Automatically use utf8mb3_general_ci collation for mysql

The index size is too big in case utf8mb4 is used as encoding
for MySQL database. We already had `sql_engine_collation_for_ids`
configuration to allow the id fields to use different collation,
but the user had to set it up manually in case of a failure to
create a db and it was not obvious, not discoverable and rather
clumsy.

Since this is really only a problem with MySQL the easy solution
is to force this parameter to utf8mb3_general_ci for all mysql
databases. It has no negative consequences, really as all
relevant IDs are ASCII anyway.

Related: #17603

airflow/config_templates/config.yml

@@ -81,9 +81,10 @@
     - name: sql_engine_collation_for_ids
       description: |
         Collation for ``dag_id``, ``task_id``, ``key`` columns in case they have different encoding.
-        This is particularly useful in case of mysql with utf8mb4 encoding because
-        primary keys for XCom table has too big size and ``sql_engine_collation_for_ids`` should
-        be set to ``utf8mb3_general_ci``.
+        By default this collation is the same as the database collation, however for ``mysql`` and ``mariadb``
+        the default is ``utf8mb3_general_ci`` so that the index sizes of our index keys will not exceed
+        the maximum size of allowed index when collation is set to ``utf8mb4`` variant
+        (see https://github.com/apache/airflow/pull/17603#issuecomment-901121618).
       version_added: 2.0.0
       type: string
       example: ~

airflow/config_templates/default_airflow.cfg

@@ -63,9 +63,10 @@ sql_alchemy_conn = sqlite:///{AIRFLOW_HOME}/airflow.db
 sql_engine_encoding = utf-8
 
 # Collation for ``dag_id``, ``task_id``, ``key`` columns in case they have different encoding.
-# This is particularly useful in case of mysql with utf8mb4 encoding because
-# primary keys for XCom table has too big size and ``sql_engine_collation_for_ids`` should
-# be set to ``utf8mb3_general_ci``.
+# By default this collation is the same as the database collation, however for ``mysql`` and ``mariadb``
+# the default is ``utf8mb3_general_ci`` so that the index sizes of our index keys will not exceed
+# the maximum size of allowed index when collation is set to ``utf8mb4`` variant
+# (see https://github.com/apache/airflow/pull/17603#issuecomment-901121618).
 # sql_engine_collation_for_ids =
 
 # If SqlAlchemy should pool database connections.

airflow/models/base.py

@@ -44,6 +44,19 @@ def get_id_collation_args():
     if collation:
         return {'collation': collation}
     else:
+        # Automatically use utf8mb3_general_ci collation for mysql
+        # This is backwards-compatible. All our IDS are ASCII anyway so even if
+        # we migrate from previously installed database with different collation and we end up mixture of
+        # COLLATIONS, it's not a problem whatsoever (and we keep it small enough so that our indexes
+        # for MYSQL will not exceed the maximum index size.
+        #
+        # See https://github.com/apache/airflow/pull/17603#issuecomment-901121618.
+        #
+        # We cannot use session/dialect as at this point we are trying to determine the right connection
+        # parameters, so we use the connection
+        conn = conf.get('core', 'sql_alchemy_conn', fallback='')
+        if conn.startswith('mysql') or conn.startswith("mariadb"):
+            return {'collation': 'utf8mb3_general_ci'}
         return {}
 
 

docs/apache-airflow/howto/set-up-database.rst

@@ -154,14 +154,18 @@ In the example below, a database ``airflow_db`` and user  with username ``airflo
 
 .. code-block:: sql
 
-   CREATE DATABASE airflow_db CHARACTER SET utf8 COLLATE utf8_general_ci;
+   CREATE DATABASE airflow_db CHARACTER SET utf8 COLLATE utf8mb4_unicode_ci;
    CREATE USER 'airflow_user' IDENTIFIED BY 'airflow_pass';
    GRANT ALL PRIVILEGES ON airflow_db.* TO 'airflow_user';
 
 
 .. note::
 
-   The database must use a UTF-8 character set
+   The database must use a UTF-8 character set. A small caveat that you must be aware of is that utf8 in newer versions of MySQL is really utf8mb4 which
+   causes Airflow indexes to grow too large (see https://github.com/apache/airflow/pull/17603#issuecomment-901121618). Therefore as of Airflow 2.2
+   all MySQL databases have ``sql_engine_collation_for_ids`` set automatically to ``utf8mb3_general_ci`` (unless you override it). This might
+   lead to a mixture of collation ids for id fields in Airflow Database, but it has no negative consequences since all relevant IDs in Airflow use
+   ASCII characters only.
 
