Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Olap scalability


Published on

Published in: Technology

Olap scalability

  1. 1. Online Analytical Processing of LargeDistributed Databases Luc Boudreau Lead Engineer, Pentaho Corporation
  2. 2. "its all about data movement and operating on that data on the fly"
  3. 3. a relational database
  4. 4. Relational Databases● Static schema● Minimized redundancy● Referential integrity● Transactional
  5. 5. Classic RDBMS internals● "Shared Everything" paradigm PLANNER / SCHEDULER● Private Planner PROCESSOR PROCESSOR PROCESSOR● Multiple private processors● Multiple private data stores
  6. 6. What RDBMS are for● Operational data● Normalized models● Static typed data
  7. 7. What RDBMS are NOT for● "Full Scan" Aggregated Computations● Multi-dimensional queries (think pivot)● Unstructured data
  8. 8. OK so hows thatdifferent from Big Data platforms?
  9. 9. Big Data - More than a buzzword(although sometimes its hard to tell...) Big Data is not a product. It is an architecture.
  10. 10. Big Data - More than a buzzword(although sometimes its hard to tell...) A schema-less distributed storage and processing model for data.
  11. 11. Big Data● Schema less ○ Programmatic queries ○ "Map" of MapReduce● High Redundancy ○ Distributed processing ○ "Reduce" of MapReduce
  12. 12. Big Data● No referential integrity● Non transactional● High latency
  13. 13. Classic Big Data internals● "Share nothing" paradigm SCHEDULER● Push the processing closer to the data PROCESSOR PROCESSOR PROCESSOR● The query defines the schema
  14. 14. What Big Data is for● Unstructured data keep everything● Distributed file system great for archiving● Data is fixed only the process evolves
  15. 15. What Big Data is for● Ludicrous amounts of data keep everything, remember?● Made on the cheap each processing unit is commodity hardware
  16. 16. What Big Data is NOT for● Low latency applications arbitrary exploration of the data is close to impossible● End-users writing code is easy. writing good code is hard.● Replacing your operational DB
  17. 17. Some more limitations● No structured query language exploration is tedious● Accuracy & Exactitude the burden is put on the end user / query designer● No query optimizer cannot optimize at runtime. does exactly what you tell it to.
  18. 18. why is this so similar to NoSQL?
  19. 19. First, defining NoSQL...● NoSQL: The thing named after what it lacks which has as many definitions as there are products. (which usually turns out to be some sort of key-value store)
  20. 20. Why "NoSQL"? Why all the hate?!● Historical reasons ○ Wrong technological choices ○ Blind faith in RDBMS scalability ○ General wishful thinking and voodoo magic
  21. 21. Why "NoSQL"? Why all the hate?!● "SQL" itself was never the issue● NoSQL projects are implementing SQL- like query languages
  22. 22. bringing structuredqueries to Big Data
  23. 23. Current efforts● Straight SQL implementations Greenplum: Straight SQL on top of Big Data Hive JDBC: A hybrid of DSL & SQL● The Splunk approach SQL with missing columns● Runtime query optimizers Optiq framework: SQL with Big Data federated sources
  24. 24. isnt there something better than SQL for analytics?
  25. 25. Online AnalyticalProcessing (OLAP)
  26. 26. Widely used. Little known.● Your favorite corporate dashboards● Google Analytics & other ad-hoc tools
  27. 27. Analytics centric language● Multidimensional Expressions (MDX) a powerful query language for analytics● Forget about rows and columns as many axis as you need● Slice & dice start from everything - progressively focus only on relevant data
  28. 28. Business domain driven● Hierarchical view of a multidimensional universe
  29. 29. An exampleWhat are my total sales for the current year, per month, for male customers? with member [Measures].[Accumulated Sales] as Sum(YTD(), [Measures].[Store Sales]) select {[Measures].[Accumulated Sales]} on columns, {Descendants([Time].[1997], [Time].[Month])} on rows from [Sales] where ([Customer].[Gender].[M])
  30. 30. how does that work?
  31. 31. Analytics data modelization● A denormalized model for performance the data is modelized for read operations - not write● High redundancy because sometimes more is better
  32. 32. The Star model
  33. 33. The Snowflake model
  34. 34. different OLAP servers. Different beasts.
  35. 35. Relational OLAP (ROLAP)● Backed by a relational database think of a MDX to SQL bridge. the aggregated data can be cached in-memory or on-disk.● Relies heavily on the RDBMS performance figures out at runtime the proper optimizations
  36. 36. Memory OLAP (MOLAP)● Loads everything in RAM● Relies on an efficient ETL platform
  37. 37. Other OLAP● On-disk aggregated data files Think SAS. Cubes are compiled into data files on disk.● Simple Bridges Converts MDX straight to SQL, with limited support of MDX syntax.
