Your SlideShare is downloading. ×
Releasing fast code - The DevOps approach
Upcoming SlideShare
Loading in...5

Thanks for flagging this SlideShare!

Oops! An error has occurred.


Introducing the official SlideShare app

Stunning, full-screen experience for iPhone and Android

Text the download link to your phone

Standard text messaging rates apply

Releasing fast code - The DevOps approach


Published on

Agile makes you Develop faster, DevOps also makes you Deploy faster but how do you make your Application faster? …

Agile makes you Develop faster, DevOps also makes you Deploy faster but how do you make your Application faster?
Many currently used Performance Management practices don’t work anymore as they are too time consuming. It takes a new approach to track performance in Continuous Integration, get more value out of Load Testing and leverage production data for performance optimization.
We will show you real world examples on how the new DevOps approach can work.

Published in: Technology

  • Be the first to comment

  • Be the first to like this

No Downloads
Total Views
On Slideshare
From Embeds
Number of Embeds
Embeds 0
No embeds

Report content
Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

No notes for slide
  • Agile developmentallowstoreact on changedrequirements (technicalandbusiness) fasterReducestheriskofworkingtoolong in thewrongdirectionProvides a „Potential ReleasableProduct“ after every Sprint/IterationIncludesTesting Practices intothe Development LifecycleAutomatesmanytasks such asBuild, FunctionalTestingandprogressreporting (Burn-Down Chart, …)In short, betterquality, lessrisk, fasterreaction time
  • Let‘slookathow a typical Sprint/Iteration works …
  • Atevery time weknowwherewe stand in termsoffunctionality – until -
  • Wegettotestingorproductionwherewe find out thatweeitherhavetogo back and fix problemsidentifiedunderloadorwhenwehavemajorissues in a prodenvironment
  • The resultisthatwe will miss our Goal andwehadwrongestimates
  • LoadTestingandProduction. Often a separate organization.Firstlythismeanswe do not leveragetheadvantageof agile fully. We do not deliverthespeedofchangetothebusinessandthustothecustomer.In a normal shelfproductthismightbefine, ifyouoperate a SAAS oreComercesiteitreducesyourabilitytocompete.Apart fromthatthereare also severalproblemswiththehardsplitbetweenthetwosides. Twooftenmore separate organizationsleadto…
  • Functionality vs. StabilitySeparate teams, different mindset, different incentives. Operationsispaidtokeepthignsstableandrunningwhiledevelopmentistoaddnewfeatures. The funny thingis, thatoperationsisdrivenbycustomerdemands but does not wanttoaddbusinessfunctionalitytobetercompete.Dev on theothersidewantstoaddnewfunctionality all the time, but they lack theexperiencewiththecustomer. They lack the real worldproblems.
  • Distributed Teams makeithardtocoordinateMakescommunication all thathearder. Timezones, langugae different culture in additionto an already different mindset
  • Even worse, ifwewanttocommunicatesomething, thetechnologyis not compatible. Different monitoring, analysistools
  • All ofthisleadsto …I am done – nowyoudeployit
  • Not a surprise – isit?
  • holistic approach to software deliveryPromoteColocationofteamsinsteadofspreadingthemaroundtheworldimaginetheguyrunninganddeployingyourappsitsnexttoyou. Thatwouldmake a lotstuffeasierOneToolsetReduces a lotofoverheadconvertingdata. Reducesmaintaince time. Andyouwouldmakesuretoreuseconfigurationofsaymonitoringfromautomatictest in prod, andfromautomaticbuildandtesttodeploythestufffor real.Agile DevlieryandDeploymentDeliverythe agile promisedirectlytothebusinessandcustomer. React in dayswithnewfeatures. Therearebigonesthat do that. (google, facebooktonametwo)ONE TEAM thatisresponsiblefor a Product (FromDevtoOps)Ifthereis a problem,theguyresponsibleisyou
