This is probably better covered in the Ceph architecture documentation [1]. [1] http://docs.ceph.com/docs/master/architecture/#smart-daemons-enable-hyperscale -- Jason Dillaman ----- Original Message ----- > From: "louis" <louisfang2013@xxxxxxxxx> > To: "Jason Dillaman" <dillaman@xxxxxxxxxx> > Cc: ceph-users@xxxxxxxxxxxxxx, "Ric Wheeler" <rwheeler@xxxxxxxxxx> > Sent: Thursday, March 10, 2016 6:08:13 PM > Subject: 回复：Re: 回复：Re: how ceph osd handle ios sent from crashed > ceph client > Thanks, if primary osd crashed after it sent journal write to replication > osds. What will happen to this io on replication osds? > 发自网易邮箱大师 > 在2016年03月10日 23:14， Jason Dillaman 写道: > > If the IO request has already reached the OSD when the client crashes, the > > OSD will have no way to know that and will still apply the op. It will > > notice the client is gone when it attempts to send the ack message back to > > the client. Similarly, the client could also crash after the OSD has > > applied > > the op and send the ack message -- but before the ack message was > > translated > > back into a completion for the librbd user. > > > -- > > > Jason Dillaman > > > ----- Original Message ----- > > > > From: "louis" <louisfang2013@xxxxxxxxx> > > > > To: "Ric Wheeler" <rwheeler@xxxxxxxxxx> > > > > Cc: "Jason Dillaman" <dillaman@xxxxxxxxxx>, ceph-users@xxxxxxxxxxxxxx > > > > Sent: Thursday, March 10, 2016 9:11:41 AM > > > > Subject: 回复：Re: how ceph osd handle ios sent from crashed > > > ceph > > > > client > > > > Thanks, will ceph sod drop the io if finding ceph client crashed?or > > > continue > > > > to handle those iOS until completed callback? > > > > 发自网易邮箱大师 > > > > 在2016年03月10日 13:15， Ric Wheeler 写道: > > > > > On 03/08/2016 08:09 PM, Jason Dillaman wrote: > > > > > > > > > > librbd provides crash-consistent IO. It is still up to your > > > > > application > > > > > > to > > > > > > provide its own consistency by adding barriers (flushes) where > > > > > necessary. > > > > > > If you flush your IO, once that flush completes you are guaranteed > > > > > that > > > > > > your previous IO is safely committed to disk. > > > > > > > > > > > > > > > > > > > Jeff Moyer wrote a good article on how applications can manage data > > > > > durability > > > > > > > > > for lwn.net a few years back - still worth reading: > > > > > > > > > https://lwn.net/Articles/457667 > > > > > > > > > This is more focused on applications on top of file systems, but is > > > > still > > > > > > > > > relevant for applications running on block devices. > > > > > > > > > Regards, > > > > > > > > > Ric > > > > > _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
 |