Oct 2011 CHADNUG Presentation on Hadoop
- 1. October 2011 – CHADNUG – Chattanooga, TNWhat is Hadoop?Josh Patterson | Sr Solution Architect
- 2. Who is Josh Patterson?josh@cloudera.comTwitter: @jpatanoogaMaster’s Thesis: self-organizing mesh networks Published in IAAI-09: TinyTermite: A Secure Routing AlgorithmConceived, built, and led Hadoop integration for openPDC project at Tennessee Valley Authority (TVA)Led team which designed classification techniques for time series and Map ReduceOpen source work at http://openpdc.codeplex.comhttps://github.com/jpatanoogaTodaySr. Solutions Architect at Cloudera
- 3. OutlineLet’s Set the StageStory Time: Hadoop and the SmartgridThe Enterprise and HadoopUse Cases and Tools3
- 4. “After the refining process, one barrel of crude oil yielded more than 40% gasoline and only 3% kerosene, creating large quantities of waste gasoline for disposal.”--- Excerpt from the book “The American Gas Station”Data Today: The Oil Industry Circa 19004
- 7. Digital universe grew by 62% last year to 800K petabytes and will grow to 1.2 “zettabytes” this yearObstacles to Leveraging DataCopyright 2010 Cloudera Inc. All rights reserved7Data comes in many shapes and sizes: relational tuples, log files, semistructured textual data (e.g., e-mail)
- 11. Customers are moving some of it to tape or cold storage, throwing it away because “it doesn’t fit”
- 13. Similar to the oil industry in 1900A Need for a Platform in an Evolving LandscapeNeed ability to look at true distribution of dataPreviously impossible due to scaleNeed lower cost of analysisAd Hoc analysis now more open and flexibleNeed Greater Flexibility, “BI Agility”Less restrictive than SQL-only systemsSpeed @ Scale is the new Killer AppResults in that previously took 1 day to process can gain new value when created in 10 minutes.Copyright 2010 Cloudera Inc. All rights reserved8
- 14. Story Time: Hadoop and the Smartgrid9“We’re gonna need a bigger boat.”--- Roy Scheider, “Jaws”
- 15. NERC Sensor Data CollectionopenPDC PMU Data Collection circa 2009 120 Sensors
- 18. Housed in HadoopNERC Wanted High-Res Smartgrid DataStarted openPDC project @ TVAhttp://openpdc.codeplex.com/We used Hadoop to store and process smartgrid (PMU) time series datahttps://openpdc.svn.codeplex.com/svn/Hadoop/Current%20Version/Copyright 2011 Cloudera Inc. All rights reserved
- 19. Major Themes From openPDCVelocity of incoming dataWhere to put the data?Wanted to scale out, not upWanted linear scalability in cost vs sizeWanted system robust in the face of HW failureNot fans of vendor lock-inWhat can we realistically expect from analysis and extraction at this scale?How long does it take to scan a Petabyte @ 40MB/s?
