I guess I don’t understand why in some cases we only have these link-to files and not the new data files. Nothing has been overwritten, unless it was by gluster.
How can I tell if a file shows as the initial size, but isn’t actually using disk space?
What should the contents of one of these link-to files look like if I try to open it with less or some other program?
From: Nithya Balachandran [mailto:nbalacha@xxxxxxxxxx]
Sent: Tuesday, November 15, 2016 10:55 AM
To: Kevin Leigeb <kevin.leigeb@xxxxxxxx>
Subject: Re: Gluster File Abnormalities
On 15 November 2016 at 22:15, Nithya Balachandran <nbalacha@xxxxxxxxxx> wrote:
On 15 November 2016 at 21:59, Kevin Leigeb <kevin.leigeb@xxxxxxxx> wrote:
Nithya -
Thanks for the reply, I will send this at the top to keep the thread from getting really ugly.
We did indeed copy from the individual bricks in an effort to speed up the copy. We had one rsync running from each brick to the mount point for the new cluster. As stated, we skipped all files with size 0 so that stub files wouldn’t be copied. Some files with permissions of 1000 (equivalent to ---------T) were larger than 0 and were also copied.
I’m mostly trying to ascertain why such files would exist (failed rebalance?) and what we can do about this problem.
Rebalance creates these linkto files as targets for the file migration - while the file size displayed would be nonzero, the actual disk usage for such files would be zero.
Missed one crucial point - the rebalance sets the size of these linkto files to that of the original file.
The rebalance process first creates these files and then checks if it is possible to migrate the file. If not, it just leaves it there as it is not using up any space and will be ignored by gluster clients. It is not a failed rebalance - sometimes some files are not migrated because of space constraints and they are considered to have been skipped.
This behaviour was modified as part of http://review.gluster.org/#
/c/12347 . We now reset the size to 0.
I'm afraid, if those files were overwritten by their linkto files, the only way forward would be to restore from a backup.
Regards,
Nithya
Thanks,
Kevin
From: Nithya Balachandran [mailto:nbalacha@xxxxxxxxxx]
Sent: Tuesday, November 15, 2016 10:21 AM
To: Kevin Leigeb <kevin.leigeb@xxxxxxxx>
Cc: gluster-users@xxxxxxxxxxx
Subject: Re: Gluster File Abnormalities
Hi kevin,
On 15 November 2016 at 20:56, Kevin Leigeb <kevin.leigeb@xxxxxxxx> wrote:
All -
We recently moved from an old cluster running 3.7.9 to a new one running 3.8.4. To move the data we rsync’d all files from the old gluster nodes that were not in the .glusterfs directory and had a size of greater-than zero (to avoid stub files) through the front-end of the new cluster.
Did you rsync via the mount point or directly from the bricks?
However, it has recently come to our attention that some of the files copied over were already “corrupted” on the old back-end. That is, these files had permissions of 1000 (like a stub file) yet were the full size of the actual file.
Does this correspond to a file permission of ___T when viewed using ls? If yes, these are dht linkto files. They were possibly created during a rebalance and left behind because the file was skipped. They should be ignored when accessing the gluster volume via the mount point.
In some cases, these were the only copies of the file that existed at all on any of the bricks, in others, another version of the file existed that was also full size and had the proper permissions. In some cases, we believe, these correct files were rsync’d but then overwritten by the 1000 permission version resulting in a useless file on the new cluster.
This sounds like you were running rsync directly on the bricks. Can you please confirm if that is the case?
These files are thought by the OS to be binaries when trying to open them using vim, but they are actually text files (or at least were originally). We can cat the file to see that it has a length of zero and so far that is our only reliable test to find which files are indeed corrupted (find . -type f | xargs wc -l). With nearly 50 million files on our cluster, this is really a non-starter because of the speed.
Has anyone seen this issue previously? We’re hoping to find a solution that doesn’t involve overthinking the problem and thought this might be a great place to start.
Let me know if there’s any info I may have omitted that could be of further use.
Thanks,
Kevin
Thanks,
Nithya
______________________________
_________________
Gluster-users mailing list
Gluster-users@xxxxxxxxxxx
http://www.gluster.org/mailman/listinfo/gluster-users
_______________________________________________ Gluster-users mailing list Gluster-users@xxxxxxxxxxx http://www.gluster.org/mailman/listinfo/gluster-users
