On Tue, Jul 24, 2012 at 3:54 PM, Dietmar Maurer <dietmar at proxmox.com> wrote: >> If node membership change is detected, object recovery (object >> replacement) start on >> all nodes. If we delay starting recovery and reject I/O requests until it, all >> nodes cannot accept I/Os during the time and it causes EIO to VMs. > > I fully understand that. but you can delay recovering of 'unused' object (only > handle prio_oids, and do other oids later)? Probably, yes, and it's simpler than introducing write journal to data objects. It would be difficult to handle node failure whose write journal is not flushed. Thanks, Kazutaka |