Original URL: http://www.theregister.co.uk/2013/10/25/facebook_reveals_700tb_of_tiered_ram_and_flash_power_graph_search/

Facebook reveals 700TB of tiered RAM and flash power Graph Search

The Social NetworkTM lifts skirts, shows glimpses of MySQL, Hadoop and Unicorn

By Simon Sharwood

Posted in Data Centre, 25th October 2013 05:58 GMT

Facebook has revealed a little more detail about how it built Graph Search, the feature it took live a couple of weeks back to enable better searching of the billion posts it now collects each day.

The reveal came in a blog post, as is the fashion at this time, that explains building better search across posts is hard because Facebook has “70 different kinds of data we sort and index on, many of them specific to certain types of posts.”

That's all managed by MySQL, which passes data to a HBase cluster before Hadoop gets to work. At this point “Unicorn”, Facebook's name for the tech powering its search engine, gets involved and the result is a colossal index with a very substantial cache that the company tries to treat as gently as possible.

Keeping the cache happy is not a trivial task, so the company has devised the following tiering regime:

“However, storing more than 700 terabytes in RAM imposes a large amount of overhead, as it involves maintaining an index that is spread across many racks of machines. The performance cost of having these machines coordinate with each other drove the Unicorn team to look into new solutions for serving the posts index. The solution we decided on involves storing the majority of the index on solid-state flash memory. We managed to preserve performance by carefully separating out the most frequently accessed data structures and placing those in RAM.”

RAM as tier zero and flash as tier one are suggested as the recipe for fast access to data by just about the whole storage and big data industries, so Facebook's infrastructure will please many a sales and marketing person and probably make its way into lots of slideware before too much time passes.

Perhaps a little more interesting is the post's statement that Graph Search was built “in the past year by a few dozen engineers”. Plenty of lesser projects take more time and more people, suggesting the most interesting thing IT pros and their managers can learn from Graph Search might be how the project was managed and the hardware designed and implemented, rather than the configurations chosen. ®