Evolutionary Database Design.docx

Evolutionary Database Design.docx

《Evolutionary Database Design.docx》由会员分享，可在线阅读，更多相关《Evolutionary Database Design.docx（9页珍藏版）》请在冰豆网上搜索。

EvolutionaryDatabaseDesign

EvolutionaryDatabaseDesign

Overthelastfewyearswe'vedevelopedanumberoftechniquesthatallowadatabasedesigntoevolveasanapplicationdevelops.Thisisaveryimportantcapabilityforagilemethodologies.Thetechniquesrelyonapplyingcontinuousintegrationandautomatedrefactoringtodatabasedevelopment,togetherwithaclosecollaborationbetweenDBAsandapplicationdevelopers.Thetechniquesworkinbothpre-productionandreleasedsystems.

January2003

MartinFowler

PramodSadalage

Translations:

JapaneseRussian

Contents

∙DealingwithChange

oLimitations

∙ThePractices

oDBAscollaboratecloselywithdevelopers

oEverybodygetstheirowndatabaseinstance

oDevelopersfrequentlyintegrateintoasharedmaster

oAdatabaseconsistsofschemaandtestdata

oAllchangesaredatabaserefactorings

oAutomatetherefactorings

oAutomaticallyUpdateallDatabaseDevelopers

oClearlyseparatealldatabaseaccesscode

∙Variations

oKeepingmultipledatabaselineages

oYoudon'tneedaDBA

∙ToolstoHelp

∙FurtherStepsandFurtherInformation

Inthelastfewyears,we'veseentheriseofanewbreedofsoftwaremethodologies,theagilemethodologies.Thesemakesomenewandsignificantdemandsondatabasedesign.Oneofthemostcentralofthesedemandsistheideaofevolutionarydesign.Onanagileprojectyouassumethatyoucannotfixtherequirementsofthesystemup-front.Asaresulthavingadetaileddesignphaseatthebeginningofaprojectbecomesimpractical.Thedesignofthesystemhastoevolvethroughthevariousiterationsofthesoftware.Agilemethods,inparticularextremeprogramming（XP）,haveanumberofpracticesthatmakethisevolutionarydesignpractical.

Manypeoplehavequestionedwhetherevolutionarydesigncanbeappliedtoasystemwithalargedatabasecomponent.Indeedmanypeopletoldusthatitwasimpossible-atroublingthoughtasThoughtWorksembarkedonalargedatabase-orientedprojectusingmanyagileandXPtechniques.

Thisarticledescribesthepracticesthatwe'veusedtoallowustodothisimpossiblething.Wewon'tsaythatwe'vecompletelysolvedthedatabaseevolutionproblem,butwedothinkwe'vedemonstratedasetoftechniquesthatmanypeoplewillfinduseful.

DealingwithChange

Oneoftheprimaryfeaturesofagilemethodsistheirattitudetowardschange.Mostofthethinkingaboutsoftwareprocessisaboutunderstandingrequirementsearly,signingoffontheserequirements,usingtherequirementsasabasisfordesign,signingoffonthat,andthenproceedingwithconstruction.Thisisaplan-drivencycle,oftenreferredto（usuallywithderision）asthewaterfallapproach

Suchapproacheslooktominimizechangesbydoingextensiveup-frontwork.Oncetheearlyworkisdone,changescausesignificantproblems.Asaresultsuchapproachesrunintotroubleifrequirementsarechanging,andrequirementschurnisabigproblemforsuchprocesses.

Agileprocessesapproachchangedifferently.Theyseektoembracechange,allowingchangestooccurevenlateinadevelopmentproject.Changesarecontrolled,buttheattitudeoftheprocessistoenablechangeasmuchaspossible.Partlythisisinresponsetotheinherentinstabilityofrequirementsinmanyprojects,partlyitistobettersupportdynamicbusinessenvironmentsbyhelpingthemchangewiththecompetitivepressures.

Inordertomakethiswork,youneedadifferentattitudetodesign.Insteadofthinkingofdesignasaphase,whichismostlycompletedbeforeyoubeginconstruction,youlookatdesignasanon-goingprocessthatisinterleavedwithconstruction,testing,andevendelivery.Thisisthecontrastbetweenplannedandevolutionarydesign.Oneofthevitalcontributionsofagilemethodsisthattheyhavecomeupwithpracticesthatallowevolutionarydesigntoworkinacontrolledmanner.Soinsteadofthecommonchaosthatoftenhappenswhendesignisn'tplannedup-front,thesemethodsprovidetechniquestocontrolevolutionarydesignandmakethempractical.

Animportantpartofthisapproachisiterativedevelopment,whereyouruntheentiresoftwarelife-cyclemanytimesduringthelifeofaproject.Agileprocessesruncompletelifecyclesineachiteration,completingtheiterationwithworking,tested,integratedcodeforasmallsubsetoftherequirementsofthefinalproduct.Theseiterationsareshort,usuallyrunningbetweenaweekandacoupleofmonths,withapreferencetowardsshorteriterations.

Whilethesetechniqueshavegrowninuseandinterest,oneofthebiggestquestionsishowtomakeevolutionarydesignworkfordatabases.Mostpeopleconsiderthatdatabasedesignissomethingthatabsolutelyneedsup-frontplanning.Changingthedatabaseschemalateinthedevelopmenttendstocausewide-spreadbreakagesinapplicationsoftware.Furthermorechangingaschemaafterdeploymentresultinpainfuldatamigrationproblems.

