[eglitobias]

On RabbitMQ 4.3 a quorum queue cancels a timed out consumer instead of closing the channel. Hutch re-subscribed on the same channel, which loses messages: the broker keeps routing deliveries to the cancelled tag, the client has already unregistered it, so Bunny drops the frame with no consumer for tag .... Those messages sit unacked under a consumer that no longer exists, and only closing the channel gets them back.

handle_cancellation now replaces the channel (close, reopen, re-declare the exchange) and re-subscribes every consumer. On its own thread, because closing the channel shuts down the work pool the callback runs on.

Reproduced with a quorum queue, x-consumer-timeout 5s, handler blocked 20s: four dropped deliveries per run before, none after. channel_recovery_spec.rb passed most of the time anyway, because the next timeout tick usually requeued them. Only 4.3 is affected, earlier versions close the channel themselves.

Also cut that spec's second wait_for from 90s to 15s.

[michaelklishin]

@eglitobias I have found a few more issues and known flaky test patterns, will push an update once I'm done with them

[michaelklishin]

I'll submit a follow-up PR.