- 21. Move complex and relational data into a single repository
- 27. Mine data first, govern later MapReduceHadoop Distributed File System (HDFS)
- 28. What Hadoop doesNetworks industry standard hardware nodes together to combine compute and storage into scalable distributed system
- 38. This aspect is part of Cloudera’s value addHDFS: Hadoop Distributed File SystemBlock Size = 64MBReplication Factor = 3Cost/GB is a few ¢/month vs $/month
- 39. MapReduceIn simple terms, it’s an application with 2 functionsMap FunctionThink massively parallel “group by”Reduce FunctionThink “aggregation + processing”Not hard to writeCan be challenging to refactor existing algorithmsDesigned to work hand-in-hand with HDFSMinimizes disk seeks, operates at “transfer rate” of disk
- 40. Speed @ ScaleScenario1 million sensors, collecting sample / 5 min5 year retention policyStorage needs of 15 TBProcessingSingle Machine: 15TB takes 2.2 DAYS to scanMapReduce@ 20 nodes: Same task takes 11 Minutes
- 41. MapReduce ToolsPigProcedural language compiled into MRHiveSQL-like language compiled into MRMahoutCollection of data mining algorithms for HadoopStreamingAbility to write MR with tools such as python, etc
- 42. “What if you don’t know the questions?” --- Forrester ReportThe Enterprise and Hadoop20
- 43. Apache Hadoop in Production©2011 Cloudera, Inc. All Rights Reserved.21How Apache Hadoop fitsinto your existing infrastructure.CUSTOMERSANALYSTSBUSINESS USERSOPERATORSENGINEERSWeb ApplicationManagement ToolsIDE’sBI / AnalyticsEnterprise ReportingEnterprise Data WarehouseOperational Rules EnginesLogsFilesWeb DataRelational Databases
- 44. What the Industry is DoingMicrosoft ships CTP of Hadoop Connectors for SQL Server and Parallel Data Warehouse(based on Cloudera’sSqoop)http://blogs.technet.com/b/dataplatforminsider/archive/2011/08/25/microsoft-ships-ctp-of-hadoop-connectors-for-sql-server-and-parallel-data-warehouse.aspxOracleAnnouncments at Oracle Open WorldConnector to Hadoop allowing data flow into OracleHadoop Accelerator for ExalogicIn-memory processing for MapReduceETL using HadoopIntegrated analytics on Oracle and HadoopCopyright 2010 Cloudera Inc. All rights reserved22
- 45. Integrating With the Enterprise IT Ecosystem23Drivers, language enhancements, testingFile System MountUI FrameworkSDKFUSE-DFSHUEHUE SDKWorkflowSchedulingMetadataSqoop* frame-work, adaptersAPACHE OOZIEAPACHE OOZIEAPACHE HIVEMore coming…Data IntegrationFast Read/Write AccessLanguages / CompilersAPACHE PIG, APACHE HIVEAPACHE FLUME, APACHE SQOOPAPACHE HBASECoordinationAPACHE ZOOKEEPERPackaging, testing*Sqoop supports JDBC, MySQL, Postgres, HSQLDB
- 46. Forester Report World Economic Forum Declared that data is a new asset classBig data is an applied science project in most companiesMajor potential constraint is not the cost of the computing technology but the skilled people needed to carry out these projectsthe data scientistsWhat if you don’t know the questions? Big data is all about exploration without preconceived notionsNeed tools to ask questions to understand the right questions to askMuch of the software is based on open-source Hadoophttp://blogs.forrester.com/brian_hopkins/11-09-30-big_data_will_help_shape_your_markets_next_big_winnersCopyright 2010 Cloudera Inc. All rights reserved24
- 47. Ever been recommended a friend on Facebook?Ever been recommended a product on Amazon?Ever used the homepage at Yahoo?What Can Hadoop Do For Me?25
- 48. Problems Addressed With HadoopText MiningIndex BuildingSearch EnginesGraph CreationTwitter, FacebookPattern RecognitionNaïve Bayes ClassificationRecommendation EnginesPredictive ModelsRisk AssessmentCopyright 2010 Cloudera Inc. All rights reserved26
- 49. A Few Named ExamplesAnalyze search terms and subsequent user purchase decisions to tune search results, increase conversion ratesDigest long-term historical trade data to identify fraudulent activity and build real-time fraud preventionModel site visitor behavior with analytics that deliver better recommendations for new purchasesContinually refine predictive models for advertising response rates to deliver more precisely targeted advertisementsReplace expensive legacy ETL system with more flexible, cheaper infrastructure that is 20 times fasterCorrelate educational outcomes with programs and student histories to improve resultsCopyright © 2011, Cloudera, Inc. All Rights Reserved.27
