Re: possibly useful info re: FSFS corruption

----- Thomas Beale <Thomas.Beale@OceanInformatics.biz> wrote:
[snip]
> Our binary files are all less than 6Mb (maybe 1% 5.5Mb, 10% PDFs @
> 1.5Mb, 25% FrameMaker files @ 250kb), and we don't do that many binary
> files like this in the total repository, which is mostly software or
> document sources. Having used SVN for about 18 months in our community
> (maybe 30 committers, but hundreds of readers), I think this problem is
> the only bug-related problem in the whole thing (i.e. I wouldn't mind
> changing or adding a couple of design features),so fixing it will
> definitely add an aura of near invincibility to SVN. I think that our
> mix of files is not unreasonable or unexpected for SVN users, so the
> symptoms are unfortunately going to keep occurring. I wish I could
> provide more detailed evidence to help the developers here (and I will
> if I see any).

That's certainly a reasonable mix, although working with binary files shouldn't be a limiting factor in any way.

>
> I guess a test server which is pounded by commits of say 5Mb+ binary
> files should help reproduce the problem. Best of luck with finding
> it.

I've done that on a dual processor machine, using a modified version of the stress test script for nearly a week... not one single failure. I'm even using APR 0.9.6. The closest I've seen to a reproduction recipe is one user used tc to fuzz the connection reliablity (introduced latencies and dropped packets). He was able to obtain a corrupt revision out of that... unforunately, I still wasn't able to cause it to happen. :-(

FWIW, I used 5, 10, and 15MB binary files, with no troubles during the stress testing.

-John

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@subversion.tigris.org
For additional commands, e-mail: dev-help@subversion.tigris.org
Received on Wed Jul 12 16:43:09 2006