[sheepdog] [PATCH 0/4] variable vnodes specified by free space

Liu Yuan namei.unix at gmail.com
Wed Aug 8 09:35:48 CEST 2012


On 08/08/2012 02:51 PM, levin li wrote:
> I tested it yesterday, and find that it doesn't pass the
> path(98241c50d23d99047c37a606a845f5f0bccbbf50), I've not had time to
> dig more deeply into it.

It seems that we never sort it out when we introduced delayed list in
the code. Even latest patch set from Christoph fails to address problem,
we still can't pass the scripts in sdog-tests/cluster/*.sh

In the assumption that we can't easily get to know who is the node with
highest epoch in a crashed cluster, then I am suspecting if the delayed
list is really useful for practical useage: we will asks all other nodes
to leave when mastership transfers, then we don't get any benefit than
old simple code without delayed list, because most chances we will see a
mastership transfer happen.

Consider with lots of complexity added without much benefit, I am asking
if we would better remove delayed list and go back to the old simple
code yet practically turn out the same effect for a crashed cluster.

Thanks,
Yuan



More information about the sheepdog mailing list