1597 EDW Big Data Analytics Kimball

8/12/2019 1597 EDW Big Data Analytics Kimball

1/33

The Evolving Role of theEnterprise Data Warehouse inthe Era of Big Data nalytics

AKimballGroupWhitePaper

ByRalphKimball


2/33

Table of ContentsExecutiveSummary......................................................................................................1

AbouttheAuthor...........................................................................................................1

Introduction..................................................................................................................2

Dataisanassetonthebalancesheet.....................................................................3

Raisingthecurtainonbigdataanalytics.....................................................................4

Usecasesforbigdataanalytics...............................................................................4

Makingsenseofbigdataanalyticusecases...........................................................7

Bigdataanalyticssystemrequirements......................................................................9

Extendedrelationaldatabasemanagementsystems............................................10

MapReduce/Hadoopsystems................................................................................13

HowMapReduceworksinHadoop........................................................................14

ToolsfortheHadoopenvironment.........................................................................16

Featureconvergenceinthecomingdecade..........................................................18

Reusableanalytics.................................................................................................20

Complexeventprocessing(CEP)..........................................................................20

Datawarehouseculturalchangesinthecomingdecade..........................................21

Sandboxes..............................................................................................................21

Lowlatency.............................................................................................................22

Continuousthirstformoreexquisitedetail.............................................................22

Lighttouchdatawaitsforitsrelevancetobeexposed ...........................................23

Simpleanalysisofallthedatatrumpssophisticatedanalysisofsomeofthedata.23

Datastructuresshouldbedeclaredatquerytime,notatdataloadtime ...............24

TheEDWsupportingbigdataanalyticsmustbemagnetic,agile,anddeep .........24

Theconflictbetweenabstractionandcontrol.........................................................24

Datawarehouseorganizationchangesinthecomingdecade...................................25

Technicalskillsetsrequired..................................................................................25

Neworganizationsrequired....................................................................................26

Newdevelopmentparadigmsrequired...................................................................27

Lessonsfromtheearlydatawarehousingera.......................................................28

Analyticsinthecloud..............................................................................................29

WhitherEDW?.........................................................................................................29

Acknowledgements....................................................................................................31

References.................................................................................................................31

TheEvolvingRoleoftheEnterpriseDataWarehouseintheEraofBigDataAnalytics


3/33

TheEvolvingRoleoftheEnterpriseDataWarehouseintheEraofBigDataAnalytics1

Executive SummaryInthiswhitepaper,wedescribetherapidlyevolvinglandscapefordesigningan

enterprisedatawarehouse(EDW)tosupportbusinessanalyticsintheeraof"big

data.Wedescribethescopeandchallengesofbuildingandevolvingaverystable

andsuccessfulEDWarchitecturetomeetnewbusinessrequirements.These

includeextremeintegration,semi-andun-structureddatasources,petabytesofbehavioralandimagedataaccessedthroughMapReduce/Hadoopaswellas

massivelyparallelrelationaldatabases,andthenstructuringtheEDWtosupport

advancedanalytics.Thispaperprovidesdetailedguidancefordesigningand

administeringthenecessaryprocessesfordeployment.Thiswhitepaperhasbeen

writteninresponsetoalackofspecificguidanceintheindustryastohowtheEDW

needstorespondtothebigdataanalyticschallenge,andwhatnecessarydesign

elementsareneededtosupportthesenewrequirements.

About the AuthorRalphKimballfoundedtheKimballGroup.Sincethemid1980s,hehasbeenthedatawarehouse/businessintelligence(DW/BI)industrysthoughtleaderonthe

dimensionalapproachandtrainedmorethan10,000ITprofessionals.PriortoworkingatMetaphorandfoundingRedBrickSystems,Ralphco-inventedtheStarworkstationatXeroxsPaloAltoResearchCenter(PARC).RalphhashisPh.D.inElectricalEngineeringfromStanfordUniversity.TheKimballGroupisthesourcefordimensionalDW/BIconsultingandeducation,consistentwithourbest-sellingToolkitbookseries,DesignTips,andaward-winningarticles.Visitwww.kimballgroup.comformoreinformation.


4/33


IntroductionWhatisbigdata?Itsbignessisactuallynotthemostinterestingcharacteristic.Big

dataisstructured,semistructured,unstructured,andrawdatainmanydifferent

formats,insomecaseslookingtotallydifferentthanthecleanscalarnumbersandtext

wehavestoredinourdatawarehousesforthelast30years.Muchbigdatacannotbe

analyzedwithanythingthatlookslikeSQL.Butmostimportant,bigdataisaparadigmshiftinhowwethinkaboutdataassets,wheredowecollectthem,howdo

weanalyzethem,andhowdowemonetizetheinsightsfromtheanalysis.Thebig

datarevolutionisaboutfindingnewvaluewithinandoutsideconventionaldata

sources.Anadditionalapproachisneededbecausethesoftwareandhardware

environmentsofthepasthavenotbeenabletocapture,manage,orprocessthenew

formsofdatawithinreasonabledevelopmenttimesorprocessingtimes.Weare

challengedtoreorganizeourinformationmanagementlandscapetoextenda

remarkablystableandsuccessfulEDWarchitecturetothisneweraofbigdata

analytics.

Inreadingthiswhitepaperpleasebearinmindthattheconsistentviewofthisauthorhasalwaysbeenthatthe"datawarehouse"comprisesthecompleteecosystemfor

extracting,cleaning,integratinganddeliveringdatatodecisionmakers,andtherefore

includestheextract-transform-load(ETL)andbusinessintelligence(BI)functions

consideredasoutsideofthedatawarehousebymoreconservativewriters.This

authorhasalwaystakentheviewthatdatawarehousinghasaverycomprehensive

roleincapturingallformsofenterprisedata,andthenpreparingthatdataforthemost

effectiveusebydecision-makersallacrosstheenterprise.Thiswhitepapertakesthe

aggressiveviewthattheenterprisedatawarehouseisonthevergeofaveryexciting

newsetofresponsibilities.ThescopeoftheEDWwillincreasedramatically.

Also,inthiswhitepaper,althoughweconsistentlyusethetermETLtodescribethemovementofdatawithintheenterprisedatawarehouse,theconventionaluseofthis

termdoesnotdojusticetothemuchlargerresponsibilityofmovingdataacross

networksandbetweensystemsandbetweenprofoundlydifferentprocessesinthe

worldofbigdataanalytics.ETLisaportionofamuchlargertechnologycalleddata

integration(DI).SincewehaveusedETLconsistentlyinourbooksandclassesfor

manyyears,wewillkeepthatterminologyinthispaper,bearinginmindthatETLis

meantinthelargersenseofDI.

Thiswhitepaperstandsbackfromthemarketplaceasitexistsinearly2011to

highlighttheclearlyemergingnewtrendsbroughtbythebigdatarevolution.Anda

revolutionitis.AsJamesMarkarian,Informatica'sExecutiveVicePresidentandChiefTechnologyOfficer,remarked:"thedatabasemarkethasfinallygotteninteresting

again."Becausemuchofthenewbigdatatoolsandapproachesareversion1or

evenversion0developments,thelandscapewillcontinuetochangerapidly.However

thereisgrowingawarenessinthemarketplacethatnewkindsofanalysisarepossible

andthatkeycompetitors,especiallye-commerceenterprises,arealreadytaking

advantageofthenewparadigm.Thiswhitepaperisintendedtobeaguidetohelp

businessintelligence,datawarehousingandinformationmanagementprofessionals


5/33


andmanagementteamsunderstandandprepareforbigdataasacomplementary

extensiontotheircurrentEDWarchitecture.

Dataisanassetonthebalancesheet

Enterprisesincreasinglyrecognizethatdataitselfisanassetthatshouldappearon

thebalancesheetinthesamewaythattraditionalassetsfromthemanufacturingage

suchasequipmentandlandhavealwaysappeared.Thereareseveralwaysto

determinethevalueofthedataasset,including

costtoproducethedata

costtoreplacethedataifitislost

revenueorprofitopportunityprovidedbythedata

revenueorprofitlossifdatafallsintocompetitorshands

legalexposurefromfinesandlawsuitsifdataisexposedtothewrongparties

Butmoreimportantthanthedataitself,enterpriseshaveshownthatinsightsfrom

datacanbemonetized.Whenane-commercesitedetectsanincreaseinfavorableclickthroughsfromanexperimentaladtreatment,thatinsightcanbetakentothe

bottomlineimmediately.Thisdirectcause-and-effectiseasilyunderstoodby

management,andananalyticresearchgroupthatconsistentlydemonstratesthese

insightsislookeduponasastrategicresourcefortheenterprisebythehighestlevels

ofmanagement.Thisgrowthinbusinessawarenessofthevalueofdata-driven

insightsisrapidlyspreadingoutwardfromthee-commerceworldtovirtuallyevery

businesssegment.

Datawarehousing,ofcourse,hasbeendemonstratingthevalueofdata-driven

insightsforatleast20years.Butuntilquiterecentlydatawarehousinghasbeen

focusedonhistoricaltransactiondata.Duringthepastdecadefrom2000to2009,threemajorseismicshiftsoccurredindatawarehousing.Thefirst,earlyinthe

decade,wasthedecisiveintroductionoflowlatencyoperationaldataintothedata

warehousetogetherwiththeexistinghistoricaldata.Ofcourse,manyofthesenew

operationaldatausecasesbenefitedfromreal-timedata,insomecasesdemanding

instantaneousdelivery.Thesecondseismicshiftgrowingincreasinglythroughoutthe

decadewasthegatheringofcustomerbehaviordata,whichnotonlyincluded

traditionaltransactionssuchaspurchasesandclickthroughsbutaddedhuge

volumesof"subtransactions"thatrepresentedmeasurableeventsleadinguptothe

transactionsthemselves.Forexample,allthewebpageeventsacustomerengaged

inpriortothefinaltransactioneventbecamearecordofcustomerbehavior."Good

paths"throughthesewebpageeventhistoriesgavelotsofinsightintoproductive(i.e.,

monetizable)customerbehavior.

Thethirdseismicevent,whichisgatheringenormousmomentumaswetransitioninto

thecurrentdecade,istheextractionofproductpreferencesandcustomers

sentimentsfromsocialmedia,especiallythemassivequantitiesofmachine-

generatedunstructureddatageneratedbythenewbusinessparadigmsofdot-com

companies.Itisthisfinalseismicshiftthathaspushedmanyenterprisesintolooking

seriouslyatunstructureddataforthefirsttime,andasking"howonearthdowe


6/33


analyzethisstuff?"Thepointhereisnotthatunstructureddataissomenewthing

recentlydiscovered,butrathertheanalysisofunstructureddatahasgone

mainstreamjustrecently.

