Вы находитесь на странице: 1из 26

Hadoop meets Mature BI:

Where the rubber meets the road for Data Scientists


Michael Hiskey
Futurist, + Product Evangelist
VP, Marketing & Business Development Kognitio
@Kognitio @mphnyc #MPP_R #OANYC

The Data Scientist


Sexiest job of the 21st Century?

@Kognitio

@mphnyc

#MPP_R #OANYC

Key Concept: Graduation


Projects will need to Graduate from the Data Science Lab and become part of Business as Usual
@Kognitio @mphnyc #MPP_R #OANYC

Demand for the Data Scientist


Organizational appetite for tens, not hundreds

@Kognitio

@mphnyc

#MPP_R #OANYC

Dont be a Railroad Stoker!


Highly skilled engineering required but the world innovated around them.

@Kognitio

@mphnyc

#MPP_R #OANYC

Business Intelligence
Straddle IT and Business
Numbers Tables Charts Indicators Time - History - Lag Access - to view (portal) - to data - to depth - Control/Secure Consumption - digestion

with ease and simplicity More granularity

Richer data model Faster

Lower latency Self service


@Kognitio @mphnyc #MPP_R #OANYC

What has changed?


More connected-users?

More-connected users?
@Kognitio @mphnyc #MPP_R #OANYC

According to one estimate, mankind created 150 exabytes of data in 2005 (billion gigabytes)

In 2010 this was 1,200 exabytes

Data flow

Data Variety

@Kognitio

@mphnyc

#OANYC

What?
New value comes from your existing data

Respondents were asked to choose up to two descriptions about how their organizations view big data from the choices above. Choices have been abbreviated, and selections have been normalized to equal 100%. n=1144
Source: IBM Institute for Business Value/Said Business School Survey

@Kognitio

@mphnyc

#OANYC

20th Century Fox

@Kognitio

@mphnyc

#MPP_R #OANYC

Hadoop ticks many but not all the boxes

aaa aaaa aa aa a aa aa a a a aa aa aaaa a aa aa


@Kognitio @mphnyc #MPP_R #OANYC

Null storage concerns


No need to triage No need to pre-process No need to align to schema
@Kognitio @mphnyc #MPP_R #OANYC

The drive for deeper understanding


Analytical Complexity

Behaviour modelling Statistical Analysis Clustering

Machine learning algorithms

Dynamic Simulation

Dynamic Interaction Fraud detection Reporting & BPM Campaign Management

#MPP_R
Technology/Automation

@Kognitio

@mphnyc

#OANYC

Hadoop just too slow for interactive BI!


while hadoop shines as a processing

platform, it is painfully slow as a query tool

loss of trainof-thought
@Kognitio @mphnyc #MPP_R #OANYC

Analytics needs low latency, no I/O wait


High speed in-memory processing

@Kognitio

@mphnyc

#MPP_R #OANYC

Analytical Platform: Reference Architecture


Application & Client Layer
All BI Tools All OLAP Clients Excel

Analytical Platform Layer

Near-line Storage (optional)

Reporting

Persistence Layer

Hadoop Clusters

Cloud Storage

Enterprise Data Warehouses

Legacy Systems

The Future
Predictive Analytics Advanced Analytics

Big Data

In-memory

Data Scientists

Logical Data Warehouse


@Kognitio @mphnyc #MPP_R #OANYC

connect www.kognitio.com linkedin.com/companies/kognitio tinyurl.com/kognitio NA: +1 855 KOGNITIO EMEA: +44 1344 300 770 twitter.com/kognitio youtube.com/kognitio

Hadoop meets Mature BI: Where the rubber meets the road for Data Scientists
The key challenge for Data Scientists is not the proliferation of their roles, but the ability to graduate key Big Data projects from the Data Science Lab and production-ize them into their broader organizations. Over the next 18 months, "Big Data' will become just "Data"; this means everyone (even business users) will need to have a way to use it - without reinventing the way they interact with their current reporting and analysis. To do this requires interactive analysis with existing tools and massively parallel code execution, tightly integrated with Hadoop. Your Data Warehouse is dying; Hadoop will elicit a material shift away from price per TB in persistent data storage.
@Kognitio @mphnyc #MPP_R

Wanted
Dead or Alive

The No SQL Posse

SQL

The new bounty hunters: Drill Impala Pivotal Stinger

Its all about getting work done

Tasks evolving: Used to be simple fetch of value Then was calc dynamic aggregate

Now complex algorithms!

Behind the numbers

select receives sum(sales) total_sales Trans_Year, Num_Trans, dept, (sum(sales) SALEDATE DATE, DOW INTEGER, ROW_ID INTEGER, PRODNO INTEGER, DAILYSALES partition by PRODNO order by PRODNO, ROW_ID from sales_history count(distinct summary Account_ID) Num_Accts, sales_fact sends ( R_OUTPUT varchar ) where year = 2006 and month = -055 and region=1; sum(count( distinct Account_ID)) over (partition by-05Trans_Year Where period between date 01 2006 and date 31 2006 isolate partitions script S'endofr( # Simple R script to Total_Spend, run a linear fit on daily sales cast(sum(total_spend)/1000 as int) group by dept prod1<-read.csv(file=file("stdin"), header=FALSE,row.names cast(sum(total_spend)/1000 as int) / count(distinct Account_ID having sum(sales) > 50000; colnames(prod1)<-c("DOW","ID","PRODNO","DAILYSALES") rank() over (partition dim1<-dim(prod1) by Trans_Year order by count(distinct A daily1<-aggregate(prod1$DAILYSALES, list(DOW = prod1$DOW), rank() over (partition by Trans_Year order by sum(total_spend) daily1[,2]<-daily1[,2]/sum(daily1[,2]) from( select Account_ID, basesales<-array(0,c(dim1[1],2)) basesales[,1]<-prod1$ID Extract(Year from Effective_Date) Trans_Year, basesales[,2]<-(prod1$DAILYSALES/daily1[prod1$DOW+1,2]) @Kognitio @mphnyc #MPP_R count(Transaction_ID) Num_Trans, colnames(basesales)<-c("ID","BASESALES")
fit1=lm(BASESALES ~ ID,as.data.frame(basesales))

create external script LM_PRODUCT_FORECAST environment

rsint

For once technology is on our side


First time we have full triumvirate of
Excellent Computing power Unlimited storage Fast Networks

now that RAM is cheap!

@Kognitio

@mphnyc

#MPP_R

Hadoop is

Lots of these
Hadoop inherently disk oriented

Not so many of these


Typically low ratio of CPU to Disk

@Kognitio

@mphnyc

#MPP_R

Вам также может понравиться