Task Instance and Retry

A task instance represents a task in a particular DAG Run; mapped tasks also have a map index. Its state records execution progress, while attempts describe retries or re-execution of that instance.

For an ordinary failing task with retries=2, the retry allowance permits an initial attempt and up to two more attempts. Delay and timeout settings control parts of this process, not whether an external effect already happened. A failure after a database commit can leave committed data behind.

Retry with stable identity and a destination rule that handles repeated effects. Reusing an interval key for a fixed replacement is one example; changed input versions and overlapping writers need further protection. Clearing instances for re-execution is an explicit operation whose upstream and downstream scope should be checked.

See Apache Airflow Fundamentals: DAGs, Tasks, and Scheduling for worked examples.

Further reading

• Airflow 3.1.0 Documentation, Tasks