Big Data Analysis Patterns - TriHUG 6/27/2013

Big Data
Analysis Patterns
TriHUG
6/27/2013
1

whoami
•

Brad Anderson

•

Solutions Architect at MapR (Atlanta)

•

ATLHUG co-chair

•

NoSQL East Conference 2009

•

“boorad” most places (twitter, github)

•

banderson@maprtech.com
2

Big Data is not new!
but the tools are.

5

The Good News in Big Data:

“Simple algorithms and lots of data
trump complex models”

Halevy, Norvig, and Pereira, Google
IEEE Intelligent Systems
6

The Challenge: So Many Solutions!
What solutions fit your business problem?
For example, do you need…



Apache Mahout?



Storm?



Apache Solr/Lucene?



Apache HBase (or MapR M7)?



Apache Drill (or Impala?)



d3.js or Tableau?



Node.js


7

Apache Hadoop?

Titan?
7

Ask a Different Question
It may be more useful to better define the problem by asking some
of these questions:



How large is the data to be queried? (the analysis volume)



What time frame is appropriate for your query response?



How fast is data arriving? (bursts or continuously?)



Are queries by sophisticated users?



Are you looking for common patterns or outliers?



8

How large is the data to be stored?

How are your data sources structures?

8

Picking the Best Solution
Your responses to these questions can help you better:


define the problem



recognize the analysis pattern to which it belongs



guide the choice of solutions to try

But first, here’s a quick review of a few of the technologies you
might choose, and then we will focus on three of the questions as a
part of the landscape.

9

9

Apache Solr/Lucene
Solr/Lucene is a powerful search engine used for flexible, heavily
indexed queries including data such as


Full text



Geographical data



Statistically weighted data

Solr is a small data tool that has flourished in a big data world

10

Apache Mahout
Mahout provides a library of scalable machine learning algorithms
useful for big data analysis based on Hadoop or other storage
systems.

Mahout algorithms mainly are used for


Recommendation (collaborative filtering)



Clustering



Classification

Mahout can be used in conjunction with solutions such as Solr: You
might use Mahout to create a co-occurrence data base that could
then be queried using a search tool such as Solr

11

Apache Drill


Google Dremel clone



Pluggable Query Languages
–
–



Pluggable Storage Backends
–
–
–



Starts with ANSI SQL 2003
Hive, Pig, Cascading, MongoQL, …
Hadoop, Hbase
MongoDB (BSON)
RDBMS?

Bypasses MapReduce

12

Storm


Realtime Stream Computation Engine



Horizontal Scalability



Guaranteed Data Processing



Fault Tolerance



Higher level abstraction over:
–

–



Message Queues
Worker Logic

“The Hadoop of Realtime”

13

Titan


Distributed Graph Database



Property Graph



Pluggable Backend Storage
–
–
–



Search Integrated
–
–



HBase or M7
Cassandra
Berkeley DB
Solr/Lucene
Elastic Search

Faunus
–
–

Graph traversals on subset
In-memory
14

Using the Answers to Guide Your Choices
For simplicity, let’s focus in on the first three questions:


How large is the data to be stored?



How large is the data to be queried? (the analysis volume)



What time frame is appropriate for your query response?

15

Big Data Decision Tree
How big is your data?
<10 GB

mid
?

?

A

Single element
at a time

>200 GB

What size queries?
One pass
over 100%

B

Response time?

C

Big storage

Multiple passes
over big chunks

Streaming

< 100s
(human scale)
D
16

throughput
not response
E

Use Cases
Company
 Data Shape
 Technique(s)
 Business Value


17

Telecommunications Giant

ETL Offload
20

Telecommunications






Data Shape

Lots of Data
Lots of Queries across Large Sets
Throughput important

21

Telecommunications

Techniques
Analytics

ETL

22

Telecommunications

Techniques

+
ETL (Hadoop)

Analytics (Teradata)
23

Telecommunications

Business Value

24

Credit Card
Issuer

Data Shape








Customer Purchase History (big)
Merchant Designations
Merchant Special Offers
Throughput important
Recommendations
26

Search Abuse

Techniques
A Recommendation Engine with Mahout and Solr/Lucene

History matrix
One row per user
One column per thing
27

Techniques
Recommendation based on
cooccurrence
Cooccurrence gives item-item
mapping
One row and column per thing
28

Techniques
Cooccurrence matrix can also be
implemented as a search index

29

Techniques
Complete
history

Cooccurrence
(Mahout)

SolR
SolR
Indexer
Solr
Indexer
indexing

Item metadata

Index
shards

30

20 Hrs  3 Hrs

Techniques
User
history

SolR
SolR
Indexer
Solr
Indexer
search

Web tier

8Hrs  3 Min

Item metadata

Index
shards

31

Techniques
Hadoop
Purchase
History

Export
(4 hrs)

App
App

Merchant
Information

Recommendation
Engine Results
(Mahout)

Presentation
Data Store
(DB2)

App
App

Merchant
Offers

App

Import
(4 hrs)
32

Techniques
Hadoop
Purchase
History
Merchant
Information

Recommendation
Engine Results
(Mahout)

Index
Update
(3 min)

App
App

Recommendation
Search Index
(Solr)

App
App

Merchant
Offers

App

33

Waste & Recycling Leader

Idle Alerts
35

Data Shape
Truck Geolocation Data
– 20,000 trucks
– 5 sec interval (arriving quickly)
 Landfill Geographic Boundaries


36

Techniques
Realtime Stream Computation
(Storm)

Truck
Geolocation

Data

Hadoop
Storage

Immediate
Alerts

Batch Computation
(MapReduce)

Tax Reduction
Reporting

Shortest Path
Graph Algorithm
(Titan)

Route
Optimization

37

Beverage Company

Social Engagement Application

39

Data Shape

Tweets, FB Messages
 Person, Activity links
 Graph Traversal


40

Consumer Activity Graph
Wal*Mart.com
Ebay
Shopping.com
Sam’s
Ebay Motors
Dollar General
StubHub
CVS

41

Toys R Us

Techniques
Property Graph
(Titan)

Social
Activity
Stream
Key/Value Store
(MapR M7)

42

Graph Traversal
(Faunus)

Big Data Analysis Patterns - TriHUG 6/27/2013

More Related Content

What's hot

Viewers also liked

Similar to Big Data Analysis Patterns - TriHUG 6/27/2013

More from boorad

Recently uploaded

Big Data Analysis Patterns - TriHUG 6/27/2013