It uses regular expression part.+c\d{3} to find uncompact parquet. But I think only file written by Spark can have this feature.