Hi all! I believe we have found a small bug in the

metrics_anomaly_score

model around those lines. Attempt to explain: If

bucket_start

is null, then these window functions will randomly sort the rows for each partition and the rolling aggregations (avg, count etc.) will produce a different/not-consistent result for each row every time the code runs. I think adding

bucket_end

to the ordering would fix the issue. To confirm, it seems that this is smth that you have faced before and you added a comment in this macro used for test. Happy to open a github issue if that helps, thanks! 🙂