  • Solution ProposalsCenteredaroundOneToolsetwehave different continuoustasks
  • Automatic, Detect Change – reuseyourexisting Unit TestsBE AWARE thatweare not necesarilymeasuring Performance/Execution Time here -> doesntmake sense forshortrunning Unit Tests. On thosewefocus on theinternals, e.g.: Database Statements andseehowtheychangeover time (regression)Perfearly, insidethecycle. This wayitispartofthesprint not outside. Itisautomted so itdoes not introduce additional effort.
  • Besidesfocusing on theinternaldynamicsofthetestedcomponentswecanruncertainsetsof real performancetests in CIThis exampleshowshowdynaTraceinternallyrunsperformancetests on criticalfeature on everybuildExplaintheidentifiedproblems …
  • In ordertorunperformancetests in CI itisimporttofollowsomegroundrulesUnit testsneedtohave a certainexecution time lengthMakesurethatyouhave a stableenvironment -> tip: runseveralwarm-upteststoensureenvironmentisstable -> thenrun real test
  • In Java/.NET itisimportanttoexclude volatile externalfactors such asGarbageCollection Time
  • Themoreinterestingusecase in CI isArchitecture Validation wherewehave a verycloselookatwhatisgoing on withinthetestedcomponents.Byanalyzingthedynamiccodeexecutionwecanvalidatecertainrulesthataredefinedbythe Software Architects, e.g.: Onlymakeasycnhronousremotingcalls, only X DB Calls per transaction, …Havingthetools in placeallowyoutoautomaticallyverifytheserulesagainstyourunitorfunctionaltests
  • Weseemoreandmoreexternalcodebeingused – whichis a goodthing – but canhave a bigimpact on performanceifwe do not understandtheinternals
  • Toomucheffortforscriptmaintenance, settingupthetestenvironmentCostsfortestingtoolsandtestenvironmentTest cycletakeslongerthanreserved in thesprint. Feedback todevscomeswhentheyarealready in thenextsprint
  • Sample Data: smalldatabaseswith sample datais not realistic. Canttestcaching, access time todatabase, …OverloadedAgents: LoadTestingAgentsare just usedtoproduceloadwithoutverifyingiftheycanproducetherequestedloadWrong Think Times: Think Times areeither not setcorrectlytoreflect real usersortheyareavoidedtoincreaseload -> thats not realisticeitherInfrastructure Issues: doesthe Test Infrastructure provideenoughbandwidth, resources, … to handle theload? Isthereanythingelserunning on theinfrastructurethatmayinterfere?No Error Detection: Missing Content Verificationleadstoincorrectresults. A Page thatresponds super fast but alwaysyresponds „pleasecome back later“ is a problemNo Server Side Data: veryoftenpeopleforgettocollectserversidemetrics such as CPU, Network, I/O, Logfiles, …
  • classical loadtestingtest environment that tries to be as close to production as possible (size/users/data vs. reality)test scripts with transactions that we expect (but we know we shouldexpecttheunexpected. Never happens)TRIES TO REBUILD TO REAL WORLD -> HUGE EFFORT!!Result: no real guarantee that what has been tested really works in prodProblems:Testing on not frozen Code Base -> Results of Testing are available not before mid or end of NEXT Sprint -> Results are outdated!HArd to get real life Production Load BehaviorMore changes in application lead to more effort in testing (script maintenance, regression analysis, ...)Size ofenvironmentNumberof UsersTransaction Mix
