Intro to Hadoop

Presented by: Jon Bloom
Senior Consultant, Agile Bay, Inc.

Jon Bloom
Blog: http://www.bloomconsultingbi.com
Twitter: @sqljon
Email: jbloom@AgileBay.com
Certification: Microsoft Certified Solutions Associate MCSA
SQL Server 2012
Customers & Partners

w w w . a g i l e b a y. c o m

Session Agenda
 What is Hadoop?
 Version
 1.0
 2.0
 Demo:

Hadoop
 Apache Foundation
 Open Source
 Batch Processing
 Parallel, Reliable, Scalable
 Distributed Stores 3 copies
 Commodity Hardware
 Large Unstructured Data Sets
 Eventually Consistent

What is Hadoop
 Ecosystem
 Comprised of multiple Projects
• MapReduce
• Hive
• Pig
• Scoop
• Oozie
• Flume
• ZooKeeper
• Tez
• Mahout
• HBase
• Ambari
• Impala

Hadoop v1.0
 2004 Yahoo
 Doug Cutting (Cloudera)
• MapReduce
• Written 100% in Java
• Mappers
• Splits Rows into Chunks
• Reducers
• Aggregates the Chunks
• HDFS
• Distributed File System
• Java code is complex

Reason for Hadoop
 Data gets ingested into HDFS
 Java MapReduce Jobs run
 Parse out the Data
 Creates Output files
 Jobs can be re-run against Output files
 Run algorithms
 Handle Large, Complex Data Sets
 Look for “Insights”
 Raw Data (CSV, TXT, Binary, XML)

Name Nodes
 The “Brains” of Hadoop
 “Master” Server
 Single Point of Failure

Data Nodes
 “Slaves”
 Commodity Hardware
 Stores the Actual Data
 Runs Java Jar Files

Secondary Named
Nodes
 Redundant Server
 Serves as Backup
 Typically on own Server

Job Tracker
 Keeps track of “Jobs” Resources
 Heartbeat

Task Tracker
 Keeps track of “Task” Resources
 Communicates with Job Tracker

Ingest Data
 When thinking about Hadoop, we think of
data. How to get data into HDFS and how
to get data out of HDFS. Luckily, Hadoop
has some popular processes to accomplish
this.

SQOOP
 SQOOP was created to move data back and forth
easily from an External Database or flat file into
HDFS or HIVE. There are some standard commands
for moving data by Importing and Exporting
data. When data is moved to HDFS, it creates files
on the HDFS folder system. Those folders can be
partitioned in a variety of ways. Data can be
appended to the files through SQOOP jobs. And
you can add a WHERE clause to pull just certain
data, for example, just bring in data from yesterday,
run the SQOOP job daily to populate Hadoop.

Hive
 Once data gets moved to Hadoop HDFS, you
can add a layer of HIVE on top which
structures the data into relational
format. Once applied, the data can be queried
by HIVE SQL. If creating a table, in the HIVE
database schema, you can create an External
table which is basically a metadata layer pass
through which points to the actual data. So if
you drop the External table, the data remains
in tact.

PIG
 In addition, you can use a Hadoop language
called PIG (not making this up), to massage
the data into a structure series of steps, a
form of ETL.

MapReduce
 HIVE and PIG allow easier access to the data
 However, they still get translated to M/R

ODBC
 From HIVE SQL, the tables are exposed to
ODBC to allow data to be accessed via
Reports, Databases, ETL, etc.
So as you can see from the basic description
above, if you can move data back and forth
easily between Hadoop and your Relational
Database (or flat files).

Connect to Data
 Once data is stored in HDW, it can be
consumed by users via HIVE ODBC or
Microsoft PowerBI, Tableau, Qlikview or
SAP HANA or a variety of other tools sitting
on top of the data layer, including Self
Service tools.

HCatalog
 Sometimes when developing, users don't know
where data is stored. And sometimes the data
can be stored in a variety of formats, because
HIVE, PIG and Map Reduce can have separate
data model types. So HCatalog was created to
alleviate some of the frustration. It's a table
abstraction layer, meta data service and a
shared schema for Pig, Hive and M/R. It
exposes info about the data to applications.

HBase
 Hbase allows a separate database to allow
random read/write access to the HDFS data,
and surprisingly it too sits with the HDFS
cluster. Data can be ingested to HBASE and
interpreted On Read, which Relational
Databases do not offer.

Accumulo
 A High performance Data Storage and
retrieval system with cell-level access
control, similar to Google’s “Big Table”
design.

OOZIE
 A Java Web application used to schedule
Hadoop jobs. Combines multiple jobs
sequentially into one logical unit of work.

Flume
 Distributed, reliable and available service for
efficiently collection, aggregating and
moving large amounts of streaming data
into HDFS (fault tolerant).

