Yes,I believe the observed race condition is to blame for at least a part
of the problem. If this hypothesis holds, then if would mean this is independent
from your patch, though. We'e probably should bring it up in a different issue
and find a way to demonstrate it with a regular version of roundup.
Note that this race condition is in rdbms_common.py,
but back_mysql.py overrides the function and seems to deal 
at least with one concurrent write situation.

It would mean that some decisions based on auditors or reactors may be wrong
in a concurrent edit situation for postgresql.

Still I believe this code is strange:
            current = node.get(propname, None)
            if value == current:
                del propvalues[propname]
                continue
            journalvalues[propname] = current

shouldn't it be 
  journalvalues[propname] = value 
instead of "current"?