Abstract. Data grids, large scale web applications generating dynamic content and database service providing pose significant scalability challenges to database engines. Replication is the most common solution but it involves difficult trade-offs. The most

Ganymed:ScalableReplication

forTransactionalWebApplications

ChristianPlattnerandGustavoAlonso

DepartmentofComputerScience

SwissFederalInstituteofTechnology(ETHZ)

ETHZentrum,CH-8092Z¨urich,Switzerland

{plattner,alonso}@inf.ethz.ch

Abstract.Datagrids,largescalewebapplicationsgeneratingdynamiccontentanddatabaseserviceprovidingposesigni cantscalabilitychal-lengestodatabaseengines.Replicationisthemostcommonsolutionbutitinvolvesdi culttrade-o s.Themostdi cultoneisthechoicebetweenscalabilityandconsistency.Commercialsystemsgiveupconsistency.Re-searchsolutionstypicallyeithero eracompromise(limitedscalabilityinexchangeforconsistency)orimposelimitationsonthedataschemaandtheworkload.InthispaperweintroduceGanymed,adatabasereplica-tionmiddlewareintendedtoprovidescalabilitywithoutsacri cingcon-sistencyandavoidingthelimitationsofexistingapproaches.Themainideaistouseanoveltransactionschedulingalgorithmthatseparatesupdateandread-onlytransactions.TransactionscanbesubmittedtoGanymedthroughaspecialJDBCdriver.Ganymedthenroutesupdatestoamainserverandqueriestoapotentiallyunlimitednumberofread-onlycopies.Thesystemguaranteesthatalltransactionsseeaconsistentdatastate(snapshotisolation).Inthepaperwedescribetheschedul-ingalgorithm,thearchitectureofGanymed,andpresentanextensiveperformanceevaluationthatprovesthepotentialofthesystem.

1Introduction

TraditionallyusedinpredictableandstaticenvironmentslikeERP(EnterpriseResourcePlanning),data-basemanagementsystemshadtofacenewchallengesinthelastfewyears.Webservices,applicationserviceprovidingandgridcom-putingrequirehigherscalabilityandavailability.Databasereplicationiswidelyusedtoachievethosegoals.Theproblemwithreplicationisthatexistingcom-mercialsolutionsdonotbehavelikesingleinstancedatabases.Often,achoicemustbemadebetweenscalabilityandconsistency.Iffullconsistencyisrequired,thepriceisalossinscalability[1].

RecentresearchintheareaofTransactionalWebApplicationshasledtomanyalternativeproposals,buttheyallsu erfromavarietyofproblems.Sys-temsthatputdataontheedgeofthenetwork,[2,3,4],areabletoreduceresponsetimesbutgiveupconsistency.Otherapproaches,[5,6,7],workonthemiddle-warelayerandneedapredeclarationoftheaccesspatternofalltransactionsto