 We rely on more strict ANSI SQL settings for MySQL in order to have sane defaults.
 Make sure to have specified ``explicit_defaults_for_timestamp=1`` option under ``[mysqld]`` section

scripts/ci/docker-compose/backend-mysql.yml

@@ -21,7 +21,6 @@ services:
     environment:
       - BACKEND=mysql
       - AIRFLOW__CORE__SQL_ALCHEMY_CONN=mysql://root@mysql/airflow?charset=utf8mb4
-      - AIRFLOW__CORE__SQL_ENGINE_COLLATION_FOR_IDS=utf8mb3_general_ci
       - AIRFLOW__CELERY__RESULT_BACKEND=db+mysql://root@mysql/airflow?charset=utf8mb4
       - AIRFLOW__CORE__EXECUTOR=LocalExecutor
     depends_on:

scripts/in_container/entrypoint_ci.sh

@@ -39,9 +39,6 @@ echo
 echo "Airflow home: ${AIRFLOW_HOME}"
 echo "Airflow sources: ${AIRFLOW_SOURCES}"
 echo "Airflow core SQL connection: ${AIRFLOW__CORE__SQL_ALCHEMY_CONN:=}"
-if [[ -n "${AIRFLOW__CORE__SQL_ENGINE_COLLATION_FOR_IDS=}" ]]; then
-    echo "Airflow collation for IDs: ${AIRFLOW__CORE__SQL_ENGINE_COLLATION_FOR_IDS}"
-fi
 
 echo
 

tests/sensors/test_base.py

@@ -19,13 +19,15 @@
 
 import unittest
 from datetime import timedelta
+from unittest import mock
 from unittest.mock import Mock, patch
 
 import pytest
 from freezegun import freeze_time
 
 from airflow.exceptions import AirflowException, AirflowRescheduleException, AirflowSensorTimeout
 from airflow.models import DagBag, TaskInstance, TaskReschedule
+from airflow.models.base import get_id_collation_args
 from airflow.models.dag import DAG
 from airflow.operators.dummy import DummyOperator
 from airflow.sensors.base import BaseSensorOperator, poke_mode_only
@@ -655,3 +657,42 @@ def test_poke_mode_only_bad_poke(self):
         sensor = DummyPokeOnlySensor(task_id='foo', mode='poke', poke_changes_mode=True, dag=self.dag)
         with pytest.raises(ValueError):
             sensor.poke({})
+
+
+class TestCollation(unittest.TestCase):
+    @mock.patch.dict(
+        'os.environ',
+        AIRFLOW__CORE__SQL_ALCHEMY_CONN='postgres://host/the_database',
+    )
+    def test_collation_empty_on_non_mysql(self):
+        assert {} == get_id_collation_args()
+
+    @mock.patch.dict(
+        'os.environ',
+        AIRFLOW__CORE__SQL_ALCHEMY_CONN='mysql://host/the_database',
+    )
+    def test_collation_set_on_mysql(self):
+        assert {"collation": "utf8mb3_general_ci"} == get_id_collation_args()
+
+    @mock.patch.dict(
+        'os.environ',
+        AIRFLOW__CORE__SQL_ALCHEMY_CONN='mysql+pymsql://host/the_database',
+    )
+    def test_collation_set_on_mysql_with_pymsql(self):
+        assert {"collation": "utf8mb3_general_ci"} == get_id_collation_args()
+
+    @mock.patch.dict(
+        'os.environ',
+        AIRFLOW__CORE__SQL_ALCHEMY_CONN='mysql://host/the_database',
+        AIRFLOW__CORE__SQL_ENGINE_COLLATION_FOR_IDS='ascii',
+    )
+    def test_collation_override_on_non_mysql(self):
+        assert {"collation": "ascii"} == get_id_collation_args()
+
+    @mock.patch.dict(
+        'os.environ',
+        AIRFLOW__CORE__SQL_ALCHEMY_CONN='postgres://host/the_database',
+        AIRFLOW__CORE__SQL_ENGINE_COLLATION_FOR_IDS='ascii',
+    )
+    def test_collation_override_on_mysql(self):
+        assert {"collation": "ascii"} == get_id_collation_args()