  38. 38. how do they compare?
  39. 39. (there are no straight answers, sorry)
  40. 40. Where the data lives matters Location Speed (ns) L1 Cache Reference 0.5 Branch Mispredict 5 L2 Cache Reference 7 Mutex lock/unlock 25 Main memory reference 100 Compress 1K bytes w/ cheap algorithm 3000 Send 2K bytes over 1 Gbps network 20 000 Read 1 MB sequentially from memory 250 000 Round trip within same datacenter 500 000 Disk seek 10 000 000 Read 1 MB sequentially from disk 20 000 000 Send packet CA -> Netherlands -> CA 150 000 000
  41. 41. Optimizing for CPU● Java NIO blocks use extremely compact chunks of 64 bits.● Primitive types use "int" instead of "Integer"● BitKeys because they are naturally CPU friendly
  42. 42. Optimizing for memory● Hard limits on the heap space must pay attention to the total memory usage.● Inherent limitations there can only be so many individual pointers on heap.
  43. 43. Optimizing for networking● Payload optimization batching. deltas.● Manageability turning nodes on & off.
  44. 44. Optimizing for disk● Concurrent access must carefully manage disk IO.● Inherently slooooow
  45. 45. how to deal with these issues?
  46. 46. a scalable indexing strategy
  47. 47. Cache indexing● Linear performance is not good enough as N grows, full scanning takes O(n)● The rollup combinatorial problem as the cache grows, reuse becomes tedious
  48. 48. The rollup combinatorial problem Gender Country Sales M USA 7 M CANADA 8 F USA 4 F CANADA 2 Country Sales USA 11 CANADA 10
  49. 49. The rollup combinatorial problem Gender Country Sales Gender Country Sales City Sales M USA 7 F USA 5 Montreal 6 M CANADA 8 Quebec 1 Age Country Sales Ottawa 8 Age Country Cost 16 - 25 USA 2 Vancouver 2 41 - 56 USA 5 26 - 40 CANADA 3 Toronto 5 26 - 40 USA 5 Country Sales ? ? ? ?
  50. 50. PoSet & BitKeys● Represent the levels / values as bitkeys because bitkeys are fast, remember?● The PartiallyOrderedSet a hierarchical hash set where elements might or might not be related to one another.
  51. 51. PoSet & BitKeys● An example application finding all primes in a set of integers
  52. 52. a scalable threading model
  53. 53. Concurrent cache access● Usage of phases peek -> load -> rinse & repeat● A scalable threading model thread safety without locks and blocks
  54. 54. A scalable threading model● Do things once. Do them right. the actor pattern
  55. 55. a scalable cachemanagement strategy
  56. 56. Operating by deltas● All part of a whole implicit relation between the dimensions● Why deltas are necessary reducing IO
  57. 57. Cache management● A data block is a complex object Schema:[FoodMart] Checksum:[9cca66327439577753dd5c3144ab59b5] Cube:[Sales] Measure:[Unit Sales] Axes:[ {time_by_day.the_year=(*)} {time_by_day.quarter=(Q1, Q2)} {product_class.product_family=(Bread, Soft Drinks)}] Excluded Regions:[ {time_by_day.quarter=(Q1)} {time_by_day.the_year=(1997)}] Compound Predicates:[] ID:[9c8ba4ec39678526f4100506994c384183cd205d19dd142eae76a9fb1d74cab7]
  58. 58. a scalable sharing strategy
  59. 59. Shared Caches● OLAP and key-value stores dont like each other OLAP requires a complex key. a hash is insufficient.● Remember the "deltas" strategy? partially invalidating a block of data would break the hash
  60. 60. Data grids & OLAP● Well suited for OLAP caches supports "rich" keys● Distributed and redundant if a node goes offline, the cache data is not lost● In-memory grids are fast multiplies the available heap space
  61. 61. a case study
  62. 62. Advertising data analysis Interactive behavioral targeting of advertising in real time
  63. 63. Advertising data analysis● Low latency the end users dont want to wait for MapReduce jobs● Scalability a huge factor were talking petabytes of data here
  64. 64. Advertising data analysis● Queries are not static we cant tell upfront what will be computed● Deployed in datacenters worldwide the hashing strategy must allow "smart" data distribution● Almost all open source
  65. 65. Monitoring & ETL Designer Client App Management olap4j Load Balancer OLAP XML/A Cache olap4j Logs ETL Analytical OLAP Logs DBBig ETLDataStore Logs ETL Logs Message ETL Queue
  66. 66. Client App● A query olap4j - UI sends MDX to a SOAP service. - load balancer dispatches the query. - OLAP layer uses its data sources and aggregates. Load - query is answered Balancer OLAP XML/A Cache olap4j Analytical OLAP DB