Raising the curtain on big data analyticsUsecasesforbigdataanalytics

Bigdataanalyticsusecasesarespreadinglikewildfire.Hereisasetofusecases

reportedrecently,includingabenchmarksetof"Hadoop-able"usecasesproposed

byJeffHammerbacher,ChiefScientistforCloudera.Followingthesebrief

descriptionsisatablesummarizingthesalientstructureandprocessing

characteristicsofeachusecase.Notethatnoneoftheseusecasescanbesatisfied

withscalarnumericdata,norcananybeproperlyanalyzedbysimpleSQL

statements.Allofthemcanbescaledintothepetabyterangeandbeyondwith

appropriatebusinessassumptions.

Search ranking.Allsearchenginesattempttoranktherelevanceofawebpagetoasearchrequestagainstallotherpossiblewebpages.Googlespagerankalgorithmis,

ofcourse,theposterchildforthisusecase.

Ad tracking. E-commercesitestypicallyrecordanenormousriverofdataincludingeverypageeventineveryusersession.Thisallowsforveryshortturnaroundof

experimentsinadplacement,color,size,wording,andotherfeatures.Whenan

experimentshowsthatsuchafeaturechangeinanadresultsinimprovedclick

throughbehavior,thechangecanbeimplementedvirtuallyinrealtime. Location and proximity tracking. ManyusecasesaddpreciseGPSlocationtracking,togetherwithfrequentupdates,inoperationalapplications,securityanalysis,

navigation,andsocialmedia.Preciselocationtrackingopensthedoorforan

enormousoceanofdataaboutotherlocationsnearbytheGPSmeasurement.These

otherlocationsmayrepresentopportunitiesforsalesorservices.

Causal factor discovery.Point-of-saledatahaslongbeenabletoshowuswhenthesalesofaproductgoessharplyupordown.Butsearchingforthecausalfactorsthat

explainthesedeviationshasbeen,atbest,aguessinggameoranartform.The

answersmaybefoundincompetitivepricingdata,competitivepromotionaldata

includingprintandtelevisionmedia,weather,holidays,nationaleventsincluding

disasters,andvirallyspreadopinionsfoundinsocialmedia.Seethenextusecaseaswell.

Social CRM.Thisusecaseisoneofthehottestnewareasformarketinganalysis.TheAltimeterGrouphasdescribedaveryusefulsetofkeyperformanceindicatorsfor

socialCRMthatincludeshareofvoice,audienceengagement,conversationreach,

activeadvocates,advocateinfluence,advocacyimpact,resolutionrate,resolution

time,satisfactionscore,topictrends,sentimentratio,andideaimpact.Thecalculation

oftheseKPIsinvolvesin-depthtrollingofahugearrayofdatasources,especially

unstructuredsocialmedia.


7/33


Document similarity testing.Twodocumentscanbecomparedtoderiveametricofsimilarity.Thereisalargebodyofacademicresearchandtestedalgorithms,for

examplelatentsemanticanalysis,thatisjustnowfindingitswaytodrivingmonetized

insightsofinteresttobigdatapractitioners.Forexample,asinglesourcedocument

canbeusedasakindofmultifacetedtemplatetocompareagainstalargesetof

targetdocuments.Thiscouldbeusedforthreatdiscovery,sentimentanalysis,and

opinionpolls.Forexample:"findallthedocumentsthatagreewithmysourcedocumentonglobalwarming."

Genomics analysis: e.g., commercial seed gene sequencing.Afewmonthsagothecottonresearchcommunitywasthrilledbyagenomesequencingannouncementthat

statedinpart"Thesequencewillserveacriticalroleasthereferenceforfuture

assemblyofthelargercottoncropgenome.Cottonisthemostimportantfibercrop

worldwideandthissequenceinformationwillopenthewayformorerapidbreeding

forhigheryield,betterfiberqualityandadaptationtoenvironmentalstressesandfor

insectanddiseaseresistance.ScientistRyanRappstressedtheimportanceof

involvingthecottonresearchcommunityinanalyzingthesequence,identifyinggenes

andgenefamiliesanddeterminingthefuturedirectionsofresearch.(SeedQuest,Sept22,2010).Thisusecaseisjustoneexampleofawholeindustrythatisbeing

formedtoaddressgenomicsanalysisbroadly,beyondthisexampleofseedgene

sequencing.

Discovery of customer cohort groups.Customercohortgroupsareusedbymanyenterprisestoidentifycommondemographictrendsandbehaviorhistories.Weareall

familiarwithAmazon'scohortgroupswhentheysayothercustomerswhoboughtthe

samebookasyouhavealsoboughtthefollowingbooks.Ofcourse,ifyoucansell

yourproductorservicetoonememberofacohortgroup,thenalltherestmaybe

reasonableprospects.Cohortgroupsarerepresentedlogicallyandgraphicallyas

links,andmuchoftheanalysisofcohortgroupsinvolvesspecializedlinkanalysisalgorithms.

In-flight aircraft status.Thisusecaseaswellasthefollowingtwousecasesaremadepossiblebytheintroductionofsensortechnologyeverywhere.Inthecaseofaircraft

systems,in-flightstatusofhundredsofvariablesonengines,fuelsystems,hydraulics,

andelectricalsystemsaremeasuredandtransmittedeveryfewmilliseconds.The

valueofthisusecaseisnotjusttheengineeringtelemetrydatathatcouldbe

analyzedatsomefuturepointintime,butdrivesreal-timeadaptivecontrol,fuel

usage,partfailureprediction,andpilotnotification.

Smart utility meters.Itdidn'ttakelongforutilitycompaniestofigureoutthatasmartmetercanbeusedformorethanjustthemonthlyreadoutthatproducesthe

customersutilitybill.Bydrasticallycrankingupthefrequencyofthereadoutstoas

muchasonereadoutpersecondpermeteracrosstheentirecustomerlandscape,

manyusefulanalysescanbeperformedincludingdynamicload-balancing,failure

response,adaptivepricing,andlonger-termstrategiesforincentingcustomersto

utilizetheutilitymoreeffectively(eitherfromthecustomerspointofvieworthe

utility'spointofview!)


8/33


9/33


Data bag exploration.Therearemanysituationsincommercialenvironmentsandintheresearchcommunitieswherelargevolumesofrawdataarecollected.One

examplemightbedatacollectedaboutstructurefires.Beyondthepredictable

dimensionsoftime,place,primarycauseoffire,andrespondingfirefighters,there

maybeawealthofunpredictableanecdotaldatathatatbestcanbemodeledasa

disorderlycollectionofnamevaluepairs,suchas"contributingweather=lightning.

Anotherexamplewouldbethelistingofallrelevantfinancialassetsforadefendantinalawsuit.Againsuchalistislikelytobeadisorderlycollectionofnamevaluepairs,

suchas"sharedrealestateownership=condominium.Thelistofexampleslikethis

isendless.Whattheyhaveincommonistheneedtoencapsulatethedisorderly

collectionofnamevaluepairswhichisgenerallyknownasa"databag.Complex

databagsmaycontainbothnamevaluepairsaswellasembeddedsubdatabags.

Thechallengeinthisusecaseistofindacommonwaytoapproachtheanalysisof

databagswhenthecontentofthedatamayneedtobediscoveredafterthedatais

loaded.

Thefinaltwousecasesareoldandvenerableexamplesthatevenpredatedata

warehousingitself.Butnewlifehasbeenbreathedintotheseusecasesbecauseoftheexcitingpotentialofultra-atomiccustomerbehaviordata.

Loan risk analysis and insurance policy underwriting.Inordertoevaluatetheriskofaprospectiveloanoraprospectiveinsurancepolicy,manydatasourcescanbe

broughtintoplayrangingfrompaymenthistories,detailedcreditbehavior,

employmentdata,andfinancialassetdisclosures.Insomecasesthecollateralfora

loanortheinsureditemmaybeaccompaniedbyimagedata.

Customer churn analysis.Enterprisesconcernedwithchurnwanttounderstandthepredictivefactorsleadinguptothelossofacustomer,includingthatcustomers

detailedbehavioraswellasmanyexternalfactorsincludingtheeconomy,lifestageandotherdemographicsofthecustomer,andfinallyrealtimecompetitiveissues.

Makingsenseofbigdataanalyticusecases

Certainlythepurposeofdevelopingthislistofusecasesistoconvincethereader

thattheusecasescomeinallshapesandsizesandformats,andrequiremany

specializedapproachestoanalyze.Upuntilveryrecentlyalltheseusecasesexisted

asseparateendeavors,ofteninvolvingspecialpurposebuiltsystems.Buttheindustry

awarenessofthe"bigdataanalyticschallenge"ismotivatingeveryonetolookforthe

architecturalsimilaritiesanddifferencesacrossalltheseusecases.Anygiven

enterpriseisincreasinglylikelytoencounteroneormoreoftheseusecases.That

realizationisdrivingtheinterestinsystemarchitecturesthataddressesthebigdata

analyticsprobleminageneralway.Pleasestudythefollowingtable.


10/33


Thesheerdensityofthistablemakesitclearthatsystemstosupportbigdata

analyticshavetolookverydifferentthantheclassicrelationaldatabasesystemsfrom

the1980sand1990s.TheoriginalRDBMSswerenotbuilttohandleanyofthe

requirementsrepresentedascolumnsinthistable!

Searchranking X X X X X X

Adtracking X X X X X X X X

Location&proximity X X X X X

Causaldiscovery X X X X X X X

SocialCRM X X X X X X X X

Documentsimilarity X X X X X X X

Genomicanalysis X X X X X

Cohortgroups X X X X X X

In-flightenginestatus X X X X X X

Smartutilitymeters X X X X X X

Buildingsensors X X X X X X X X

Satelliteimages X X X X

CATscans X X X X X X

Financialfraud X X X X X X X X X

Hackingdetection X X X X X X X X X

Gamegestures X X X X X X X X

Bigscience X X X X X X X X X

Databagexploration X X X X X X

Riskanalysis X X X X X X X X

Churnanalysis X X X X X X X

Vector,

matrix

,or

complex

structure

Free

text

Image

orbinary

data

D

ata

bags

Iter

ative

logic

or

complex

branching

Advanced

ana

lytic

routines

Rap

idly

repeated

me

asurements

Extreme

low

latency

Access

to

all

data

required


11/33


Big data analytics system requirementsBeforediscussingtheexcitingnewtechnicalandarchitecturaldevelopmentsofthe

2010s,let'ssummarizetheoverallrequirementsforsupportingbigdataanalytics,

keepinginmindthatwearenotrequiringasinglesystemorasinglevendor's

technologytoprovideablanketsolutionforeveryusecase.Fromtheperspectiveof

2011,wehavetheluxuryofstandingbackfromalltheseusecasesgatheredinthelastfewyears,andwearenowinapositiontosurroundtherequirementswithsome

confidence.

