The archiver stages all db traffic to per-key data files, so if the DB goes away the data can later be flushed. When the archiver shuts down it makes a new connection to the db and flushes all the files. But it flushes all of them inside a single transaction, and deletes the cache files as it reads them: if anything goes wrong in the transaction the whole flush is rolled back, but the files are then gone and all pending data is lost.

Move the file deletion to a step after the transaction has been successfully committed.