On Fri, 13 Jun 2014, Charles 'Boyo wrote: > Aha! Thanks Sage. > > I completely get it now. So I can use a ramdisk provided it is always > flushed to disk during shutdowns and I never have unplanned outages > right?
i.e., never! :) > Does this hard OSD consistency also explain why eventual consistency at > the RADOS level was "designed" out? Having all OSDs in a replicated set > writing at least to their journal causes latency concerns with my > WAN-type links. The consistency of the local data is important so that after a restart or failure we can rely on accurate metadata to quickly resync without having to walk the data on the disk. It's not directly related to the distributed consistency model, though. sage > > Charles > ------Original Message------ > From: Sage Weil > To: Charles 'Boyo > Cc: [email protected] > Subject: Re: [ceph-users] Disabling OSD journals, parallel reads and eventual > consistency for RBD > Sent: Jun 13, 2014 01:31 > > On Fri, 13 Jun 2014, Charles 'Boyo wrote: > > Hello Sage. > > > > I'm running xfs and crashes are rarely enough. When they do happen, I > > would rather just rebuild the entire cluster than bother with fsck > > anyway. > > I mean any unclean/abrupt shutdown of ceph-osd, not an XFS error. Like a > power failure. If you disable the journal, a simple 'reboot -f -n' would > corrupt the data Ceph has stored on the disk. It has nothing to do with > whether XFS is consistent (it is crash safe), but whether the data Ceph is > writing into XFS is consistent with itself (it requires the ceph journal > for crash safety). > > > So can you show me how to turn off journalling using the xfs FileStore > > backend? :) > > You could point the journal at tmpfs. But, you *will* lose data if you > do that! > > sage > > > > > Charles > > ------Original Message------ > > From: Sage Weil > > To: Charles 'Boyo > > Cc: [email protected] > > Subject: Re: [ceph-users] Disabling OSD journals, parallel reads and > > eventual consistency for RBD > > Sent: Jun 13, 2014 00:29 > > > > On Thu, 12 Jun 2014, Charles 'Boyo wrote: > > > Hello list. > > > > > > Is it possible, or will it ever be possible to disable the OSD's > > > journalling activity? > > > > > > I understand it is risky and has the potential for data loss but in my > > > use case, the data is easily re-built from scratch and I'm really > > > bothered about the reduced throughput "wasted" on journalling. > > > > We are investigating alternative backends that do not require the journal. > > There are also half-baked plans to avoid the double-writes when running on > > btrfs. You definitely do not want to disable the journal or stick it on a > > ramdisk for ext4/xfs, though, or you will corrupt the OSD's data on crash. > > > > sage > > > > Sent from my BlackBerry? wireless handheld from Glo Mobile. > > Sent from my BlackBerry? wireless handheld from Glo Mobile. _______________________________________________ ceph-users mailing list [email protected] http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
