数据库系统概念全套配套课件PPT ch26.ppt 53页

'Chapter26:AdvancedTransactionProcessing Chapter26:AdvancedTransactionProcessingTransaction-ProcessingMonitorsTransactionalWorkflowsHigh-PerformanceTransactionSystemsMainmemorydatabasesReal-TimeTransactionSystemsLong-DurationTransactionsTransactionmanagementinmultidatabasesystems TransactionProcessingMonitorsTPmonitorsinitiallydevelopedasmultithreadedserverstosupportlargenumbersofterminalsfromasingleprocess.Provideinfrastructureforbuildingandadministeringcomplextransactionprocessingsystemswithalargenumberofclientsandmultipleservers.Provideservicessuchas:PresentationfacilitiestosimplifycreatinguserinterfacesPersistentqueuingofclientrequestsandserverresponsesRoutingofclientmessagestoserversCoordinationoftwo-phasecommitwhentransactionsaccessmultipleservers.SomecommercialTPmonitors:CICSfromIBM,PathwayfromTandem,TopEndfromNCR,andEncinafromTransarc TPMonitorArchitectures TPMonitorArchitectures(Cont.)Processperclientmodel-insteadofindividualloginsessionperterminal,serverprocesscommunicateswiththeterminal,handlesauthentication,andexecutesactions.MemoryrequirementsarehighMultitasking-highCPUoverheadforcontextswitchingbetweenprocessesSingleprocessmodel-allremoteterminalsconnecttoasingleserverprocess.Usedinclient-serverenvironmentsServerprocessismulti-threaded;lowcostforthreadswitchingNoprotectionbetweenapplicationsNotsuitedforparallelordistributeddatabases TPMonitorArchitectures(Cont.)Many-serversingle-routermodel-multipleapplicationserverprocessesaccessacommondatabase;clientscommunicatewiththeapplicationthroughasinglecommunicationprocessthatroutesrequests.IndependentserverprocessesformultipleapplicationsMultithreadserverprocessRunonparallelordistributeddatabaseManyservermany-routermodel-multipleprocessescommunicatewithclients.Clientcommunicationprocessesinteractwithrouterprocessesthatroutetheirrequeststotheappropriateserver.Controllerprocessstartsupandsupervisesotherprocesses. DetailedStructureofaTPMonitor DetailedStructureofaTPMonitorQueuemanagerhandlesincomingmessagesSomequeuemanagersprovidepersistentordurablemessagequeueingcontentsofqueuearesafeevenifsystemsfails.Durablequeueingofoutgoingmessagesisimportantapplicationserverwritesmessagetodurablequeueaspartofatransactiononcethetransactioncommits,theTPmonitorguaranteesmessageiseventuallydelivered,regardlessofcrashes.ACIDpropertiesarethusprovidedevenformessagessentoutsidethedatabaseManyTPmonitorsprovidelocking,loggingandrecoveryservices,toenableapplicationserverstoimplementACIDpropertiesbythemselves. ApplicationCoordinationUsingTPMonitorsATPmonitortreatseachsubsystemasaresourcemanagerthatprovidestransactionalaccesstosomesetofresources.TheinterfacebetweentheTPmonitorandtheresourcemanagerisdefinedbyasetoftransactionprimitivesTheresourcemanagerinterfaceisdefinedbytheX/OpenDistributedTransactionProcessingstandard.TPmonitorsystemsprovideatransactionalremoteprocedurecall(transactionalRPC)interfacetotheirserviceTransactionalRPCprovidescallstoencloseaseriesofRPCcallswithinatransaction.UpdatesperformedbyanRPCarecarriedoutwithinthescopeofthetransaction,andcanberolledbackifthereisanyfailure. WorkflowSystems TransactionalWorkflowsWorkflowsareactivitiesthatinvolvethecoordinatedexecutionofmultipletasksperformedbydifferentprocessingentities.Withthegrowthofnetworks,andtheexistenceofmultipleautonomousdatabasesystems,workflowsprovideaconvenientwayofcarryingouttasksthatinvolvemultiplesystems.Exampleofaworkflowdeliveryofanemailmessage,whichgoesthroughseveralmailssystemstoreachdestination.Eachmailerperformsatasks:forwardingofthemailtothenextmailer.Ifamailercannotdelivermail,failuremustbehandledsemantically(deliveryfailuremessage).Workflowsusuallyinvolvehumans:e.g.loanprocessing,orpurchaseorderprocessing. ExamplesofWorkflows