[Gluster-devel] Problem with little write/read files in Gluster 3.0.0

Gordan Bobic gordan at bobich.net
Thu Jan 28 16:49:48 UTC 2010

Anand Avati wrote:
> On Thu, Jan 28, 2010 at 4:44 PM, Gordan Bobic <gordan at bobich.net> wrote:
>>> Thanks for your answer Gordan. Do you have more information about
>>> 3.0.2 ?
>> None, other than that Avati said a couple of days ago that it will be
>> available "very quickly".
>>> Do you think the 3.0.1 could be a temporary solution or I have to wait ?
>> I don't think I could in good conscience recommend using a release labeled
>> by the developers as DOA, even if it seems better than a previous non-DOA
>> release.
> 3.0.2rc1 is out and has been working fine in our tests thus far.
> Please feel free to give it a try. It is very likely to become 3.0.2.

The virtual memory usage seems to be reasonably sane again, and resident 
memory usage actually seems to be a significant improvement on even 2.0.9.

BTW, I just found out that my gcc problem with gluster root was due to 
gcc (and cpp) having gotten corrupted along with a number of header 
files from glibc-headers and kernel-headers, in the usual file content 
swap clobbering way I described before. My best guess is that this 
happened when I ran yum update recently, as all of the mentioned 
packages got updated. Worryingly, this happened whan I was running 
2.0.9, which indicates that there is a serious data corruption bug in 
it, and it's not even particularly related to open files (the chances of 
those particular header files all being open are pretty close to 0). 
What is the scope for a write getting mis-written to a wrong file due to 
an unexpected disconnect?

I think the evidence isn't just circumstantial any more - this 
corruption issue is happening with downright frightening regularity. I'd 
like to hope that 3.0.2rc1 solved this problem, but unless the cause is 
identified, and verified to not exist in the 3.0.x branch I think I'll 
have to limit my use of glusterfs to strictly non-live systems.


More information about the Gluster-devel mailing list