On Fri, 4 Jun 2010, Andre Noll wrote:
> On Wed, Jun 02, 11:30, Sage Weil Wrote
> > > 3. OSD:
> > > Again, I suppose we could just kill the daemon, but that'd leave
> > > holes in the data placement which doesn't seem to be very elegant.
> > > Setting the device weight to 0 in the crushmap works, but trying to
> > > remove a device entriely produces strange results. Could you shed some
> > > light on this?
> >
> > There are a few ways to go about it. Simply marking the osd 'out' ('ceph
> > osd out #') will work, but may not be optimal depending on how the crush
> > map is set up. The default crush maps use the 'straw' bucket type
> > everywhere, which deals with addition/removal optimally, so taking the
> > additional step of removing the item from the crush map will keep things
> > tidy and erase all trace of the osd.
>
> BTW: It would be nice to be able to _replace_ an OSD without having
> a time window during which there is less redundancy.
>
> IOW, something like the "clone failing disk" feature of some hardware
> raids or LVM's pvmove. The idea is to first mirror all data on two
> disks/PVs/OSDs and to kick out one only after mirroring is complete.
You basically get this by marking the osd 'out' but not 'down', e.g.,
$ ceph osd out 23 # mark out osd23
The data on osd23 isn't removed until the pg is fully replicated/migrated
to it's new location.
sage
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
