Suppose that there is a requisition on which 3 different nodes have the exact same IP marked as primary (i.e. snmp-primary='P'). Keep in mind that the 3 nodes may be a combination of discovered and requisitioned nodes, and could belong to different requisitions (it doesn't matter how they have been added to OpenNMS).
Collectd is not going to work properly on this scenario, and the RRDs are going to contain invalid and wrong data. This also affects threshold processing.
The solution is clean the requisitions (and the discovered nodes if necessary) to be sure that each primary IP address is being used once.
After fixing the problem and verifying that the IP is associated with only one node on the database (i.e. ipinterface table), the ILR is still reporting a bad collection count for the IP in question, which means Collectd is not rescheduling the data collection after fixing the problem.
The current solution is to restart OpenNMS, after fixing a duplicate IP problem.