As was noticed in
CASSANDRA-7317, using '-pr' alongside '-local' for repair doesn't really work properly, and disabling the combination was definitively the right short time fix. However, the main goal of '-pr' is to make it easy to repair a full cluster without doing any duplication of work. Doing the same only within a data-center is obviously desirable.
I think a reasonably simple solution would be modify the behavior of '-pr' when it's limited to only one DC. If applied to nodeX in dcY, instead of repairing only the "primary" range of nodeX for the whole ring, we'll repair that range but also all ranges that are "primary" for a node not in dcY and for which nodeX is the first node of dcY found in ring order. Basically we'll ensure that running 'repair -local -pr' on every nodes of a given DC will repair all ranges for the nodes of that DC without repairing the same range twice.