It’s an optimization technique. hive.fetch.task.conversion property can (FETCH task) minimize latency of mapreduce overhead. When queried SELECT, FILTER, LIMIT queries, this property skip mapreduce and using FETCH task. As a result Hive can execute query without run mapreduce task.
By default it’s value “minimal”. Which optimize: SELECT STAR, FILTER on partition columns, LIMIT queries only, where as another value is “more” which optimize : SELECT, FILTER, LIMIT only (+TABLESAMPLE, virtual columns).