Beyond RAID

This article advises the use of RAID in all servers:

"But on the server, RAID is a definite must."

I submit it as a warning to startups who value their time and money: treat unqualified technology advice with skepticism.

RAID has costs in time and complexity. In some cases, you will spend that money because the alternatives cost more. A write-heavy MySQL server is constrained by IOPS. Ignoring SSD for a moment, the two ways to get more IOPS are faster spindles and more spindles. This is why RAID 10 on 15K SAS disks wind up in database servers. Once you have exhausted other solutions and are truly IOPS-bound, the cost of adding expensive disks to your database server is likely much less than the alternatives.

Let's take a case where RAID does not make sense. On our site, our webservers are predominantly CPU bound. Aside from reading code and writing temporary files, the only disk access on these servers is HTTP logs. Every night, we sync the logs, process them, and store the on S3. RAID would at best save one day's worth of logs in case of catastrophic disk failure. If we could not tolerate the loss of even a few minutes of logs, there are a variety of software solutions that would accomplish this goal without spending money on RAID controllers and disks. Our MogileFS storage nodes also do not benefit from RAID, where we save many thousands of dollars a month by pushing availability up the stack into inexpensive software. (Brad and the Danga crew, we thank you.)

Wow, a big week for Atwood-- first he discovers floating point arithmetic, now RAID. Somebody wake me when he hits "recursion"...