Agile Data Science: Building Hadoop Analytics Applications

Agile Data Science

January 2014

Agile Analytics Applications with Hadoop

2

About Me…Bearding.•Bearding is my #1 natural talent.•I’m going to beat this guy. •Seriously.•Salty Sea Beard•Fortified with Pacific Ocean Minerals

2

3

Agile Data Science: The Book

A philosophy.Not the only way, but it’s a really good way!

Code: ‘AUTHD’ – 50% off

3

4

We Go Fast, But Don’t Worry!•Download the slides - click the links - read examples!•If it’s not on the blog (Hortonworks, Data Syndrome), it ’s in the book!

•Order now: http://shop.oreilly.com/product/0636920025054.do

4

5

Agile Application Development: Check•LAMP stack mature•Post-Rails frameworks to choose from•Enable rapid feedback and agility

+ NoSQL

5

6

Data Warehousing

6

7

Scientific Computing / HPC

Tubes and Mercury (Old School) Cores and Spindles (New School)

UNIVAC and Deep Blue both fill a warehouse. We’re back!

7

‘Smart Kid’ Only: MPI, Globus, etc. Until Hadoop

8

Data Science?Application

Development Data Warehousing

Scientific Computing / HPC

8

9

Data Center as Computer

“A key challenge for architects of WSCs is to smooth out these discrepancies in a cost efficient manner.” Click here for a paper on operating a ‘data center as computer.’

9

Warehouse Scale Computers and Applications

10

Hadoop to the Rescue!•Easy to use (Pig, Hive, Cascading)•CHEAP: 1% the cost of SAN/NAS•A department can afford its own Hadoop cluster!•Dump all your data in one place: Hadoop DFS•Silos come CRASHING DOWN!• JOIN like crazy!•ETL like whoa!•An army of mappers and reducers at your command•OMGWTFBBQ ITS SO GREAT! I FEEL AWESOME!

10

11

NOW WHAT?

11

12

Analytics Apps: It takes a Team•Broad skill-set•Nobody has them all•Inherently collaborative

12

13

Data Science Team•3-4 team members with broad, diverse skill-sets that overlap

•Transactional overhead dominates at 5+ people•Expert researchers: lend 25-50% of their time to teams•Creative workers. Like a studio, not an assembly line•Total freedom... with goals and deliverables.•Work environment matters most

13

14

How To Get Insight Into Product•Back-end has gotten THICKER

•Generating $$$ insight can take 10-100x app dev

•Timeline disjoint: analytics vs agile app-dev/design•How do you ship insights efficiently?•Can you collaborate on research vs developer timeline?

14

15

The Wrong Way - Part One“We made a great design. Your job is to predict the future for it.”

15

16

The Wrong Way - Part Two“What is taking you so long to reliably predict the future?”

16

17

The Wrong Way - Part Three“The users don’t understand what 86% true means.”

17

18

The Wrong Way - Part FourGHJIAEHGIEhjagigehganb!!!!!RJ(@J?!!

18

19

The Wrong Way - ConclusionInevitable Conclusion

Plane Mountain

19

20

Reminds me of... the waterfall model

:(

20

21

Chief ProblemYou can’t design insight in analytics applications.

You discover it.

You discover by exploring.

21

22

-> StrategySo make an app for exploring your data.

Which becomes a palette for what you ship.

Iterate and publish intermediate results.

22

23

Data Design•Not the 1st query that = insight, it’s the 15th, or 150th

•Capturing “Ah ha!” moments

•Slow to do those in batch…

•Faster, better context in an interactive web application.

•Pre-designed charts wind up terrible. So bad.

•Easy to invest man-years in wrong statistical models

•Semantics of presenting predictions are complex

•Opportunity lies at intersection of data & design

23

24

How Do We Get Back to Agile?

24

25

Statement of Principles(Then Tricks With Code)

25

26

Setup An Environment Where:•Insights repeatedly produced•Iterative work shared with entire team•Interactive from day Zero•Data model is consistent end-to-end•Minimal impedance between layers•Scope and depth of insights grow•Insights form the palette for what you ship•Until the application pays for itself and more

26

27

Snowballing Audience

27

28

Value Document > Relation

Most data is dirty. Most data is semi-structured or unstructured. Rejoice!28

29

Value Document > Relation

Note: Hive/ArrayQL/NewSQL’s support of documents/array types blur this distinction.

