Intro to Hadoop Presentation at Carnegie Mellon - Silicon Valley

© 2010 – 2015 Cloudera, Inc. All Rights Reserved

Introduc=on to Apache Hadoop and its Ecosystem Mark Grover | Intro to Cloud Compu=ng, Carnegie Mellon SV github.com/markgrover/hadoop-‐intro-‐fast

© Copyright 2010-‐2014 Cloudera, Inc. All rights reserved.


About Me •  CommiNer on Apache Bigtop, commiNer and PPMC member on Apache Sentry (incuba=ng). •  Contributor to Apache Hadoop, Hive, Spark, Sqoop, Flume. •  SoUware developer at Cloudera • @mark_grover • www.linkedin.com/in/grovermark


Co-‐author O’Reilly book

• @hadooparchbook •  hadooparchitecturebook.com •  To be released early 2015


About the Presenta=on…

• What’s ahead •  Fundamental Concepts •  HDFS: The Hadoop Distributed File System •  Data Processing with MapReduce •  Demo •  Conclusion + Q&A


Fundamental Concepts Why the World Needs Hadoop


What’s the craze about Hadoop?

•  Volume • More and more data being generated • Machine generated data increasing

•  Velocity •  Data coming it at higher speed

•  Variety •  Audio, video, images, log files, web pages, social network connec=ons, etc.


We Need a System that Scales

•  Too much data for tradi=onal tools •  Two key problems

•  How to reliably store this data at a reasonable cost •  How to we process all the data we’ve stored


What is Apache Hadoop?

•  Scalable data storage and processing •  Distributed and fault-‐tolerant •  Runs on standard hardware

•  Two main components •  Storage: Hadoop Distributed File System (HDFS) •  Processing: MapReduce

• Hadoop clusters are composed of computers called nodes •  Clusters range from a single node up to several thousand nodes


How Did Apache Hadoop Originate?

• Heavily influenced by Google’s architecture •  Notably, the Google Filesystem and MapReduce papers

• Other Web companies quickly saw the benefits •  Early adop=on by Yahoo, Facebook and others

2002 2003 2004 2005 2006

Google publishes MapReduce paper

Nutch rewritten for MapReduce

Hadoop becomesLucene subproject

Nutch spun off from Lucene

Google publishes GFS paper


Comparing Hadoop to Other Systems

• Monolithic systems don’t scale • Modern high-‐performance compu=ng systems are distributed

•  They spread computa=ons across many machines in parallel • Widely-‐used used for scien=fic applica=ons •  Let’s examine how a typical HPC system works


Architecture of a Typical HPC System

Storage System

Compute Nodes

Fast Network



Storage System

Compute Nodes

Step 1: Copy input data

Fast Network



Storage System

Compute Nodes

Step 2: Process the data

Fast Network



Storage System

Compute Nodes

Step 3: Copy output data

Fast Network


You Don’t Just Need Speed…

•  The problem is that we have way more data than code

$ du -ks code/ 1,087 $ du –ks data/ 854,632,947,314


You Need Speed At Scale

Storage System

Compute Nodes

Bottleneck


Hadoop Design Fundamental: Data Locality

•  This is a hallmark of Hadoop’s design •  Don’t bring the data to the computa=on •  Bring the computa=on to the data

• Hadoop uses the same machines for storage and processing •  Significantly reduces need to transfer data across network


Other Hadoop Design Fundamentals

• Machine failure is unavoidable – embrace it •  Build reliability into the system

•  “More” is usually beNer than “faster” •  Throughput maNers more than latency


The Hadoop Distributed Filesystem

HDFS


HDFS: Hadoop Distributed File System

•  Inspired by the Google File System •  Reliable, low-‐cost storage for massive amounts of data

•  Similar to a UNIX filesystem in some ways •  Hierarchical •  UNIX-‐style paths (e.g., /sales/alice.txt) •  UNIX-‐style file ownership and permissions


HDFS: Hadoop Distributed File System

•  There are also some major devia=ons from UNIX filesystems • Highly-‐op=mized for processing data with MapReduce

•  Designed for sequen=al access to large files •  Cannot modify file content once wriNen

•  It’s actually a user-‐space Java process •  Accessed using special commands or APIs

• No concept of a current working directory


Copying Local Data To and From HDFS

•  Remember that HDFS is dis=nct from your local filesystem •  hadoop fs –put copies local files to HDFS •  hadoop fs –get fetches a local copy of a file from HDFS

$ hadoop fs -put sales.txt /reports

Hadoop Cluster

Client Machine

$ hadoop fs -get /reports/sales.txt


HDFS Demo

•  I will now demonstrate the following 1.  How to list the contents of a directory 2.  How to create a directory in HDFS 3.  How to copy a local file to HDFS 4.  How to display the contents of a file in HDFS 5.  How to remove a file from HDFS


A Scalable Data Processing Framework

Data Processing with MapReduce


What is MapReduce?

