Hadoop introduction

Training (Day 1)

Introduction

Big-dataFour parameters:

Velocity: Streaming data and large volume data movement.Volume: Scale from terabytes to zettabytes.Variety: Manage the complexity of multiple relational and non-relational data types and schemas.Voracity: Produced data has to be consumed fast before it becomes meaningless.

Not just internet companiesBig Data Shouldnt Be a SiloMust be an integrated part of enterprise information architecture

Data >> Information >> Business ValueRetail By combining data feeds on inventory, transactional records, social media and online trends, retailers can make real-time decisions around product and inventory mix adjustments, marketing and promotions, pricing or product quality issues.

Financial Services By combining data across various groups and services like financial markets, money manager and lending, financial services companies can gain a comprehensive view of their individual customers and markets.

Government By collecting and analyzing data across agencies, location and employee groups, the government can reduce redundancies, identify productivity gaps, pinpoint consumer issues and find procurement savings across agencies.

Healthcare Big data in healthcare could be used help improve hospital operations, provide better tracking of procedural outcomes, help accelerate research and grow knowledge bases more rapidly. According to a 2011 Cleveland Clinic study, leveraging big data is estimated to drive $300B in annual value through an 8% systematic cost savings.

Single-core, single processorSingle-core, multi-processor

Single-core

Multi-core, single processorMulti-core, multi-processorMulti-core

Cluster of processors (single or multi-core) with shared memoryCluster of processors with distributed memory

Cluster

Broadly, the approach in HPC is to distribute the work across a cluster of machines, which access a shared file-system, hosted by a SAN.

Grid of clusters

Embarrassingly parallel processing

MapReduce, distributed file system

Cloud computing

Pipelined Instruction level

Concurrent Thread level

Service Object level

Indexed File level

Mega Block level

Virtual System Level

Data size: small

Data size: large

Reference: Bina Ramamurthy 2011

Processing Granularity

How to Process BigData?

Need to process large datasets (>100TB)Just reading 100TB of data can be overwhelmingTakes ~11 days to read on a standard computerTakes a day across a 10Gbit link (very high end storage solution)On a single node (@50MB/s) 23daysOn a 1000 node cluster 33min

ExamplesWeb logs;RFID; sensor networks; social networks; social data (due to the social data revolution), Internet text and documents; Internet search indexing; call detail records; astronomy, atmospheric science, genomics, biogeochemical, biological, and other complex and/or interdisciplinary scientific research; military surveillance; medical records; photography archives; video archives; and large-scale e-commerce.

Not so easy

Moving data from storage cluster to computation cluster is not feasible

In large clustersFailure is expected, rather than exceptional. In large clusters, computers fail every dayData is corrupted or lostComputations are disruptedThe number of nodes in a cluster may not be constant. Nodes can be heterogeneous.

Very expensive to build reliability into each applicationA programmer worries about errors, data motion, communicationTraditional debugging and performance tools dont apply

Need a common infrastructure and standard set of tools to handle this complexityEfficient, scalable, fault-tolerant and easy to use

Why is Hadoop and MapReduce needed?

The answer to this questions comes from another trend in disk drives: seek time is improving more slowly than transfer rate.

Seeking is the process of moving the disks head to a particular place on the disk to read or write data.

It characterizes the latency of a disk operation, whereas the transfer rate corresponds to a disks bandwidth.

If the data access pattern is dominated by seeks, it will take longer to read or write large portions of the dataset than streaming through it, which operates at the transfer rate.

Why is Hadoop and MapReduce needed?

On the other hand, for updating a small proportion of records in a database, a traditional B-Tree (the data structure used in relational databases, which is limited by the rate it can perform seeks) works well.

For updating the majority of a database, a B-Tree is less efficient than MapReduce, which uses Sort/Merge to rebuild the database.

MapReduce can be seen as a complement to an RDBMS.

MapReduce is a good fit for problems that need to analyze the whole dataset, in a batch fashion, particularly for ad hoc analysis.

Why is Hadoop and MapReduce needed?

