Big Data Platforms: An Overview
109 likes•28,301 views
The document provides an overview of big data platforms and the concept of 'big data,' which refers to datasets that are too large for traditional database tools to manage effectively. It explains different types of big data platforms, particularly focusing on NoSQL technologies, their advantages, disadvantages, and the CAP theorem that outlines the trade-offs in database design. The document also categorizes various NoSQL approaches such as key-value stores, document stores, and columnar databases, detailing their functionality and use cases.
Big Data Platforms: An Overview
- 1. Big Data Platforms: An Overview C. Scyphers Chief Technical Architect Daemon Consulting, LLC
- 2. What Is “Big Data”? • Big Data is not simply a huge pile of information • A good starting place is the following thought: “Big Data describes datasets so large they become very difficult to manage with traditional database tools.”
- 3. What Is A Big Data Platform? Putting it simply, it is any platform which supports those kind of large datasets.
- 4. It doesn’t have to be cutting edge technology.
- 5. Lots of legacy technologies can address the problem.
- 6. If only at a sizeable cost.
- 7. SQL Of the new technologies, the most promising are from the “NoSQL” family.
- 8. What Is “NoSQL”? SQL A family of non-relational data storage technologies
- 12. Usually with less cost than more traditional approaches.
- 13. Some of these technologies are new and innovative
- 14. Others have been around for decades.
- 15. NoSQL Does Not Mean “SQL Is Bad” When the trend was just starting, “NoSQL” was coined. It’s unfortunate, because it implies antagonism towards SQL.
- 16. NoSQL Means “Not Only SQL” RELATIONAL RELATIONAL NON- NoSQL is a complement to a traditional RDBMS, not necessarily as a replacement of them.
- 17. Why Won’t SQL Do?
- 18. Scale is very hard without ridiculous expense
- 19. SQL can get very complex, very quickly
- 20. Changing a schema for a large production system is both risky and expensive
- 21. Throughput can be a challenge
- 22. How Does NoSQL Do It?
- 23. Scale is achieved through a shared-nothing architecture, removing bottlenecks
- 24. Schemaless design means change becomes much less risky and significantly cheaper
- 25. Most solutions use simple RESTful interfaces
- 26. NoSQL is based upon a better understanding of data storage, usually referred to as the “CAP Theorem”
- 27. The CAP Theorem Grossly simplified (with apologies to Brewer): A database can be • Consistent (All clients see the same data) • Available (All clients can find some available node) • Partition-Tolerant (the database will continue to function even if split into disconnected sets – e.g. a network disruption) Pick Any Two.
- 28. CAP In Practice • Consistent & Available (no Partition Tolerance) • Either single machines or single site clusters. • Typically uses 2 phase commits
- 29. CAP In Practice • Consistent & Partition Tolerant (no Availability) • Some data may be inaccessible, but the remainder is available and consistent • Sharding is an example of this implementation Customers Customers Customers A-F G-R S-Z
- 30. CAP In Practice • Available & Partition Tolerant (no Consistency) • Some data may be inaccurate; a conflict resolution strategy is required. • DNS is an example of this, as well as standard master-slave replication
- 31. CAP From A Vendor POV • C-A (no P) – this is generally how most RDBMS vendors operate • C-P (no A) – this is how many RDBMS’ attempt to address scale without incurring large costs • A-P (no C) – this is how most NoSQL approaches solve the problem
- 32. ACID vs BASE Traditional Databases NoSQL Databases Tend Are ACID Compliant To Be BASE Compliant Atomicity – either the entire transaction Basically completes or none of it does Consistent – any transaction will take the Available database from one consistent state to another, with no broken constraints Isolation – changes do not affect other users Scalable until committed Durability – committed transactions can be Eventually consistent recovered in case of system failure Eventually consistent is the key phrase here
- 33. SQL Strengths Very well known technology
- 36. Large talent pool from which to choose
- 37. Ad hoc operations common, if not encouraged
- 38. NoSQL Strengths Built to address massive scale
- 40. While remaining highly available
- 41. And handling unstructured data
- 42. NoSQL Pros/Cons Pros Cons • Schema Evolution • Querying the data is • Horizontal Scalability much harder • Simple Protocols • Paradigm Shift • Security is a big issue • May or may not support data types (BLOBs, spatial) • Generally, uniqueness cannot be enforced
- 43. A Disclaimer Before We Continue • I am not an expert on every possible Big Data Platform • There are hundreds of them; these are the ones I consider the leaders in the field and recommend • If you have a favorite, please let me know and I’ll update the deck for next time • The internal details on how these systems work are rather complex; I would prefer to take those questions offline
- 44. Flavors Of NoSQL The major four divisions of NoSQL are: • Key-Value • Document Store • Columnar • Other
- 45. Key-Value • At a very high level, key-value works essentially by pairing a index token (a key) with a data element (a value). • Both index token and the data value can be of any structure. • Such a pairing is arbitrary and up to the developer of the system to determine.
- 46. A Key-Value Example “John Smith”, “100 Century Dr. Alexandria VA 22304” “John Doe”, “16 Kozyak Street, Lozenets District, 1408 Sofia Bulgaria” In both examples, the key is a name and the value is an address. However, the structure of the address differs between the two.