• MapReduce is a programming model •  It’s a way of processing data •  You can implement MapReduce in any language


Understanding Map and Reduce

•  You supply two func=ons to process data: Map and Reduce • Map: typically used to transform, parse, or filter data •  Reduce: typically used to summarize results

•  The Map func=on always runs first •  The Reduce func=on runs aUerwards, but is op=onal

•  Each piece is simple, but can be powerful when combined


MapReduce Benefits

•  Scalability •  Hadoop divides the processing job into individual tasks •  Tasks execute in parallel (independently) across cluster

•  Simplicity •  Processes one record at a =me

•  Ease of use •  Hadoop provides job scheduling and other infrastructure •  Far simpler for developers than typical distributed compu=ng


MapReduce in Hadoop

• MapReduce processing in Hadoop is batch-‐oriented • A MapReduce job is broken down into smaller tasks

•  Tasks run concurrently •  Each processes a small amount of overall input

• MapReduce code for Hadoop is usually wriNen in Java •  This uses Hadoop’s API directly

•  You can do basic MapReduce in other languages •  Using the Hadoop Streaming wrapper program •  Some advanced features require Java code


MapReduce Example in Python

•  The following example uses Python •  Via Hadoop Streaming

•  It processes log files and summarizes events by type •  I’ll explain both the data flow and the code


Job Input

• Here’s the job input

•  Each map task gets a chunk of this data to process •  Typically corresponds to a single block in HDFS

2013-06-29 22:16:49.391 CDT INFO "This can wait"2013-06-29 22:16:52.143 CDT INFO "Blah blah blah"2013-06-29 22:16:54.276 CDT WARN "This seems bad"2013-06-29 22:16:57.471 CDT INFO "More blather"2013-06-29 22:17:01.290 CDT WARN "Not looking good"2013-06-29 22:17:03.812 CDT INFO "Fairly unimportant"2013-06-29 22:17:05.362 CDT ERROR "Out of memory!"


#!/usr/bin/env python import sys levels = ['TRACE', 'DEBUG', 'INFO', 'WARN', 'ERROR', 'FATAL'] for line in sys.stdin: fields = line.split() level = fields[3].upper() if level in levels: print "%s\t1" % level

1 2 3 4 5 6 7 8 9

10 11 12 1314

Python Code for Map Func=on

If it matches a known level, print it, a tab separator, and the literal value 1 (since the level can only occur once per line)

Read records from standard input. Use whitespace to split into fields.

Define list of known log levels

Extract “level” field and convert to uppercase for consistency.


Output of Map Func=on

•  The map func=on produces key/value pairs as output

INFO 1INFO 1WARN 1INFO 1WARN 1INFO 1ERROR 1


The “Shuffle and Sort”

• Hadoop automa9cally merges, sorts, and groups map output •  The result is passed as input to the reduce func=on • More on this later…


ERROR 1INFO 1INFO 1INFO 1INFO 1WARN 1WARN 1

Shuffle and Sort

Map Output Reduce Input


Input to Reduce Func=on

•  Reduce func=on receives a key and all values for that key

•  Keys are always passed to reducers in sorted order

•  Although not obvious here, values are unordered



Python Code for Reduce Func=on

#!/usr/bin/env python import sys previous_key = None sum = 0 for line in sys.stdin: key, value = line.split() if key == previous_key: sum = sum + int(value) # continued on next slide

1 2 3 4 5 6 7 8 9

10 11 12 13

Ini=alize loop variables

Extract the key and value passed via standard input

If key unchanged, increment the count


Python Code for Reduce Func=on

# continued from previous slide else: if previous_key: print '%s\t%i' % (previous_key, sum) previous_key = key sum = 1 print '%s\t%i' % (previous_key, sum)

14 15 16 17 18 19 20 21 22 Print data for the final

key

If key changed, print data for old level

Start tracking data for the new record


Output of Reduce Func=on

•  Its output is a sum for each level

ERROR 1INFO 4WARN 2


Recap of Data Flow

ERROR 1INFO 4WARN 2



Map input

Map output Reduce input Reduce output

2013-06-29 22:16:49.391 CDT INFO "This can wait"2013-06-29 22:16:52.143 CDT INFO "Blah blah blah"2013-06-29 22:16:54.276 CDT WARN "This seems bad"2013-06-29 22:16:57.471 CDT INFO "More blather"2013-06-29 22:17:01.290 CDT WARN "Not looking good"2013-06-29 22:17:03.812 CDT INFO "Fairly unimportant"2013-06-29 22:17:05.362 CDT ERROR "Out of memory!"

Shuffle and sort


How to Run a Hadoop Streaming Job

•  I’ll demonstrate this now…


Open Source Tools that Complement Hadoop

The Hadoop Ecosystem


The Hadoop Ecosystem

•  "Core Hadoop" consists of HDFS and MapReduce •  These are the kernel of a much broader plauorm

• Hadoop has many related projects •  Some help you integrate Hadoop with other systems •  Others help you analyze your data

•  These are not considered “core Hadoop” •  Rather, they’re part of the Hadoop ecosystem • Many are also open source Apache projects


Visual Overview of a Complete Workflow Import Transaction Data

from RDBMSSessionize WebLog Data with Pig

Analyst uses Impala forbusiness intelligence

Sentiment Analysis on Social Media with Hive

Hadoop Cluster with Impala

Generate Nightly Reports using Pig, Hive, or Impala

Build product recommendations for

Web site


Key Points

• We’re genera=ng massive volumes of data •  This data can be extremely valuable •  Companies can now analyze what they previously discarded

• Hadoop supports large-‐scale data storage and processing •  Heavily influenced by Google's architecture •  Already in produc=on by thousands of organiza=ons •  HDFS is Hadoop's storage layer • MapReduce is Hadoop's processing framework

• Many ecosystem projects complement Hadoop •  Some help you to integrate Hadoop with exis=ng systems •  Others help you analyze the data you’ve stored


Highly Recommended Books

Author: Tom White ISBN: 1-‐449-‐31152-‐0

Author: Eric Sammer ISBN: 1-‐449-‐32705-‐2


Ques=ons?

•  Thank you for aNending! •  I’ll be happy to answer any addi=onal ques=ons now… • Demo and slides at github.com/markgrover/hadoop-‐intro-‐fast •  TwiNer: mark_grover •  Survey page: =ny.cloudera.com/mark

Intro to Hadoop Presentation at Carnegie Mellon - Silicon Valley

Engineering

blah blah

rights reserved

rights reserved

rights reserved

rights reserved

rights reserved

hadoop ecosystem

apache hadoop