In #1438 we added logic to disable the heartbeat thread while processing group rebalances. Heartbeats are then reenabled via a callback on the final join group success future, which is triggered after the final syncgroup request succeeds and the member assignment data is received.

However, there is a race condition: (1) if the join_group() call times out while the join/sync requests are in flight, (2) the join/sync future succeeds after the timeout while the consumer is not calling join_goup(), (3) the future enables the heartbeat thread, (4) the next call to join_group() disables the heartbeat thread before finding that the future has succeeded. In this case the consumer will be in a stable group membership but will not send any heartbeat requests.

The fix here is to only disable the heartbeat thread on the first call to join_group().

Fix #2610

I also moved heartbeat-related logs from the kafka.coordinator logger to the kafka.coordinator.heartbeat logger.