docs: clarify Redis maintenance for CeleryExecutor (#67393)

Author: Will-thom

providers/celery/docs/celery_executor.rst

@@ -95,6 +95,27 @@ Some caveats:
 - Tasks can consume resources. Make sure your worker has enough resources to run ``worker_concurrency`` tasks
 - Queue names are limited to 256 characters, but each broker backend might have its own restrictions
 
+Redis maintenance
+-----------------
+
+When Redis is used as the Celery broker, Airflow uses it for transient Celery messages: queued task
+commands, task acknowledgement data, and other broker-side state. Airflow task history, Dag runs,
+connections, Variables, and XCom records are stored in the Airflow metadata database and should be
+maintained with database tooling such as ``airflow db clean``. Cleaning the metadata database does not
+clean Redis, and cleaning Redis does not clean Airflow's metadata database.
+
+Do not flush or delete keys from the Redis database used by Airflow while schedulers or Celery workers
+are running. Removing broker keys while tasks are queued, running, or being acknowledged can lose
+pending work or make task state temporarily inconsistent. If you need to discard stale broker data
+after an outage, first stop Airflow schedulers and Celery workers, make sure there are no tasks that
+you intend to preserve in ``queued`` or ``running`` state, back up the Redis database if it is
+persistent, and then remove only the Redis database or keyspace configured in ``[celery] broker_url``.
+
+If you configured Redis as the Celery ``result_backend``, the same maintenance window should also
+account for result backend keys. Deleting those keys can remove Celery task result information that the
+scheduler may still query. For production deployments, Airflow recommends a database-backed result
+backend so broker cleanup and task-result storage are handled separately.
+
 See :doc:`apache-airflow:administration-and-deployment/modules_management` for details on how Python and Airflow manage modules.
 
 Architecture