LoanProcessingWorkflowInthepast,workflowswerehandledbycreatingandforwardingpaperformsComputerizedworkflowsaimtoautomatemanyofthetasks.Butthehumansstillplayrolee.g.,inapprovingloans. TransactionalWorkflowsMustaddressfollowingissuestocomputerizeaworkflow.Specificationofworkflows-detailingthetasksthatmustbecarriedoutanddefiningtheexecutionrequirements.Executionofworkflows-executetransactionsspecifiedintheworkflowwhilealsoprovidingtraditionaldatabasesafeguardsrelatedtothecorrectnessofcomputations,dataintegrity,anddurability.E.g.:Loanapplicationshouldnotgetlostevenifsystemfails.Extendtransactionconceptstothecontextofworkflows.Stateofaworkflow-consistsofthecollectionofstatesofitsconstituenttasks,andthestates(i.e.,values)ofallvariablesintheexecutionplan. WorkflowSpecificationStaticspecificationoftaskcoordination:Tasksanddependenciesamongthemaredefinedbeforetheexecutionoftheworkflowstarts.Canestablishpreconditionsforexecutionofeachtask:tasksareexecutedonlywhentheirpreconditionsaresatisfied.Definedpreconditionsthroughdependencies:Executionstatesofothertasks.“taskticannotstartuntiltasktjhasended”Outputvaluesofothertasks.“taskticanstartiftasktjreturnsavaluegreaterthan25”Externalvariables,thataremodifiedbyexternalevents.“tasktimustbestartedwithin24hoursofthecompletionoftasktj” WorkflowSpecification(Cont.)DynamictaskcoordinationE.g.,Electronicmailroutingsysteminwhichthetexttobescheduleforagivenmailmessagedependsonthedestinationaddressandonwhichintermediateroutersarefunctioning. Failure-AutomicityRequirementsUsualACIDtransactionalrequirementsaretoostrong/unimplementableforworkflowapplications.However,workflowsmustsatisfysomelimitedtransactionalpropertiesthatguaranteeaprocessisnotleftinaninconsistentstate.Acceptableterminationstates-everyexecutionofaworkflowwillterminateinastatethatsatisfiesthefailure-atomicityrequirementsdefinedbythedesigner.Committed-objectivesofaworkflowhavebeenachieved.Aborted-validterminationstateinwhichaworkflowhasfailedtoachieveitsobjectives.Aworkflowmustreachanacceptableterminationstateeveninthepresenceofsystemfailures. ExecutionofWorkflowsWorkflowmanagementsystemsinclude:Scheduler-programthatprocessworkflowsbysubmittingvarioustasksforexecution,monitoringvariousevents,andevaluationconditionsrelatedtointertaskdependenciesTaskagents-controltheexecutionofataskbyaprocessingentity.Mechanismtoquerytostateoftheworkflowsystem. WorkflowManagementSystemArchitecturesCentralized-asingleschedulerschedulesthetasksforallconcurrentlyexecutingworkflows.usedinworkflowsystemswherethedataisstoredinacentraldatabase.easiertokeeptrackofthestateofaworkflow.Partiallydistributed-hasone(instanceofa)schedulerforeachworkflow.Fullydistributed-hasnoscheduler,butthetaskagentscoordinatetheirexecutionbycommunicatingwitheachothertosatisfytaskdependenciesandotherworkflowexecutionrequirements.usedinsimplestworkflowexecutionsystemsbasedonelectronicmail WorkflowSchedulerIdeallyschedulershouldexecuteaworkflowonlyafterensuringthatitwillterminateinanacceptablestate.ConsideraworkflowconsistingoftwotasksS1andS2.Letthefailure-atomicityrequirementbethateitherbothorneitherofthesubtransactionsshouldbecommitted.SupposesystemsexecutingS1andS2donotprovideprepared-to-commitstatesandS1orS2donothavecompensatingtransactions.Itisthenpossibletoreachastatewhereonesubtransactioniscommittedandtheotheraborted.Bothcannotthenbebroughttothesamestate.Workflowspecificationisunsafe,andshouldberejected.Determinationofsafetybytheschedulerisnotpossibleingeneral,andisusuallylefttothedesigneroftheworkflow. RecoveryofaWorkflowEnsurethatisafailureoccursinanyoftheworkflow-processingcomponents,theworkfloweventuallyreachesanacceptableterminationstate.Failure-recoveryroutinesneedtorestorethestateinformationofthescheduleratthetimeoffailure,includingtheinformationabouttheexecutionstatesofeachtask.Logstatusinformationonstablestorage.Handoffoftasksbetweenagentsshouldoccurexactlyonceinspiteoffailure.Problem:Repeatinghandoffonrecoverymayleadtoduplicateexecutionoftask;notrepeatinghandoffmayleadtotasknotbeingexecuted.Solution:Persistentmessagingsystems