I do see some mentions in a paper about statistically evaluating watermarks

I'm using rabbitmq, so I will likely need something custom instead of the source-based watermark generators flink has (for Kafka for example)