Introduction

The DAG_airflow_db_cleaner.py DAG periodically cleans some records of the Airflow metadata database. The goal is to keep airflow fast and avoid performance issues.

Setup

1. In your metadata database, create a new user and give to it drop and read permissions on the log, task_instance, job and dag_run tables in the airflow database.

2. In a secure directory, create a file containing your new user's password.

3. Alter the DAG_airflow_db_cleaner.py file and change the line of the first occurrence of the passFile variable with the password path.
   It should look like this:

# Path to the password file
passFile = open("/my/safe/directory/user.passfile", "r")

4. Next, change the session variable giving your connection informations to the conn function. The first parameter is the user; the next is a password variable called passwd that was defined with the result of your passfile path, so you don't need to do anything here; then you need to inform the host; next, the connection port; and finally the Airflow database.
   It should look like this:

session = conn('user',passwd,'localhost','3306','airflow')

5. Finally, when defining your DAG DAG_airflow_db_cleaner = DAG( .... ) you can adjust the start_date parameter.
   If everything is ok, it is done!

How it works?

By default, every 15 days the DAG performs the following cleanings:

• Table log: all the logs stored before the current date will be deleted.
• Table task_instance: all records older than 15 days will be deleted, but not the last one.
• Table job: records with running estate, latest_heartbeat >= 15 days and start_date with more than 15 days, will be deleted. These kind of records can be generated by Airflow issues, and they never changes their estate.
• Table dag_run: all the dagruns with running estate and start_date with more than 15 days will be deleted. As on the job case, this is also to avoid endless dagruns.

Feel free to change these "rules" inside the clean(...) function of each table model, add new models, etc.