[sheepdog] zookeeper driver join panic

[Liu Yuan]

On Fri, Jun 28, 2013 at 07:25:51PM +0800, Kai Zhang wrote:
> 
> On Jun 28, 2013, at 11:41 AM, Liu Yuan <namei.unix at gmail.com> wrote:
> 
> > The current master election process is too complicted to me, the payback is to
> > solve a very unusal corner startup case. I guess it needs time to get stablized,
> > Is this really worthwhile? We should reconsider the old method, since it is quit
> > simple and well tested. What do you think?
> 
> To be honest, I have an opposite opinion. 
> 
> -  I don't think the current master election process is too complicated.
>     There is only one simple rule: sheep with the least sequential number become master.
>     And this is also a standard way that is suggested officially.
>     And I believe there are many distributed systems are using the same way for master election.
> 
> -  It is not an unusual case, every time when network is crowd for a long time, it will be sure to happen.
> 
> -  The current zookeeper driver has been tested under several situations.
>     In addition, I will also start to do more testings as soon as possible.
>     I think I can manage to fix any problem about the stability.
> 
> -  I don't think 'quit' is a good manner in a production environment. We should do our best
>     to protect them from quit. (This is because a 'quit' will be aware by Guest OS)
> 
> -  Based on our testing results, the previous zookeeper driver is not quite stable, at least, not as stable as corosync.
> 
> -  Maybe the complexity is caused by the current implementation. I think we can clean code if needed.
>     However, I prefer the dirty code which can do right things than a clean code which contains bugs.
>     The original intention of the patch is that previous implementation has fatal error.

Okay, I was not convinced to restore to old master election, just a concern that
current master eletion isn't as simple as old ones then harder to maintain. 
Anyway, it will get stable as time goes by.

Thanks
Yuan