CSVDataSet uses identityHashCode to make "unique" id for thread group and thread: https://github.com/apache/jmeter/blob/aa210850885837e7f9f58c49e0a5f503a7eb4cca/src/components/org/apache/jmeter/config/CSVDataSet.java#L166-170 In fact it does not always work, since identityHashCodes could return duplicate values for different objects. Even codes for thread group and thread could be the same. Something more unique should be used to implement "share mode".
This issue has been migrated to GitHub: https://github.com/apache/jmeter/issues/3589