  • Rerun, additional dataAdding log statemetnsandstatistics in code, errorprone.First, automatethestuff.Second, captureenoughfromthebeginning. (traces, BTM)Dynamic Capture all the time.Overhead discussion 5-10Benefits of Agile and DevOpsShorter iterations with fewer changes leads to reduced effort in testingfewer script maintenance changes, higher quality smoke test, easier regression analysis,more focused testing on the areas that have changed
  • Makesurewecaptureenoughinfrastructuredata – not just responsetimes. This helpsustoget a quick overviewoftheproblematicareas
  • In ordertoidentifyproblematiccomponentsorservicesitsrecommendedtorun different loadandseewhichcomponentsscaleandwhichdont
  • You do not onlyrunoneloadtest in a devcycle. Yourunmanyandalwayswanttoseewhathaschanged.Itisthereforeimportanttocomparetestresults. Critical hereisthatwehaveteststhatwecancompareandthatwehavedatathatiscomparable
  • Inthe end itisthedeveloperthatneedstolookattheproblem. The more granular transactionaldatawecanprovidetheeasiertoanalyzeand fix theproblem.ProfilingandTracing Solutions becamevery powerful andallowdeeptracing also duringload.
  • Real environment, real validityLesscost, lessmaintaince,lesseffort.Yoursystem will bemoststableandyou will learnperformancemanagement.Test during off-hours/off-daysorduringmaintainencecyclesOff peaktimesnewversiontestloadtestingPart ofproductionwithnewversionShadow production, trafficduplication
  • During off hourswetakesomeoftheproductionmachines out ofthecluster, installthetestversion on it, configuretheloadbalancerandrunloadagainstitusing a loadtestingtoolThis allowsustotest on real hardware
  • This is not alwayspossible – depending on thechanges in thesoftware. This howeverisgreattotestpatches on howtheybehave. Inboundtrafficisduplicated. Wecanverifyifweseeanyfunctionalproblemsaswellashowtheperformancecomparestothe „current“ system
  • In thiscasewe update a setofmachineswith a newversion/patch. The loadbalancerisconfiguredto route certainuserstothenewmachinestotest out howthisworks.
  • Testing real numberofusers, with real loadand on yourproductionwith real contentSimulatenumberofusersasyoucouldn‘tyourselfReal end userperformanceIfyouknowwhereyouruserbaseis, testfromthereSee localdifferencesAt a fractionofthecost, ist on demand.Letthemmaintaintheirscripts, concentrate on yourproductLeverageserviceslikekeynoteandgomes. Simulate real userload, and not like in loadtesting just transactionload.provisioning in the cloud -> dynamic resources -> cater for millions of users
  • Knowwhat‘sreallygoing onReal problemsNoreproductionOptimzebased on what‘s realNoguess – just factsDirectfeedbacktodevUseProductiontoanalyzetheperformance. Reactto real problems. Noreproduction.Tune andoptimzebased on real worldtransactionanduserload.We solve the "Black Box View" of the test departmentWe dont know the insight of the app -> so what shall we look for?We dont know how the real prod evnrionment looks like -> so how shall we know if it works in prod or notAPM in Produktion --> fast feedback to dev (same team anyway?), agile cycles allow formsquick fixesless change, smaller errors, faster solution
  • The real environmentisalways different thanwhatyouhadtested. Thereforeitisimportanttoanalyzethesystemandseehowtransactionsflowandwheretheymighthave a problemThis allowsyoutofocus on thoseservices/componentsthathave a problem in real life
  • Analyzingthebusinessimpactofyourappsperformanceisimportanttoprioritizeyourtroubleshootingactivities
  • Kreis mit pfeilen, mitte ein tool, communicationbarrier, objectivness, lessmaintaincePerformance analysisPerformance testingPerformance monitoringvalidation
  • Agile + DevOps + CAPMChanged Testing Mindest -> WE ARE NEVER DONE -> WE CONSTANTLY IMPROVE
  • Transcript

