Oh, so the way it currently works clearml-serving will push the data in real-time into Prometheus (you can control the stats/input/out), then you can build the anomaly detection in grafana (for example alerts on histograms over time is out-of-the-box, and clearml creates the histograms overtime).
Would you also need access to the stats data in Prometheus ? or are you saying you need to process it before it gets there?

Is there any references to setup this? I'm not familiar about advanced features in clearml-serving