Some workloads e.g. Criteo1TB spent 75% of the total runtime time on evaluations. We should re-analyze and potentially reconfigure the intervals.