Server Suggestion Request

[Gavin McCullagh]

Hi,

On Mon, 30 Jun 2008, Daniel Kahn Gillmor wrote:

> On Mon 2008-06-30 05:20:18 -0400, Gavin McCullagh wrote:
> 
> > -- Consistency --
> >
> > The trouble with RAID is what happens when you get a sudden system
> > shutdown (eg due to a power failure) while you're writing to the
> > disk array.  Suppose the data is written to one drive correctly, but
> > not yet to the other.  When the system resumes, it's very difficult
> > for the system to determine which disk (if either) is correct.  So,
> > you end up with some ambiguous blocks where depending on which disk
> > is read, you can get a different answer.  An FSCK may not fix or
> > even detect the issue.
> >
> > Most software RAID (ZFS excepted) suffers from the above issue and
> > some cheaper hardware RAID cards do too.
> 
> Is this still true with modern mdadm?  I thought that the RAID
> superblock on modern systems contained information about
> "last-written" times, to enable recovery from exactly this problem.
> I'm not a RAID guru, though: if anyone could point me toward the
> relevant documentation, i'd be grateful.

I'm not aware that this has changed, but I could be wrong and would be
delighted to find I am.  Would "last-written" times be enough to recover
from this?  I would have thought you'd need a full-blown journal.

> At any rate, RAID or not, for important machines you probably want to
> have a UPS prevent sudden power failures, and you should also have a
> real backup strategy.  

Indeed, though if you're on a budget, UPS is not always affordable and it
won't prevent every sort of interruption (like a kernel panic).

> RAID (protection against hardware failures) is not a substitute for
> backups (protection against human error).

Absolutely.  RAID keeps uptime, not backups.

Gavin