- 47. Document Store • Document stores extend the key-value paradigm into values with multiple attributes. • The document values tend to be semi-structured data (XML, JSON, et al) but can also be Word or PDF documents.
- 48. A Document Store Example “John Smith”, “<address><street>100 Century Dr.</street> <city>Alexandria</city> <state>VA</state> <postalCode>22304</postalCode> </address>” “John Doe”, “{ “address”: { “street”: “16 Kozyak Street” “district”: “Lozenets, 1408” “city”: “Sofia” “country”: “Bulgaria” } }”
- 49. Columnar Family • Usually has “rows” and “columns” • Or, at least, their logical equivalents • Not a traditional, “pure” column store • More of a hybridized approach leveraging key-value pairs • A key with many values attached
- 50. The Others • Hierarchical Databases • LDAP, Active Directory • Graph Databases • Neo4j, Flock DB, InfiniteGraph • XML • MarkLogic • Object Oriented Databases • Versant • Lotus Notes • HPCC (LexisNexis)
- 51. Key-Value Recap Pairing a index token (a key) with a data element (a value)
- 52. Key-Value Pro/Con Pros Cons • Schema Evolution • Packing & unpacking each key • Horizontal Scalability • Keys typically are not related • Simple Protocols to each other • Works well for volatile data • The entire value must be • High throughput, typically returned, not just a part of it optimized for reads or writes • Security tends to be an issue • Keys become meaningful • Hard to support reporting, rather than arbitrary analytics, aggregation or • Application logic defines ordered values object model • Generally does not support updates in place • Application logic defines object model
- 53. Where Did Key-Value Come From? The concept is quite old, but most people trace the lineage back to Amazon and the Dynamo paper.
- 54. Dynamo Amazon devised the Dynamo engine as a way to address their scalability issues in a reliable way. • Communication between nodes is peer to peer (P2P) • Replication occurs with the end client addressing conflict resolution • Quorum Reads/Writes • Always writable (Hinted Handoff) • Eventually Consistent
- 55. Eventually Consistent • Rather than expending the runtime resources to ensure that all nodes are aware of a change before continuing, Dynamo uses an eventually consistent model. • In this model, a subset of nodes are changed • Those nodes then inform their neighbors until all nodes are changed (grossly simplifying).
- 56. Can I Use Dynamo? No. It’s an Amazon only internal product. However, AWS S3 is largely based upon it. Amazon did announce a DynamoDB offering for their AWS customers. While it’s probably the same, I cannot guarantee that it is.
- 57. Riak • Riak is a key-value database largely modeled after the Dynamo model. • Open source (free) with paid support from Basho. • Main claims to fame: • Extreme reliability • Performance speed
- 58. Riak Pro/Con Pros Cons • All nodes are equal – no • Not meant for small, discrete single point of failure and numerous datapoints. • Horizontal Scalability • Getting data in is great; • Full Text Search getting it out, not so much • RESTful interface (and HTTP) • Security is non-existent: • Consistency level tunable on “Riak assumes the internal each operation environment is trusted” • Secondary indexes available • Conflict resolution can bubble • Map/Reduce (JavaScript & up to the client if not careful. Erlang only) • Erlang is fast, but it’s got a serious learning curve.
- 59. Riak Users
- 60. Redis • Redis is a key-value in-memory datastore. • Open source (free) with support from the community. • Main claims to fame: • Fast. So very, very fast. • Transactional support • Best for rapidly changing data
- 61. Redis Pro/Con Pros Cons • Transactional support • Entirely in memory • Blob storage • Master-slave replication • Support for sets, lists and (instead of master-master) sorted sets • Security is non-existent: • Support for Publish-Subscribe designed to be used in (Pub-Sub) messaging trusted environments • Robust set of operators • Does not support encryption • Support can be hard to find
- 62. Redis Users
- 63. Voldemort • Voldemort is a key-value in-memory database built by LinkedIn. • Open source (free) with support from the community • Main claims to fame: • Low latency • Highly Available • Very fast reads
- 64. Voldemort Pro/Con Pros Cons • Highly customizable – each • Versioning means lots of disk layer of the stack can be space being used. replaced as needed • Does not support range • Data elements are versioned queries during changes • No complex query filters • All nodes are independent – • All joins must be done in no single point of failure code • Very, very fast reads • No foreign key constraints • No triggers • Support can be hard to find
- 65. Voldemort Users
- 66. Key/Value “Big Vendors” • Microsoft Azure Table Storage • Oracle NoSQL • BerkleyDB (Oracle)
- 67. Document Store Recap Document stores store an index token with a grouping of attributes in a semi- structured document
- 68. Document Store Pro/Con Pros Cons • Tends to support a more • The entire value must be complex data model than returned, not just a part of it key/value • Security tends to be an issue • Good at content • Joins are not available within management the database • Usually supports multiple • No foreign keys indexes • Application logic defines • Schemaless (can be nested) object model • Typically low latency reads • Application logic defines object model
- 69. CouchDB • CouchDB is a document store database. • Open source (free), part of the Apache foundation with paid support available from several vendors. • Main claims to fame: • Simple and easy to use • Good read consistency • Master-master replication