RecoveryofaWorkflow(Cont.)Persistentmessages:messagesarestoredinpermanentmessagequeueandthereforenotlostincaseoffailure.DescribedindetailinChapter19(DistributedDatabases)Beforeanagentcommits,itwritestothepersistentmessagequeuewhatevermessagesneedtobesentout.Thepersistentmessagesystemmustmakesurethemessagesgetdeliveredeventuallyifandonlyifthetransactioncommits.Themessagesystemneedstoresendamessagewhenthesiterecovers,ifthemessageisnotknowntohavereacheditsdestination.Messagesmustbeloggedinstablestorageatthereceivingendtodetectmultiplereceiptsofamessage. HighPerformanceTransactionSystems High-PerformanceTransactionSystemsHigh-performancehardwareandparallelismhelpimprovetherateoftransactionprocessing,butareinsufficienttoobtainhighperformance:DiskI/Oisabottleneck—I/Otime(10milliseconds)hasnodecreasedataratecomparabletotheincreaseinprocessorspeeds.Paralleltransactionsmayattempttoreadorwritethesamedataitem,resultingindataconflictsthatreduceeffectiveparallelismWecanreducethedegreetowhichadatabasesystemisdiskboundbyincreasingthesizeofthedatabasebuffer. Main-MemoryDatabaseCommercial64-bitsystemscansupportmainmemoriesoftensofgigabytes.Memoryresidentdataallowsfasterprocessingoftransactions.Disk-relatedlimitations:Loggingisabottleneckwhentransactionrateishigh.Usegroup-committoreducenumberofoutputoperations(Willstudytwoslidesahead.)Iftheupdaterateformodifiedbufferblocksishigh,thediskdata-transferratecouldbecomeabottleneck.Ifthesystemcrashes,allofmainmemoryislost. Main-MemoryDatabaseOptimizationsToreducespaceoverheads,main-memorydatabasescanusestructureswithpointerscrossingmultiplepages.Indiskdatabases,theI/Ocosttotraversemultiplepageswouldbeexcessivelyhigh.Noneedtopinbufferpagesinmemorybeforedataareaccessed,sincebufferpageswillneverbereplaced.Designquery-processingtechniquestominimizespaceoverhead-avoidexceedingmainmemorylimitsduringqueryevaluation.Improveimplementationofoperationssuchaslockingandlatching,sotheydonotbecomebottlenecks.Optimizerecoveryalgorithms,sincepagesrarelyneedtobewrittenouttomakespaceforotherpages. GroupCommitIdea:Insteadofperformingoutputoflogrecordstostablestorageassoonasatransactionisreadytocommit,waituntillogbufferblockisfull,oratransactionhasbeenwaitingsufficientlylongafterbeingreadytocommitResultsinfeweroutputoperationspercommittedtransaction,andcorrespondinglyahigherthroughput.However,commitsaredelayeduntilasufficientlylargegroupoftransactionsarereadytocommit,oratransactionhasbeenwaitinglongenough-leadstoslightlyincreasedresponsetime.Abovedelayacceptableinhigh-performancetransactionsystemssincelogbufferblockswillfillupquickly. Real-TimeTransactionSystemsInsystemswithreal-timeconstraints,correctnessofexecutioninvolvesbothdatabaseconsistencyandthesatisfactionofdeadlines.Harddeadline–SeriousproblemsmayoccuriftaskisnotcompletedwithindeadlineFirmdeadline-Thetaskhaszerovalueifitcompletedafterthedeadline.Softdeadline-Thetaskhasdiminishingvalueifitiscompletedafterthedeadline.Thewidevarianceofexecutiontimesforreadandwriteoperationsondiskscomplicatesthetransactionmanagementproblemfortime-constrainedsystemsmain-memorydatabasesarethusoftenusedWaitsforlocks,transactionaborts,contentionforresourcesremainasproblemsevenifdataisinmainmemoryDesignofareal-timesysteminvolvesensuringthatenoughprocessingpowerexiststomeetdeadlinewithoutrequiringexcessivehardwareresources. LongDurationTransactionsTraditionalconcurrencycontroltechniquesdonotworkwellwhenuserinteractionisrequired:Longduration:DesigneditsessionsareverylongExposureofuncommitteddata:E.g.,partialupdatetoadesignSubtasks:supportpartialrollbackRecoverability:oncrashstateshouldberestoredevenforyet-to-becommitteddata,souserworkisnotlost.Performance:fastresponsetimeisessentialsousertimeisnotwasted. Long-DurationTransactionsRepresentasanestedtransactionatomicdatabaseoperations(read/write)atalowestlevel.Iftransactionfails,onlyactiveshort-durationtransactionsabort.Activelong-durationtransactionsresumeonceanyshortdurationtransactionshaverecovered.Theefficientmanagementoflong-durationwaits,andthepossibilityofaborts.Needalternativestowaitsandaborts;alternativetechniquesmustensurecorrectnesswithoutrequiringserializability. ConcurrencyControlCorrectnesswithoutserializability:Correctnessdependsonthespecificconsistencyconstraintsforthedatabases.Correctnessdependsonthepropertiesofoperationsperformedbyeachtransaction.Usedatabaseconsistencyconstraintsastosplitthedatabaseintosubdatabasesonwhichconcurrencycanbemanagedseparately.Treatsomeoperationsbesidesreadandwriteasfundamentallow-leveloperationsandextendconcurrencycontroltodealwiththem. ConcurrencyControl(Cont.)Anon-conflict-serializableschedulethatpreservesthesumofA+B NestedandMultilevelTransactionsAnestedormultileveltransactionTisrepresentedbyasetT={t1,t2,...,tn}ofsubtransactionsandapartialorderPonT.AsubtransactiontiinTmayabortwithoutforcingTtoabort.Instead,Tmayeitherrestartti,orsimplychoosenottorunti.Ifticommits,thisactiondoesnotmaketi,permanent(unlikethesituationinChapter15).Instead,ti,commitstoT,andmaystillabort(orrequirecompensation)ifTaborts.AnexecutionofTmustnotviolatethepartialorderP,i.e.,ifanedgetitiappearsintheprecedencegraph,thentitimustnotbeinthetransitiveclosureofP. NestedandMultilevelTransactions(Cont.)Subtransactionscanthemselvesbenested/multileveltransactions.Lowestlevelofnesting:standardreadandwriteoperations.Nestingcancreatehigher-leveloperationsthatmayenhanceconcurrency.Typesofnested/multileveltransactions:Multileveltransaction:subtransactionofTispermittedtoreleaselocksoncompletion.Saga:multilevellong-durationtransaction.Nestedtransaction:locksheldbyasubtransactiontiofTareautomaticallyassigntoToncompletionofti. ExampleofNestingRewritetransactionT1usingsubtransactionsTaandTbthatperformincrementordecrementoperations:T1consistsofT1,1,whichsubtracts50fromAT1,2,whichadds50toBRewritetransactionT2usingsubtransactionsTcandTdthatperformincrementordecrementoperations:T2consistsofT2,1,whichsubtracts10fromBT2,2,whichadds10toANoorderingisspecifiedonsubtransactions;anyexecutiongeneratesacorrectresult. CompensatingTransactionsAlternativetoundooperation;compensatingtransactionsdealwiththeproblemofcascadingrollbacks.Insteadofundoingallchangesmadebythefailedtransaction,actionistakento“compensate”forthefailure.Consideralong-durationtransactionTirepresentingatravelreservation,withsubtransactionsTi,1,whichmakesairlinereservations,Ti,2whichreservesrentalcars,andTi,3whichreservesahotelroom.Hotelcancelsthereservation.InsteadofundoingallofTi,thefailureofTi,3iscompensatedforbydeletingtheoldhotelreservationandmakinganewone.Requiresuseofsemanticsofthefailedtransaction. ImplementationIssuesForlong-durationtransactionstosurvivesystemcrashes,wemustlognotonlychangestothedatabase,butalsochangestointernalsystemdatapertainingtothesetransactions.Loggingofupdatesismademorecomplexbyphysicallylargedataitems(CADdesign,documenttext);undesirabletostorebotholdandnewvalues.Twoapproachestoreducingtheoverheadofensuringtherecoverabilityoflargedataitems:Operationlogging.Onlytheoperationperformedonthedataitemandthedata-itemnamearestoredinthelog.Loggingandshadowpaging.Useloggingfromsmalldataitems;useshadowpagingforlargedataitems.Onlymodifiedpagesneedtobestoredinduplicate. TransactionManagementinMultidatabaseSystemsTransactionmanagementiscomplicatedinmultidatabasesystemsbecauseoftheassumptionofautonomyGlobal2PL-eachlocalsiteusesastrict2PL(locksarereleasedattheend);lockssetasaresultofaglobaltransactionarereleasedonlywhenthattransactionreachestheend.GuaranteesglobalserializabilityDuetoautonomyrequirements,sitescannotcooperateandexecuteacommonconcurrencycontrolschemeE.g.,nowaytoensurethatalldatabasesfollowstrict2PLSolutions:provideverylowlevelofconcurrentexecution,oruseweakerlevelsofconsistency