Sorry, but currently I strongly not recommended multi-master.
We have a big troubles in the same case, but most high load.
I am currently working on the two major issues:
- syncrepl seems unreliable, survives no more than 3-5 min under a
stress test (ITS#7968).
- fix of ITS#7904 does not affect syncrepl, thereby MDB_MAP_FULL may
occur in case of LMDB backed (no ITS now).
But another couple seem be forever:
- replication is significantly slowdown if are more than 3 nodes.
- replication may never be completed until all ldap-write activities
are stopped.