Implement a recovery mechanism (probably around here) that checks if the pipeline errored out and restarts the pipeline after the backoff delay, unless the returned error is a fatal error.

Make sure to listen if the context gets cancelled so a pipeline can stop even while waiting for the backoff retry. Hint: see how we do it in the http processor.

Also be careful because the tomb in pipeline.Instance will be alive and will prevent the pipeline to be restarted until it's dead.