I can easily see myself failing to catch this type of bug, especially if when you run it locally, the latency on jobs from enqueue to finish is aboue 5ms, whereas it probably fluctuates in production from a few ms to 5 minutes. It probably passed QA when latency was low.

If the desire is to mostly keep this architecture, the flag in the DB for "has a streaming account linked" needs to not be a boolean, and then you could have a third state besides "Ready to link" and "Link": 'Pending unlink' which would cause the UI to ask the user to stand by until the streaming site confirms the unlinking. Mildly inconvenient for the 0.1% of people who need to unlink just to immediately re-link, but better than buggy.