[jira] [Created] (AIRFLOW-3045) Duplicate entry error with MySQL when update task_instances

["Haotian Wu (JIRA)" <ji...@apache.org>]

Haotian Wu created AIRFLOW-3045:
-----------------------------------

             Summary: Duplicate entry error with MySQL when update task_instances
                 Key: AIRFLOW-3045
                 URL: https://issues.apache.org/jira/browse/AIRFLOW-3045
             Project: Apache Airflow
          Issue Type: Bug
          Components: database
    Affects Versions: 1.10.0
            Reporter: Haotian Wu
            Assignee: Haotian Wu


h3. How to reproduce
# Setup apach-airflow==1.10.0 with MySQL, bring up both webserver and scheduler.
# Add a DAG and it becomes running but none of the task will be actually executed.
# Manually trigger another run for the same dag, airflow scheduler will crash with error {{sqlalchemy.exc.IntegrityError: (_mysql_exceptions.IntegrityError) (1062, "Duplicate entry 'xxxxxxx-yyyyyy-YYYY-MM-DD ...' for key 'PRIMARY'")}}

h3. The Reason

In Airflow-1.10.0, execution_date field of task_instance is changed from DateTime to Timestamp. However, in MySQL first Timestamp column in a table is declared with {{ON UPDATE CURRENT_TIMESTAMP}} clause. Database in MySQL will look like below after {{airflow initdb}}.

| Field           | Type          | Null | Key | Default              | Extra |
| task_id         | varchar(250)  | NO   | PRI | NULL                 |       |
| dag_id          | varchar(250)  | NO   | PRI | NULL                 |       |
| execution_date  | timestamp(6)  | NO   | PRI | CURRENT_TIMESTAMP(6) |  on update CURRENT_TIMESTAMP(6)      |

# When a task_instance is updated from state NULL to state "scheduled", its execution_date is also reset to current timestamp automatically. 
# task_instance is linked to a given dag_run by same execution_date, so changed execution_date means task_instance is no longer linked to any known dag_run.
# If there are more than one dag_run for the same dag, multiple task_instance with same <task_id, dag_id> will be "unlinked" to their dag_run. Airflow scheduler will try to update them to state NULL and thus try to update them to the same <task_id, dag_id, execution_date> primary key.




--
This message was sent by Atlassian JIRA
(v7.6.3#76005)