![MapReduce
• “Single Threaded” MapReduce:
cat input/* | map | sort | reduce > output
• Map program parses the input and emits [key,value] pairs
• Sort by key
• Reduce computes output from values with same key
Reduce
Map Sort
• Extrapolate to PB of data on thousands of nodes
© 2010 Journey Dynamics Ltd
15](https://image.slidesharecdn.com/hadoop-accu2010-100416011416-phpapp01/85/Introduction-to-Hadoop-ACCU2010-16-320.jpg)
Introduction to Hadoop - ACCU2010
- 1. MapReduce with Apache Hadoop Analysing Big Data April 2010 Gavin Heavyside gavin.heavyside@journeydynamics.com
- 2. About Journey Dynamics • Founded in 2006 to develop software technology to address the issues of congestion, fuel efficiency, driving safety and eco-driving • Based in the Surrey Technology Centre, Guildford, UK • Analyse large amounts (TB) of GPS data from cars, vans & trucks • TrafficSpeedsEQ® - Accurate traffic speed forecasts by hour of day and day of week for every link in the road network • MyDrive® - Unique & sophisticated system that learns how drivers behave Drivers can improve fuel economy Insurance companies can understand driver risk Navigation devices can improved route choice & ETA Fleet managers can monitor their fleet to improve safety & eco-driving © 2010 Journey Dynamics Ltd 2
- 3. Big Data • Data volumes increasing • NYSE: 1TB new trade data/day • Google: Processes 20PB/day (Sep 2007) http://tcrn.ch/agYjEL • LHC: 15PB data/year • Facebook: several TB photos uploaded/day © 2010 Journey Dynamics Ltd 3
- 4. “Medium” Data • Most of us arenʼt at Google or Facebook scale • But: data at the GB/TB scale is becoming more common • Outgrow conventional databases • Disks are cheap, but slow • 1TB drive - £50 • 2.5 hours to read 1TB at 100MB/s © 2010 Journey Dynamics Ltd 4
- 5. Two Challenges • Managing lots of data • Doing something useful with it © 2010 Journey Dynamics Ltd 5
- 6. Managing Lots of Data • Access and analyse any or all of your data • SAN technologies (FC, iSCSI, NFS) • Querying (MySQL, PostgreSQL, Oracle) ➡ Cost, network bandwidth, concurrent access, resilience ➡ When you have 1000s of nodes, MTBF < 1 day © 2010 Journey Dynamics Ltd 6
- 7. Analysing Lots of Data • Parallel processing • HPC • Grid Computing • MPI • Sharding ➡ Too big for memory, specialised HW, complex, scalability ➡ Hardware reliability in large clusters © 2010 Journey Dynamics Ltd 7
- 8. Apache Hadoop • Reliable, scalable distributed computing platform • HDFS - high throughput fault-tolerant distributed file system • MapReduce - fault-tolerant distributed processing • Runs on commodity hardware • Cost-effective • Open source (Apache License) © 2010 Journey Dynamics Ltd 8
- 9. Hadoop History • 2003-2004 Google publishes MapReduce & GFS papers • 2004 Doug Cutting add DFS & MapReduce to Nutch • 2006 Cutting joins Yahoo!, Hadoop moves out of Nutch • Jan 2008 - top level Apache project • April 2010: 95 companies on PoweredBy Hadoop wiki • Yahoo!, Twitter, Facebook, Microsoft, New York Times, LinkedIn, Last.fm, IBM, Baidu, Adobe "The name my kid gave a stuffed yellow elephant. Short, relatively easy to spell and pronounce, meaningless, and not used elsewhere: those are my naming criteria. Kids are good at generating such. Googol is a kid's term" Doug Cutting © 2010 Journey Dynamics Ltd 9
- 10. Hadoop Ecosystem • HDFS • MapReduce • HBase • ZooKeeper • Pig • Hive • Chukwa • Avro © 2010 Journey Dynamics Ltd 10
- 11. Anatomy of a Hadoop Cluster Namenode Tasktracker Tasktracker Tasktracker Tasktracker Datanode Rack 1 Datanode Datanode Datanode JobTracker Tasktracker Tasktracker Tasktracker Tasktracker Datanode Datanode Rack n Datanode Datanode © 2010 Journey Dynamics Ltd 11
- 12. HDFS • Reliable shared storage • Modelled after GFS • Very large files • Streaming data access • Commodity Hardware • Replication • Tolerate regular hardware failure © 2010 Journey Dynamics Ltd 12
- 13. HDFS • Block size 64MB • Default replication factor = 3 1 2 HDFS 1 2 3 4 5 3 2 3 4 5 1 4 3 4 5 1 2 5 © 2010 Journey Dynamics Ltd 13
- 14. HDFS • Block size 64MB • Default replication factor = 3 1 2 HDFS 1 2 3 4 5 3 2 3 4 5 1 4 3 4 5 1 2 5 © 2010 Journey Dynamics Ltd 13
- 15. MapReduce • Based on 2004 Google paper • Concepts from Functional Programming • Used for lots of things within Google (and now everywhere) • Parallel Map => Shuffle & Sort => Parallel Reduce • Easy to understand and write MapReduce programs • Move the computation to the data • Rack-aware • Linear Scalability • Works with HDFS, S3, KFS, file:// and more © 2010 Journey Dynamics Ltd 14
- 16. MapReduce • “Single Threaded” MapReduce: cat input/* | map | sort | reduce > output • Map program parses the input and emits [key,value] pairs • Sort by key • Reduce computes output from values with same key Reduce Map Sort • Extrapolate to PB of data on thousands of nodes © 2010 Journey Dynamics Ltd 15
- 17. MapReduce • Distributed Example sort Split 0 HDFS Map copy merge sort Split 1 part 0 HDFS Map Reduce HDFS merge part 1 Reduce HDFS sort Split n HDFS Map © 2010 Journey Dynamics Ltd 16
