We hit exactly this, offline feature store on Redshift, ~200-column denormalized table fed by domain models. Adding a feature meant NULLs for history, and full recompute was too expensive, so we landed on: recompute the domain for the backfill range into a staging table, then UPDATE only the new columns into prod (domain table + wide table). It works but it's all hand-rolled. Curious how your pivot/long-table approach held up at scale.