On 11 September 2014 08:47, John Spray <john.spray at redhat.com> wrote: > I do think this is something we could think about building a tool for: > lots of people will have comparatively tiny quantities of metadata so > full dumps would be a nice thing to have in our back pockets. Reminds > me of the way Lustre people used LVM snapshots for their metadata. > Some users have set maintenance windows in which even a purely offline > metadata dump tool could be useful. This is the sort of thing I was thinking of. Operationally we would have full tape backups of the filesystems (sans 24 hours) - but restoring petabytes of data from tape is such an exercise you really want other types of insurance, especially where problems with the MDS can be expected. Something like an atomic snapshot of both data and metadata pools, where you could then export a copy of the metadata snapshot out of the operational cluster. Then if the MDS went haywire and corrupted things there'd be a relatively lightweight path to a rollback recovery. Though I say this with complete ignorance of how the MDS works... -- Cheers, ~Blairo
 |