Thedevelopmentofbigdataanalyticshasreachedapointwhereitneedsanoverall

missionstatementandidentityindependentofalistofusecases.Manyofushave

livedthroughearlierinstantiationsofadvancedanalyticsthatwentbythenamesof

advancedstatistics,artificialintelligenceanddatamining.Noneoftheseearlier

wavesbecameacoherentthemethattranscendedtheindividualexamples,as

compellingasthoseexampleswere.

Hereisanattempttostepbackanddefinethecharacteristicsofbigdataanalyticsatthehighestlevels.Inthefollowing,theterm"UDF"isusedinthebroadestsenseof

anyuserdefinedfunctionorprogramoralgorithmthatmayappearanywhereinthe

end-to-endanalysisarchitecture.

Inthecoming2010sdecade,theanalysisofbigdatawillrequireatechnologyor

combinationoftechnologiescapableof:

scalingtoeasilysupportpetabytes(thousandsofterabytes)ofdata

beingdistributedacrossthousandsofprocessors,potentiallygeographically

unaware,andpotentiallyheterogeneous

subsecondresponsetimeforhighlyconstrainedstandardSQLqueries embeddingarbitrarilycomplexuser-definedfunctions(UDFs)within

processingrequests

implementingUDFsinawidevarietyofindustry-standardprocedural

languages

assemblingextensivelibrariesofreusableUDFscrossingmostorallofuse

cases

executingUDFsas"relationscans"overpetabytesizeddatasetsinafew

minutes

supportingawidevarietyofdatatypesgrowingtoincludeimages,waveforms,

arbitrarilyhierarchicaldatastructures,anddatabags

loadingdatatobereadyforanalysis,atveryhighrates,atleastgigabytesper

second

integratingdatafrommultiplesourcesduringtheloadprocessatveryhigh

rates(GB/sec)

loadingdatabeforedeclaringordiscoveringitsstructure

executingcertainstreaminganalyticqueriesinrealtimeonincomingload

data


12/33


updatingdatainplaceatfullloadspeeds

joiningabillionrowdimensiontabletoatrillionrowfacttablewithoutpre-

clusteringthedimensiontablewiththefacttable

schedulingandexecutionofcomplexmulti-hundrednodeworkflows

beingconfiguredwithoutbeingsubjecttoasinglepointoffailure

failoverandprocesscontinuationwhenprocessingnodesfail supportingextrememixedworkloadsincludingthousandsofgeographically

dispersedon-lineusersandprogramsexecutingavarietyofrequestsranging

fromadhocqueriestostrategicanalysis,andwhileloadingdatainbatchand

streamingfashion

Twoarchitectureshaveemergedtoaddressbigdataanalytics:extendedRDBMS,

andMapReduce/Hadoop.Thesearchitecturesarebeingimplementedascompletely

separatesystemsandinvariousinterestinghybridcombinationsinvolvingboth

architectures.Wewillstartbydiscussingthearchitecturesseparately.

Extendedrelationaldatabasemanagementsystems

Allofthemajorrelationaldatabasemanagementsystemvendorsareaddingfeatures

toaddressbigdataanalyticsfromasolidrelationalperspective.Thetwomost

significantarchitecturaldevelopmentshavebeentheovertakingofthehighendofthe

marketwithmassivelyparallelprocessing(MPP),andthegrowingadoptionof

columnarstorage.WhenMPPandcolumnarstoragetechniquesarecombined,a

numberofthesystemrequirementsintheabovelistcanstarttobeaddressed,

including:

scalingtosupportexabytes(thousandsofpetabytes)ofdata

beingdistributedacrosstensofthousandsofgeographicallydispersedprocessors

subsecondresponsetimeforhighlyconstrainedstandardSQLqueries

updatingdatainplaceatfullloadspeeds

beingconfiguredwithoutbeingsubjecttoasinglepointoffailure

failoverandprocesscontinuationwhenprocessingnodesfail

Additionally,RDBMSvendorsareaddingsomecomplexuser-definedfunctions

(UDF's)totheirsyntax,butthekindofgeneralpurposeprocedurallanguage

computingrequiredbybigdataanalyticsisnotbeingsatisfiedinrelational

environmentsatthistime.

Inasimilarvein,RDBMSvendorsareallowingcomplexdatastructurestobestored

inindividualfields.Thesekindofembeddedcomplexdatastructureshavebeen

knownas"blobs"formanyyears.It'simportanttounderstandthatrelational

databaseshaveahardtimeprovidinggeneralsupportforinterpretingblobssince

blobsdonotfittherelationalparadigm.AnRDBMSindeedprovidessomevalueby

hostingtheblobsinastructuredframework,butmuchofthecomplexinterpretation

andcomputationontheblobsmustbedonewithspeciallycraftedUDFs,orBI


13/33


applicationlayerclients.Blobsarerelatedtodatabagsdiscussedelsewhereinthis

paper.SeethesectionentitledDatastructuresshouldbedeclaredatquerytime.

MPPimplementationshaveneversatisfactorilyaddressedthe"bigjoinissuewherea

billionrowdimensiontableisattemptedtobejoinedtoatrillionrowfacttablewithout

resortingtoclusteredstorage.Thebigjoincrisisoccurswhenanadhocconstraintis

placedagainstthedimensiontableresultinginapotentiallyverylargesetofdimensionkeysthatmustbephysicallydownloadedintoeveryoneofthephysicalsegmentsof

thetrillionrowfacttablestoredseparatelyintheMPPsystem.Sincethedimension

keysarescatteredrandomlyacrosstheseparatesegmentsofthetrillionrowfacttable,

itisveryhardtoavoidalengthydownloadstepoftheverylargedimensiontableto

everyoneofthefacttablestoragepartitions.Tobefair,theMapReduce/Hadoop

architecturehasnotbeenabletoaddressthebigjoinproblemeither.

Columnardatastoragefitstherelationalparadigm,andespeciallydimensionally

modeleddatabases,verywell.Besidesthesignificantadvantageofhighcompression

ofsparsedata,columnardatabasesallowaverylargenumberofcolumnscompared

torow-orienteddatabases,andplacelittleoverheadonthesystemwhencolumnsareaddedtoanexistingschema.ThemostsignificantAchilles'heel,atleastin2011,is

theslowloadingspeedofdataintothecolumnarformat.Althoughimpressiveload

speedimprovementsarebeingannouncedbycolumnardatabasevendors,theyhave

stillnotachievedthegigabytes-per-secondrequirementlistedabove.

ThestandardRDBMSarchitectureforimplementinganenterprisedatawarehouse

basedondimensionalmodelingprinciplesissimpleandwellunderstood,asshownin

Figure1.Recallthatthroughoutthiswhitepaper,theEDWisdefinedinthe

comprehensivesensetoincludeallbackroomandfrontroomprocessesincluding

ETL,datapresentation,andBIapplications.

Figure 1. The standard RDBMS based architecture for an enterprise data warehouseSource: The Data Warehouse Lifecycle Toolkit, 2ndedition, Kimball et al. (2008)


14/33


InthisstandardEDWarchitecturetheETLsystemisamajorcomponentthatsits

betweenthesourcesystemsandthepresentationserversthatareresponsiblefor

exposingalldatatobusinessintelligenceapplications.Inthisview,theETLsystem

addssignificantvaluebycleaning,conforming,andarrangingthedataintoaseriesof

dimensionalschemaswhicharethenstoredphysicallyinthepresentationserver.A

crucialelementofthisarchitectureisthepreparationofconformeddimensionsinthe

ETLsystemthatservesasthebasisofintegrationfortheBIapplications.Itisthestrongconvictionofthisauthorthatdeferringthebuildingofthedimensional

structuresandtheissuesofintegrationuntilquerytimeisthewrongarchitecture.

Sucha"deferredcomputation"approachrequiresanundulyexpensivequery

optimizertocorrectlyquerycomplexnon-dimensionalmodelseverytimeaqueryis

presented.Thecalculationofintegrationatqueryprocessingtimegenerallyrequires

complexapplicationlogicintheBItoolswhichalsomighthavetobeexecutedfor

everyquery.

TheextendedRDBMSarchitecturetosupportbigdataanalyticspreservesthe

standardarchitecturewithanumberofimportantadditions,shownbelowinFigure2

withlargearrows:

Figure 2. The extended RDBMS based architecture for an enterprise data warehouseThefactthatthehigh-levelenterprisedatawarehousearchitectureisnotmaterially

changedbytheintroductionofnewdatastructures,oragrowinglibraryofspecially

crafteduser-definedfunctions,orpowerfulprocedurallanguage-basedprograms

actingaspowerfulBIclients,isthecharmoftheextendedRDBMSapproachtobig

dataanalytics.ThemajorRDBMSplayersareabletomarshaltheirenormouslegacy

ofmillionsoflinesofcode,powerfulgovernancecapabilities,andsystemstabilitybuilt

overdecadesofservingthemarketplace.


15/33


However,itistheopinionofthisauthorthattheextendedRDBMSsystemscannotbe

theonlysolutionforbigdataanalytics.Atsomepoint,tackingonnon-relationaldata

structuresandnon-relationalprocessingalgorithmstothebasic,coherentRDBMS

architecturewillbecomeunwieldyandinefficient.TheSwissArmyknifeanalogy

comestomind.Anotheranalogyclosertothetopicistheprogramminglanguage

PL/1.Originallydesignedasanoverarching,multipurpose,powerfulprogramming

languageforallformsofdataandallapplications,itultimatelybecameabloatedandsprawlingcorpusthattriedtodotoomanythingsinasinglelanguage.Sincethe

heydayofPL/1therehasbeenawonderfulevolutionofmorenarrowlyfocused

programminglanguageswithmanynewconceptsandfeaturesthatsimplycouldn'tbe

tackedontoPL/1afteracertainpoint.Relationaldatabasemanagementsystemsdo

somanythingssowellthatthereisnodangerofsufferingthesamefateasPL/1.The

bigdataanalyticsspaceisgrowingsorapidlyandinsuchexcitingandunexpected

newdirectionsthatalighterweight,moreflexibleandmoreagileprocessing

frameworkinadditiontoRDBMSsystemsmaybeareasonablealternative.

MapReduce/Hadoopsystems

MapReduceisaprocessingframeworkoriginallydevelopedbyGoogleintheearly

2000sforperformingwebpagesearchesacrossthousandsofphysicallyseparated

machines.TheMapReduceapproachisextremelygeneral.CompleteMapReduce

systemscanbeimplementedinavarietyoflanguagesalthoughthemostsignificant

implementationisinJava.MapReduceisreallyaUDF(userdefinedfunction)

executionframework,wherethe"F"canbeextraordinarilycomplex.Originally

targetedtobuildingGoogle'swebpagesearchindex,aMapReducejobcanbe

definedforvirtuallyanydatastructureandanyapplication.Thetargetprocessorsthat

actuallyperformtherequestedcomputationcanbeidentical(a"cluster"),orcanbea

heterogeneousmixofprocessortypes(a"grid").Thedataineachprocessorupon

whichtheultimatecomputationisperformedcanbestoredinadatabase,ormorecommonlyinafilesystem,andcanbeinanydigitalformat.

ThemostsignificantimplementationofMapReduceisApacheHadoop,knownsimply

asHadoop.Hadoopisanopensource,top-levelApacheproject,withthousandsof

contributorsandawholeindustryofdiverseapplications.Hadooprunsnativelyonits

owndistributedfilesystem(HDFS)andcanalsoreadandwritetoAmazonS3and

others.Conventionaldatabasevendorsarealsoimplementinginterfacestoallow

Hadoopjobstoberunovermassivelydistributedinstancesoftheirdatabases.

AswewillseewhenwegiveabriefoverviewofhowaHadoopjobworks,bandwidth

betweentheseparateprocessorscanbeahugeissue.HDFSisaso-called"rackaware"filesystembecausethecentralnamenodeknowswhichnodesresideonthe

samerackandwhichareconnectedbymorethanonenetworkhop.Hadoopexploits

therelationshipbetweenthecentraljobdispatcherandHDFStosignificantlyoptimize

amassivelydistributedprocessingtaskbyhavingdetailedknowledgeofwheredata

actuallyresides.Thisalsoimpliesthatacriticalaspectofperformancecontrolisco-

locatingsegmentsofdataonactualphysicalhardwarerackssothattheMapReduce

communicationcanbeaccomplishedatbackplanespeedsratherthanslowernetwork

speeds.Notethatremotecloud-basedfilesystemssuchasAmazonS3and


16/33


CloudStoreare,bytheirnature,unabletoprovidetherackawarebenefit.Ofcourse,

cloud-basedfilesystemshaveanumberofcompellingadvantageswhichwe'll

discusslater.

HowMapReduceworksinHadoop

AMapReducejobissubmittedtoacentralizedJobTracker,whichinturnschedules

partsofthejobtoanumberofTaskTrackernodes.Although,ingenerala

TaskTrackermayfailanditstaskcanbereassignedbytheJobTracker,the

JobTrackerisasinglepointoffailure.IftheJobTrackerhalts,theMapReducejob

mustberestartedorberesumedfromintermediatesnapshots.

AMapReducejobisalwaysdividedintotwodistinctphases,mapandreduce.The

overallinputtoaMapReducejobisdividedintomanyequalsizedsplits,eachof

whichisassignedamaptask.Themapfunctionisthenappliedtoeachrecordin

eachsplit.Forlargejobs,thejobtrackerschedulesthesemaptasksinparallel.The

overallperformanceofaMapReducejobdependssignificantlyonachievinga

balanceofenoughparallelsplitstokeepmanymachinesbusy,butnotsomany

parallelsplitsthattheinterprocesscommunicationofmanagingallthesplitsbogs

downtheoveralljob.WhenMapReduceisrunovertheHDFSfilesystem,atypical

defaultsplitsizeis64MBofinputdata.

Asthenamesuggests,themaptaskisthefirsthalfoftheMapReducejob.Eachmap

taskproducesasetofintermediateresultrecordswhicharewrittentothelocaldiskof

themachineperformingthemaptask.ThesecondhalfoftheMapReducejob,the

reducetask,mayrunonanyprocessingnode.Theoutputsofthemappers(nodes

runningmaptasks)aresortedandpartitionedinsuchawaythattheseoutputscanbe

transferredtothereducers(nodesrunningthereducetask).Thefinaloutputsofthe

reducerscomprisethesortedandpartitionedresultssetoftheoverallMapReduce

job.InMapReducerunningoverHDFS,theresultssetiswrittentoHDFSandis

replicatedforreliability.

InFigure3,weshowthistaskflowforaMapReducejobwiththreemappernodes

feedingtworeducernodes,byreproducingfigure2.3fromTomWhite'sbook,

Hadoop,TheDefinitiveGuide,2ndEdition,(O'Reilly,2010).


17/33


Figure 3. An example MapReduce jobInTomWhite'sbook,asimpleMapReducejobisdescribedwhichweextend

somewhathere.Supposethattheoriginaldatabeforethesplitsareappliedconsists

ofaverylargenumber(perhapsbillions)ofunsortedtemperaturemeasurements,one

perrecord.Suchmeasurementscouldcomefrommanythousandsofautomatic

sensorslocatedaroundtheUnitedStates.Thesplitsareassignedtotheseparate

mappernodestoequalizeasmuchaspossiblethenumberofrecordsgoingtoeach

node.Theactualformofthemapperinputsarekey-valuepairs,inthiscasea

sequentialrecordidentifierandthefullrecordcontainingthetemperature

measurementsaswellasotherdata.Thejobofeachmapperissimplytoparsetherecordspresentedtoitandextracttheyear,thestate,andthetemperature,which

becomesthesecondsetofkey-valuepairspassedfromthemappertothereducer.

Thejobofeachreduceristofindthemaximumreportedtemperatureforeachstate,

andeachdistinctyearintherecordspassedtoit.Eachreducerisresponsiblefora

state,soinordertoaccomplishthetransfer,theoutputofeachmappermustbe

sortedsothatthekey-valuepairscanbedispatchedtotheappropriatereducers.In

thiscasetherewouldbe50reducers,oneforeachstate.Thesesortedblocksare

thentransferredtothereducersinastepwhichisacriticalfeatureoftheMapReduce

architecture,whereitiscalledthe"shuffle.

Noticethattheshuffleinvolvesatruephysicaltransferofdatabetweenprocessing

nodes.Thismakesthevalueoftherackawarefeaturemoreobvious,sincealotof

dataneedstobemovedfromthemapperstothereducers.Thecleverreadermay

wonderifthisdatatransfercouldbereducedbyhavingthemapperoutputscombined

sothatmanyreadingsfromasinglestateandyeararegiventothereducerasa

singlekey-valuepairratherthanmany.Theanswerisyes,andHadoopprovidesa

combinerfunctiontoaccomplishexactlythisend.


18/33


Eachreducerreceivesalargenumberofstate/year-temperaturekey-valuepairs,and

findsthemaximumtemperatureforagivenyear.Thesemaximumtemperaturesfor

eachyeararethefinaloutputfromeachreducer.

Thisapproachcanbescaledmoreorlessindefinitely.ReallyseriousMapReduce

jobsrunningonHDFSmayhavehundredsorthousandsofmappersandreducers,

processingpetabytesofinputdata.

AtthispointtheappealoftheMapReduce/Hadoopapproachshouldbeclear.There

arevirtuallynorestrictionsontheformoftheinputstotheoveralljob.Thereonly

needstobesomerationalbasisforcreatingsplitsandreadingrecords,inthiscase

therecordidentifierinTomWhite'sexample.Actuallogicinthemappersandthe

reducerscanbeprogrammedinvirtuallyanyprogramminglanguageandcanbeas

simpleastheaboveexample,ormuchmorecomplicatedUDFs.Thereadershould

beabletovisualizehowsomeofthemorecomplexusecases(e.g.,comparisonof

satelliteimages)describedearlierinthepapercouldfitintothisframework.

ToolsfortheHadoopenvironment

Whatwehavedescribedthusfaristhecoreprocessingcomponentwhen

MapReduceisrunintheHadoopenvironment.Thisisroughlyequivalentto

describingtheinnerprocessingloopinarelationaldatabasemanagementsystem.In

bothcasesthere'salotmoretothesesystemstoimplementacompletefunctioning

environment.ThefollowingisabriefoverviewoftypicaltoolsusedinaMapReduce/

Hadoopenvironment.Wegroupthesetoolsbyoverallfunction.TomWhite'sbook,

mentionedabove,isanexcellentstartingpointforunderstandinghowthesetoolsare

used.

Gettingdatainandgettingdataout

ETLplatforms--ETLplatforms,withtheirlonghistoryofimportingand

exportingdatatorelationaldatabases,providespecificinterfacesformoving

dataintoandoutofHDFS.Theplatform-basedapproach,ascontrastedwith

handcoding,providesextensivesupportformetadata,dataquality,

documentation,andavisualstyleofsystembuilding.

SqoopSqoop,developedbyCloudera,isanopensourcetoolthatallows

importingdatafromarelationalsourcetoHDFSandexportingdatafrom

HDFStoarelationaltarget.DataimportedbySqoopintoHDFScanbeused

bothbyMapReduceapplicationsandHBaseapplications.HBaseisdescribed

below.

ScribeScribe,developedatFacebookandreleasedasopensource,isused

toaggregatelogdatafromalargenumberofWebservers.

FlumeFlume,developedbyCloudera,isadistributedreliablestreamingdata

collectionservice.ItusesacentralconfigurationmanagedbyZookeeperand

supportstunablereliabilityandautomaticfailoverandrecovery.


19/33


Programming

Low-levelMapReduceprogramming--primarycodeformappersand

reducerscanbewritteninanumberoflanguages.Hadoop'snativelanguage

isJavabutHadoopexposesAPIsforwritingcodeinotherlanguagessuchas

RubyandPython.AninterfacetoC++isprovided,whichisnamedHadoop

Pipes.ProgrammingMapReduceatthelowestlevelobviouslyprovidesthemostpotentialpower,butthislevelofprogrammingisverymuchlike

assemblylanguageprogramming.Itcanbeverylaborious,especiallywhen

attemptingtodoconceptuallysimpletaskslikejoiningtwodatasets.

