> Op 21 september 2016 om 17:23 schreef Iain Buclaw <ibuc...@gmail.com>:
> On 20 September 2016 at 19:27, Gregory Farnum <gfar...@redhat.com> wrote:
> > In librados getting a stat is basically equivalent to reading a small
> > object; there's not an index or anything so FileStore needs to descend its
> > folder hierarchy. If looking at metadata for all the objects in the system
> > efficiently is important you'll want to layer an index in somewhere.
> > -Greg
> >
> Yeah, that's not particularly good to hear.  Is this slowness also
> inherent in list_nobjects too?  It looks like I can iterate all
> objects at a rate no faster than 25K per second.  No chance at
> speeding this up either by having two or more instances starting at
> different pg offsets.

RADOS has no index of objects. Everything is done using calculation. So when 
listing objects you basically have to go to each primary OSD for all PGs in a 
pool and ask them what objects are in the pool/PG.

> For this particular operation, it's only looking for orphaned objects.
> This wouldn't be needed if a mechanism for TTLs existed and set on all
> objects.  But that would mean finding out how RGW gets away with it,
> and I assume with another very large index and actively keeping track
> of all set destruction times.

RGW indeed keeps a index, but that limits it in the amount of objects you can 
store in a single bucket. Yes, bucket sharding helps, but the limit for a RGW 
bucket is still way lower then for a RADOS pool.


> -- 
> Iain Buclaw
> _______________________________________________
> ceph-users mailing list
> ceph-users@lists.ceph.com
> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
ceph-users mailing list

Reply via email to