7. Recent NOSQL database products Columnar or Extensible record Google BigTable HBase Cassandra HyperTable SimpleDB Document Store CouchDB MongoDB Lotus Domino Graph DB Neo4j FlockDB InfiniteGraph Key/Value Store Mnesia Memcached Redis Tokyo Cabinet Dynamo Project Voldemort Dynomite Riak
8.
9.
10.
11.
12.
13. BigTable’s Data Model Google’s Bigtable is essentially a massive, distributed 3-D spreadsheet. It doesn’t do SQL, there is limited support for atomic transactions, nor does it support the full relational database model. In short, in these and other areas, the Google team made design trade-offs to enable the scalability and fault-tolerance Google apps require. - Robin Harris, StorageMojo (blog), 2006-09-08 t 6 t 5 t 3 name contents: anchor:cnnsi.com ... anchor:my.look.ca ... “ com.cnn.www” “ CNN” ... “ CNN.com” ... “ <html>...” “ <html>...” “ <html>...”
14.
15.
16.
17.
18.
19. Dynamo data partitioning and replication Virtual node Host “node” Host “node” Virtual node Virtual node Virtual node Virtual node Virtual node Virtual node . . Hash ring using consistent hashing Host “node” Virtual node Virtual node Virtual node Virtual node 4 4 3 Item Hashes to this spot coordinator node replicas
20.
21.
22.
23. The Gold Rush Columnar or Extensible record Google BigTable HBase Cassandra HyperTable SimpleDB Document Store CouchDB MongoDB Lotus Domino Graph DB Neo4j FlockDB InfiniteGraph Key/Value Store Mnesia Memcached Redis Tokyo Cabinet Dynamo Project Voldemort Dynomite Riak Hibari
This talk really comes out of my attempt to orient myself in this space. Background is in monitoring distributed systems, concerned with scalable collection and data analysis. But also want to know what I can use for semi-structured data “in the small”.
Where it applies, the distinction between relatively fixed schemas and dynamic ones is more technically significant than what query syntax is used to access the data, as has been shown by a number of products that provided a dialect of SQL as an alternative query language either alongside or on top of their native syntax.
PICK -- MultiValue (aka PICK) databases are developed at TRW in 1965. M[umps] -- According to comment from Scott Jones M[umps] is developed at Mass General Hospital in 1966. It is a programming language that incorporates a hierarchical database with B+ tree storage. IBM IMS -- IBM IMS, a hierarchical database, is developed with Rockwell and Caterpillar for the Apollo space program in 1966. ISM -- InterSystems develops the ISM product family succeeded by the Open M product, all M[umps] implementations. See comment from Scott Jones below. ANSI M -- M[umps] is approved as a ANSI standard language in 1977. AT&T DBM -- in 1979 Ken Thompson creates DBM which is released by AT&T. At it's core it is a file-based hash. TDBM -- TDBM supporting atomic transactions NDBM -- NDBM was the Berkeley version of DBM supporting having multiple databases open at the same time. SDBM -- SDBM - another clone of DBM mainly for licensing reasons. GT.M -- GT.M is the first version of a key-value store with focus on high performance transaction processing. It is open sourced in 2000. BerkeleyDB -- BerkeleyDB is created at Berkeley in the transition from 4.3BSD to 4.4BSD. Sleepycat software is started as a company in 1996 when Netscape needed new features for BerkeleyDB. Later acquired by Oracle which still sell and maintain BerkeleyDB. Lotus Domino -- Lotus Notes or rather the server part, Lotus Domino, which really is a document database has it's initial release in 1989, now sold by IBM. It has evolved a lot from the early versions and is now a full office and collaboration suite. GDBM -- GDBM is the Gnu project clone of DBM Mnesia -- Mnesia is developed by Ericsson as a soft real-time database to be used in telecom. It is relational in nature but does not use SQL as query language but rather Erlang itself. Cache -- InterSystems CachÈ launched in 1997 and is a hybrid so-called post-relational database. It has object interfaces, SQL, PICK/MultiValue and direct manipulation of data structures. It is a M[umps] implementation. See Scott Jones comment below for more on the history of InterSystems Metakit -- Metakit is started in 1997 and is probably the first document oriented database. Supports smaller datasets than the ones in vogue nowadays. Neo4j -- Graph database Neo4j is started in 2000. db4o -- db4o an object database for java and .net is started in 2000 QDBM -- QDBM is a re-implementation of DBM with better performance by Mikio Hirabayashi. Memcached -- Memcached is started in 2003 by Danga to power Livejournal. Memcached isn't really a database since it's memory-only but there is soon a version with file storage called memcachedb. Infogrid graph DB -- Infogrid graph database is started as closed source in 2005, open sourced in 2008 CouchDB -- CouchDB is started in 2005 and provides a document database inspired by Lotus Notes. The project moves to the Apache Foundation in 2008. Google BigTable -- Google BigTable is started in 2004 and the research paper is released in 2006. JackRabbit -- JackRabbit is started in 2006 as an implementation of JSR 170 and 283. Tokyo Cabinet -- Tokyo Cabinet is a successor to QDBM by (Mikio Hirabayashi) started in 2006 Dynamo -- The research paper on Amazon Dynamo is released in 2007. MongoDB -- The document database MongoDB is started in 2007 as a part of a open source cloud computing stack and first standalone release in 2009. Cassandra -- Facebooks open sources the Cassandra project in 2008 Voldemort -- Project Voldemort is a replicated database with no single point-of-failure. Started in 2008. Dynomite -- Dynomite is a Dynamo clone written in Erlang. Terrastore -- Terrastore is a scalable elastic document store started in 2009 Redis -- Redis is persistent key-value store started in 2009 Riak -- Riak Another dynamo-inspired database started in 2009. HBase -- HBase is a BigTable clone for the Hadoop project while Hypertable is another BigTable type database also from 2009. Vertexdb -- Vertexdb another graph database is started in 2009 Term: NOSQL -- Eric Evans of Rackspace, a committer on the Cassandra project, introduces the term NoSQL often used in the sense of Not only SQL to describe the surge of new projects and products.