- 70. CouchDB Pro/Con Pros Cons • Very simple API for • The simple API for development development is somewhat • MVCC support for read limited consistency • No foreign keys • Full Map/Reduce support • Conflict resolution devolves • Data is versioned to the application • Secondary indexes supported • Versioning requires extensive • Some security support disk space • RESTful API, JSON support • Versioning places large load • Materialized views with on I/O channels incremental update support • Replication for performance, not availability
- 71. CouchDB Users
- 72. MongoDB • MongoDB is a document store database. • Open source (free) with paid support available from 10Gen. • Main claims to fame: • Index anything • Ad hoc query support • SQL like operations (not SQL syntax)
- 73. MongoDB Pro/Con Pros Cons • Auto-sharding • Does not support JSON: BSON • Auto-failover instead • Update in place • Master-slave replication • Spatial index support • Has had some growing pains • Ad hoc query support (e.g. Foursquare outage) • Any field in Mongo can be • Not RESTful by default indexed • Failures require a manual • Very, very popular (lots of database repair operation production deployments) (similar to MySQL) • Very easy transition from SQL • Replication for availability, not performance
- 74. MongoDB Users
- 75. Document Store “Big Vendors” • Lotus Domino
- 76. Columnar Family Recap • A key with many values attached • Usually presenting as “rows” and “columns” • Or, at least, their logical equivalents
- 77. Columnar Pro/Con Pros Cons • Tend to have some level of • Is much less efficient when rudimentary security support processing many columns • Usually include a degree of simultaneously versioning • Joins tend to not be • Can be more efficient than supported row databases when • Referential integrity not processing a limited number available of columns over a large amount of rows
- 78. Where Did Columnar Come From? The concept has been around for a while, but most people trace the NoSQL lineage back to Google.
- 79. BigTable Google devised the BigTable engine as a way to address their search related scalability issues in a reliable way. • Data is organized through a set of keys: • Row • Column • Timestamp • A hybrid row/column store with a single master • Versioning is handled through the time key • Tablets are a dynamic partition of a sequence of rows – supports very efficient range scans • Columns can be grouped into column families • Column families can have access control
- 80. Can I Use BigTable? No. It’s a Google only internal product. However, quite a few open source products are built upon the concepts.
- 81. Cassandra • Cassandra is a hybrid of Big Table built on Dynamo infrastructure • Open source (free), built by Facebook with paid support available from several vendors. • Main claims to fame: • An Apache project • Very, very fast writes • Spans multiple datacenters
- 82. Cassandra Pro/Con Pros Cons • Designed to span multiple • No joins datacenters • No referential integrity • Peer to peer communication • Written in Java – quite between nodes complex to administer • No single point of failure and configure • Always writeable • Last update wins • Consistency level is tunable at run time • Supports secondary indexes • Supports Map/Reduce • Supports range queries
- 83. Cassandra Users
- 84. HBase • Hbase is a columnar database built on top of the Hadoop environment. • Open source (free) with paid support from numerous vendors • Main claims to fame: • Ad hoc type abilities • Easy integration with Map/Reduce
- 85. HBase Pro/Con Pros Cons • Map/Reduce support • Secondary indexes generally • More of a CA approach and not supported an AP • Security is non-existent • Supports predicate push • Requires a Hadoop down for performance gains infrastructure to function • Automatic partitioning and rebalancing of regions • Data is stored in a sorted order (not indexed) • RESTful API • Strong and vibrant ecosystem
- 86. HBase Users
- 87. Hadoop • Hadoop is not a columnar store as such. • Rather, Hadoop is a massively parallel data processing engine • Main claims to fame: • Specializes in unstructured data • Very flexible and popular
- 88. Hadoop Pro/Con Pros Cons • While written in Java, almost • Large amounts of disk space any language can leverage and bandwidth required Hadoop • Paradigm shift for IT staff • Runs on commodity servers • Quality talent is highly in • Horizontally scalable demand and expensive • Very fast and powerful • Security is non-existent • Where Map/Reduce • Name node is a single point originated of failure • Ample support from vendors • More or less only supporting • “Helper” languages like Hive batch processing and Pig • Not user friendly to anyone • Strong and vibrant ecosystem other than developers
- 89. Hadoop Users Plus lots, lots more
- 90. Columnar “Big Vendor” • EMC Greenplum • Teradata Aster In so far as both of these solutions are grafting Map/Reduce into a (more or less) SQL environment
- 91. Which One Do I Use Where? • Key-Value for (relatively) simple, volitile data • Document store for more complex data • Columnar for analytical processing • RBDMS for traditional processing – particularly where a lazy consistency is not acceptable • Point Of Sale, for example
- 92. Questions?
- 93. @scyphers Additional Information At http://www.daemonconsulting.net/BDC-FOSE-2012 Daemon Consulting, LLC http://www.daemonconsulting.net/ Specializing In The Hard Stuff