HighlevelMapReduceprogramming--ApachePig,orsimplyPig,isaclient-

sideopen-sourceapplicationprovidingahighlevelprogramminglanguagefor

processinglargedatasetsinMapReduce.Theprogramminglanguageitselfis

calledPigLatin.Hiveisanalternativeapplicationdesignedtolookmuchmore

likeSQL,andisusedfordatawarehousingusecases.Whenemployedforthe

appropriateusecases,PigandtheHiveprovideenormousprogramming

productivitybenefitsoverlow-levelMapReduceprogramming,oftenbya

factorof10ormore.PigandHivelifttheapplicationdeveloper sperspective

upfrommanagingthedetailedmapperandreducerprocessestomoreofan

applicationsfocus.

IntegrateddevelopmentenvironmentMapReduce/Hadoopdevelopment

needstomovedecisivelyawayfrombarehandcodingtobeadoptedby

mainstreamITshops.Anintegrateddevelopmentenvironmentfor

MapReduce/Hadoopneedstoincludeeditorsforsourcecode,compilers,tools

forautomatingsystembuilds,debuggers,andaversioncontrolsystem.

Integratedapplicationenvironmentanevenhigherlayeraboveanintegrated

developmentenvironmentcouldbecalledanintegratedapplication

environment,wherecomplexreusableanalyticroutinesareassembledintocompleteapplicationsviaagraphicaluserinterface.Thiskindofenvironment

mightbeabletouseopensourcealgorithmssuchasprovidedbytheApache

MahoutprojectwhichdistributesmachinelearningalgorithmsonHadoop

platform.

Cascading--Cascadingisanothertoolthatisanabstractionlayerforwriting

complexMapReduceapplications.ItisbestdescribedasathinJavalibrary

typicallyinvokedfromcommandlinetobeusedasaqueryAPIandprocess

scheduler.ItisnotintendedtobeacomprehensivealternativetoPigorHive.

HBase--HBaseisanopen-source,nonrelational,columnorienteddatabase

thatrunsdirectlyonHadoop.ItisnotaMapReduceimplementation.AprincipaldifferentiatorofHBasefromPigorHive(MapReduce

implementations)istheabilitytoprovidereal-timereadandwriterandom-

accesstoverylargedatasets.

Oozie--Oozieisaserver-basedworkflowenginespecializedinrunning

workflowjobswithactionsthatexecuteHadoopjobs,suchasMapReduce,

Pig,Hive,Sqoop,HDFSoperations,andsub-workflows .


20/33


ZooKeeperZooKeeperisacentralizedconfigurationmanagerfordistributed

applications.ZookeepercanbeusedindependentlyofHadoopaswell.

Administering

EmbeddedHadoopadminfeaturesHadoopsupportsacomprehensive

runtimeenvironmentincludingeditlog,safemodeoperation,auditlogging,

filesystemcheck,datanodeblockverifier,datanodeblockdistribution

balancer,performancemonitor,comprehensivelogfiles,metricsfor

administrators,countersforMapReduceusers,metadatabackup,data

backup,filesystembalancer,commissioninganddecommissioningnodes.

JavamanagementextensionsastandardJavaAPIformonitoringand

managingapplications

GangliaContextanopensourcedistributedmonitoringsystemforverylarge

clusters

Featureconvergenceinthecomingdecade

ItissafetosaythatrelationaldatabasemanagementsystemsandMapReduce/

Hadoopsystemswillincreasinglyfindwaystocoexistgracefullyinthecoming

decade.Butthesystemshavedistinctcharacteristics,asdepictedinthefollowing

table:

IntheupcomingdecadeRDBMSswillextendtheirsupportforhostingcomplexdata

typesas"blobs,andwillextendAPIsforarbitraryanalyticroutinestooperateonthe

contentsofrecords.MapReduce/Hadoopsystems,especiallyHive,willdeepentheir

supportforSQLinterfacesandfullersupportofthecompleteSQLlanguage.But

neitherwilltakeoverthemarketforbigdataanalyticsexclusively.Asremarked

earlier,RDBMSscannotprovide"relational"semanticsformanyofthecomplexuse

casesrequiredbybigdataanalytics.Atbest,RDBMSswillproviderelationalstructure

surroundingthecomplexpayloads.

RelationalDBMSs MapReduce/Hadoop

Proprietary,mostly Opensource

Expensive Lessexpensive

Datarequiresstructuring Datadoesnotrequirestructuring

Greatforspeedyindexedlookups Greatformassivefulldatascans

Deepsupportforrelationalsemantics Indirectsupportforrelationalsemantics,e.g.Hive

Indirectsupportforcomplexdatastructures Deepsupportforcomplexdatastructures

Indirectsupportforiteration,complexbranching

Deepsupportforiteration,complexbranching

Deepsupportfortransactionprocessing Littleornosupportfortransactionprocessing


21/33


Similarly,MapReduce/HadoopsystemswillnevertakeoverACID-compliant

transactionprocessing,orbecomesuperiortoRDBMSsforindexedqueriesonrow

andcolumnorientedtables.

Asthispaperisbeingwritten,significantadvancesarebeingmadeindeveloping

hybridsystemsusingbothrelationaldatabasetechnologyandMapReduce/Hadoop

technology.Figure4illustratestwoprimaryalternatives.ThefirstalternativedeliversthedatadirectlyintoaMapReduce/Hadoopconfigurationforprimarynon-relational

analysis.Aswehavedescribed,thisanalysiscanrangethefullgamutfromcomplex

analyticalroutinestosimplesortingthatlookslikeaconventionalETLstep.Whenthe

MapReduce/Hadoopstepiscomplete,theresultsareloadedintoanRDBMSfor

conventionalstructuredqueryingwithSQL.

ThesecondalternativeconfigurationloadsthedatadirectlytoanRDBMS,evenwhen

theprimarydatapayloadsarenotconventionalscalarmeasurements.Atthatpoint

twoanalysismodesarepossible.Thedatacanbeanalyzedwithspeciallycrafted

user-definedfunctions,effectivelyfromtheBIlayer,orpassedtoadownstream

MapReduce/Hadoopapplication.

Inthefutureevenmorecomplexcombinationswilltiethesearchitecturesmore

closelytogether,includingMapReducesystemswhosemappersandreducersare

actuallyrelationaldatabases,andrelationaldatabasesystemswhoseunderling

storageconsistsofHDFSfiles.

Figure 4. Alternative hybrid architectures using both RDBMS and Hadoop.ItwillprobablybedifficultforITorganizationstosortoutthevendorclaimswhichwill

almostcertainlyclaimthattheirsystemsdoeverything.Insomecasestheseclaims


22/33


are"objectionremovers"whichmeansthattheyareclaimsthathaveagrainoftruth

tothem,andaremadetomakeyoufeelgood,butdonotstanduptoscrutinyina

competitiveandpracticalenvironment.Buyerbeware!

Reusableanalytics

Uptothispointwehavebeggedtheissueofwheredoesallthespecialanalytic

softwarecomefrom.Bigdataanalyticswillneverprosperifeveryinstanceisa

customcodedsolution.BoththeRDBMSandtheopen-sourcecommunities

recognizethisandtwomaindevelopmentthemeshaveemerged.High-endstatistical

analysisvendors,suchasSAS,havedevelopedextensiveandproprietaryreusable

librariesforawiderangeofanalyticapplications,includingadvancedstatistics,data

mining,predictiveanalytics,featuredetection,linearmodels,discriminantanalysis,

andmanyothers.Theopensourcecommunityhasanumberofinitiatives,themost

notableofwhichareHadoop-MLandApacheMahout.QuotingfromHadoop- MLs

website:

Hadoop-ML(is)aninfrastructuretofacilitatetheimplementationofparallel

machinelearning/datamining(ML/DM)algorithmsonHadoop.Hadoop-ML

hasbeendesignedtoallowforthespecificationofbothtask-parallelanddata-

parallelML/DMalgorithms.Furthermore,itsupportsthecompositionof

parallelML/DMalgorithmsusingbothserialaswellasparallelbuildingblocks

--thisallowsonetowritereusableparallelcode.Theproposedabstraction

easestheimplementationprocessbyrequiringtheusertoonlyspecify

computationsandtheirdependencies,withoutworryingaboutscheduling,

datamanagement,andcommunication.Asaconsequence,thecodesare

portableinthattheuserneverneedstowriteHadoop-specificcode.This

potentiallyallowsonetoleveragefutureparallelizationplatformswithout

rewritingone'scode.

ApacheMahoutprovidesfreeimplementationsofmachinelearningalgorithmson

Hadoopplatform.

Complexeventprocessing(CEP)

Complexeventprocessing(CEP)consistsofprocessingeventshappeninginsideand

outsideanorganizationtoidentifymeaningfulpatternsinordertotakesubsequent

actioninrealtime.Forexample,CEPisusedinutilitynetworks(electrical,gasand

water)toidentifypossibleissuesbeforetheybecomedetrimental.TheseCEP

deploymentsallowforreal-timeinterventionforcriticalnetworkorinfrastructure

situations.ThecombinationofdeepDWanalyticsandCEPcanbeappliedinretail

customersettingstoanalyzebehaviorandidentifysituationswhereacompanymay

loseacustomerorbeabletosellthemadditionalproductsorservicesatthetimeof

theirdirectengagement.Inbanking,sophisticatedanalyticsmighthelptoidentifythe

10mostcommonpatternsoffraudandCEPcanthenbeusedtowatchforthose

patternssotheymaybethwartedbeforealoss.

Atthetimeofthiswhitepaper,CEPisnotgenerallythoughtofaspartoftheEDW,

butthisauthorbelievesthattechnicaladvancesincontinuousqueryprocessingwill


23/33


causeCEPandEDWtosharedataandworkmorecloselytogetherinthecoming

decade.

Data warehouse cultural changes in the coming decadeTheenterprisedatawarehousemustabsolutelystayrelevanttothebusiness.Asthe

valueandthevisibilityofbigdataanalyticsgrows,thedatawarehousemust

encompassthenewculture,skills,techniques,andsystemsrequiredforbigdata

analytics.

