Scalable Social Architectures by Biren Gandhi
5 likes822 views
The document discusses the scalability challenges and architectural considerations for social applications, emphasizing that one size does not fit all and highlighting the importance of persistence, caching, and sharding techniques. It outlines various scaling strategies for different components, including web, app, and database tiers, and addresses the complexities of handling traffic patterns and deployment disciplines. Overall, it conveys that successful scaling is a continuous process requiring careful planning, monitoring, and adaptation.
Scalable Social Architectures by Biren Gandhi
- 1. Scalable Social ArchitecturesThink Early, Think Often, Think Throughbirengandhihttp://unhub.com/biren
- 2. Scalability by the numberssource: developeranalytics.com (as of Dec 1, 2010)Scalable Social Architectures
- 3. Scalability by the numbers500M active users, 50% log-in EVERY DAYsource: hitwise.comScalable Social Architectures
- 4. Isn’t that great?Exponential Growth is AWESOME or PAINFUL Scalable Social Architectures
- 5. Scaling ChallengesScalable Social ArchitecturesThink EarlyThink OftenThink Through
- 6. Scaling ChallengesAnd the biggest challenge is …One Size Does NOT Fit AllScalable Social Architectures
- 7. Social Web ArchitecturesScalable Social ArchitecturesFirewallCachingWeb and App TierPublic NetworkLB/ProxyDB TierPlatform API Calls“Social” Part
- 8. Social Game ArchitecturesScalable Social ArchitecturesNot “nice to have”, but “must have”CDNUpload/MissFirewallCachingWeb and App TierPublic NetworkLB/ProxyDB TierPlatform API Calls
- 9. Gaming Stack SummaryScalable Social Architectures
- 10. Obvious Challenges PersistenceCachingWeb ServerClient PerceptionDataScalable Social Architectures
- 11. PersistenceRDBMS (mysql)Data modelingDe-normalization is your friend SlavingShardingNoSQLHbase, Cassandra, MongoDB, Redis, …MembaseGraph DatabasesNeo4j, InfiniteGraph, …Traffic PatternsRead/Write MixConsistency ModelsScalable Social Architectures
- 12. MysqlSharding ExampleApp TierApp TierM0M1M0M1MasterS0S1S0S1S2S3SlaveS’0S’1S’2S’3
- 13. CachingIn-process (MMO)Java vs. Scripting LanguagesAPCMostly read-only global dataPriming after releasesSerialization MemcacheFunctional (vertical) ShardingStatistical (horizontal) ShardingGranularity & Compression of Stored ObjectsScalable Social Architectures
- 14. Memcached ExampleKey Format: <function>:<version>:value (e.g. user:1:123 or better u:1:123)Scalable Social ArchitecturesHorizontalShardingVerticalSharding
- 15. Memcached ExampleKey Format: <function>:<version>:value (e.g. user:1:123 or better u:1:123)Scalable Social ArchitecturesHorizontalShardingVerticalShardinguid % nfeature?
- 16. Web TierEasy to Scale (Stateless)Cloud-friendly auto-grow policiesSticky sessions?H/W or S/W Load BalancersHAProxy, Zeus for Cloud environmentsDNS could be a bottleneckRound Robin DNSTech ChoicesApache, Nginx + FastCGINode.js on the server sideScalable Social Architectures
- 17. Client TierCDNStatic assets (images, JS, CSS etc.)Invalidation/update schemes (versioning)UGC might be little challengingFlash (or Graphic Engines)Watch out for frame ratesUsers’ machines are not as powerfulJavaScriptFrameworks (jQuery, Prototype, etc.)Tons of great advice on minifying, compressing, obfuscatingScalable Social Architectures
- 18. DataVersioningCoexistence of old and new dataRollback during catastrophic eventsIntegrityChecksum (costly)Field-level Validation (not cheap either)Gaming ContentLightweight CMS (integrated in art pipeline)Scalable Social Architectures
- 19. Not-so-obvious ChallengesData Center vs. CloudNetwork Port SaturationCPU UtilizationConnection BloatRuntime ConfigurationProgramming ModelsDeployment DisciplineScalable Social Architectures
- 20. Not-so-obvious ChallengesData Center vs. CloudVirtualization is good, but slowLittle control over optimization High failure ratesNetwork SaturationTraffic going across (DB/Memcache boxes)CPU UtilizationOne core is overworked Scalable Social ArchitecturesCloud Perf. 1/5 to 1/8 of Data Center
- 21. Not-so-obvious ChallengesConnection BloatDatabase/MemcachedPersistent ConnectionsProxies (mysql-proxy, moxi, memagent etc.)Runtime ConfigurationControlled rollout of featuresFire-fighting defenseHome-grown, Apache ZooKeeper, … Scalable Social Architectures
- 22. Not-so-obvious ChallengesProgramming ModelsExternal platform APIsFailure semanticsMake async, as much as possibleGraceful subsystem failuresOrganize all users data into same shardsThings WILL fail when you least expect them toConstant Performance EvaluationProfiling Tools (xhprof, xdebug etc. for PHP)HipHop CompilerPHP Extensions for Common CodeHeterogeneous languages for subcomponents Scalable Social Architectures
- 23. Deployment DisciplinesTools Hudson/Bamboo (deployment)Munin/Nagios (monitoring)Staging environmentAs close to production as possibleUser Downtime404 PagesTransparent HandlingCache PrimingDB failures after every deploymentDeployment Time1000+ serversP2P solutionsScalable Social Architectures
- 24. SummaryMetrics-driven OptimizationsYou can’t improve something you can’t measureEvery smart person has an ‘opinion’Leverage automated monitoring toolsLet machines work while humans restAny system is only as robust as its weakest linkIt’s an ONGOING process – a journey …better get used to it Think Early, Think Often, Think ThroughScalable Social Architectures