Pulp

In Automation Hub the task galaxy_ng.app.tasks.synclist.curate_synclist_repository failed due to redis failure. However resource reservation for that task remained in the database blocking entire tasking system (if workers number = 1).

https://gist.github.com/cutwater/4ec7960f0eac2793ca17a78723dca75d

Environment:

pulpcore 3.7.1

pulp-ansible 0.4.3

galaxy-ng 1326eb5f1679880b68e05a48d4377def7c72a95b

Workers number: 1

Analysis¶

After review, the failure scenario goes like this:

1. The tasking code itself runs to completion
2. RQ attempts to notify Redis the task is completed (in the RQ registry) in its (handle_job_success)[https://github.com/rq/rq/blob/master/rq/worker.py#L925].
3. Interacting with Redis raised an exception at this line: https://github.com/rq/rq/blob/master/rq/worker.py#L932
4. This fatal exception raised and handled by Pulp's handle_job_failure handler implementation which records the exception (how we know this) and also marks the task as failed
5. Also when Redis became unavailable, it forgot the tasks it was storing in memory which includes the _release_resources that pairs with the now failed task and is intended to release the locks
6. The worker never died so other lock cleanup processes never occurred.
7. Tasks backup and eventually a sysadmin restarts the processes
8. The cleanup code in mark_worker_offline is triggered, but since the task is already at FAILED, this line does not issue it's cancellation which would release the locks
9. The locks are never released....

Solution¶

Add in code to mark_worker_offline that will ensure all locks for a worker being cleaned up are released even if the a task failed and its _release_resources was never delivered. This should occur after the cancellation for all tasks in "completed" states.