29

30

Relational Data = Legacy Format•Why JOIN? Storage is fundamentally cheap!

•Duplicate that JOIN data in one big record type!

•ETL once to document format on import, NOT every job

•Not zero JOINs, but far fewer JOINs

•Semi-structured documents preserve data’s actual structure

•Column compressed document formats beat JOINs!

30

31

Value Imperative > Declarative•We don’t know what we want to SELECT.

•Data is dirty - check each step, clean iteratively.•85% of data scientist’s time spent munging. ETL.•Imperative is optimized for our process.•Process = iterative, snowballing insight•Efficiency matters, self optimize

31

32

Value Dataflow > SELECT

32

33

Ex. Dataflow: ETL + Email Sent Count

(I can’t read this either. Get a big version here.)33

34

Value Pig > Hive (for app-dev)•Pigs eat ANYTHING•Pig is optimized for refining data, as opposed to consuming it

•Pig is imperative, iterative•Pig is dataflows, and SQLish (but not SQL)•Code modularization/re-use: Pig Macros•ILLUSTRATE speeds dev time (even UDFs)•Easy UDFs in Java, JRuby, Jython, Javascript•Pig Streaming = use any tool, period.•Easily prepare our data as it will appear in our app.•If you prefer Hive, use Hive.

Actually, I wish Pig and Hive were one tool. Pig, then Hive, then Pig, then Hive. See: HCatalog for Pig/Hive integration.

34

35

Localhost vs Petabyte Scale:Same Tools•Simplicity essential to scalability: highest level tools we can•Prepare a good sample - tricky with joins, easy with documents•Local mode: pig -l /tmp -x local -v -w•Frequent use of ILLUSTRATE•1st: Iterate, debug & publish locally•2nd: Run on cluster, publish to team/customer•Consider skipping Object-Relational-Mapping (ORM)•We do not trust ‘databases,’ only HDFS @ n=3•Everything we serve in our app is re-creatable via Hadoop.

35

36

Data-Value Pyramid

Climb it. Do not skip steps. See here.36

37

0/1) Display Atomic Records On The Web

37

38

0.0) Document - Serialize Events•Protobuf•Thrift•JSON•Avro - I use Avro because the schema is onboard.

38

39

0.1) Documents Via Relation ETLenron_messages = load '/enron/enron_messages.tsv' as (

message_id:chararray,

sql_date:chararray,

from_address:chararray,

from_name:chararray,

subject:chararray,

body:chararray);

 enron_recipients = load '/enron/enron_recipients.tsv' as ( message_id:chararray, reciptype:chararray, address:chararray,

name:chararray);

 split enron_recipients into tos IF reciptype=='to', ccs IF reciptype=='cc', bccs IF reciptype=='bcc';

headers = cogroup tos by message_id, ccs by message_id, bccs by message_id parallel 10;

with_headers = join headers by group, enron_messages by message_id parallel 10;

emails = foreach with_headers generate enron_messages::message_id as message_id,

CustomFormatToISO(enron_messages::sql_date, 'yyyy-MM-dd HH:mm:ss') as date,

TOTUPLE(enron_messages::from_address, enron_messages::from_name) as from:tuple(address:chararray,

name:chararray), enron_messages::subject as subject,

enron_messages::body as body,

headers::tos.(address, name) as tos,

headers::ccs.(address, name) as ccs,

headers::bccs.(address, name) as bccs;

store emails into '/enron/emails.avro' using AvroStorage(

Example here.39

40

0.2) Serialize Events From Streamsclass GmailSlurper(object): ...  def init_imap(self, username, password):    self.username = username    self.password = password    try:      imap.shutdown()    except:      pass    self.imap = imaplib.IMAP4_SSL('imap.gmail.com', 993)    self.imap.login(username, password)    self.imap.is_readonly = True ...  def write(self, record):    self.avro_writer.append(record) ...  def slurp(self):    if(self.imap and self.imap_folder):      for email_id in self.id_list:        (status, email_hash, charset) = self.fetch_email(email_id)        if(status == 'OK' and charset and 'thread_id' in email_hash and 'froms' in email_hash):          print email_id, charset, email_hash['thread_id']          self.write(email_hash)

