Is the table salted? If the SALT_BUCKETS property was ever changed during the life of the table, this could lead to duplicates. You could get the same row key with a different salt byte. If not salted, are there any global, mutable secondary indexes. A number of issues have been fixed which will be in 4.12 release that lead to out-of-sync issues between the data and index table under high load (in particular when the same set of rows are frequently changing).
If you could file a JIRA with more information (Phoenix version, HBase version, DDL, indexes, etc), that'd be much appreciated. We can discuss more there.