  67. 67. ● An update - Strategy #1 - the ETL process updates the analytical DB. - a cache delta is sent to a message queue. - OLAP processes the message. - OLAP uses its index to spot the regions to invalidate. - aggregated cache is updated incrementally. OLAP Cache Logs ETL Analytical OLAP Logs DB Big ETL Data Store Logs ETL Logs Message ETL Queue
  68. 68. ● An update - Strategy #2 - ETL updates the analytical DB. - ETL acts directly on the OLAP cache. - OLAP processes events from its cache. - OLAP updates its index OLAP Cache Logs ETL Analytical OLAP Logs DB Big ETL Data Store Logs ETL Logs ETL
  69. 69. a stack built on open standards (get ready, the next slide will hurt your brains)
  70. 70. Java Client App load balancer Client App olap4j-xmla olap4j-xmla HTTP (XMLA) olap4j server olap4j server olap4j server olap4j olap4j olap4j jdbc jdbc jdbc JDBC connection connection connection pool pool pool jdbc jdbc jdbc olap4j impl olap4j impl olap4j impl Mondrian Mondrian Mondrian server server server manager manager manager Java Mondrian Mondrian Mondrian cache cache cache manager manager manager infinispan infinispan infinispanUDP (Hot Rod) infinispan data grid
  71. 71. the UI
  72. 72. Yahoo! Cocktails● A Node.js implementation runs on Manhattan JS hosted execution Client App● Mojito client application framework● Works both online / offline
  73. 73. the OLAP service
  74. 74. olap4j-xmla / olap4j-server Client App olap4j● JDBC for OLAP extension to JDBC. became the de facto standard. Load● A Java toolkit for OLAP Balancer - MDX parser / validator - a rich type system / MDX object model - driver specification - programmatic query models XML/A - olap4j to XMLA bridge olap4j
  75. 75. the OLAP layer
  76. 76. Mondrian● Developed by Pentaho Corp. used worldwide. pure java. open source. OLAP● Highly extensible exposes many APIs & SPIs for enterprise integration.● ROLAP / MOLAP hybrid uses the best of whats available.● Extensible MDX parser new MDX functions can be created for specific business domains.
  77. 77. the OLAP cache
  78. 78. Stuff that didnt work● memcached ○ doesnt have an index. ○ enforces random TTLs. OLAP Cache ○ a hash key is not enough● simple Java collections
  79. 79. Infinispan● Developed for JBoss AS well tested. OLAP Cache● UDP Multicast nodes can join and leave the cluster as needed.● Can distribute the processing jobs can be distributed and ran on the nodes.● Serializes rich objects the contents can be read from APIs.
  80. 80. the analytical DB layer
  81. 81. Oracle● Cluster of instances partitioned Oracle nodes Analytical● Why Oracle? DB because their DBAs are good enough with Oracle to get it to run properly under such a load
  82. 82. Other options● An analytical oriented DB use of Vectorwise, Vertica, MonetDB, Greenplum, ... Analytical● Column stores DB Column stores scale marvelously and are well suited for analytics
  83. 83. the Big Data layer
  84. 84. Big Data Layer● Homebrew Java MapReduce Logs ETL Logs Big ETL● 42 000 nodes Data Store Logs ETL Logs● ETL processes managed ETL with Pig● A keynote in itself (see the resources at the end for a keynote from Scott Burke, Senior VP of Yahoo!)
  85. 85. some numbers
  86. 86. Final processing capacity● Big Data layer ○ 140 petabytes ○ 500 users ○ 42 000 nodes ○ 10 000 000 hours of CPU time usage per day ○ 100 000 000 000 records per day
  87. 87. Final processing capacity● Analytical DB layer ○ 50 terabytes ○ 100s of tables (heavy use of the snowflake schema) ○ 1 000 000 000 new rows per day
  88. 88. Final processing capacity● OLAP layer ○ 10s of Mondrian instances ○ 10s of cubes ○ 100s of dimensions ○ 1 000s of levels ○ 1 000 000s of members per level ○ 1 000 000 000s of facts per day
  89. 89. skunkworks(future stuff you might care about)
  90. 90. Mondrian over Googles BigQuery● Big Data as a service upload CSVs & other formats to a ad-hoc cluster● No code required MapReduce jobs usually require you to code them
  91. 91. Pentaho Instaview● Interactive data discovery for Big Data fully integrated ETL / OLAP. all you need is a URL and a user / password.● A rich UI environment for data drag & drop. full OLAP support. mobile.● Open source
  92. 92. resources Mondrian - The open source analytics engine olap4j - The open standard for OLAP in Java olap4j.orgInfinispan - The distributed data grid platform Burke, SVP Advertising & Data @ Yahoo! Keynote of Hadoop Summit 2012
  93. 93. resources Pentaho
  94. 94. big thanks On Twitter: @luclemagnifiqueOn the blogosphere: