Вовченко Алексей, к.т.н., с.н.с. ВМК МГУ ИПИ РАН

Zettabytes Petabytes ABC Sharding A B C Id Fn Ln Addr 1 Fred Jones Liberty, NY 2 John Smith??????

122+ NoSQL Database Offerings Today! 4 Dominant Flavors The Buzz Regular machine failure, data center outages, and network service interruptions happen frequently Need is higher volume, fewer features Existing RDBMS do not automatically manage the distribution of data over available hardware. Sharding solutions over RDBMS introduce large overhead High-scale RDBMS too expensive for increased data volume Need for a flexible data model Need for a low-latency, low-overhead API to access data Need to scale-out on cheap commodity hardware Increase use of distributed analytics

Simple Key Value Stores Simplest NoSQL Store, provides low-latency writes but single key/value access Stores data as hash table of keys where every key maps to an opaque binary object Easily scales across many machines, does not support other data types Ideal for apps that require massive amounts of simple data like sensor data or rapidly changing data such as stock quotes Use-cases: apps that require massive amounts of simple data (sensor, web ops), apps that require rapidly changing data (stock quotes), Caching. Examples : MemcaheD, Dynamo Document Stores Represents rich, hierarchical data structures, reducing need for multi-table joins Structure of the documents need not be known a priori, can be variable, and evolve instantly, but Query can understand the contents of the document Applications: rapid ingest and delivery for evolving schemas and web-based objects. Examples. MongoDB, couchdb (couchbase) Column Family Manages structured data, with multiple-attribute access Columns are grouped together in column-families/groups. Each storage block contains data from only one column/column set to provide data locality for hot columns Column groups defined a-priori, but supports variable schema within a column group Scale using replication, multi-node distribution for high availability and easy failover. Optimized for writes (writes faster than reads) Applications: High throughput verticals (activity feeds, message queues). Caching. Web ops. Examples. HBase, Cassandra, BigTable, Amazon Dynamo Graph Store Uses nodes, relationships between nodes, and key-value properties Accesses data using graph traversal, navigating from start nodes to related nodes according to graph algorithms Faster for associative data sets Uses schema-less, bottoms-up model for capturing ad-hoc and rapidly changing data Common Model : RDF Applications: storing and reasoning on complex and connected data, e.g. inferencing applications in healthcare, government, telecom,oil, perform closure on social networking graphs Examples : Neo4J, DB2

Row A Row B Row C Column A Column B Column C X Note: Column Families contain Columns with time stamped versions. Columns only exist when inserted (i.e. sparse) Row A Column A Integer Row A Column B Value Row B Column B Long Timestamp1 Row B Row C Long Timestamp2 Row C Column C Huge URL Family 1 Family 2

11111 22222

Data Layout Transactions Query Language Security Indexes HBase A sparse, distributed, persistent multidimensional sorted map ACID Support on Single Row Only get/put/scan only unless combined with Hive or other technology Authentication /Authorization Row-Key only or special table RDBMS Row or Column Oriented Yes SQL Authentication /Authorization Throughput Millions of Queries per Sec Thousands of Queries per Sec Maximum Database Size PBs Yes TBs

Given this sample RDBMS table SSN primary key Last Name First Name Account Number Type of Account Timestamp 01234 Smith John abcd1234 Checking 20120118 01235 Johnson Michael wxyz1234 Checking 20120118 01235 Johnson Michael aabb1234 Checking 20111123

Row key Value (CF, Column, Version, Cell) 01234 info: { lastname : Smith, firstname : John } acct: { checking : abcd1234 } 01235 info: { lastname : Johnson, firstname : Michael } acct: { checking : wxyz1234 @ts=2012, checking : aabb1234 @ts=2011}

info Column Family Row Key Column Key Timestamp Cell Value 01234 info:fname 1330843130 John 01234 info:lname 1330843130 Smith 01235 info:fname 1330843345 Michael 01235 info:lname 1330843345 Johnson acct Column Family Row Key Column Key Timestamp Cell Value 01234 acct:checking 1330843130 abcd1234 01235 acct:checking 1330843345 wxyz1234 01235 acct:checking 1330843239 aabb1234 Key Key/Value Row Column Family Column Qualifier Timestamp Value

Key Key/Value Row Column Family Column Qualifier Timestamp Value

HBase API Region Servers Master HFile Memstore Write-Ahead Log ZooKeeper HDFS

The data is sharded Each shard contains all the data in a key-range Rows Region Server 1 Region Server 2 Region Server 3 Table Logical View A.. K.. T.. Z Keys:[A-D] Region Keys:[E-H] Region Keys:[S-V] Region Keys:[N-R] Region Keys:[I-M] Region Keys:[W-Z] Region Auto-Sharded Regions

HLog Region Server Region Client Store StoreFile MemStore StoreFile StoreMemStore StoreFile HFile HFile HFile

HLog HLog HLog External Gateway Clients atch Clients: Thrift Client Rest Client Avro Client Other Clients: Hive HBase Thrift Server Rest Server Avro Server jaql Pig Cluster HTable HTable HTable Region Server Region Server Region Server JRuby MapReduce Store MemStore StoreFile HFile Store MemStore StoreFile HFile Store StoreFile MemStore HFile AsyncHBase HTable Java Client HTable Thrift Server Thrift Client External API Clients

Key Key/Value Row Column Family Column Qualifier Timestamp Value

/////////////////////////////////////An Example///////////////////////////////////// //Create an HBase Table import hbase(*); T = hbasestring('test', schema { key: string, f1?: {*:string}, f2?: {*:string}, f3?: {*:string} }, replace=true); //Define a Row data = [ { key: "1", f1: { Name: "Bruce Brown"}, f2: { Address: "BigData Blvd, Washington DC 20001" }, f3: {Email:"brownb@us.ibm.com"} }, { key: "2", f1: { Name: "John Smith"}, //Put the data into the HBase Table data -> write(t); f3: {Email: "jsmith@us.ibmcom"} } ]; Speeds Development of HBase Applications!

InfoSphere BigInsights Visualization & Exploration Development Tools Advanced Engines Connectors Workload Optimization Administration & Security IBM-certified Apache Hadoop

HBase GUI for Queries Integration with BigSheets for Advanced Analytics Advanced Support and Analytics through Jaql

DB2/ InfoSphere Warehouse / Netezza tools/apps DB2 / InfoSphere Warehouse BigInsights (HBase) HDFS

*http://www-01.ibm.com/software/data/bigdata/

Process Heap Description NameNode 8 GB About 1 GB of heap for every 100TB of raw data stored, or per million files / inodes Secondary NameNode 8 GB Applies the edits in memory, and therefore needs about the same amount as the NameNode JobTracker 2 GB Moderate requirements HBase Master 4 GB Usually lightly loaded, moderate requirements only DataNode 1 GB Moderate requirements TaskTracker 1 GB Moderate requirements HBase RegionServer 12 GB Majority of available memory, while leaving enough room for the operating system (for the buffer cache), and for the Task Attempts processes Tasks Attempts 1 GB (ea.) Multiply by the maximum number you allow for each ZooKeeper 1 GB Moderate requirements

NoSQL -- Your Ultimate Guide to the Non - Relational Universe! http://nosql-database.org/links.html Brewer's CAP Theorem posted by Julian Browne, January 11, 2009. http://www.julianbrowne.com/article/viewer/brewers-cap-theorem HBase Schema design: http://hbase.apache.org/book/schema.html HBase in Action by Nick Dimiduk, Amandeep Khurana