On Fri, Jun 01, 2012 at 11:54:00PM +0800, Liu Yuan wrote: > Okay, I comfired this scenario, I think it is regression to let other > nodes join in when cluster is waiting for original ones. I don't think it's a regression - the sheepdog join code always just counted the number of nodes to see if it can start a cluster. That doesn't mean I agree that it's a good behaviour, but it's defintively how things have been as far as I can look back. Independent of fixing a nasty crash I think this patch series also is a great cleanup as it makes sure all of recovery simply operates on the in-memory data structures. I plan to look into the join code a bit more soon, and hopefully implement a better algorithm to set the cluster live than now. |