ProcessingDiagnosisQueries:A Principledand ScalableApproach

Many popular Web sites suffer occasional user- visible problems suchasslowresponses, blankpagesorerror messages beingdisplayed, itemsnotbeingaddedtoshopping carts, database slowdowns, andothers. Suchdeviations ofsystems fromdesired behavior, orfailures, cancauseuserdissatisfaction andconsiderable lossofrevenue. Thescale, complexity, and dynamics ofmodernsystems makeithardtotrackdownthe causeoffailures manually. We address thisproblem through a new classofdeclarative queries, called diagnosis queries, thata systemadministrator orusercanposetopinpoint the causeofa failure. We describe how diagnosis queries are specified oversystem-monitoring data, andthechallenges faced bycurrent techniques toprocess these queries. We develop and evaluate anewalgorithm, basedonacombination ofclustering andclassification, toprocess diagnosis queries automatically, efficiently, andwithgoodaccuracy.