Why Your Airflow DAG Runs a Day Late (It Doesn't)
A daily DAG labelled Monday starts on Tuesday. That is Airflow working as designed, and it is also the default that changed in Airflow 3. Here is what the logical date means and what moves when you upgrade.
The run that is dated yesterday
Schedule a DAG with 0 0 * * * on Airflow 2 and look at the run that starts on Tuesday 3 November 2026 at 00:00. Its logical date is Monday 2 November. The ds variable in its templates reads 2026-11-02. Nothing is broken, and restarting the scheduler will not change it.
The same DAG file on Airflow 3 starts at the same moment with a logical date of Tuesday 3 November, so ds reads 2026-11-03.
What it does to a pipeline
Take a task that loads one day of data with WHERE event_date = '{{ ds }}'. On Airflow 2 the Tuesday run loads Monday, which is the only full day available at midnight on Tuesday. The job was written around that behaviour, usually without anyone writing it down.
Upgrade to Airflow 3 and change nothing else. The Tuesday run now asks for Tuesday. At 00:00 on Tuesday there is no Tuesday data yet, so the task succeeds on zero rows, and Monday is never loaded by any run. The DAG is green every day and the table is a day behind, or empty, depending on how the downstream models treat a missing partition.
The mechanism
Airflow 2 gives a cron string the CronDataIntervalTimetable. A run belongs to a period, and it starts when that period closes. The daily run for Monday covers Monday 00:00 up to Tuesday 00:00, so it starts on Tuesday at 00:00. Its logical date is the start of the period, which is why it reads Monday. data_interval_start and data_interval_end are the two ends of that window.
Airflow 3 changed two defaults. create_cron_data_intervals is now False, so a cron string gets the CronTriggerTimetable, which treats the schedule as a clock. The run for Tuesday 00:00 starts on Tuesday 00:00, its logical date is that same instant, and its data interval starts and ends there. catchup_by_default is also False now, so a DAG that used to backfill every missed period when it was unpaused creates only the most recent run.
The timetables themselves barely moved. We ran the real timetable classes from Airflow 2.11.2 and 3.4.0 over 3,580 combinations of cron expression, timezone, start date and catchup setting. With the same timetable and catchup chosen explicitly, the two versions agree on all but 12 cases, and all 12 sit right after a daylight saving change. What changed is which timetable you get when you do not choose.
Daylight saving moves runs too, because cron is evaluated on the wall clock of the DAG timezone. A DAG scheduled for 30 2 * * * in America/New_York has no 02:30 on Sunday 8 March 2026, because the clock jumps from 02:00 to 03:00. Airflow runs it at 03:30 that day and at 02:30 again on Monday.
What to do about it
If your tasks read ds, data_interval_start or data_interval_end, decide before the upgrade which behaviour you want and say so in the DAG. Pass CronDataIntervalTimetable explicitly, or set create_cron_data_intervals = True, to keep the Airflow 2 dates. Set catchup on each DAG instead of leaning on the default. If you would rather move to the trigger timetable, change the tasks to compute the day they want from the trigger time, and do it in the same change.
To see what a specific schedule does on either version, the Airflow schedule explainer lists each run with its start time, logical date and data interval. It runs in the browser and its answers are tested against those same Airflow classes.
A DAG that looks a day late on Airflow 2 is on time: the date on the run names the period it covers, and the run starts when that period ends.