( , , windowLength, slideInterval,[numTasks])
countByValueAndWindow(windowLength, slideInterval, [numTasks])
- DStream
- PairDStreamFunctions
- JavaDStream
- JavaPairDStream
- DStream
API Docs

print()
saveAsTextFile(prefix, [suffix])
foreachRDD( )
saveAsHadoopFiles(prefix, [suffix])

Dev Ops Training

More Related Content

What's hot (20)

Similar to Dev Ops Training (20)

More from Spark Summit (20)

Recently uploaded (20)

Dev Ops Training