Netflix's Siddharth Anand spoke at Silicon Valley Cloud Computing group and gave a technical presentation on how Netflix moved from RDBMS to NoSQL, and it sheds some light on how the Netflix website uses the cloud.
The first 10 minutes are presentations from sponsors, but the talk and long Q&A session are very interesting if you're curious how Netflix.com works.
Fascinating!
In the mid nineties when the Internet was blooming there wasn't much controversy about what to do with the damn thing. It was obvious to everyone that Video on Demand (VOD) was where the money was. Alas Oracle, which spent a lot of time and money trying to create an RDBMS for video, failed. There would be no VOD - at least for a while.
The Web is in effect a fall back usage cobbled together when VOD proved to be too hard a nut to crack.
One secret revealed here is that they just gave up on relational systems altogether. An RDBMS is in a sense a traffic cop or regulations enforcer but these services come at a cost in performance. As hardware got bigger and faster the inefficiencies of relational mattered less and less.
When I first used Oracle it was very controversial. Many old timers would tell me that such a thing could never be fast enough for full size regular businesses. They argued for network DBMSs that stored path dependencies. In those old fashion technologies certain queries were pre-selected. This made these software products fast for routine look ups but hopeless for ad hoc queries.
It was common then for big shops to not allow queries against the main production database (typically IMS or IDMS). The staff would make excerpts or copies into an RDBMS that could be used for queries. At the time IBM called these Data Centers. They tried to sell you a smaller 370 type machine that would run one of their relational or query friendly products. They marketed these things also as an alternative to using a PC and a spreadsheet.
Later of course RDBMSs on newer hardware became fast enough to serve as an enterprise production database. But apparently the VOD problem that had once stumped Oracle could only be solved by tossing out the whole idea of being relational altogether.
Fascinating!
Posted by: PatB | February 21, 2011 at 03:18 PM
Thanks for the Heads Up, @PatB. I wasn't sure I wanted to take over an hour to watch this video but your comments have me very interested. CA-IDMS (Cullinet), IBM DL/I and DB2 are in my career background as well. I say something like "JCL" or "RJE" now a days in an IT meeting and people just blank stare. We use to take pride in having to IPL only once a year. Talk about your five 9's of availability...
Posted by: dAVe | February 21, 2011 at 04:16 PM
@PatB, thanks for your thoughtful comment. The basis for this talk was a white paper at http://practicalcloudcomputing.com/post/1267489138/netflixcloudstorage
@dAVe the actual talk runs from 0:10 to ~0:55, so is only about 40-50 minutes.
Posted by: R39132 | February 21, 2011 at 07:23 PM