Re: [rdiff-backup-users] Hash doesn't match recorded hash

[Alex Schuster]

Maarten Bezemer writes:

> On Tue, 15 Nov 2011, Dominic Raferd wrote:
> 
>> On 15/11/2011 00:23, Alex Schuster wrote:
> 
>>> Warning: Hash da39a3ee5e6b4b0d3255bfef95601890afd80709 of winXPPro-0.log
>>> doesn't match recorded hash db9943d0284382131b553cef380aae99871b0f2e!
>>>
>>> Any idea what has happened?
> 
>> Vmware files for a running vm may change while being backed up by 
>> rdiff-backup, in which case you don't have a consistent backup. Maybe this 
>> is 
>> why you have this inconsistent hash warning.
> 
> If the contents changed during the rdiff-backup run, it bails out and 
> doesn't save the (inconsistent) increment. So that shouldn't happen.

And I amusing a backup script which takes a LVM snapshot of my /home
partition before rdiff-backupping it. Sorry Dominic for not mentioning this.
Restoring other VMs seems to work, although I did not test many. But I
cannot restore ANY backup of this special VM I need, I tried about six
of my twenty backups.
Well, it would be really nice to be able to restore this VM, but if not,
I can (and have to) live with that. But I really would like to know why
this happened, and if I can trust other backups I do.
I successfully restored this VM in the past already, but that was an
earlier snapshot I have deleted meanwhile.

> My guess would be a RAM problem, which is more likely to cause corruption 
> in huge vmware disk images than it is for smaller regular files.

Right. But the system is running stable otherwise. I am on Gentoo Linux,
where the whole system is compiled from source, and RAM errors tend to
show up as weird, unreproducible compilation errors there. Especialle
gcc is a good test for this, because it is compiled twice in the build
process, and the two results are compared bitwise. But I compiled gcc
for 24 times since I made that backup, without a problem.

> Another known cause for this type of corruption is a Via KT400a chipset 
> when running with DDR400. Memtest won't find any errors, but repeatedly 
> md5summing the same large file will result in wrong checksums every now 
> and then ( < 5% of the cases, so try some 100 times... )
> But that kind of hardware should have been retired these days anyway.

I don't have this hardware, but thanks for the hint. Oh, and I also
don't have one of those Samsung SpinPoint drives that write an empty
block whenever the drive's identification is checked with hdparm.

        Wonko