Scrape your own gmail in Python and Ruby.40

41

0.3) ETL Logslog_data = LOAD 'access_log' USING org.apache.pig.piggybank.storage.apachelog.CommongLogLoader AS (remoteAddr, remoteLogname, user, time, method, uri, proto, bytes);

41

42

1) Plumb Atomic Events->Browser

(Example stack that enables high productivity)42

43

1.1) Cat Avro Serialized Eventsme$ cat_avro ~/Data/enron.avro

{ u'bccs': [], u'body': u'scamming people, blah blah', u'ccs': [], u'date': u'2000-08-28T01:50:00.000Z', u'from': {u'address': u'[email protected]', u'name': None}, u'message_id': u'<1731.10095812390082.JavaMail.evans@thyme>', u'subject': u'Re: Enron trade for frop futures', u'tos': [ {u'address': u'[email protected]', u'name': None} ]}

Get cat_avro in python, ruby43

44

1.2) Load Events in Pigme$ pig -l /tmp -x local -v -w

grunt> enron_emails = LOAD '/enron/emails.avro' USING AvroStorage();grunt> describe enron_emailsemails: { message_id: chararray, datetime: chararray, from:tuple(address:chararray,name:chararray) subject: chararray, body: chararray, tos: {to: (address: chararray,name: chararray)}, ccs: {cc: (address: chararray,name: chararray)}, bccs: {bcc: (address: chararray,name: chararray)}} 

44

45

1.3) ILLUSTRATE Events in Piggrunt> illustrate enron_emails ---------------------------------------------------------------------------| emails | | message_id:chararray | | datetime:chararray | | from:tuple(address:chararray,name:chararray) | | subject:chararray | | body:chararray | | tos:bag{to:tuple(address:chararray,name:chararray)} || ccs:bag{cc:tuple(address:chararray,name:chararray)} || bccs:bag{bcc:tuple(address:chararray,name:chararray)} |---------------------------------------------------------------------------| | | <1731.10095812390082.JavaMail.evans@thyme> | | 2001-01-09T06:38:00.000Z | | ([email protected], J.R. Bob Dobbs) | | Re: Enron trade for frop futures | | scamming people, blah blah | | {([email protected],)} | | {} | | {} |

Upgrade to Pig 0.10+45

46

1.4) Publish Events to a ‘Database’

pig -l /tmp -x local -v -w -param avros=enron.avro \ -param mongourl='mongodb://localhost/enron.emails' avro_to_mongo.pig

/* MongoDB libraries and configuration */

register /me/mongo-hadoop/mongo-2.7.3.jarregister /me/mongo-hadoop/core/target/mongo-hadoop-core-1.1.0-SNAPSHOT.jarregister /me/mongo-hadoop/pig/target/mongo-hadoop-pig-1.1.0-SNAPSHOT.jar

/* Set speculative execution off to avoid chance of duplicate records in Mongo */set mapred.map.tasks.speculative.execution falseset mapred.reduce.tasks.speculative.execution falsedefine MongoStorage com.mongodb.hadoop.pig.MongoStorage(); /* Shortcut */

/* By default, lets have 5 reducers */set default_parallel 5avros = load '$avros' using AvroStorage();store avros into '$mongourl' using MongoStorage();

Full instructions here.

Which does this:

From Avro to MongoDB in one command:

46

47

1.5) Check Events in ‘Database’$ mongo enronMongoDB shell version: 2.0.2connecting to: enron

show collectionsEmailssystem.indexes

>db.emails.findOne({message_id: "<1731.10095812390082.JavaMail.evans@thyme>"}){

"_id" : ObjectId("502b4ae703643a6a49c8d180"),"message_id" : "<1731.10095812390082.JavaMail.evans@thyme>","date" : "2001-01-09T06:38:00.000Z","from" : { "address" : "[email protected]", "name" : "J.R. Bob Dobbs" },"subject" : Re: Enron trade for frop futures,"body" : "Scamming more people...","tos" : [ { "address" : "connie@enron", "name" : null } ],"ccs" : [ ],"bccs" : [ ]

