>From http://pythonhosted.org/airflow/faq.html:

*What’s the deal with ``start_date``?*

start_date is partly legacy from the pre-DagRun era, but it is still
relevant in many ways. When creating a new DAG, you probably want to set a
global start_date for your tasks usingdefault_args. The first DagRun to be
created will be based on the min(start_date) for all your task. From that
point on, the scheduler creates new DagRuns based on your schedule_interval and
the corresponding task instances run as your dependencies are met. When
introducing new tasks to your DAG, you need to pay special attention to
start_date, and may want to reactivate inactive DagRuns to get the new task
to get onboarded properly.

We recommend against using dynamic values as start_date, especially
datetime.now() as it can be quite confusing. The task is triggered once the
period closes, and in theory an @hourly DAG would never get to an hour
after now as now() moves along.

Previously we also recommended using rounded start_date in relation to your
schedule_interval. This meant an @hourly would be at 00:00 minutes:seconds,
a @daily job at midnight, a @monthlyjob on the first of the month. This is
no longer required. Airflow will not auto align the start_dateand the
schedule_interval, by using the start_date as the moment to start looking.

You can use any sensor or a TimeDeltaSensor to delay the execution of tasks
within the schedule interval. While schedule_interval does allow specifying
a datetime.timedelta object, we recommend using the macros or cron
expressions instead, as it enforces this idea of rounded schedules.

When using depends_on_past=True it’s important to pay special attention to
start_date as the past dependency is not enforced only on the specific
schedule of the start_date specified for the task. It’ also important to
watch DagRun activity status in time when introducing new
depends_on_past=True, unless you are planning on running a backfill for the
new task(s).

Also important to note is that the tasks start_date, in the context of a
backfill CLI command, get overridden by the backfill’s command start_date.
This allows for a backfill on tasks that havedepends_on_past=True to
actually start, if it wasn’t the case, the backfill just wouldn’t start.

On Tue, Aug 9, 2016 at 7:44 AM, הילה ויזן <[email protected]> wrote:

> Hi,
>
> We're experiencing a strange problem with the start_date configuration in
> Airflow.
>
> When we first ran the DAGs, we defined the start_date as 'datetime.now()',
> which at the time was 01/08/2016. This worked fine. A week afterwards, we
> changed the DAGs to a specific newer date - 08/08/2016, and reset all of
> the tasks. After resetting the Airflow and all of the DAGs *we are still
> seeing the tasks running from original date (01/08)*. Why is this
> happening?
>
> We don't understand why the tasks are still using the old date. Is there a
> cache/DB/persistent file that the DAG reads on startup that overrides our
> definition? Is it maybe Celery? We really would appreciate your input
> because we are totally stuck.
>
> We use airflow version 1.7.1.3 with postgress as the backend DB.
> In addition, we run in CeleryExecutor mode with rabbitMQ as Celery backend.
>
> Thank you,
> Hila
>

Reply via email to