TransactionManagementLocaltransactionsareexecutedbyeachlocalDBMS,outsideoftheMDBSsystemcontrol.Globaltransactionsareexecutedundermultidatabasecontrol.Localautonomy-localDBMSscannotcommunicatedirectlytosynchronizeglobaltransactionexecutionandthemultidatabasehasnocontroloverlocaltransactionexecution.localconcurrencycontrolschemeneededtoensurethatDBMS’sscheduleisserializableincaseoflocking,DBMSmustbeabletoguardagainstlocaldeadlocks.needadditionalmechanismstoensureglobalserializability Two-LevelSerializabilityDBMSensureslocalserializabilityamongitslocaltransactions,includingthosethatarepartofaglobaltransaction.Themultidatabaseensuresserializabilityamongglobaltransactionsalone-ignoringtheorderingsinducedbylocaltransactions.2LSRdoesnotensureglobalserializability,however,itcanfulfillrequirementsforstrongcorrectness.1.Preserveconsistencyasspecifiedbyagivensetofconstraints2.GuaranteethatthesetofdataitemsreadbyeachtransactionisconsistentGlobal-readprotocol:Globaltransactionscanread,butnotupdate,localdataitems;localtransactionsdonothaveaccesstoglobaldata.Therearenoconsistencyconstraintsbetweenlocalandglobaldataitems. Two-LevelSerializability(Cont.)Local-readprotocol:Localtransactionshavereadaccesstoglobaldata;disallowsallaccesstolocaldatabyglobaltransactions.Atransactionhasavaluedependencyifthevaluethatitwritestoadataitematonesitedependsonavaluethatitreadforadataitemonanothersite.Forstrongcorrectness:Notransactionmayhaveavaluedependency.Global-read-write/local-readprotocol:Localtransactionshavereadaccesstoglobaldata;globaltransactionsmayreadandwritealldata;Noconsistencyconstraintsbetweenlocalandglobaldataitems.Notransactionmayhavevaluedependency. GlobalSerializabilityEvenifnoinformationisavailableconcerningthestructureofthevariousconcurrencycontrolschemes,averyrestrictiveprotocolthatensuresserializabilityisavailable.Transaction-graph:agraphwithverticesbeingglobaltransactionnamesandsitenames.Anundirectededge(Ti,Sk)existsifTiisactiveatsiteSk.Globalserializabilityisassurediftransaction-graphcontainsnoundirectedcycles. EnsuringGlobalSerializabilityEachsiteSihasaspecialdataitem,calledticketEverytransactionTjthatrunsatsiteSkwritestotheticketatsiteSiEnsuresglobaltransactionsareserializedateachsite,regardlessoflocalconcurrencycontrolmethod,solongasthemethodguaranteeslocalserializabilityGlobaltransactionmanagerdecidesserialorderingofglobaltransactionsbycontrollingorderinwhichticketsareaccessedHowever,aboveprotocolresultsinlowconcurrencybetweenglobaltransactions. EndofChapter26 Figure26.01 Figure26.02 Figure26.03 Figure26.04 Figure26.05 Extraslides WeakLevelsConsistencyUsealternativenotionsofconsistencythatdonotensureserializability,toimproveperformance.Degree-twoconsistencyavoidscascadingabortswithoutnecessarilyensuringserializability.Unliketwo-phaselocking,S-locksmaybereleasedatanytime,andlicksmaybeacquiredatanytime.X-locksbereleaseduntilthetransactioneithercommitsoraborts. ExampleSchedulewithDegree-TwoConsistencyNonserializableschedulewithdegree-twoconsistency(Figure20.5)whereT3readsthevalueifQbeforeandafterthatvalueiswrittenbyT4.T3T4lock-S(Q)read(Q)unlock(Q)lock-S(Q)read(Q)unlock(Q)lock-X(Q)read(Q)write(Q)unlock(Q) CursorStabilityFormofdegree-twoconsistencydesignedforprogramswritteningeneral-purpose,record-orientedlanguages(e.g.,Pascal,C,Cobol,PL/I,Fortran).Ratherthanlockingtheentirerelation,cursorstabilityensuresthatThetuplethatiscurrentlybeingprocessedbytheiterationislockedinsharedmode.Anymodifiedtuplesarelockedinexclusivemodeuntilthetransactioncommits.Usedonheavilyaccessedrelationsasameansofincreasingconcurrencyandimprovingsystemperformance.Useislimitedtospecializedsituationswithsimpleconsistencyconstraints.'