Slashdot Mirror
World's Largest Databases Ranked
prostoalex writes "Winter Corp. has summarized its findings of the annual TopTen competition, where the world's largest and most hard-working (in terms of load) databases are ranked. The results are in, and this year the contestants were ranked on size, data volume, number of rows and peak workload. I wrote up a brief summary of the top three winners in each category for those too lazy to browse the interactive WinterCorp chart."
You're off by 3 orders of magnitude. The largest is 30TB.
Doesn't Google use 'big files' rather than a database for storing all its data?
g hemawat.pdf which describes the Google filesystem.
see http://www.cs.rochester.edu/sosp2003/papers/p125-
Even if Google qualified, which it probably doesn't due to the methods it uses for its data storage, if I read the article properly the database vendors are responsible for naming the participants.
Since Google's stuff seems to be developed in-house, they don't have a major database vendor to nominate them.
Game... blouses.
In order to qualify for the TopTen program consideration, any commercial production database implementation was required to feature a minimum of 500GB of data for Microsoft Corp.'s Windows and NT platforms and 1TB of data for all other platforms.
You can find the link to the article yourself but
It seems that they are comparing relational databases. Search engines use proprietary databases which, among other things, do not allow for live insertion of records, SQL commands, etc. As for data volume, Google (or Yahoo or MSN, for that matter) are probably in the ballpark. The average html page is around 10k. Google probably stores at least 10^9 raw web pages in their cache(that's 10 TB alone) plus a lot of meta information about links to-from many others.
See charts for twitter trends on Trendistic
Lastly, in the Windows OTLP category HP servers were used by 7 of 10 organizations, and Microsoft SQL Server was the DBMS choice for seven respondents.
Neither WindowsNT, nor MS SQL are generally a choice for the top databases. In fact, to make the entry in this list, a Windows-Database was required to be only half as big as databases on other platforms:
In order to qualify for the TopTen program consideration, any commercial production database implementation was required to feature a minimum of 500 GB of data for Microsoft Corp.'s Windows and NT platforms and 1 TB of data for all other platforms
ms
I have none, nada, zip experience in big databases.
:-)
S'okay, I have plenty
But it surprised me that the peak workloads were measured in 100s of concurrent queries. If I had to make a wild guess, I would have guessed 10s of thousands. My blessed ignorance destroyed.
You would typically see tens of thousands (or more) of concurrent connections to a middleware layer - like Tuxedo - which would then multiplex them down to hundreds of connections to the database. This is because there is a lot of latency in establishing a connection, in fact logging in often takes an order of magnitude longer than running an actual query, yet few users submit transactions nonstop. So there is no sense in maintaining tens of thousands of expensive user contexts on the DB server, and there is no sense in requiring intermittent (relatively speaking) users to log out after a short idle period. Middleware does nothing but manage concurrent user contexts, and it can do so very efficiently. A database can't, because it tries to preallocate as much context as it can, and that doesn't match real-world usage patterns, and anyway, database vendors concentrate on their SQL engines and leave middleware vendors to manage the rest.
Of course, if you are a big database vendor, you probably also sell middleware, but there's no-one who tries to bundle the two into one, any more than you'd want a web server to have its own filesystem.
They don't have to put all their data into one database, though ...,
when the shit hits the fan, it is not equally spread
IMS is the database that was used to keep track of things for the moonshot. It is an IBM product. It is hierarchical as opposed to relational. Because of this it can do certain things very quickly, though in general it isn't as flexible as say DB2. Because it has been around so long, applications where having a DB was really important tend to have bought IMS a long time ago and developed systems around it. If your system is old enough, large enough and still works well for you there is no need to migrate to relational. Most of the world's financial transactions pass through an IMS system at some point. It is very stable and has uptimes that measure in years if not decades by now.
Because of this I am surprised that it is not on the list. There are really big IMS databases out there that run a lot of transactions. Because it isn't relational there is some bigotry against it and it is ignored in the popular press.
Lasers Controlled Games!
does anybody believe that the "SMP" used in reference to the French Telecom DB means "symbol manipulation program" rather than "symmetric multiprocessing"? how are we supposed to take seriously a study (or at least a report about the study) where they just look up acronyms with no understanding?
but they're not running a big Oracle or IBM style database, they're using a content management and static file system.
Typical Microsoft calling their product something generic that should apply to any SQL server. Almost like calling a product .. Windows.
It was originally called Sybase SQL Server but was later picked up by MS who adapted the name. Typical /. objectivity.
What about visa/mastercard/american express?
IMHO some of them didn't want to be in that list.
If you look at "database size", number 4 is listed as anonymous. They probably aren't too interested in telling everyone what database and platform they are using for storing very critical data with.
