Meta scale kognitio hadoop webinar

Webinar: Make Big Data Easy
with the Right tools and talent
- MetaScale Expertise and Kognitio Analytics
Accelerate Hadoop for Organizations Large and Small

October 2012

Today’s webinar

• 45 minutes with 15 minutes Q&A
• We will email you a link to the slides
• Feel free to use the Q & A feature

Agenda

Presenters
• Opening introduction
Dr. Phil Shelley
• MetaScale Expertise CEO, MetaScale
– Case study – Sears CTO, Sears Holdings

Holdings Roger Gaskell
• Kognitio Analytics CTO
Kognitio
– Hadoop acceleration
explained Host
• Summary Michael Hiskey
VP Marketing & Business
• Q&A Development
Kognitio

Big Data < > Hadoop

Big Data is high volume, velocity and variety
information assets that demand cost-effective,
innovative forms of information processing for
enhanced insight and decision-making

 Volume (not only) size
 Velocity (speed of Input / Output)
 Variety (lots of data sources)
 Value – not the SIZE of your data,
but what you can DO with it!

OK, so you’ve decided to put data in Hadoop...

Now what?

Dr. Phil Shelley
CEO – MetaScale
CTO Sears Holdings

Where Did We Start?
 Issues with meeting production schedules
 Multiple copies of data, no single point of truth
 ETL complexity, cost of software and cost to manage
 Time take to setup ETL data sources for projects
 Latency in data, up to weeks in some cases
 Enterprise Data Warehouses unable to handle load
 Mainframe workload over consuming capacity
 IT Budgets not growing – BUT data volumes escalating

Why Hadoop?

Traditional
Databases &
Warehouses

Hadoop

Enterprise Integration
 Data Sourcing
 Connecting to Legacy source systems

 Loaders and tools (speed considerations)

 Batch or near-real time

 Enterprise Data Model
 Establish a model and enterprise data strategy early

 Data Transformations
 The End of ETL as we know it

 Data re-use
 Drive re-use of data

 Single point of truth is now a possibility

 Data Consumption and user Interaction
 Consume data in-place wherever possible

 Move data only if you have to

 Exporting to legacy systems can be done, but it duplicates data

 Loaders and tools (speed considerations)

 How will your users interact with the data

Rethink Everything
The way you capture data
The way you store data
The structure of your data
The way you analyze data
The costs of data storage
The size of your data
What you can analyze
The speed of analysis
The skills of your team
The way user interact with data

The Learning from our Journey

• Big Data tools are here and ready for the Enterprise

• An Enterprise Data Architecture model is essential

• Hadoop can handle Enterprise workload
 To reduce strain on legacy platforms

 To reduce cost
 To bring new business opportunities

• Must be part of an overall data strategy

• Not to be underestimated

• The solution must be an Eco-System
 There has to be a simple way to consume the data

Page 12

Hadoop Strengths & Weaknesses?

• Cost effective platform
• Powerful / fast data processing environment
• Good at standard reporting
• Flexibility: Programmable, Any data type
• Huge scalability

• Barriers to entry: lots of engineering and coding
• High on-going coding requirements
• Difficult to access with standard BI/analytical tools
• Ad hoc complex analytics difficult
• Too slow for interactive analytics

What is an “In-memory” Analytical Platform?
• DBMS where all of the data of interest or specific portions of the data
have been permanently pre-loaded into random access memory (RAM)
• Not a large cache
– Data is held in structures that take advantage of the properties of
RAM – NOT copies of frequently used disk blocks
– The databases query optimiser knows at all times exactly which
data is in memory (and which is not)

In-Memory Analytical Database Mangement
Not a large cache:
• No disk access during query execution
– Temporary tables in RAM
– Results sets in RAM
• In-Memory means in high speed RAM
– NOT slow flash-based SSDs that mimic
mechanical disks

For more information:
• Gartner: “Who's Who in In-Memory DBMSs��
Roxanne Edjlali, Donald Feinberg
10 Sept 2012 www.gartner.com/id=2151315

Why In-memory: RAM is Faster Than Disk (Really!)
Actually, this only part of the story
Analytics completely change the workload
workload
characteristics on the database
Simple reporting and transactional processing
filtering
is all about “filtering” the data of interest
Analytics is all about complex “crunching”
crunching
of the data once it is filtered
CPU Crunching needs processing power and
cycles consumes CPU cycles
Storing data on physical disks severely limits the
storing
rate at which data can be provided to the CPUs
Accessing data directly from RAM allows
access
much more CPU power to be deployed

Analytics is about through Data
CPU cycle-intensive & CPU-bound
“CRUNCHING”
Analytical
Joins Functions
Aggregations

Sorts Grouping

• To understand what is happening in the data
More complex More pronounced
analytics = this becomes
• Analytical platforms are therefore CPU-bound
– Assume disk I/O speeds not a bottleneck
– In-memory removes the disk I/O bottleneck

