We recently merged a PR that clears some variables that were causing a memory leak in the bidirectional streaming classes. It seems like after getting that fix in, it exposed a deeper issue: the background thread isn't exiting as expected.

Quick Fix

When this happens, it's leading to a null pointer exception here, now that the _on_response can be empty. So as a quick fix, we should add a None check there

Deeper Issue

The root issue is that the BackgroundConsumer's background thread is getting stuck calling _on_response while being torn down. In the case of Firestore, it looks like the deadlock is the fault of the Watch class, which tries to re-enter a lock. But I think we should consider changing the BackgroundConsumer to close quicker, without doing that final on_response call at all. Is it really necessary to call that callback while being closed?

More Context

googleapis/python-firestore#985 (comment)

Let me know if you want me to submit a PR