}

47

48

1.6) Publish Events on the Webrequire 'rubygems'require 'sinatra'require 'mongo'require 'json'

connection = Mongo::Connection.newdatabase = connection['agile_data']collection = database['emails']

get '/email/:message_id' do |message_id| data = collection.find_one({:message_id => message_id}) JSON.generate(data)end

48

49

1.6) Publish events on the web

49

50

One-Liner to Transition Stack

50

51

What’s the Point?•A designer can work against real data.•An application developer can work against real data.•A product manager can think in terms of real data.•Entire team is grounded in reality!•You’ll see how ugly your data really is.•You’ll see how much work you have yet to do.•Ship early and often!•Feels agile, don’t it? Keep it up!

51

52

1.7) Wrap Events with Bootstrap

<link href="/static/bootstrap/docs/assets/css/bootstrap.css" rel="stylesheet"></head><body><div class="container" style="margin-top: 100px;"> <table class="table table-striped table-bordered table-condensed"> <thead> {% for key in data['keys'] %} <th>{{ key }}</th> {% endfor %} </thead> <tbody> <tr> {% for value in data['values'] %} <td>{{ value }}</td> {% endfor %} </tr> </tbody> </table></div></body>

Complete example here with code here.52

53

1.7) Wrap Events with Bootstrap

53

54

Refine. Add Links Between Documents.

Not the Mona Lisa, but coming along... See: here54

56

1.8) List Links to Sorted Events

mongo enron> db.emails.ensureIndex({message_id: 1})> db.emails.find().sort({date:0}).limit(10).pretty() { { "_id" : ObjectId("4f7a5da2414e4dd0645d1176"), "message_id" : "<CA+bvURyn-rLcH_JXeuzhyq8T9RNq+YJ_Hkvhnrpk8zfYshL-wA@mail.gmail.com>", "from" : [ ...

pig -l /tmp -x local -v -wemails_per_user = foreach (group emails by from.address) { sorted = order emails by date;last_1000 = limit sorted 1000;generate group as from_address, emails as emails; };store emails_per_user into '$mongourl' using MongoStorage();

Use Pig, serve/cache a bag/array of email documents:

Use your ‘database’, if it can sort.

56

57

1.8) List Links to Sorted Documents

57

58

1.9) Make It SearchableIf you have list, search is easy with ElasticSearch and Wonderdog.../* Load ElasticSearch integration */register '/me/wonderdog/target/wonderdog-1.0-SNAPSHOT.jar';register '/me/elasticsearch-0.18.6/lib/*';define ElasticSearch com.infochimps.elasticsearch.pig.ElasticSearchStorage();

emails = load '/me/tmp/emails' using AvroStorage();store emails into 'es://email/email?json=false&size=1000' using ElasticSearch('/me/elasticsearch-0.18.6/config/elasticsearch.yml', '/me/elasticsearch-0.18.6/plugins');

curl -XGET 'http://localhost:9200/email/email/_search?q=hadoop&pretty=true&size=1'

Test it with curl:

ElasticSearch has no security features. Take note. Isolate.58

59

2) Create Simple Charts

59

60

2) Create Simple Tables and Charts

60

61

2) Create Simple Charts•Start with an HTML table on general principle.•Then use nvd3.js - reusable charts for d3.js•Aggregate by properties & displaying is first step in entity resolution•Start extracting entities. Ex: people, places, topics, time series•Group documents by entities, rank and count.•Publish top N, time series, etc.•Fill a page with charts.•Add a chart to your event page.

61

62

2.1) Top N (of Anything) in Pigpig -l /tmp -x local -v -w

top_things = foreach (group things by key) { sorted = order things by arbitrary_rank desc;top_10_things = limit sorted 10;generate group as key, top_10_things as top_10_things;};store top_n into '$mongourl' using MongoStorage();

Remember, this is the same structure the browser gets as json.This would make a good Pig Macro.