Solr
 Open Source platform for searches of data
stored in HDFS Hadoop including full text
search and near real time indexing.

Streaming
 And you can receive Steaming Data.

HUE
 Open Source Web Interface
 Aggregates most common components into
single web interface
 View HDFS File Structure
 Simplify user experience

WebHDFS
 A REST API
 Interface to expose complete File System
 Provides Read & Write access
 Supports all HDFS parameters
 Allows remote access via many languages
 Uses Kerbos for Authentication

Monitor
 There's Zookeeper which is a centralized
service to keep track of things. A high
performance coordination service for
distributed applications.

Machine Learning
 In addition, you could apply MAHOUT
Machine Learning algorithms to you
Hadoop cluster for Clustering, Classification
and Collaborative Filtering. And you can
run Statistical language analysis with a
language called Revolution Analytic R
version of Hadoop R.

Machine Learning
 Clustering
 Similarities between data points in Clusters
 Classification
 Learns from existing categories to assign
unassigned categories
 User Based Recommendations
 Predict future behavior based on user
preferences and behavior

Hadoop 2.0
 And with the latest Hadoop 2.0, there's the addition
of YARN which is a new layer that sits between
HDFS2 and the application layers. Although HDFS
Map Reduce was originally designed as the sole
batch oriented approach to getting data from HDFS,
it's no longer the sole way. HIVE SQL has been sped
up through Impala which completely bypasses Map
Reduce and the Stinger initiative which sits atop
Tez. Tez has ability to compress data with column
stores which allows the interaction to be sped up.

YARN
 Allows the separation of MapReduce layers
of Service and Framework
 Resource Manager
 Application Manager
 Node Manager
 Containers
 Separates Resources

YARN
 Traditional MapReduce
 Expensive
 Original M/R spawned many process
 Wrote to Disk intermediate data
 Sort / Shuffle
 Now we have Applications
 M/R, Tez, Giraff, Spark, Storm, etc.
 Compiled down to a lower level
 Single Strand w/ More Complexity

Tez
 Generalized data flow programming
framework, built on Hadoop YARN for batch
and interactive use cases, such as Pig, HIVE
and other frameworks. It has the potential
to replace the MapReduce execution engine.

Impala
 Cloudera Impala is runs massively parallel
processing (MPP) SQL query engine that
runs natively in Hadoop.
 Allows data querying without the need for
data movement or transformation
 It by-passes MapReduce

Graph
 And Girage, which allows Hadoop the ability
to process Graph connections between
nodes.

Ambari
 Ambari allows Hadoop Cluster
administration and has an API layer for 3rd
party tools to hook into.

Spark
 And Spark, provides a simple and expressive
programming model that supports ETL,
Machine Learning, stream processing and
graph computation.

Knox
 Provides a single point of authentication
and access to Hadoop services. Specifically
for Hadoop users who access the cluster data
and execute jobs, operators who control
access and manage the cluster.

Falcon
 Framework for simplifying data management
and pipeline processing in Hadoop. Enables
users to automate the movement and
processing of datasets for ingest, pipelines,
disaster recovery and data retention use cases.
It simplifies data management by removing
complex coding (out of the box).

More Apache
Projects
 Apache Kafka
 Next Generation Distributed Messaging
System
 Apache Avro
 Data Serialization System
 Apache Chukwa
 Data Collection System for Monitoring large
distributed systems

Cloud
 You can run your Hybrid Data Warehouse in
the Cloud with Microsoft Azure Blobstorage
HDInsight or Amazon Web Services.

On Premise
 You can run On Premise with IBM
Infosphere BigInsights, Cloudera,
Hortonworks and MapR.

Hybrid Data
Warehouse
 You can build a Hybrid Data Warehouse. As
Data Warehousing is a concept, a
documented framework to follow with
guidelines and rules. And storing the data
in Hadoop and Relational Databases is
typically known as a Hybrid Data
Warehouse.

BI vs. Hadoop
 Hadoop not a replacement of BI
 Extends BI capabilities
 BI = Scale up to 100s of Gigabytes
 Hadoop = From 100s of Gygabytes to Terabytes
(1,000s og Gygabytes) and Terabytes (1,000,000
Gigabytes)

Where’s Hadoop
Headed?
 Transactional Data?
 More Real Time?
 Integrate with Traditional Data Warehouses?
 Hadoop for the Masses?
 Artificial Intelligence?
 Turing Test
 Neural Networks
 Internet of Things

Blog: www.bloomconsultingbi.com
Twitter: @sqljon
Linked-in: http://www.linkedin.com/in/BloomConsultingBI
Email: jbloom@agilebay.com

Intro to Hadoop

Recommended

More Related Content

What's hot (19)

Viewers also liked (7)

Similar to Intro to Hadoop (20)

More from Jonathan Bloom (7)

Recently uploaded (20)

Intro to Hadoop