- 18. MapReduce can be good for: • “Embarrassingly Parallel” problems • Semi-structured or unstructured data • Index generation • Log analysis • Statistical analysis of patterns in data • Image processing • Generating map tiles • Data Mining • Much, much more © 2010 Journey Dynamics Ltd 17
- 19. MapReduce is not be good for: • Real-time or low-latency queries • Some graph algorithms • Algorithms that canʼt be split into independent chunks • Some types of joins* • Not a replacement for RDBMS * Can be tricky to write unless you use an abstraction e.g. Pig, Hive © 2010 Journey Dynamics Ltd 18
- 20. Writing MapReduce Programs • Java • Pipes (C++, sockets) • Streaming • Frameworks, e.g. wukong(ruby), dumbo(python) • JVM languages e.g. JRuby, Clojure, Scala • Cascading.org • Cascalog • Pig • Hive © 2010 Journey Dynamics Ltd 19
- 21. Streaming Example (ruby) • mapper.rb • reducer.rb © 2010 Journey Dynamics Ltd 20
- 22. Pig • High level language for writing data analysis programs • Runs MapReduce jobs • Joins, grouping, filtering, sorting, statistical functions • User-defined functions • Optional schemas • Sampling • Pig Latin similar to imperative language, define steps to run © 2010 Journey Dynamics Ltd 21
- 23. Pig Example © 2010 Journey Dynamics Ltd 22
- 24. Hive • Data warehousing and querying • HiveQL - SQL-like language for querying data • Runs MapReduce jobs • Joins, grouping, filtering, sorting, statistical functions • Partitioning of data • User-defined functions • Sampling • Declarative syntax © 2010 Journey Dynamics Ltd 23
- 25. Hive Example © 2010 Journey Dynamics Ltd 24
- 26. Getting Started • http://hadoop.apache.org • Cloudera Distribution (VM, source, rpm, deb) • Elastic MapReduce • Cloudera VM • Pseudo-distributed cluster © 2010 Journey Dynamics Ltd 25
- 27. Learn More • http://hadoop.apache.org • Books • Mailing Lists • Commercial Support & Training, e.g. Cloudera © 2010 Journey Dynamics Ltd 26
- 28. Related • Cassandra 0.6 has Hadoop integration - run MapReduce jobs against data in Cassandra • NoSQL DBs with MapReduce functionality include CouchDB, MongoDB, Riak and more • RDBMS with MapReduce include Aster, Greenplum, HadoopDB and more © 2010 Journey Dynamics Ltd 27
- 29. Gavin Heavyside gavin.heavyside@journeydynamics.com www.journeydynamics.com © 2010 Journey Dynamics Ltd 28
Editor's Notes
- #4: Andrei Alexandrescu - 1300ish photos/second at facebook
- #7: Use only subset/sample of data A server might have a MTBF of a few years With thousands you can expect failures at least daily System needs to be handle HW failures
- #9: Explain what we mean by commodity hardware 8-core, 16-24GB ram, 4x1TB disks, gig ethernet
- #10: 95 Companies not exhaustive!
- #11: Zookeeper - distributed synchronisation service Locking, race conditions. Used for HBase distributed column-oriented database Chukwa is large-scale log collection & analysis tools Avro - like protocol buffers - move towards default IPC mechanism in Hadoop
- #12: Mention secondary namenode Namenode is point of failure - not highly available Mention lots of disks - JBOD: don&#x2019;t stripe, don&#x2019;t RAID Rack-aware - splits will be processed local to the HDFS blocks where possible. Pseudo distributed cluster can run all services on single machine DN/TT on lots of machines, NN, SNN, JT depend on cluster size.
- #13: Optimised for large (GB) files 64MB block size not efficient for lots of small files -> concatenate them
- #15: Text data Binary data (protocol buffers, avro) Custom input/output formats
- #17: Compare with previous example, splits on each node, shuffle to reduce nodes Talk about partitioning Reducers have to wait until the map and sort/shuffle has finished Combiners can run on Tasktracker Can be single Reducer or many
- #18: Mention Speculative Execution Forward index: list of words per document Inverted Index: list of documents containing word Once you &#x201C;get&#x201D; it, you start to see how lots of problems can be parallelised Map Tiles at Google - Tech Talk on YouTube
- #19: Algorithm has a strict data dependency between iterations? Might not be efficient to parallelise with MapReduce
- #20: JVM languages can access the Hadoop Java classes Streaming works in any language that can read/write stdin/stdout Cascalog brand new; only heard about it last night
- #21: Implements a map-side hash join
- #22: SAMPLE command samples random parts of data SPLIT AVG, COUNT, CONCAT, MIN, MAX, SUM Custom input/output formats
- #24: Hive developed at Facebook, contributed back to community SQL is a declarative language, you specify an outcome, and the query planner generates a sequence of steps. Ad-hoc queries using basic SQL Custom input/output formats
- #25: Create Table Load Data Select/Join/Group
- #27: Cloudera do developer training and sysadmin training Cloudera do commercial support Other providers on Hadoop homepage
- #28: Cassandra is an alternative to HBase - seems to have more momentum behind it. Plenty of hobbyist/small-scale MapReduce frameworks out there