Overthecourseofthelastthreeyearswe'vebeeninvolvedinalargeproject（calledAtlas）thathasusedevolutionarydatabasedesignandmadeitwork.Theprojectinvolvedalmost100peopleinmultiplesitesworld-wide（US,Australia,andIndia）.Itisaroundhalfamillionlinesofcodeandhasover200tables.Thedatabaseevolvedduringayearandahalfofinitialdevelopmentandcontinuestoevolveeventhoughit'sinproductionformultiplecustomers.Duringthisprojectwestartedwithiterationsofamonth,butafterafewmonthschangedtotwoweekiterationswhichworkedbetter.Thetechniqueswedescribeherearetheonesthatwe（ormoreaccuratelyPramod）usedtomakethiswork.

Sincethatprojectgotgoingwe'vespreadthesetechniquesovermoreofourprojects,gainingmoreexperiencefrommorecases.We'vealsofoundinspiration,ideas,andexperiencefromotheragileprojects.

Limitations

Beforewediveintothetechniques,it'simportanttostatethatwehaven'tsolvedalltheproblemsofevolutionarydatabasedesign.Inparticular:

∙Wedevelopedanapplicationdatabaseforasingleapplicationratherthananintegrationdatabasethattriestointegratemultipledatabases.

∙Wedon'thavetokeeptheproductiondatabasesup24/7

Wedon'tconsidertheseproblemstobeinherentlyunsolvable,afterallmanypeoplebelievedwecouldn'tsolvethisone.Butuntilwedo,wewon'tclaimwecansolvethemeither.

ThePractices

Ourapproachtoevolutionarydatabasedesigndependsonahandfulofimportantpractices.

DBAscollaboratecloselywithdevelopers

Oneofthetenetsofagilemethodsisthatpeoplewithdifferentskillsandbackgroundsneedtocollaborateverycloselytogether.Theycan'tcommunicatemainlythroughformalmeetingsanddocuments.Insteadtheyneedtobeouttalkingwitheachotherandworkingwitheachotherallthetime.Everybodyisaffectedbythis:

analysts,PMs,domainexperts,developers...andDBAs.

EverytaskthatadeveloperworksonpotentiallyneedsaDBA'shelp.BoththedevelopersandtheDBAneedtoconsiderwhetheradevelopmenttaskisgoingtomakeasignificantchangetothedatabaseschema.IfsothedeveloperneedstoconsultwiththeDBAtodecidehowtomakethechange.Thedeveloperknowswhatnewfunctionalityisneeded,andtheDBAhasaglobalviewofthedataintheapplication.

TomakethishappentheDBAhastomakehimselfapproachableandavailable.Makeiteasyforadevelopertojustpopoverforafewminutesandasksomequestions.MakesuretheDBAsanddeveloperssitclosetoeachothersotheycaneasilygettogether.EnsurethatapplicationdesignsessionsareknownaboutsotheDBAcanpopineasily.InmanyenvironmentsweseepeopleerectingbarriersbetweentheDBAandapplicationdevelopmentfunctions.Thesebarriersmustcomedownforanevolutionarydatabasedesignprocesstowork.

Everybodygetstheirowndatabaseinstance

Evolutionarydesignrecognizesthatpeoplelearnbytryingthingsout.Inprogrammingtermsdevelopersexperimentwithhowtoimplementacertainfeatureandmaymakeafewattemptsbeforesettlingdowntoapreferredalternative.Databasedesigncanbelikethattoo.Asaresultit'simportantforeachdevelopertohavetheirownsandboxwheretheycanexperiment,andnothavetheirchangesaffectanyoneelse.

ManyDBAexpertsseemultipledatabasesasanathema,toodifficulttoworkinpractice,butwe'vefoundthatyoucaneasilymanageahundredorsodatabaseinstances.Thevitalthingistohavetotoolstoallowyoutomanipulatedatabasesmuchasyouwouldmanipulatefiles.

Developersfrequentlyintegrateintoasharedmaster

Althoughdeveloperscanexperimentfrequentlyintheirownarea,it'simportanttobringthedifferentapproachesbacktogetheragainfrequently.Anapplicationneedsasharedmasterdatabasethatallworkflowsfrom.Whenadeveloperbeginsatasktheycopythemasterintotheirownworkspace,manipulate,andthenintegratetheirchangesbackintothemaster.Asaruleofthumbeachdevelopershouldintegrateonceaday.

Let'stakeanexamplewhereMikestartsadevelopmenttaskat10am（assumingheactuallycomesinthatearly）.Aspartofthistaskheneedstochangethedatabaseschema.Ifthechangeiseasy,likeaddingacolumn,hejustdecideshowtomakethechangehimself,Mikealsomakessurethecolumnhewantstoadddoesnotalreadyexistinthedatabase,withthehelpofthedatadictionary（discussedlater）.Ifit'smorecomplicatedthenhegrabstheDBAandtalksoverthelikelychangeswithhim.

Oncehe'sreadytobeginhetakesacopyofthedatabasemasterandcanmodifyboththedatabaseschemaandcodefreely.Ashe'sinasandboxanychangeshemakesdon'timpactanyoneelse's.Atsomepoint,sayaround3pm,he'sprettycomfortablethatheknowswhatthedatabasechangeneedstobe,ev