On Thu, Apr 28, 2005 at 11:44:59AM -0400, Jan Harkes wrote:
> This was caused by a subtle error in the file-resolution code in 6.0.9,
> where we would always propage the data from the server with index 1
> instead of the dominant server (which is index i). As a result there is
> a 50% chance with double replication that everything is perfectly find,

Actually, just thinking a bit more about this.

This problem should never hit for doubly replicated servers, it only
affect 3 or more replicas and then only if the dominant copy is located
on a server with index > 1.

The code selects the first valid server as the dominant index (i.e. 0)
and then walks through the list and checks if any of the VVs is strictly
higher. If that is the case it will update the dominant index, which was
incorrectly being set to one.

However for 2 replicas this is perfectly fine, which is probably why it
took about 7 weeks before anyone noticed this, I typically don't test
with more than 2 replicas. The problem only happens with 3 or more
replicas and if the dominant copy happens to be neither server 0 or
server 1.

Jan