I only think ECC is worth using when the server requires it. Wikipedia:
Error detection and correction in computer systems seems to go in and out of fashion. Seymour Cray famously said "parity is for farmers" when asked why he left this out of the CDC 6600. He included parity in the CDC 7600, and reputedly said "I learned that a lot of farmers buy computers."
I can't find a definitive source on the internet, other than nebulous claims of one bit error per month per gigabyte, which is patently ridiculous; servers would be crashing left and right all over the world if this was remotely true.
Some highlights from a MetaFilter thread from actual server admins:
I think ECC is cool stuff, but I've had servers both with and without it, and I've never had its presence or absence do anything, either way.
I understand the purpose of ECC RAM, but not the point. I mean, I've never noticed any issue resulting from cosmic ray bit flipping. Even on personal compute/compile servers with multiple year uptimes. Not to say that bits didn't flip, but they certainly didn't matter.
In my experience running farms of a few thousands machines here and there, you're more likely to have Ext3 silently puke all over you than to have an ECC-correctable problem.
Personally, I think ECC is a bit cargo-cultish, but it's a reasonable insurance policy on a big beefy server as long as the cost premium isn't too high.