By default sensors block a worker and just sleep between pokes. This is very inefficient, especially when there are many long-running sensors.
There is a hacky workaroud by setting a small timeout value and a high retry number. But that has drawbacks:
- Errors raised by sensors are hidden and the sensor retries too often
- The sensor is retried in a fixed time interval (with optional exponential backoff)
- There are many attempts and many log files are generated
I'd like to propose an explicit reschedule mechanism:
- A new "reschedule" flag for sensors, if set to True it will raise an AirflowRescheduleException that causes a reschedule.
- AirflowRescheduleException contains the (earliest) re-schedule date.
- Reschedule requests are recorded in new `task_reschedule` table and visualized in the Gantt view.
- A new TI dependency that checks if a sensor task is ready to be re-scheduled.
- This change is backward compatible. Existing sensors behave like before. But it's possible to set the "reschedule" flag.
- The poke_interval, timeout, and soft_fail parameters are still respected and used to calculate the next schedule time.
- Custom sensor implementations can even define the next sensible schedule date by raising AirflowRescheduleException themselves.
- Existing TimeSensor and TimeDeltaSensor can also be changed to be rescheduled when the time is reached.
- This mechanism can also be used by non-sensor operators (but then the new ReadyToRescheduleDep has to be added to deps or BaseOperator).
Design decisions and caveats:
- When handling AirflowRescheduleException the `try_number` is decremented. That means that subsequent runs use the same try number and write to the same log file.
- Sensor TI dependency check now depends on `task_reschedule` table. However only the BaseSensorOperator includes the new ReadyToRescheduleDep for now.
Open questions and TODOs:
- Should a dedicated state `UP_FOR_RESCHEDULE` be used instead of setting the state back to `NONE`? This would require more changes in scheduler code and especially in the UI, but the state of a task would be more explicit and more transparent to the user.
- Add example/test for a non-sensor operator
- Document the new feature