Sandboxes

Forexample,bigdataanalysisencouragesexploratorysandboxesfor

experimentation.Thesesandboxesarecopiesorsegmentsofthemassivedatasets

beingsourcedbytheorganization.Individualanalystsorverysmallgroupsare

encouragedtoanalyzethedatawithaverywidevarietyoftools,rangingfromserious

statisticaltoolslikeSAS,MatlaborR,topredictivemodels,andmanyformsofadhocqueryingandvisualizationthroughadvancedBIgraphicalinterfaces.Theanalyst

responsibleforagivensandboxisallowedtodoanythingwiththedata,usinganytool

theywant,evenifthetoolstheyusearenotcorporatestandards.Thesandbox

phenomenonhasenormousenergybutitcarriesasignificantrisktotheIT

organizationandEDWarchitecturebecauseitcouldcreateisolatedandincompatible

stovepipesofdata.Thispointisamplifiedinthesectiononorganizationalchanges,

below.

Exploratorysandboxesusuallyhavealimitedtimeduration,lastingweeksoratmost

afewmonths.Theirdatacanbeafrozensnapshot,orawindowonacertainsegment

ofincomingdata.Theanalystmayhavepermissiontorunanexperimentchangingafeatureontheproductorserviceinthemarketplace,andthenperformingA/Btesting

toseehowthechangeaffectscustomerbehavior.Typically,ifsuchanexperiment

producesasuccessfulresult,thesandboxexperimentisterminated,andthefeature

goesintoproduction.Atthatpoint,trackingapplicationsthatmayhavebeen

implementedinthesandboxusingaquickanddirtyprototypinglanguage,areusually

reimplementedbyotherpersonnelintheEDWenvironmentusingcorporatestandard

tools.Inseveralofthee-commerceenterprisesinterviewedforthiswhitepaper,

analyticsandboxeswereextremelyimportant,andinsomecaseshundredsofthe

sandboxexperimentswereongoingsimultaneously.Asoneintervieweecommented,

"newlydiscoveredpatternshavethemostdisruptivepotential,andinsightsfromthem

leadtothehighestreturnsoninvestment."

Architecturally,sandboxesshouldnotbebruteforcecopiesofentiredatasets,or

evenmajorsegmentsofthesedatasets.Indimensionalmodelingparlance,the

analystneedsmuchmorethanjustafacttabletoruntheexperiment.Ataminimum

theanalystalsoneedsoneormoreverylargedimensiontables,andpossibly

additionalfacttablesforcomplete"drillacross"analysis.If100analystsarecreating

bruteforcecopyversionsofthedataforthesandboxestherewillbeenormous

wastingofdiskspaceandresourcesforalltheredundantcopies.Rememberthatthe


24/33


largestdimensiontables,suchascustomerdimensions,canhave500millionrows!

Therecommendedarchitectureforaserioussandboxenvironmentistobuildeach

sandboxusingconformed(shared)dimensionswhichareincorporatedintoeach

sandboxasrelationalviews,ortheirequivalentunderHadoopapplications.

Lowlatency

Anelementarymistakewhengatheringbusinessrequirementsduringthedesignofa

datawarehouseistoaskthebusinessuseriftheywant"realtime"data.Usersare

likelytosay"ofcourse!"Althoughperhapsthisanswerhasbeensomewhatgratuitous

inthepast,agoodbusinesscasecannowbemadeinmanysituationsthatmore

frequentupdatesofdatadeliveredtothebusinesswithlowerandlowerlatenciesare

justified.BothRDBMSsandMapReduce/Hadoopsystemsstrugglewithloading

giganticamountsofdataandmakingthatdataavailablewithinsecondsofthatdata

beingcreated.Butthemarketplacewantsthis,andregardlessofatechnologists

doubtabouttherequirement,therequirementisrealandoverthenextdecadeitmust

beaddressed.

Aninterestingangleonlowlatencydataisthedesiretobeginseriousanalysisonthe

dataasitisstreamingin,butpossiblyfarbeforethedatacollectionprocesseven

terminates.Thereissignificantinterestinstreaminganalysissystemswhichallow

SQL-likequeriestoprocessthedataasitflowsintothesystem.Insomeusecases

whentheresultsofastreamingquerysurpassathreshold,theanalysiscanbehalted

withoutrunningthejobtothebitterend.Anacademiceffort,knownascontinuous

querylanguage(CQL),hasmadeimpressiveprogressindefiningtherequirements

forstreamingdataprocessingincludingcleversemanticsfordynamicallymovingtime

windowsonthestreamingdata.LookforCQLlanguageextensionsandstreaming

dataquerycapabilitiesintheloadprogramsforbothRDBMSsandHDFSdeployed

datasets.Anidealimplementationwouldallowstreamingdataanalysistotakeplacewhilethedataisbeingloadedatgigabytespersecond.

Theavailabilityofextremelyfrequentandextremelydetailedeventmeasurements

candriveinteractiveintervention.Theusecaseswherethisinterventionisimportant

spansmanysituationsrangingfromonlinegamingtoproductoffersuggestionsto

financialaccountfraudresponsestothestabilityofnetworks.

Continuousthirstformoreexquisitedetail

Analystsareforeverthirstingformoredetailineverymarketplaceobservation,

especiallyofcustomerbehavior.Forexampleeverywebpageevent(apagebeing

paintedonauser'sscreen)spawnshundredsofrecordsdescribingeveryobjecton

thepage.Inonlinegames,whereeverygestureentersthedatastream,asmanyas

100descriptorsareattachedtoeachofthesegesturemicro-events.Forinstance,ina

hypotheticalonlinebaseballgame,whenthebatterswingsatapitch,everything

describingthepositionoftheplayers,thescore,runnersonthebases,andeventhe

characteristicsofthepitch,areallstoredwiththatindividualrecord.Inbothofthese

examples,thecompletecontextmustbecapturedwithinthecurrentrecord,because

itisimpracticaltocomputethisdetailedcontextafterthefactfromseparatedata

sources.Thelessonforthecomingdecadeisthatthisthirstforexquisitedetailwill


25/33


onlygrow.Itispossibletoimaginethousandsofattributesbeingattachedtosome

micro-events,andthecategoriesandnamesoftheseattributeswillgrowin

unpredictableways.Thismakesthedatabagapproachdiscussedearlierinthepaper

muchmoreimportant.Itmeansthatpositionallydependentschemas,withthekeys

(namesofthedata)pre-declaredascolumnnamesisanunworkabledesign.

Finally,aperfecthistoricalreconstructionofinterestingeventssuchaswebpageexposuresneedstobemorethanjustalistofattributesonthewebpagewhenitwas

displayed,evenifthatlistisenormouslydetailed.Aperfecthistoricalreconstructionof

thewebpageneedstobeseenthroughamultimediauserinterface,i.e.,abrowser.

Lighttouchdatawaitsforitsrelevancetobeexposed

Lighttouchdataisanaspectoftheexquisitedetaildatadescribedintheprevious

section.Forexample,ifacustomerbrowsesawebsiteextensivelybeforemakinga

purchase,agreatdealofmicro-contextisstoredinallthewebpageeventspriortothe

purchase.Whenthepurchaseismade,someofthatmicro-contextsuddenly

becomesmuchmoreimportant,andiselevatedfrom"lighttouchdata"torealdata.At

thatpointthesequenceofexposurestotheselectedproductortocompetitive

productsinthesamespacebecomespossibletobesessionized.Thesemicro-events

areprettymuchmeaninglessbeforethepurchaseevent,becausetherearesomany

conceivableandirrelevantthreadsthatwouldbedeadendsforanalysis.This

requiresoceansoflighttouchdatatobestored,waitingfortherelevanceofselected

threadsofthesemicro-eventstoeventuallybeexposed.Conventionalseasonality

thinkingsuggeststhatatleastfivequarters(15months)ofthislighttouchdataneeds

tobekeptonline.Thisisoneinstanceofaremarkmadeconsistentlyduring

interviewsforthiswhitepaperthatanalystswant"longertails"whichmeansthatthey

wantmoresignificanthistoriesthantheycurrentlyget.

Simpleanalysisofallthedatatrumpssophisticatedanalysisofsomeofthedata

Althoughdatasamplinghasneverbeenapopulartechniqueindatawarehousing,

surprisinglythearrivalofenormouspetabytesizeddatasetshasnotincreasedthe

interestinanalyzingasubsetofthedata.Onthecontrary,anumberofanalystspoint

outthatmonetizableinsightscanbederivedfromverysmallpopulationsthatcouldbe

missedbyonlysamplingsomeofthedata.Ofcoursethisisasomewhatcontroversial

point,sincethesameanalystsadmitthatifyouhave1trillionbehaviorobservation

records,youmaybeabletofindanybehaviorpatternifyoulookhardenough.

Anothersomewhatcontroversialpointraisedbysomeanalystsistheirconcernthat

anyformofdatacleaningontheincomingdatacoulderaseinterestinglow-frequency

"edgecases.Ultimatelyboththecasesofmisleadingrarebehaviorpatterns,and

misleadingcorrupteddataneedtobegentlyfilteredoutofthedata.

Assumingthatthebehaviorinsightsfromverysmallpopulationsarevalid,thereis

widespreadrecognitionthatmicro-marketingtothesmallpopulationsispossible,and

doingenoughofthiscanbuildasustainablestrategicadvantage.


26/33


Afinalargumentinfavorofanalyzingcompletedatasetsisthatthese"relationscans"

donotrequireindexesoraggregationstobecomputedinadvanceoftheanalysis.

ThisapproachfitswellwiththebasicMapReducedistributedanalysisarchitecture.

Datastructuresshouldbedeclaredatquerytime,notatdataloadtime

Anumberofanalystsinterviewedforthiswhitepapersaidthattheenormousdata

setstheyweretryingtoanalyzeneededtobeloadedinaqueryablestatebeforethe

structureandcontentofthedatasetswerecompletelyunderstood.Again,thinkingof

thedatabagkindofmarketplaceobservationwherewithinawell-structured

dimensionalmeasurementprocesstheactualobservationisadisorderlyand

potentiallyunpredictablesetofkeyvaluepairs,thestructureofthisdatabagmay

needtobediscovered,andalternateinterpretationofthestructuresmayneedtobe

possiblewithoutreloadingthedatabase.Onerespondentremarkedthatyesterdays

fringedataistomorrowswell -structureddata,implyingthatweneedexceptional

flexibilityasweexplorenewkindsofdatasources.

AkeydifferentiatorbetweentheRDBMSapproachandtheMapReduce/Hadoop

approachisthedeferralofthedatastructuredeclarationuntilquerytimeinthe

MapReduce/Hadoopsystems.AnobjectionfromtheRDBMScommunitythatforcing

everyMapReducejobtodeclarethetargetdatastructurepromotesakindofchaos

becauseeveryanalystcandotheirownthing.Butthatobjectionseemstomissthe

pointthatastandarddatastructuredeclarationcaneasilybepublishedasalibrary

modulethatcanbepickedupbyeveryanalystwhentheyareimplementingtheir

application.

