Full_Name: Quanah Gibson-Mount
Submission from: (NULL) (18.104.22.168)
If one has configured N-WAY MMR in such a fashion that only one master ever gets
the write ops (whether this is mirrormode or other methodologies), and there are
zero entries in the database with an entryCSN of the other masters in the
system, attempts to bring those other masters back up will result in a REFRESH
phase on those masters, and potentially destabilize the entire cluster (See
ITS#8789 for example).
To avoid this, we should document that best practice involves having one object
per master that is only ever modified by that master, so there is always a
minimum of one object in the database with a CSN matching that master. This
should avoid the REFRESH issue when bringing up a failed master.