Obviously, you would be crazy to not use some middleware, but things aren't as simple as any of the PR guys claim. Running queries asynchronously creates a different set of problems and complicates the entire architecture. If you look at the biggest installation, they all use middleware and most of them use Tuxedo. This includes most, if not all MS Sql Server deployments. OLEDB can't that kind of load and neither can standard COM+. Just look read the full disclosures for TPC. You'll see all the MS Sql Server tests wrapped Tuxedo with COM+. As much as Microsoft likes to slam EJB and Tuxedo being too expensive, you can't scale Sql Server without using tuxedo for really heavy deployments.
One might see a database as merely a "big file" with mechanisms to access and modify it consistently (and surely, Google has some means to ensure consistency). A big file does not disqualify for the term "database" just because it is not produced by one of {Oracle, MS-SQL, ...} or cannot be queried by the language SQL.
It is also possible to consider the Web to be a database (of Web sites). Or an XML, BibTeX, dbm, whatsoever file.
Sebastian
We have databases in our organization (Star Schema, Red Brick) where the fact tables literally have billions of rows. I'm sure there are many other organizations (especially government entities) that have huge databases not on this "list". For those interested on operating at this scale, other interesting hardware/software data mining solutions in the same vein as a Teradata are Netezza Corp's database applicances.
I used to work for a company called Epsilon Data Management[1], in Burlington MA. They've been bought since I left them a while ago, but they where the keeper of AmEx customer transaction database for data mining and direct marketing (junk mail and phone calls).
Big. 7 data silos big. Each silo holds 50k tapes, each tape was 30gb, and it usually took 4 days to load.
[1] Epsilon was originally an AmEx division, which was spun off to keep other customers happy (banks and other CC companies).
III.IIVIVIXIIVIVIIIVVIIIIXVIIIXIIIIIIIIVIIIIVVIII
How about automatically categorize, find, and download?
It exists, and its open source. Welcome to the wonderful world of porn-get.
Yup.
Methinks the character who wrote the article came across the term 'SMP', went to FOLDOC or The Jargon File, and whaddya know - the first hit returns 'Symbol Manipulation Program - Stephen Wolfram's yadda yadda yadda'.
How do they backup a database that is 94.3 TB?
I support very large Oracle databases for a living (very large meaning > 1TB), databases that must be up 24/7. Backups are done in a number of different ways:
1) Disk syncs, block by block, between disk subsystems at disparate locations, to retain multiple copies of a database in different locations. They can be synced to more than one location too, so you can have as many copies of the database as you want. Your main database is the only "hot" database, the others can be brought up and recovered if needed. We mainly use EMC disk subsystems to do this, the process is called BCV (can't remember what that stands for right now)
2) Real-time replication. One-to-one or one-to-many. All databases are "hot" at all times. This can be great for load balancing too since you can have multiple system onine at the same time. Very difficult to maintain and monitor.
Large databases just can't be put to tape anymore. Even if you did, it would take days or weeks to recover them if they failed. Disk to disk is about the only way to provide backups for really large databases.
To add to that,
Standby databases are popular when (in Oracle scenario) the archived log files from your hot production database are constantly automatically applied to the cold standby database in some different location and if something happens to the primary it takes very little time to bring the standby up.
Also Oracle hot backup is by nature incremental, you can do like one tablespace per night, dont have to do the whole database at the same time (while backing up all the archived log files). I have seen sites where last cold backup was done something like 4 or 5 years ago.
A database is any collection of data. A database management system (which is what most people erroneously call a database) is a system of programs (say Oracle/MS SQL) to maintain the data in a database.
Thanks,
--
Matt
BCV - Business Contingency Volume I think. We call it Snap backup'ing.
When we dump data, it gets dumped to a VTS (that's Virtual Tape system which is a whopping collection of disk, or DASD pretending to be loads of cartridges). Once the data is on the VTS, it then makes it's way to a selection of real MagStar drives which sit behind the VTS system.
Works quite nicely.
As someone mentioned: Business Continuance Volumes is local copy within the storage array. Sync times depend on data change rate(dirty tracks). Host does not see any performance degradation. Copies are consistent from app level down, if done right.
SRDF = Symmetrix remote data facility. is a bcv copy across a link (network, fiber, DS3/1, OCs etc...fill in the blank). Again it only copies any changed tracks....
Good stuff, this is how most of the Fininacials recovered from 9/11 so quickly...
The databases then are put to tape using the copies. when the db exceed 24 hour backup time, you use multiple copies in rotation. Usually there's a regulatory reason to go to tape, otherwise people just use disk.
"It's technical in a psychometric kind a way" -- C. Parish