Hadoop distributions

Apache HadoopApache Hadoop- based Services

for Windows Azure

Clouderas Distribution Including Apache Hadoop (CDH)

Hortonworks Data Platform

IBM InfoSphere BigInsights

Platform Symphony MapReduce

MapR Hadoop Distribution

EMC Greenplum MR (using MapRsM5 Distribution)

Zettaset Data Platform

SGI Hadoop Clusters (uses

Cloudera distribution)

Grand Logic JobServer

OceanSync Hadoop Management

Software

Oracle Big Data Appliance (uses

Cloudera distribution)

Whats up with the names?

When naming software projects, Doug Cutting seems to have been inspired by his family.

Lucene is his wifes middle name, and her maternal grandmothers first name.

His son, as a toddler, used Nutch as the all-purpose word for meal and later named a yellow stuffed elephant Hadoop.

Doug said he was looking for a name that wasnt already a web domain and wasnt trademarked, so I tried various words that were in my life but not used by anybody else. Kids are pretty good at making up words.

Hadoop features

Distributed Framework for processing and storing data generally on commodity hardware.

Completely Open Source.

Written in JavaRuns on Linux, Mac OS/X, Windows, and Solaris.Client apps can be written in various languages.

Scalable: store and process petabytes, scale by adding Hardware

Economical: 1000s of commodity machines

Efficient: run tasks where data is located

Reliable: data is replicated, failed tasks are rerun

Primarily used for batch data processing, not real-time / user facing applications

Components of Hadoop

HDFS (Hadoop Distributed File System)Modeled on GFSReliable, High Bandwidth file system that can

store TB' and PB's data.

Map-ReduceUsing Map/Reduce metaphor from Lisp languageA distributed processing framework paradigm that

process the data stored onto HDFS in key-value .

DFS

Processing Framework

Client 1 Client 2

Inputdata

Outputdata

Map

Map

Map

Reduce

Reduce

Input Map Shuffle & Sort Reduce Output

Very Large Distributed File System10K nodes, 100 million files, 10 PBLinearly scalableSupports Large files (in GBs or TBs)

EconomicalUses Commodity HardwareNodes fail every day. Failure is expected, rather than exceptional.The number of nodes in a cluster is not constant.

Optimized for Batch Processing

HDFS

HDFS Goals

Highly fault-tolerantruns on commodity HW, which can fail frequently

High throughput of data accessStreaming access to data

Large filesTypical file is gigabytes to terabytes in sizeSupport for tens of millions of files

Simple coherencyWrite-once-read-many access model

HDFS: Files and Blocks

Data OrganizationData is organized into files and directoriesFiles are divided into uniform sized large blocksTypically 128MBBlocks are distributed across cluster nodes

Fault ToleranceBlocks are replicated (default 3) to handle hardware failureReplication based on Rack-Awareness for performance and fault tolerance

Keeps checksums of data for corruption detection and recoveryClient reads both checksum and data from DataNode. If checksum fails, it tries other replicas

HDFS: Files and Blocks

High Throughput:Client talks to both NameNode and DataNodesData is not sent through the NameNode.Throughput of file system scales nearly linearly with the number of nodes.

HDFS exposes block placement so that computation can be migrated to data

HDFS Components

NameNodeManages the file namespace operation like opening, creating, renaming etc.File name to list blocks + location mappingFile metadata Authorization and authenticationCollect block reports from DataNodes on block locationsReplicate missing blocksKeeps ALL namespace in memory plus checkpoints & journal

DataNodeHandles block storage on multiple volumes and data integrity.Clients access the blocks directly from data nodes for read and writeData nodes periodically send block reports to NameNodeBlock creation, deletion and replication upon instruction from the NameNode.

name:/users/joeYahoo/myFile - blocks:{1,3}name:/users/bobYahoo/someData.gzip - blocks:{2,4,5}

Datanodes (the slaves)

Namenode (the master)

1 12

224 5

33 4 4

55

ClientMetadata

I/O

1

3

HDFS Architecture

Simple commandshdfs dfs -ls, -du, -rm, -rmr

Uploading fileshdfs dfs copyFromLocal foo mydata/foo

Downloading fileshdfs dfs - moveToLocal mydata/foo foo

hdfs dfs -cat mydata/foo

Adm