[talk] Climate Mirror

Isaac (.ike) Levy ike at blackskyresearch.net
Wed Dec 14 14:28:18 EST 2016


> On Dec 14, 2016, at 2:23 PM, Pete Wright <pete at nomadlogic.org> wrote:
> 
> 
> 
> On 12/14/16 11:15 AM, Thomas Levine wrote:
>> The data don't need to be online; save them to a redundant bunch of
>> cheap hard drives (or maybe tapes), and distribute them among lots of
>> bookshelves. They can even be slow and small hard drives pulled from old
>> computers; we need to write to each one only once, we might need to read
>> from each one once, and we otherwise only need to turn them on once
>> every couple years to make sure that they're still intact. Maintain a
>> website with a list of the datasets, the datasets' checksums, and the
>> contact information for the people with the hard drives on their
>> bookshelves.
>> 
>> Note that this is my opinion only on how this project could be
>> implemented. I don't know enough about the datasets or the likely
>> effects of geopolitics on their implementation in order to comment as to
>> whether I think the project should be implemented.
>> 
> 
> not to nit-pick but i would strongly recommend *against* using HDD's in this manner (magnetic spinning ones, or SSD ones).  Drives are not designed to reliably store data cold like this mechanically or electrically.  This is why tapes are still in use to this day - they *are* designed for cold store.  And if you do hit a bad sector it is quite possible to skip that sector and continue reading data.
> 
> This is coming from quite a bit first-hand experience where I've lost data-sets which were in cold-storage on HDD's for about a year that were totally lost, versus data on tapes which were in cold-store for around 5-7years where we had few problems recovering our assets.

Hrm.  You may have put the silver bullet in my previous post.

Pete: any thoughts on mitigating this effect by using ZFS mirrors?  Perhaps even increasing the block mirror count across disks, so even on one of the mirrored disks there are 2 mirrored blocks?

I mean, one crappy way to test this is to just do it and wait a year :P

Best,
.ike





More information about the talk mailing list