Topic
Scheduling that survives re-runs and backfills: data intervals, idempotent loads and the Airflow defaults that change between versions.
A scheduler’s job looks simple until a day has to be re-run, a backfill goes in the wrong order, or an upgrade changes what a schedule means. Most of the failures here end with every task green: the load ran, it just asked for the wrong rows. These notes run real DAGs over data whose correct answer is known and count what each design loses.
Airflow 3's @daily has a zero-length data interval, so a window-based load ran green for 30 days and loaded nothing. Backfilled backwards, watermarks lost 97%.
9 min read