On Fri, Jun 20, 2014 at 05:16:19PM +0800, Robin Dong wrote: > 2014-06-20 16:51 GMT+08:00 Liu Yuan <namei.u...@gmail.com>: > > > On Fri, Jun 20, 2014 at 04:21:27PM +0800, Robin Dong wrote: > > > From: Robin Dong <san...@taobao.com> > > > > > > After following steps: > > > > > > 1. start 6 sheep cluster, every sheep has 4 data path > > > 2. unplug 4 data path for one sheep daemon > > > > > > then we will see the sheep daemon which has no data path is also > > > in the cluster by using 'dog node list'.But a sheep daemon with > > > no data path should be a pure gateway. > > > > > > So we should return EIO after unpluging last disk which will cause > > > sheep to leave cluster. > > > > > > > sheep will leave cluster by EIO later when it tries to access disks. See > > md_handle_eio(). > > > > > Suppose you unplug all the 4 disks for maintainance and then plug back 4 > > new > > disks. We should allow this use case, so this patch looks wrong to me. > > > > By the way, what is purpose of making a sheep as pure gateway while > > running? > > > > If all disks corrupt, the sheep daemon on this node will become a pure > gateway, after this, > some one 's wrong operation (such as plug a new disk) will make all nodes > coredump.
If all disks are borken, then sheep will get EIO finally and leave cluster with the currrent code. Thanks Yuan -- sheepdog mailing list sheepdog@lists.wpkg.org http://lists.wpkg.org/mailman/listinfo/sheepdog
