Requests to the sequencer are batched heavily. If the sequencer fails, the cluster goes through a recovery and will be unavailable for 2-3 seconds and then recover.