TheEDWsupportingbigdataanalyticsmustbemagnetic,agile,anddeep

CohenandDolanintheirseminalbutsomewhatcontroversialpaperonbigdata

analyticsarguethatEDWsmustshedsomeoldorthodoxiesinordertobemagnetic,agile,anddeep.Amagneticenvironmentplacestheleastimpedimentsonthe

incorporationofnew,unexpected,andpotentiallydirtydatasources.Specifically,this

supportstheneedtodeferdeclarationofdatastructuresuntilafterthedataisloaded.

AccordingtoCohenandDolan,anagileenvironmenteschewslong-rangecareful

designandplanning!Andadeepenvironmentallowsrunningsophisticatedanalytic

algorithmsonmassivedatasetswithoutsampling,orperhapsevencleaning.We

havemadethesepointselsewhereinthiswhitepaperbutCohenandDolanspaper

isaparticularlypotent,ifunusual,argument.Readthispapertogetsomeprovocative

perspectives!AlinktoCohenandDolanspaperisprovidedinthereferencessection

attheendofthiswhitepaper.

Theconflictbetweenabstractionandcontrol

IntheMapReduce/Hadoopworld,PigandHivearewidelyregardedasvaluable

abstractionsthatallowtheprogrammertofocusondatabasesemanticsratherthan

programmingdirectlyinJava.Butseveralanalystsinterviewedforthispaper

remarkedthattoomuchabstractionandtoomuchdistancingfromwherethedata

actuallyisstoredcanbedisastrouslyinefficient.Thisseemslikeareasonable

concernwhendealingwiththeverylargestdatasets,whereabadalgorithmcould


27/33


resultinruntimesmeasuredindays.Forthebreakingwaveofthebiggestdatasets,

programmingtoolswillneedtoallowconsiderablecontroloverthestoragestrategy,

andtheprocessingapproaches,butwithoutrequiringprogrammingusingthelowest

levelcode.

Data warehouse organization changes in the coming decadeThegrowingimportanceofbigdataanalyticsamountstosomethingbetweena

midcoursecorrectionandarevolutionforenterprisedatawarehousing.Newskillsets,

neworganizations,newdevelopmentparadigms,andnewtechnologywillneedtobe

absorbedbymanyenterprises,especiallythosefacingtheusecasesdescribedinthis

paper.Noteveryenterpriseneedstojumpintothepetabyteocean,butitisthis

author'spredictionthattheupcomingdecadewillseeasteadygrowthinthe

percentageoflargeenterprisesrecognizingthevalueofbigdataanalytics.

Mostobserverswouldagreethatbigdataanalyticsfallswithin"information

management,"butthesameobserversmayquibbleaboutwhetherthisaffectsthe

"datawarehouse."Ratherthanworryingaboutwhethertheboxontheorganization

chartlabeledEDWhasresponsibilityforbigdataanalytics,wetaketheperspective

thatenterprisedatawarehousingwithoutthecapitallettersabsolutelyencompasses

bigdataanalytics.Havingsaidthat,therewillbemanydifferentorganizational

structuresandmanagementperspectivesasindustriesexpandtheirinformation

management.Thiskindoftinkeringandadjustingtothenewparadigmisnormaland

expected.Wewentthroughaverysimilarphaseinthemid1980swhendata

warehousingitselfwasanewparadigmforITandthebusiness.Manyofthemost

successfulearlydatawarehousinginitiativesstartedinthebusinessorganizations

andwereeventuallyincorporatedintothoseITorganizationsthatthenmademajorcommitmentstobeingbusinessrelevant.Itislikelythesameevolutionwilltakeplace

withbigdataanalytics.

Thechallengebeforeinformationmanagersinlargeenterprisesishowtoencourage

threeseparatedatawarehouseendeavors:conventionalRDBMSapplications,

MapReduce/Hadoopapplications,andadvancedanalytics.

Technicalskillsetsrequired

Itisworthrepeatingherethemessageoftheveryfirstsentenceofthiswhitepaper.

Petabytescaledatasetsareofcourseabigchallengebutbigdataanalysisisoften

aboutdifficultiesotherthandatavolume.Youcanhavefastarrivingdataorcomplexdataorcomplexanalyseswhichareverychallengingevenifallyouhaveare

terabytesofdata!

ThecareandfeedingofRDBMS-orienteddatawarehousesinvolvesa

comprehensivesetofskillsthatisprettywellunderstood:SQLprogramming,ETL

platformexpertise,databasemodeling,taskscheduling,systembuildingand

maintenanceskills,oneormorescriptinglanguagessuchasPythonorPerl,UNIXor

Windowsoperatingsystemskills,andbusinessintelligencetoolsskills.SQL


28/33


programming,whichisatthecoreofanRDBMSimplementation,isadeclarative

language,whichcontrastswiththemindsetoftheprocedurallanguageskillsneeded

forMapReduce/Hadoopprogramming,atleastinJava.Thedatawarehouseteam

alsoneedstohaveagoodpartnershipwithinotherareasofITincludingstorage

management,security,networking,andsupportofmobiledevices.Finally,gooddata

warehousingalsorequiresanextensiveinvolvementwiththebusinesscommunity,

andwiththecognitivepsychologyofend-users!

ThecareandfeedingofMapReduce/Hadoopdatawarehouses,includinganyofthe

bigdataanalyticsusecasesdescribedinthispaper,involvesasetofskillsthatonly

partiallyoverlaptraditionalRDBMSdatawarehouseskills.Thereinliesasignificant

challenge.Thesenewskillsincludelower-levelprogramminglanguagessuchas

Java,C++,Ruby,Python,andMapReduceinterfacesmostcommonlyavailablevia

Java.Althoughtherequirementtoprogramviaproceduralbasedlower-level

programminglanguageswillbereducedsignificantlyduringtheupcomingdecadein

favorofPig,Hive,andHBase,itmaybeeasiertorecruitMapReduce/Hadoop

applicationdevelopersfromtheprogrammingcommunityratherthanthedata

warehousecommunity,ifthedatawarehousejobapplicantslackprogrammingandUNIXskills.IfMapReduce/Hadoopdatawarehousesaremanagedexclusivelywith

opensourcetools,thenZookeeperandOozieskillswillbeneededtoo.Keepinmind

thattheopen-sourcecommunityinnovatesquickly.Hive,PigandHBasearenotthe

lastwordinhigh-levelinterfacestoHadoopforanalysis.Itislikelythatwewillsee

muchmoreinnovationinthisdecadeincludingentirelynewinterfaces.

ETLplatformprovidershaveabigopportunitytoprovidemuchofthegluethatwilltie

togetherthebigdatasources,MapReduce/Hadoopapplications,andexisting

relationaldatabases.DeveloperswithETLplatformskillswillbeabletoleveragea

greatdealoftheirexperienceandinstinctsinsystembuildingwhentheyincorporate

MapReduce/Hadoopapplications.

Finally,theanalystswhomwehavedescribedasoftenworkinginsandbox

environmentswillarrivewithaneclecticandunpredictablesetofskillsstartingwith

deepanalyticexpertise.Forthesepeopleitisprobablymoreimportanttobe

conversantinSAS,Matlab,orRthantohavespecificprogramminglanguageor

operatingsystemskills.SuchindividualstypicallywillarrivewithUNIXskills,and

somereasonableprogrammingproficiency,andmostofthesepeopleareextremely

tolerantoflearningnewcomplextechnicalenvironments.Perhapsthebiggest

challengewithtraditionalanalystsisgettingthemtorelyontheotherresources

availabletothemwithinIT,ratherthanbuildingtheirownextractanddatadelivery

pipelines.Thisisatrickybalancebecauseyouwanttogivetheanalystsunusualfreedom,butyouneedtolookovertheirshoulderstomakesurethattheyarenot

wastingtheirtime.

Neworganizationsrequired

Atthisearlystageofthebigdataanalyticsrevolution,thereisnoquestionthatthe

analystsmustbepartofthebusinessorganization,bothtounderstandthe

microscopicworkingsofthebusiness,butalsotobeabletoconductthekindofrapid

turnaroundexperimentsandinvestigationswehavedescribedinthispaper.Aswe


29/33


havedescribed,theseanalystsmustbeheavilysupportedinatechnicalsense,with

potentiallymassivecomputepoweranddatatransferbandwidth.Soalthoughthe

analystsmayresideinthebusinessorganizations,thisisagreatopportunityforITto

gaincredibilityandpresencewiththebusiness.Itwouldbeasignificantmistakeand

alostopportunityfortheanalystsandtheirsandboxestoexistasroguetechnical

outpostsinthebusinessworldwithoutrecognizingandtakingadvantageoftheirdeep

dependenceonthetraditionalITworld.

Insomeorganizationsweinterviewedforthiswhitepaper,wesawseparateanalytic

groupsembeddedwithindifferentbusinessorganizations,butwithoutverymuch

crosscommunicationorcommonidentityestablishedamongtheanalyticgroups.In

somenoteworthycases,thislackofan"analyticcommunity"ledtolostopportunities

toleverageeachother'swork,andledtomultiplegroupsreinventingthesame

approaches,andduplicatingprogrammingeffortsandinfrastructuredemandsasthey

madeseparatecopiesofthesamedata.

Werecommendthatacrossdivisionalanalyticscommunitybeestablishedmimicking

someofthesuccessfuldatawarehousecommunitybuildingeffortswehaveseeninthepastdecade.Suchacommunityshouldhaveregularcrossdivisionalmeetings,as

wellasakindofprivateLinkedInapplicationtopromoteawarenessofallthecontacts

andperspectivesandresourcesthattheseindividualscollectintheirown

investigations,andaprivatewebportalwhereinformationandnewseventsare

shared.Periodictalkscanbegiven,hopefullyinvitingmembersofthebusiness

communityaswell,andabovealltheanalyticscommunityneedsT-shirtsandmugs!

Newdevelopmentparadigmsrequired

Evenbeforethearrivalofbigdataanalytics,datawarehousinghasbeentransforming

itselftoprovidemorerapidresponsetonewopportunitiesandtobemoreintouch

withthebusinesscommunity.Someofthepracticesoftheagilesoftware

developmentmovementhavebeensuccessfullyadoptedbythedatawarehouse

community,althoughrealisticallythishasnotbeenahighlyvisibletransformation.

