We currently do backfilling of wmf_dumps.wikitext_raw_rc1 with a job that ingests data on a per year basis. Details on rationale are available at T340861: Implement a backfill job for the dumps hourly table.
The number one consumer of time for this approach is the constant re-reading of the wmf.mediawiki_wikitext_history table, which, for, say, enwiki, is ~11TB. A MERGE INTO wants to read this table 3 times, so it quickly becomes expensive time wise.
In this task we should explore whether having a temporary table that partitions and transforms wmf.mediawiki_wikitext_history to a more efficient form gains us performance.