62

63

2.2) Time Series (of Anything) in Pigpig -l /tmp -x local -v -w

/* Group by our key and date rounded to the month, get a total */things_by_month = foreach (group things by (key, ISOToMonth(datetime))generate flatten(group) as (key, month),

COUNT_STAR(things) as total;

/* Sort our totals per key by month to get a time series */things_timeseries = foreach (group things_by_month by key) { timeseries = order things by month;generate group as key, timeseries as timeseries;};

store things_timeseries into '$mongourl' using MongoStorage();

Yet another good Pig Macro.

63

64

Data Processing in Our StackA new feature in our application might begin at any layer…GREAT!

Any team member can add new features, no problemo!

I’m creative!I know Pig!

I’m creative too!

I <3 Javascript!

omghi2u!

where r my legs?

send halp

64

65

Data Processing in Our Stack... but we shift the data-processing towards batch, as we are able.

Ex: Overall total emails calculated in each layerSee real example here.

65

66

3) Exploring with Reports

66

67

3) Exploring with Reports

67

68

3.0) From Charts to Reports•Extract entities from properties we aggregated by in charts (Step 2)•Each entity gets its own type of web page

•Each unique entity gets its own web page

•Link to entities as they appear in atomic event documents (Step 1)•Link most related entities together, same and between types.•More visualizations!•Parametize results via forms.

68

69

3.1) Looks Like This:

69

70

3.2) Cultivate Common Keyspaces

70

71

3.3) Get People Clicking. Learn.•Explore this web of generated pages, charts and links!•Everyone on the team gets to know your data.•Keep trying out different charts, metrics, entities, links. •See whats interesting.•Figure out what data needs cleaning and clean it.•Start thinking about predictions & recommendations.

‘People’ could be just your team, if data is sensitive.

71

72

4) Predictions and Recommendations

72

73

4.0) Preparation•We’ve already extracted entities, their properties and relationships•Our charts show where our signal is rich•We’ve cleaned our data to make it presentable•The entire team has an intuitive understanding of the data•They got that understanding by exploring the data•We are all on the same page!

73

74

4.2) Think in Different Perspectives•Networks•Time Series / Distributions•Natural Language Processing•Conditional Probabilities / Bayesian Inference•Check out Chapter 2 of the book

74

75

4.3) Networks

75

76

4.3.1) Weighted Email Networks in Pig

76

77

4.3.2) Networks Viz with Gephi

77

78

4.3.3) Gephi = Easy

78

79

4.3.4) Social Network Analysis

79

80

4.4) Time Series & Distributions

80

81

4.4.1) Smooth Sparse Data

See here. 81

82

4.4.2) Regress to Find TrendsJRuby Linear Regression UDF Pig to use the UDF

Trend Line in your Application

82

83

4.5.1) Natural Language Processing

Example with code here and macro here.

83

84

4.5.2) NLP: Extract Topics!

84

85

4.5.3) NLP for All: Extract Topics!•TF-IDF in Pig - 2 lines of code with Pig Macros:

•http://hortonworks.com/blog/pig-macro-for-tf-idf-makes-topic-summarization-2-lines-of-pig/

•LDA with Pig and the Lucene Tokenizer: •http://thedatachef.blogspot.be/2012/03/topic-discovery-with-apache-pig-and.html

85

86

4.6) Probability & Bayesian Inference

86

87

4.6.1) Gmail Suggested Recipients

87

88

4.6.1) Reproducing it with Pig

88

89

4.6.2) Step 1: COUNT (From -> To)

89

90

4.6.2) Step 2: COUNT (From, To, Cc)/Total

P(cc | to) = Probability of cc’ing someone, given that you’ve to’d someone

90

91

4.6.3) Wait - Stop Here! It Works!

They match…

91

92

4.4) Add Predictions to Reports

92

93

5) Enable New Actions

93

94

Why Doesn’t Kate Reply to My Emails?•What time is best to catch her?•Are they too long?•Are they meant to be replied to (original content)?•Are they nice? (sentiment analysis)•Do I reply to her emails (reciprocity)?•Do I cc the wrong people (my mom)?

94

97

Thank You!•Questions & Answers

97

•Follow: @rjurney•Read the Blog: datasyndrome.com