Upgrading Airflow to a newer version

Why you need to upgrade

Newer Airflow versions can contain database migrations so you must run airflow db upgrade to upgrade your database with the schema changes in the Airflow version you are upgrading to. Don’t worry, it’s safe to run even if there are no migrations to perform.

Upgrade preparation - make a backup of DB

It is highly recommended to make a backup of your metadata DB before any migration. If you do not have a “hot backup” capability for your DB, you should do it after shutting down your Airflow instances, so that the backup of is consistent. If you did not make a backup and your migration fails, you might end-up in a half-migrated state and restoring DB from backup and repeating the migration might be the only easy way out. This can for example be caused by a broken network connection between your CLI and the database while the migration happens, so taking a backup is an important precaution to avoid problems like this.

When you need to upgrade

If you have a custom deployment based on virtualenv or Docker Containers, you usually need to run the DB upgrade manually as part of the upgrade process.

In some cases the upgrade happens automatically - it depends if in your deployment, the upgrade is built-in as post-install action. For example when you are using Helm Chart for Apache Airflow with post-upgrade hooks enabled, the database upgrade happens automatically right after the new software is installed. Similarly all Airflow-As-A-Service solutions perform the upgrade automatically for you, when you choose to upgrade airflow via their UI.

How to upgrade

In order to manually upgrade the database you should run the airflow db upgrade command in your environment. It can be run either in your virtual environment or in the containers that give you access to Airflow CLI Using the Command Line Interface and the database.

Offline SQL migration scripts

If you want to run the upgrade script offline, you can use the -r or --revision-range flag to get the SQL statements that would be executed. This feature is supported in Postgres and MySQL from Airflow 2.0.0 onward and in MSSQL from Airflow 2.2.0 onward.

Sample usage:

airflow db upgrade -r "2.0.0:2.2.0" airflow db upgrade --revision-range "e959f08ac86c:142555e44c17"

Handling migration problems

Wrong Encoding in MySQL database

If you are using old Airflow 1.10 as a database created initially either manually or with previous version of MySQL, depending on the original character set of your database, you might have problems with migrating to a newer version of Airflow and your migration might fail with strange errors (“key size too big”, “missing indexes” etc). The next chapter describes how to fix the problem manually.

Why you might get the error? The recommended character set/collation for MySQL 8 database is utf8mb4 and utf8mb4_bin respectively. However this has been changing in different versions of MySQL and you could have custom created database with a different character set. If your database was created with an old version of Airflow or MySQL, the encoding could have been wrong when the database was created or broken during migration.

Unfortunately, MySQL limits the index key size and with utf8mb4, Airflow index key sizes might be too big for MySQL to handle. Therefore in Airflow we force all the “ID” keys to use utf8 character set (which is equivalent to utf8mb3 in MySQL 8). This limits the size of indexes so that MySQL can handle them.

Here are the steps you can follow to fix it BEFORE you attempt to migrate (but you might also choose to do it your way if you know what you are doing).

Get familiar with the internal Database structure of Airflow which you might find at ERD Schema of the Database and list of migrations that you might find in Reference for Database Migrations.

  1. Make a backup of your database so that you can restore it in case of a mistake.

  2. Check which of the tables of yours need fixing. Look at those tables:

SHOW CREATE TABLE task_reschedule;
SHOW CREATE TABLE xcom;
SHOW CREATE TABLE task_fail;
SHOW CREATE TABLE rendered_task_instance_fields;
SHOW CREATE TABLE task_instance;

Make sure to copy the output. You will need it in the last step. Your dag_id, run_id, task_id and key columns should have utf8 or utf8mb3 character set set explicitly, similar to:

``task_id`` varchar(250) CHARACTER SET utf8 COLLATE utf8_bin NOT NULL,  # correct

or

``task_id`` varchar(250) CHARACTER SET utf8mb3 COLLATE utf8mb3_bin NOT NULL,  # correct

The problem is if your fields have no encoding:

``task_id`` varchar(250),  # wrong !!

or just collation set to utf8mb4:

``task_id`` varchar(250) COLLATE utf8mb4_unicode_ci DEFAULT NULL,  # wrong !!

or character set and collation set to utf8mb4

``task_id`` varchar(250) CHARACTER SET utf8mb4 COLLATE utf8mb4_bin NOT NULL,  # wrong !!

You need to fix those fields that have wrong character set/collation set.

3. Drop foreign key indexes for tables you need to modify (you do not need to drop all of them - do it just for those tables that you need to modify). You will need to recreate them in the last step (that’s why you need to keep the SHOW CREATE TABLE output from step 2.