But,inparticular,theagiledevelopmentapproachsupportsthedatawarehouseby

beingorganizedaroundsmallteamsdrivenbythebusiness,nottypicallybyIT.An

agiledevelopmenteffortalsoproducesfrequenttangibledeliveries,deemphasizes

documentationandformaldevelopmentmethodologies,andtoleratesmidcourse

correctionandtheincrementalacceptanceofnewrequirements.Themostsensitive

ingredientforsuccessofagiledevelopmentprojectsisthepersonalityandskillsofthe

businessleaderwhoultimatelyisincharge.Theagilebusinessleaderneedstobea

thoughtfulandsophisticatedobserverofthedevelopmentprocessandtherealitiesoftheinformationworld.Hopefullytheagilebusinessleaderisaprettygoodmanageras

well.

Bigdataanalyticscertainlyopensthedoortobusinessinvolvementsincethecentral

analysisisprobablydoneinthebusinessenvironmentdirectly.Butitisprobably

unlikelythattheprofessionalanalystistherightpersontobetheoverallagiledata

warehouseprojectleader.Theagileprojectleaderneedstobewellskilledin

facilitatingshorteffectivemeetings,resolvingissuesanddevelopmentchoices,


30/33


determiningthetruthofprogressreportsfromindividualdevelopers,communicating

withtherestoftheorganization,andgettingfundingforinitiatives.

Traditionaldatawarehousedevelopmenthasdiscoveredtheattractivenessof

buildingincrementallyfromamodeststart,butwithagoodarchitecturalfoundation

thatprovidesablueprintforwherefuturedevelopmentwillgo.Thisauthorhas

describedinmanypapersthetechniquesfor"gracefulmodification"ofdimensionaldatawarehouseschemas.Inadimensionallymodeleddatawarehouse,new

measurementfacts,newdimensionalattributes,andevennewdimensionscanbe

addedtoexistingdatawarehouseapplicationswithoutchanging,invalidating,or

rollingoverexistinginformationdeliverypipelinestotheendusers.Manyoftheuse

caseswehavedescribedinthispaperforbigdataanalyticssuggestthatnewfacts,

newattributes,andnewdimensionswillroutinelybecomeavailable.

Integrationofnewdatasourcesintoadatawarehousehasalwaysbeena

significantchallenge,sinceoftenthesenewdatasourcesarrivewithoutanythought

tointegrationwithexistingdatasources.Thiswillcertainlybethecasewithbigdata

analytics.Againfordimensionallymodeleddatawarehouses,thisauthorhasdescribedtechniquesforincrementalintegration,where"enterprisedimensional

attributes"aredefinedandplantedinthedimensionsoftheseparatedatasources.

Wecalltheseconformeddimensions.Thedevelopmentanddeploymentof

conformeddimensionsfitstheagiledevelopmentapproachbeautifully,sincethis

kindofintegrationcanbeimplementedonedatasourceatatime,andone

dimensionalattributeatatime,againinawaythatisnondestructivetoexisting

applications.Pleaseseethereferencessectionattheendofthiswhitepaperfor

moreinformationonconformeddimensions.

Finally,atleastoneorganizationinterviewedforthiswhitepaperhastakenagilityto

itslogicalextreme.Individualdevelopersaregivencompleteend-to-endresponsibilityforaproject,allthewayfromoriginalsourcingofthedata,through

experimentalanalysis,re-implementingtheprojectforproductionuse,andworking

withtheend-usersandtheirBItoolsinsupportivemode.Althoughthisdevelopment

approachremainsanexperiment,earlyresultsareveryinterestingbecausethese

developersfeelasignificantsenseofresponsibilityandpridefortheirprojects.

Lessonsfromtheearlydatawarehousingera

Ittookmostofthe1990sfororganizationstounderstandwhatadatawarehouse

wasandhowtobuildandmanagethosekindsofsystems.Interestingly,attheend

ofthe1990s,datawarehousingwaseffectivelyrelabeledasbusinessintelligence.

Thiswasaverypositivedevelopmentbecauseitreflectedtheneedforthebusiness

toownandtakeresponsibilityfortheusesofdata.

Theearliestdatawarehousepioneershadnochoicebuttodotheirownsystems

integration,assemblingbest-of-breedcomponents,andcopingwiththeinevitable

incompatibilitiesinissuesofdealingwithmultiplevendors.Bytheendofthe1990s,

thebestofbreedapproachgavewaytovendorstacksofintegratedproducts,a

trendwhichcontinuesuntiltoday.Atthispoint,thereareonlyafewindependent

vendorsinthedatawarehousespace,andthosevendorshavesucceededby


31/33


interfacingwithnearlyeveryconceivableformatandinterface,therebyproviding

bridgesbetweenthemorelimitedproprietaryvendorstacks.

Withthebenefitofhindsightgainedfromthetraditionaldatawarehouseexperience,

thebigdataanalyticsversionofdatawarehousingislikelytoconsolidatequite

quickly.Onlythebravestorganizationswithverystrongsoftwaredevelopmentskills

shouldconsiderrollingtheirownbigdataanalyticsapplicationsdirectlyonrawMapReduce/Hadoop.Forinformationmanagementorganizationswishingtofocuson

thebusinessissuesratherthanonthebreakingwaveofsoftwaredevelopment,a

packagedHadoopdistribution(e.g.,Cloudera)makesalotofsense.TheleadingETL

platformvendorslikelywillalsointroducepackagedenvironmentsforhandlingmany

ofthephasesofMapReduce/Hadoopdevelopment.

Analyticsinthecloud

Thiswhitepaperhasnotdiscussedcloudimplementationsofbigdataanalytics.Most

oftheenterprisesinterviewedforthiswhitepaperwerenotusingpubliccloud

implementationsfortheirproductionanalytics.Nevertheless,cloudimplementations

maybeveryattractiveinthestartupphaseforananalyticseffort.Acloudservicecan

provideinstantscalabilityduringthisstartupphase,withoutcommittingtoamassive

legacyinvestmentinhardware.Dataanalysisprojectscanbeturnedonandturned

offonshortnotice.Recallthattypicalanalyticenvironmentsmayinvolvehundredsof

separatesandboxesandparallelexperiments.

Manyoftheorganizationsinterviewedforthispaperstatedthatmatureanalytics

shouldbebroughtin-house,perhapsimplementedtechnicallyasacloudbutwithin

theconfinesoftheorganization.Ofcourse,suchanin-housecloudmayreducefears

ofsecurityandprivacybreaches(fairlyornot).

Aremotecloudimplementationraisesissuesofnetworkbandwidth,especiallyinabroadlyintegratedapplicationwithmultipleverylargedatasetsindifferentlocations.

Imaginesolvingthebigjoinproblemwhereyourtrillionrowfacttableisoutonthe

cloud,andyourbillionrowdimensiontableislocatedin-house.

Althoughthebestperformingsystemstrytoachieveathree-waybalanceamong

CPU,diskspeed,andbandwidth,mostorganizationsinterviewedforthispaper

predictedthatbandwidthwouldemergeasthenumberonelimitingfactorforbigdata

analyticssystemperformance.

WhitherEDW?

Theenterprisedatawarehousemustexpandtoencompassbigdataanalyticsaspart

ofoverallinformationmanagement.Themissionofthedatawarehousehasalways

beentocollectthedataassetsoftheorganizationandstructuretheminawaythatis

mostusefultodecision-makers.Althoughsomeorganizationsmaypersistwithabox

ontheorgchartlabeledEDWthatisrestrictedtotraditionalreportingactivitieson

transactionaldata,thescopeoftheEDWshouldgrowtoreflectthesenewbigdata

developments.InsomesensethereareonlytwofunctionsofIT:gettingthedatain

(transactionprocessing),andgettingthedataout.TheEDWisgettingthedataout.


32/33


Thebigchoicefacingshopswithgrowingbigdataanalyticsinvestmentsiswhetherto

chooseanRDBMS-onlysolution,oradualRDBMSandMapReduce/Hadoop

solution.Thisauthorpredictsthatthedualsolutionwilldominate,andinmanycases

thetwoarchitectureswillnotexistasseparateislandsbutratherwillhaverichdata

pipelinesgoinginbothdirections.Itissafetosaythatbotharchitectureswillevolve

hugelyoverthenextdecade,butthisauthorpredictsthatbotharchitectureswillshare

thebigdataanalyticsmarketplaceattheendofthedecade.

Sometimeswhenanexcitingnewtechnologyarrives,thereisatendencytoclosethe

dooronoldertechnologiesasiftheyweregoingtogoaway.Datawarehousinghas

builtanenormouslegacyofexperience,bestpractices,supportingstructures,

technicalexpertise,andcredibilitywiththebusinessworld.Thiswillbethefoundation

forinformationmanagementintheupcomingdecadeasdatawarehousingexpands

toincludebigdataanalytics.


33/33

AcknowledgmentsThisauthorisgratefulforInformaticassponsoringofthiswhitepaperandfor

providingabsolutelyno"vendorbias.Theopinionsinthiswhitepaperaresolelythe

responsibilityoftheauthor.

Anumberofsmartandknowledgeablebigdatapractitionersmadethemselvesavailableduringtheresearchphaseofthewhitepaperforinterviews.These

individualsprovidedmanyusefulinsights.Inalphabeticorderbyorganization,we

thank

AmrAwadallah,MikeOlson,Cloudera

BrianDolan,Discovix

OliverRatzesberger,eBay

AlexIgnatius,ElectronicArts

WilliamSchmarzo,EMC

AshishThusoo,Facebook

JuliannaDeLua,JohnHaddad,SanjayKrishnamurthi,RonLunasin,Informatica

NicholasWakefield,LinkedIn

DanGraham,DilipKrishna,RonKunze,Teradata

ProfessorMichaelFranklin,ComputerScienceDepartment,U.C.Berkeley

RaymieStata,Yahoo!

DanMcCaffrey,KenRudin,Zynga

ReferencesAnArchitectureforDataQuality,aKimballGroupWhitepaper:http://

vip.informatica.com/?elqPURLPage=8784

EssentialStepsfortheIntegratedEDW,aKimballGroupWhitepaper:http://

vip.informatica.com/?elqPURLPage=8785

Hadoop,TheDefinitiveGuide, 2ndEdition,TomWhite,OReilly (2011)

Hadoop-MLwebsite:http://videolectures.net/nipsworkshops09_pednault_hmli/

MADSkills:NewAnalysisPracticesforBigData,Cohen,Dolanetal,http://

db.cs.berkeley.edu/jmh/papers/madskills-032009.pdf