Both of these systems are still used. An open-source version of M, called GT.M, is available (since 2000). M is still used by the US Dept of Veterans Affairs, and also by Ameritrade (Cache’: 12B transactions a day), ING Direct, and others in the financial industry. The IBM IMS system is still very actively used today, in particular for the US Federal Reserve. According to Wikipedia, odds are good your ATM transaction hits an IMS database. Chinese banks have purchased IMS technology. IMS includes a separate “transaction management” (TM) system.
E. F. Codd’s seminal 1970 paper, “ A Relational Model of Data for Large Shared Data Banks” laid out a solid mathematical basis for databases in contrast to the hierarchical and network models of the time, relational algebra, an offshoot of first-order logic, provided a declarative means of reasoning about the data that did not depend on the implementation SQL is “loosely based” on relational algebra
This taxonomy will be explored in more detail later, the point for now is that there are several different types of datastores and a number of examples of each and, referring back to the timeline, most of these implementations have occurred in the past few years..
Corporations (once again) found themselves at the forefront of systems research. But what was that research? (Read on..)
If nothing else, being able to refer to the “CAP theorem” the next time your networked demo breaks..
In his talk, Brewer said “there is almost no work in this area”. I think that the existence of scalable (schemaless) database systems is proof that this has changed.
Pictured is Parliament, pioneers of funk!
Trivia: what major movie was about producing a script called “Chubby Rain”?
Example of a BIgTable that stores web pages (directly out of the paper). The row names are reversed URLs (so sorted rows tend to group things by the same domain) There are two column families, “contents” and “anchor” In this example, each anchor cell has one version, and the contents column has 3
Paxos is an old and well-known algorithm. The Chubby “Database” is really a set of directories with small “lockfiles”. Each tablet server gets one Chubby directory, and each of its tablets is a lockfile.
These core services included the Amazon e-commerce shopping cart.
Each virtual node is responsible for keys between itself and its predecessor on the ring. The mapping of a single node to a variable number of virtual nodes on the hash ring accounts for heterogeneity (host “power”) in the system.
The quorum is “sloppy” because R and W refer to the number of healthy nodes, which may change between the write and subsequent read of the key.
(Who knows what this is?) The picture is a close-up of a vegetable: the “ Chou Romanesco&quot; cauliflower
Particularly appropriate analogy because of the industry’s tendency to rush towards shiny new technologies! Following sections will examine each of these categories and walk through one publicly available product (or more) for each. With the exception of graph databases, which I simply haven’t taken the time to grok yet.
Both Voldemort and the next database, Riak, claim they were “inspired” by the early Dynamo paper
In the diagram, the green nodes are head; orange middle; red are tails. The white arrows are write requests, grey read requests, and red are (all) replies.
Developed by former engineers from BigTable and Dynamo projects, in heavy use at Facebook. For consistency level, zero = totally async.; Any= 1 node, including hinted handoff; Quorum = R/2+1 where R = #replicas Reads of 0 or Any don’t make sense. 0=no data, Any=wrong node; can’t do read-repairs, just the handed-off version
Has a nice Web UI called “Futon”. Yes, everything is a reclining furniture pun.
Obviously, this is at best a micro-benchmark. YCSB stands for Yahoo! Cloud Serving Benchmark
I won’t attempt to actually cover Map/Reduce, and don’t know Erlang. Instead: what impact do these databases have on data modeling efforts?