I am interested in the type of problem where you need last N values but last approx N values will not do. If you can make do with approx then just bucket them i

I am interested in the type of problem where you need last N values but last approx N values will not do. If you can make do with approx then just bucket them into 100s or whatever. Then there is no need to fo anything fancy for deletes and buckets are already aggregated. The for latency you could keep histogram counts.