    • 1. Releasing Fast CodeThe DevOps Approach to Performance
    • 2. Why Agile Development took off
    • 3. It‘s Sprint TimeStory Points Developme Testing Estimate nt Sprint Timeline Production Remaining Team Velocity
    • 4. You are in controlStory Points Developme Testing Estimate nt Sprint Timeline Production Remaining Team Velocity
    • 5. What happened?Story Points Developme Testing Estimate nt Sprint Timeline Production Remaining Team Velocity
    • 6. Missed Goals and EstimatesStory Points Estimates Missed Developme Testing Estimate nt Production Missed Remaining Goal Team Velocity
    • 7. What is the Status Quo beyond Dev?Source:
    • 8. Problem #1: Different MindsetSource:
    • 9. Problem #2: Dislocated TeamsSource:
    • 10. Problem #2: Different ToolsSource:
    • 11. Problem #4: Over the Fence AttitudeSource:
    • 12. These Problems lead to …Source:
    • 13. A potential SolutionSource:
    • 14. Real World Perf Test in Feedback CI ONECloud based Toolset Architecture ValidationTesting Test in Production Traditional Load Testing
    • 15. Performance in Continuous Integration
    • 16. Identify Performance Regressions Early
    • 17. Tip: Stable Environment
    • 18. Tip: Exclude Garbage Collection
    • 19. Architecture Validation Analyze your existing Unit Tests Validate against your Arch Rules
    • 20. Learn how foreign code works• Get insight into whats going on „under the hood“ • Understand internals of the frameworks you use • Choose the right usage for your use case
    • 21. Traditional Load Testingtakes too long in an agile process time Sprint 1 Starts with Dev and dedicated Time for Test at the End Update Test Setup Run Tests Scripts Environment Adapt Scripts Re Run Tests Feedback to Dev Testing extends sprints Consumes too many Dev Resources Sprint 2 Re Run Tests Feedback to Dev Ready for Ops Either postponed or impacted because Dev Resources are still occopuied with testing of previous sprint
    • 22. Common MistakesOnly use Sample Data No Server Side DataWrong Think Times Overloaded Load AgentsNo Error Detection Infrastructure Issues
    • 23. Load Testing needs to be „real“ Real Size and Content Real Third Party CodeReal Infrastructure Setup No Guarantees! Real Load Real Number of Users
    • 24. Minimize and automate real Load TestsDeveloping Test Run Reproduction Refine Capturing Re Run Tests Reproduction Refine Capturing Multiple Test Iterations needed to analyze Root-cause Re Run Tests Reproduction Problem Analysis Problem Solving timeDeveloping Test Run Reproduction Refine Capturing Re Run Tests Reproduction Refine Capturing •Eliminates Test Iterations •Go directly to problem analysis •Frees up resources for other proje Re Run Tests Reproduction Problem Analysis Problem Solving time
    • 25. Tip: Measure more than Response Time
    • 26. Tip: Identify Scalabilty Problems Response time Presentation Layer Presentation Layer Business Logic Presentation Layer Business Logic Business Logic Web Service/Remoting Web Service/Remoting Web Service/Remoting Data Access Data Access Data Access Minimum Load Medium Load Maximum Load
    • 27. Incresing Load Test is your friend
    • 28. Tip: Identify Regressions on Code Level
    • 29. Tip: Get as granular data as possible
    • 30. Test in your Production Environment Real Environment Real Validity Less Cost
    • 31. Scenario 1: Use Production Hardware Real Users Production Environment Virtual Users LB/FW Web Server App Server DB Route Incoming Use some machines Traffic during off hours for testing
    • 32. Scenario 2: Duplicate Traffic Real Users Production Environment LB/FW Web Server App Server DB Send incoming traffic to Prod and Testing Machines
    • 33. Scenario 3: Partial Update Real Users Production Environment LB/FW Web Server App Server DB Update some machines to the version that should be tested
    • 34. Use Cloud based Testing Services Real End User TimeCDN Proxies High ControlVolume Costs Last Mile Load Balancers
    • 35. APM in Production
    • 36. Understand your production environment
    • 37. Focus on areas with biggest impact160140120100 80 Number of Sales 60 No of Users 40 20 0 Response Time Transaction Load
    • 38. ONE set of Tools Communication Collaboration
    • 40. Michael Koppmichael.kopp@dynaTrace.com