[mythtv-users] Consequences of drive failure

[Simon Hobson]

Joseph Fry wrote:

I don't believe they use enterprise drives. Most mechanical devices prefer aconsistenttemperature, they are indifferent to hot or cold, they just don't do well with all the expansion and contraction that occurs with temperature changes. Only the solid state portion of it would prefer cool to hot, but typically the mechanical parts will fail before the solid state parts, even in aconsistentlyhot environment.

Indeed, they use (or used, but I can't see why it would have changed) commodity components. At my local LUG, one of the members got a job at Google (in their reliability engineering group) a while back, and came back to talk about his work. From the description of how they build their servers (minimalist chassis that is nothing more than a tray to hold commodity items), it sounded like they would dispense with the chassis altogether if they could persuade the components to hover in formation !

Basically, their approach is not to throw money at incremental reliability upgrades in the hardware, but to design the system so that hardware failures don't matter. EG, a storage cluster might have 20,000 servers in it, and every chunk of data will be replicated over at least 3 servers spead over a minimum of 2 racks. So taking out any server doesn't impact the cluster, nor does taking out an entire rack (eg power or 'top of rack' switch failure. Ie, the redundancy is built into the software - in the above example, their GFS filesystem.
I'm sure they do consider reliability, but I'd imagine at the "we'll choose board X because it's about the same price but has less failures that board Y" level rather than what most of us would be doing - specifying 'server grade' hardware, RAID, ECC memory, and so on.



-- 
Simon Hobson

Visit http://www.magpiesnestpublishing.co.uk/ for books by acclaimed
author Gladys Hobson. Novels - poetry - short stories - ideal as
Christmas stocking fillers. Some available as e-books.