Why does a backfill of millions of rows run in small batches from a background job, after dual writes have started, rather than as one UPDATE before them?

Answer

One big UPDATE holds locks and floods the primary (and its replicas' replication lag) while users are writing. Small, throttled batches leave room for live traffic. Starting dual writes first means every row written during the backfill already has the new shape, so the backfill only has to cover rows written before, and it ends with nothing left behind.

Review this in your daily deck → All cards in Databases