A backfill materializes many partitions at once: history after a new asset ships, or a range after a bug fix. From the asset page, choose Backfill, pick a range and, if you only want the gaps, tick Only missing and failed partitions. Each backfill gets an id, and its runs are tagged with it.
By default a backfill launches one run per partition. That isolates failures but is slow for hundreds of small days. With backfill_policy=dg.BackfillPolicy.single_run() the whole range runs in one run: context.partition_keys lists the days and context.partition_time_window spans them all. Returning a dictionary keyed by partition stores each day separately.
From the terminal, dg launch --assets ... --partition-range 2026-11-03...2026-11-06 launches a range in one run, which requires the single-run policy.
Make partitioned assets idempotent: rerunning a day must replace that day, not add to it, because backfills and retries will rerun days.
A whole range in one run, stored day by day.