In T244771: Refining is failing to refine centranoticeimpression events we had to ALTER a Hive table and then re-refine to backfill a large number of hourly datasets. After the backfill Refine completed, the Hive table's schema was correct, but previously added Hive partitions had the old schema. They had to be manually dropped and re-added to get them to pick up the new and proper table schema.

Refine already overwrites data directories when it finds that they exist; it should also forcibly re-create the Hive partition when it does this, in case the table schema has changed and needs to be propagated to the backfilled partition.