[volunteergrid2-l] Slashdot post

[Shawn Willden]

On Wed, Apr 4, 2012 at 10:54 PM, Marco Tedaldi <marco.tedaldi at gmail.com>wrote:

> On 03.04.2012 14:20, Shawn Willden wrote:
> > What if we were to just admit everyone to the grid, with the
> understanding
> > that for the first few months (say, six?) members are somewhat
> > "probationary" and can have their access revoked?  We've done the
> > new-introducer-furl dance a couple of times and it went pretty smoothly,
> so
> > I think removing someone from the grid isn't too difficult.
>
> Yes. but I'm not sure, how mucht this was just "good luck". Because for
> example I will be on vacation during the next 2 weeks. I might have
> internet access only from time to time and I'm not sure, if checking my
> tahoe-mails will be the highest priority than...
>

And as the grid grows, that will be more of a problem... but the impact of
one or two nodes not migrating quickly will be less, so I think it balances
out.


> > On a related note:  We have one member node right now which isn't meeting
> > the storage requirements.
>
> thats my node and it has already bitten me .-(
>
> >  I think we should give that node a deadline to
> > get upgraded.  If it hasn't been upgraded by the deadline, then we should
> > ask that it be removed from the grid.  How about three months?  That's a
> > deadline of July 3rd.
> >
> sounds reasonable to me. Since the prices for harddrives seem to be
> reasonable once again, this should be no big issue.
>

Great!


> This would solve the issues almost entrirely. I than have plenty of
> SATA-Ports to attach  more disks (with my actual system I would have to
> add a PCI-SATA-card which would clog the bandwith of the PCI even more).
> And I have 4 or 5 160GB sata disks which would make a nice storage for
> my images as a raid 5 and free up the space on my 1TB drive for tahoe.
>

One comment here:  I've come to the conclusion that RAID5 is evil and
dangerous.  I've been meaning to write a paper on it for years but never
gotten around to it.  Assuming you back everything important up to the grid
it would be okay, but personally I use RAID6 for everything important.

Why is RAID5 dangerous?  Because if you lose one disk, the array resyncing
process *will* uncover any latent defects in any of the remaining disks.
 The core assumption behind RAID5 reliability is that disk failures are
independent events, but the fact that resyncing a degraded array onto a new
replacement drive must read every sector of every disk makes that a
questionable assumption.  And, of coure, losing two disks from a RAID5
array means *all* of your data is gone.

This actually happened to me.  I was able to recover by telling MD to
forcibly reconstruct the degraded array, using the ordering from the e-mail
it sent me when the first disk died.  Luckily the second disk failure was
transient.  It actually occurred again during the second attempt to resync
the array, but the third time worked.

With RAID6, you would have to have two failures during resynchronization in
order for your array to lose your data.  So now I put everything important
on RAID6, have smartd configured to run periodic full-disk tests on each
drive, and use another tool (I forget what it is) to run periodic
full-volume read tests on my LVM logical volumes (which are build from the
RAID arrays).  The testing should hopefully uncover latent defects earlier,
so I'm less likely to run into them when resynchronizing to replace a
failed drive.

My machine has 9 drives in it, so drive failures aren't rare events.  I get
one every couple of years.


> Couldn't we create something like a "furl-distributor"?
> A webserver, that serves the furl as a text file. Protected by
> individual username/passowrd combinations. https.
> So on the client we could create a little cron-job that fetches that
> file with wget on a regular basis and replaces it in the config file if
> needed and restart tahoe.
>
> That way, changing the furl could be automated and the distribution to
> other members can be controlled.
>

That would be a good idea, but I think the whole approach to introduction
is about to change with the distributed-introducer patches (not sure when
they're hitting).  So it would probably make more sense to investigate how
that would work and build a solution for that context rather than the one
that's going away.

Of course, Brian's accounting work will probably give us an entirely
different (and better!) solution in the not too distant future.

-- 
Shawn
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://tahoe-lafs.org/cgi-bin/mailman/private/volunteergrid2-l/attachments/20120405/cafe7e86/attachment.html>