- 50. Packages For HadoopDataFuFrom Linkedinhttp://sna-projects.com/datafu/UDFs in Pigused at LinkedIn in many of off-line workflows for data derived products"People You May Know”"Skills”TechniquesPageRankQuantiles (median), variance, etc.SessionizationConvenience bag functionsConvenience utility functions28
- 51. Integration with LibsMix MapReduce with Machine Learning LibsWEKAKXENCPLEXMap side “groups data”Reduce side processes groups of data with Lib in parallelInvolves tricks in getting K/V pairs into libPipes, tmp files, task cache dir, etc29
- 52. Ask the right questions up front..Is the job disk bound?What is the latency requirement on the job?Does it need a sub-minute latency? (not good for Hadoop!)Does the job look at a lot or all of the data at the same time?Hadoop is good at looking at all data, complex/fuzzy joinsIs large amounts of ETL processing needed before analysis?Hadoop is good at ETL pre-processing workCan the analysis be converted into MR / Pig / Hive?Copyright 2010 Cloudera Inc. All rights reserved30
- 53. What Hadoop Not Good At in Data MiningAnything highly iterativeAnything that is extemely CPU bound and not disk boundAlgorithms that can’t be inherently parallelizedExamplesStochastic Gradient Descent (SGD)Support Vector Machines (SVM)Doesn’t mean they arent great to use
- 54. Questions? (Thank You!)Hadoop World 2011http://www.hadoopworld.com/Cloudera’s Distribution including Apache Hadoop (CDH):http://www.cloudera.comResourceshttp://www.slideshare.net/cloudera/hadoop-as-the-platform-for-the-smartgrid-at-tvahttp://gigaom.com/cleantech/the-google-android-of-the-smart-grid-openpdc/http://news.cnet.com/8301-13846_3-10393259-62.htmlhttp://gigaom.com/cleantech/how-to-use-open-source-hadoop-for-the-smart-grid/Timeseries blog articlehttp://www.cloudera.com/blog/2011/03/simple-moving-average-secondary-sort-and-mapreduce-part-1/32
- 60. Get your own copy of Cloudera Distribution for Apache Hadoop (CDH)
Editor's Notes
- Its all about the love, baby.
- Theme: they through away a lot of valuable gas and oil just like we through away data today
- Example of data production trends
- But what if some constraints changed?
- Talk about changing market dynamics of storage costWhat if some of the previously held constraints changed? Enter hadoop
- Let’s set the stage in the context of story, why we were looking at big data for time series.
- Ok, so how did we get to this point?Older SCADA systems take 1 data point per 2-4 seconds --- PMUs --- 30 times a sec, 120 PMUs, Growing by 10x factor
- Data was sampled 30 times a secondNumber of sensors (Phasor Measurement Units / PMU) was increasing rapidly was 120, heading towards 1000 over next 2 years, currently taking in 4.3 billion samples per dayCost of SAN storage became excessiveLittle analysis possible on SAN due to poor read rates on large amounts (TBs) of data
- Pool commodity servers in a single hierarchical namespace.Designed for large files that are written once and read many times.Example here shows what happens with a replication factor of 3, each data block is present in at least 3 separate data nodes.Typical Hadoop node is eight cores with 16GB ram and four 1TB SATA disks.Default block size is 64MB, though most folks now set it to 128MB
- Note: these are not simple queries, they are DEEP COMPLEX SCANS
- Note: these are not simple queries, they are DEEP COMPLEX SCANS
- Theme: they through away a lot of valuable gas and oil just like we through away data today
- Apache Hadoop is a new solution in your existing infrastructure.It does not replace any existing major existing investment.Apache brings data that you’re already generating into context and integrates it with your business.You get access to key information about how your business is operating but pulling togetherWeb and application logsUnstructured filesWeb dataRelational dataHadoop is used by your team to analyze this data and deliver it to business users directly and via existing data management technologies
- Allows vendor specific solutions to build on topCertain custom code would be written for the platform as well
- Theme: they through away a lot of valuable gas and oil just like we through away data today
- Transition into “hadoop as a platform”, but vendors are building on top of it for specific vertical challenges
- This is probably a better slide 30
- Check this against the Mahout impl