For Analytics, the CPU is King
• The key metric of any analytical platform should be GB/CPU
– It needs to effectively utilize all available cores
– Hyper threads are NOT the equivalent of cores
• Interactive/adhoc analytics:
– THINK data to core ratios ≈ 10GB data per CPU core
• Every cycle is precious – CPU cores need to used efficiently
– Techniques such as “dynamic machine code generation”

Careful – performance impact of compression:

Makes disk-based databases go faster
Makes in-memory databases go slower

Speed & Scale are the Requirements
• Memory & CPU on an individual server = NOWHERE near enough for big data
– Moore’s Law – The power of a processor doubles every two years
– Data volumes – Double every year!!
• The only way to keep up is to parallelise or scale-out
• Combine the RAM of many individual servers
• many CPU cores spread across
Many • many CPUs, housed in
• many individual computers
– Data is split across all the CPU cores
– All database operations need to be parallelised with no points of
serialisation – This is true MPP

• Every CPU core in
Every • Every server needs to efficiently involved in
• Every query

Hadoop Connectivity
Kognitio - External Tables
– Data held on disk in other systems can be seen as non-memory
resident tables by Kognitio users.
– Users can select which data they wish to “suck” into memory.
• Using GUI or scripts
– Kognitio seamlessly sucks data out of the source system
into Kognitio memory.
– All managed via SQL
Kognitio - Hadoop Connectors
– Two types
• HDFS Connector
• Filter Agent Connector
– Designed for high speed
• Multiple parallel load streams
• Demonstrable 14TB+/hour load rates

Tight Hadoop integration
HDFS Connector
• Connector defines access to hdfs file
system
• External table accesses row-based data
in hdfs
• Dynamic access or “pin” data into
memory
• Complete hdfs file is loaded into memory

Filter Agent Connector
• Connector uploads agent to Hadoop
nodes
• Query passes selections and relevant
predicates to agent
• Data filtering and projection takes
place locally on each Hadoop node
• Only data of interest in loaded into
memory via parallel load streams

Not Only SQL
Kognitio V8 External Scripts
– Run third party scripts embedded within SQL
• Perl, Python, Java, R, SAS, etc.
• One-to-many rows in, zero-to-many rows out, one to one
create interpreter perlinterp
command '/usr/bin/perl' sends 'csv' receives 'csv' ;
select top 1000 words, count(*) This reads long comments
from (external script using environment perlinterp text from customer enquiry
receives (txt varchar(32000))
sends (words varchar(100)) table, in line perl converts
script S'endofperl( long text into output
while(<>)
{ stream of words (one word
chomp(); per row), query selects top
s/[,.!_]//g;
foreach $c (split(/ /)) 1000 words by frequency
{ if($c =~ /^[a-zA-Z]+$/) { print "$cn”} } using standard SQL
}
)endofperl' aggregation
from (select comments from customer_enquiry))dt
group by 1
order by 2 desc;

Hardware Requirements for
In-memory Platforms
• Hadoop = industry standard servers
• Careful to avoid vendor lock-in
• Off the shelf, low cost, servers match
neatly with Hadoop
– Intel or AMD CPU (x86)
– No special components
• Ethernet network
• Standard OS

Benefits of an In-memory Analytical Platform
• A seamless in-memory analytical layer on top of your data
persistence layer(s):
Analytical queries that used to run in hours and minutes, now
run in minutes and seconds (often sub-second)

High query throughput = massively higher concurrency

Flexibility
• Enables greater query complexity
• Users freely interact with data
• Use preferred BI Tools (relational or OLAP)

Reduced complexity
• Administration de-skilled
• Reduced data duplication

The Learning from our Journey

• Big Data tools are here and ready for the Enterprise
• An Enterprise Data Architecture model is essential
• Hadoop can handle Enterprise workload
 To reduce strain on legacy platforms
 To reduce cost
 To bring new business opportunities
• Must be part of an overall data strategy
• Not to be underestimated
• The solution must be an Eco-System
 There has to be a simple way to consume the data

Page 26

connect contact
www.kognitio.com Michael Hiskey
Vice President
kognitio.com/blog Marketing & Business Development
Michael.hiskey@kognitio.com
twitter.com/kognitio
Phone: +1 (855) KOGNITIO
linkedin.com/companies/kognitio
Dr. Phil Shelley
facebook.com/kognitio CEO – MetaScale
CTO Sears Holdings

youtube.com/user/kognitio

Upcoming Web Briefings: kognitio.com/briefings

Meta scale kognitio hadoop webinar

Related slideshows

More Related Content

What's hot

What's hot (20)

Similar to Meta scale kognitio hadoop webinar

Similar to Meta scale kognitio hadoop webinar (20)

Meta scale kognitio hadoop webinar