Re: Passing master tag around in a multi-site master-slave system

On 06/13/12 3:25 PM, Gauthier, Dave wrote:
> One master, 4 slaves.  Can only write to the master (over WAN).  No
> write transaction can be committed until it's duplicated at all the
> slave sites.  (this, so far, is I think a standard
> requirement/request).  Now, the "master" token can get passed from one
> site to the other depending on the viability of the communications
> between the sites.  If site A was the master but went down, the
> remaing 4 should be smart enough to detect this and decide who becomes
> the new master.  If site A became isolated, it ought to detect that it
> can't communicate with the other sites and that it needs to put itself
> into read_only mode.

what if A (initial master) and B are isolated from C, D, E (maybe a
transpacific link gets broken, so continents are isolated).  does C,D,E
decide its a quorum and promotes one to master, while A decides they
lost too many peers so demotes itself to offline? or does A assume that
B is sufficient slave, and retains master status, unknowingly while C
has been promoted too?

all schemes like this are frigging complicated.

oh yeah, and a 5-way synchronous commit across a global WAN is going to
be SLOW!




--
john r pierce                            N 37, W 122
santa cruz ca                         mid-left coast