ALTER TABLE task_reschedule DROP FOREIGN KEY task_reschedule_ti_fkey;
ALTER TABLE xcom DROP FOREIGN KEY xcom_task_instance_fkey;
ALTER TABLE task_fail DROP FOREIGN KEY task_fail_ti_fkey;
ALTER TABLE rendered_task_instance_fields DROP FOREIGN KEY rtif_ti_fkey;

4. Modify your ID fields to have correct character set/encoding. Only do that for fields that have wrong encoding (here are all potential commands you might need to use):

ALTER TABLE task_instance MODIFY task_id VARCHAR(255) CHARACTER SET utf8mb3 COLLATE utf8mb3_bin;
ALTER TABLE task_reschedule MODIFY task_id VARCHAR(255) CHARACTER SET utf8mb3 COLLATE utf8mb3_bin;

ALTER TABLE rendered_task_instance_fields MODIFY task_id VARCHAR(255) CHARACTER SET utf8mb3 COLLATE utf8mb3_bin;
ALTER TABLE rendered_task_instance_fields MODIFY dag_id VARCHAR(255) CHARACTER SET utf8mb3 COLLATE utf8mb3_bin;

ALTER TABLE task_fail MODIFY task_id VARCHAR(255) CHARACTER SET utf8mb3 COLLATE utf8mb3_bin;
ALTER TABLE task_fail MODIFY dag_id VARCHAR(255) CHARACTER SET utf8mb3 COLLATE utf8mb3_bin;

ALTER TABLE sla_miss MODIFY task_id VARCHAR(255) CHARACTER SET utf8mb3 COLLATE utf8mb3_bin;
ALTER TABLE sla_miss MODIFY dag_id VARCHAR(255) CHARACTER SET utf8mb3 COLLATE utf8mb3_bin;

ALTER TABLE task_map MODIFY task_id VARCHAR(255) CHARACTER SET utf8mb3 COLLATE utf8mb3_bin;
ALTER TABLE task_map MODIFY dag_id VARCHAR(255) CHARACTER SET utf8mb3 COLLATE utf8mb3_bin;
ALTER TABLE task_map MODIFY run_id VARCHAR(255) CHARACTER SET utf8mb3 COLLATE utf8mb3_bin;

ALTER TABLE xcom MODIFY task_id VARCHAR(255) CHARACTER SET utf8mb3 COLLATE utf8mb3_bin;
ALTER TABLE xcom MODIFY dag_id VARCHAR(255) CHARACTER SET utf8mb3 COLLATE utf8mb3_bin;
ALTER TABLE xcom MODIFY run_id VARCHAR(255) CHARACTER SET utf8mb3 COLLATE utf8mb3_bin;
ALTER TABLE xcom MODIFY key VARCHAR(255) CHARACTER SET utf8mb3 COLLATE utf8mb3_bin;
  1. Recreate the foreign keys dropped in step 3.

Repeat this one for all the indexes you dropped. Note that depending on the version of Airflow you Have, the indexes might be slightly different (for example map_index was added in 2.3.0) but if you keep the SHOW CREATE TABLE output prepared in step 2., you will find the right CONSTRAINT_NAME and CONSTRAINT to use.

# Here you have to copy the statements from SHOW CREATE TABLE output
ALTER TABLE <TABLE> ADD CONSTRAINT `<CONSTRAINT_NAME>` <CONSTRAINT>

This should bring the database to the state where you will be able to run the migration to the new Airflow version.

Post-upgrade warnings

Typically you just need to successfully run airflow db upgrade command and this is all. However in some cases, the migration might find some old, stale and probably wrong data in your database and moves it aside to a separate table. In this case you might get warning in your webserver UI about the data found.

Typical message that you might see:

Airflow found incompatible data in the <original table> table in the metadatabase, and has moved them to <new table> during the database migration to upgrade. Please inspect the moved data to decide whether you need to keep them, and manually drop the <new table> table to dismiss this warning.

When you see such message, it means that some of your data was corrupted and you should inspect it to determine whether you would like to keep or delete some of that data. Most likely the data was corrupted and left-over from some bugs and can be safely deleted - because this data would not be anyhow visible and useful in Airflow. However if you have particular need for auditing or historical reasons you might choose to store it somewhere. Unless you have specific reasons to keep the data most likely deleting it is your best option.

There are various ways you can inspect and delete the data - if you have direct access to the database using your own tools (often graphical tools showing the database objects), you can drop such table or rename it or move it to another database using those tools. If you don’t have such tools you can use the airflow db shell command - this will drop you in the db shell tool for your database and you will be able to both inspect and delete the table.

How to drop the table using Kubernetes:

  1. Exec into any of the Airflow pods - webserver or scheduler: kubectl exec -it <your-webserver-pod> python

  2. Run the following commands in the python shell:

from airflow.settings import Session

session = Session()
session.execute("DROP TABLE _airflow_moved__2_2__task_instance")
session.commit()

Please replace <table> in the examples with the actual table name as printed in the warning message.

Inspecting a table:

SELECT * FROM <table>;

Deleting a table:

DROP TABLE <table>;

Migration best practices

Depending on the size of your database and the actual migration it might take quite some time to migrate it, so if you have long history and big database, it is recommended to make a copy of the database first and perform a test migration to assess how long the migration will take. Typically “Major” upgrades might take longer as adding new features require sometimes restructuring of the database.

Was this entry helpful?